From 953bd55e0170a2ed3ec4a80de2b8966f6f1b67c0 Mon Sep 17 00:00:00 2001
From: Qualcomm AI Stack Models Bot <quic_qaihm_bot@quicinc.com>
Date: Wed, 27 Mar 2024 11:24:02 -0700
Subject: [PATCH] v0.4.0r3
Update Performance Numbers
Signed-off-by: QAIHM Team <quic_qaihm_bot@quicinc.com>
---
README.md | 2 +-
.../models/_shared/ffnet_quantized/model.py | 22 --
.../models/_shared/whisper/model.py | 8 -
qai_hub_models/models/aotgan/export.py | 2 +-
qai_hub_models/models/common.py | 1 +
qai_hub_models/models/convnext_tiny/export.py | 2 +-
qai_hub_models/models/ddrnet23_slim/export.py | 2 +-
.../models/deeplabv3_resnet50/export.py | 2 +-
qai_hub_models/models/densenet121/export.py | 2 +-
.../models/detr_resnet101/export.py | 2 +-
.../models/detr_resnet101_dc5/export.py | 2 +-
qai_hub_models/models/detr_resnet50/export.py | 2 +-
.../models/detr_resnet50_dc5/export.py | 2 +-
.../models/efficientnet_b0/export.py | 2 +-
qai_hub_models/models/esrgan/export.py | 2 +-
.../models/facebook_denoiser/export.py | 2 +-
qai_hub_models/models/fastsam_s/export.py | 2 +-
qai_hub_models/models/fastsam_x/export.py | 2 +-
qai_hub_models/models/fcn_resnet50/export.py | 2 +-
.../models/ffnet_122ns_lowres/export.py | 2 +-
qai_hub_models/models/ffnet_40s/export.py | 2 +-
.../models/ffnet_40s_quantized/export.py | 2 +-
qai_hub_models/models/ffnet_54s/export.py | 2 +-
.../models/ffnet_54s_quantized/export.py | 2 +-
qai_hub_models/models/ffnet_78s/export.py | 2 +-
.../models/ffnet_78s_lowres/export.py | 2 +-
.../models/ffnet_78s_quantized/export.py | 2 +-
qai_hub_models/models/googlenet/export.py | 2 +-
.../models/googlenet_quantized/export.py | 2 +-
.../models/googlenet_quantized/model.py | 6 -
qai_hub_models/models/hrnet_pose/export.py | 6 +-
.../models/hrnet_pose_quantized/export.py | 6 +-
.../huggingface_wavlm_base_plus/export.py | 2 +-
.../huggingface_wavlm_base_plus/perf.yaml | 41 ++--
qai_hub_models/models/inception_v3/export.py | 2 +-
.../models/inception_v3_quantized/export.py | 2 +-
.../models/inception_v3_quantized/model.py | 14 --
qai_hub_models/models/lama_dilated/export.py | 2 +-
qai_hub_models/models/litehrnet/export.py | 2 +-
.../models/mediapipe_face/export.py | 13 +-
.../models/mediapipe_hand/export.py | 13 +-
.../models/mediapipe_pose/export.py | 13 +-
.../models/mediapipe_selfie/export.py | 2 +-
qai_hub_models/models/mnasnet05/export.py | 2 +-
qai_hub_models/models/mobilenet_v2/export.py | 2 +-
.../models/mobilenet_v2_quantized/export.py | 2 +-
.../models/mobilenet_v2_quantized/model.py | 14 --
.../models/mobilenet_v3_large/export.py | 2 +-
.../mobilenet_v3_large_quantized/export.py | 2 +-
.../mobilenet_v3_large_quantized/model.py | 6 -
.../models/mobilenet_v3_small/export.py | 2 +-
qai_hub_models/models/openai_clip/export.py | 13 +-
qai_hub_models/models/openai_clip/perf.yaml | 145 +++----------
qai_hub_models/models/openpose/export.py | 2 +-
qai_hub_models/models/protocols.py | 56 ++++-
.../models/quicksrnetlarge/export.py | 2 +-
.../quicksrnetlarge_quantized/export.py | 2 +-
.../models/quicksrnetlarge_quantized/model.py | 9 +
.../models/quicksrnetmedium/export.py | 2 +-
.../quicksrnetmedium_quantized/export.py | 2 +-
.../quicksrnetmedium_quantized/model.py | 9 +
.../models/quicksrnetsmall/export.py | 2 +-
.../quicksrnetsmall_quantized/export.py | 2 +-
.../models/quicksrnetsmall_quantized/model.py | 9 +
.../models/real_esrgan_general_x4v3/export.py | 2 +-
.../models/real_esrgan_x4plus/export.py | 2 +-
qai_hub_models/models/regnet/export.py | 2 +-
qai_hub_models/models/resnet101/export.py | 2 +-
.../models/resnet101_quantized/export.py | 2 +-
.../models/resnet101_quantized/model.py | 14 --
qai_hub_models/models/resnet18/export.py | 2 +-
.../models/resnet18_quantized/export.py | 2 +-
.../models/resnet18_quantized/model.py | 14 --
qai_hub_models/models/resnet50/export.py | 2 +-
qai_hub_models/models/resnext101/export.py | 2 +-
.../models/resnext101_quantized/export.py | 2 +-
.../models/resnext101_quantized/model.py | 14 --
qai_hub_models/models/resnext50/export.py | 2 +-
.../models/resnext50_quantized/export.py | 2 +-
.../models/resnext50_quantized/model.py | 14 --
qai_hub_models/models/sam/export.py | 5 +-
qai_hub_models/models/sesr_m5/export.py | 2 +-
.../models/sesr_m5_quantized/export.py | 2 +-
.../models/sesr_m5_quantized/model.py | 9 +
qai_hub_models/models/shufflenet_v2/export.py | 2 +-
.../models/shufflenet_v2_quantized/export.py | 2 +-
.../models/shufflenet_v2_quantized/model.py | 14 --
qai_hub_models/models/sinet/export.py | 2 +-
qai_hub_models/models/squeezenet1_1/export.py | 2 +-
.../models/squeezenet1_1_quantized/export.py | 2 +-
.../models/squeezenet1_1_quantized/model.py | 14 --
qai_hub_models/models/stylegan2/export.py | 2 +-
qai_hub_models/models/swin_base/export.py | 2 +-
qai_hub_models/models/swin_small/export.py | 2 +-
qai_hub_models/models/swin_tiny/export.py | 2 +-
qai_hub_models/models/trocr/export.py | 14 +-
.../models/unet_segmentation/export.py | 2 +-
qai_hub_models/models/vit/export.py | 2 +-
.../models/whisper_base_en/export.py | 5 +-
.../models/whisper_small_en/export.py | 5 +-
.../models/whisper_tiny_en/export.py | 5 +-
qai_hub_models/models/wideresnet50/export.py | 2 +-
.../models/wideresnet50_quantized/export.py | 2 +-
.../models/wideresnet50_quantized/model.py | 14 --
qai_hub_models/models/xlsr/export.py | 2 +-
.../models/xlsr_quantized/export.py | 2 +-
qai_hub_models/models/xlsr_quantized/model.py | 9 +
qai_hub_models/models/yolov6/export.py | 2 +-
qai_hub_models/models/yolov7/export.py | 2 +-
qai_hub_models/models/yolov8_det/export.py | 2 +-
qai_hub_models/models/yolov8_seg/export.py | 2 +-
qai_hub_models/utils/args.py | 4 +
qai_hub_models/utils/base_model.py | 6 +-
qai_hub_models/utils/config_loaders.py | 194 +++++++++++++-----
qai_hub_models/utils/printing.py | 2 +
qai_hub_models/utils/quantization_aimet.py | 25 ++-
116 files changed, 445 insertions(+), 496 deletions(-)
diff --git a/README.md b/README.md
index dbb402ab..d9443541 100644
--- a/README.md
+++ b/README.md
@@ -373,8 +373,8 @@ Qualcomm® AI Hub Models is licensed under BSD-3. See the [LICENSE file](../LICE
| Model | README | Torch App | Device Export | CLI Demo
| -- | -- | -- | -- | --
| | | | |
-| [TrOCR](https://aihub.qualcomm.com/models/trocr) | [qai_hub_models.models.trocr](qai_hub_models/models/trocr/README.md) | ✔️ | ✔️ | ✔️
| [OpenAI-Clip](https://aihub.qualcomm.com/models/openai_clip) | [qai_hub_models.models.openai_clip](qai_hub_models/models/openai_clip/README.md) | ✔️ | ✔️ | ✔️
+| [TrOCR](https://aihub.qualcomm.com/models/trocr) | [qai_hub_models.models.trocr](qai_hub_models/models/trocr/README.md) | ✔️ | ✔️ | ✔️
### Generative Ai
diff --git a/qai_hub_models/models/_shared/ffnet_quantized/model.py b/qai_hub_models/models/_shared/ffnet_quantized/model.py
index 68343a1c..438e6169 100644
--- a/qai_hub_models/models/_shared/ffnet_quantized/model.py
+++ b/qai_hub_models/models/_shared/ffnet_quantized/model.py
@@ -19,11 +19,8 @@
from aimet_torch.batch_norm_fold import fold_all_batch_norms
from aimet_torch.model_preparer import prepare_model
from aimet_torch.quantsim import QuantizationSimModel, load_encodings_to_sim
-from qai_hub.client import DatasetEntries
from qai_hub_models.models._shared.ffnet.model import FFNet
-from qai_hub_models.utils.base_model import SourceModelFormat, TargetRuntime
-from qai_hub_models.utils.input_spec import InputSpec
MODEL_ID = __name__.split(".")[-2]
FFNET_AIMET_CONFIG = os.path.abspath(
@@ -47,14 +44,6 @@ def __init__(
FFNet.__init__(self, ffnet_model.model)
AIMETQuantizableMixin.__init__(self, ffnet_model)
- def get_hub_compile_options(
- self, target_runtime: TargetRuntime, other_compile_options: str = ""
- ) -> str:
- compile_options = super().get_hub_compile_options(
- target_runtime, other_compile_options
- )
- return compile_options + " --quantize_full_type int8 --quantize_io"
-
@classmethod
def default_aimet_encodings(cls) -> str:
raise NotImplementedError()
@@ -88,14 +77,3 @@ def from_pretrained(
sim.model.eval()
return cls(sim)
-
- def preferred_hub_source_model_format(
- self, target_runtime: TargetRuntime
- ) -> SourceModelFormat:
- return SourceModelFormat.ONNX
-
- def get_calibration_data(
- self, target_runtime: TargetRuntime, input_spec: InputSpec | None = None
- ) -> DatasetEntries | None:
- # Do not provide calibration data
- return None
diff --git a/qai_hub_models/models/_shared/whisper/model.py b/qai_hub_models/models/_shared/whisper/model.py
index ab5db089..dffcbb07 100644
--- a/qai_hub_models/models/_shared/whisper/model.py
+++ b/qai_hub_models/models/_shared/whisper/model.py
@@ -82,14 +82,6 @@ def get_input_spec() -> InputSpec:
def from_pretrained(cls):
return Whisper.from_pretrained().encoder
- def get_hub_compile_options(
- self, target_runtime: TargetRuntime, other_compile_options: str = ""
- ) -> str:
- compile_options = super().get_hub_compile_options(
- target_runtime, other_compile_options
- )
- return compile_options + " --compute_unit gpu"
-
def get_hub_profile_options(
self, target_runtime: TargetRuntime, other_profile_options: str = ""
) -> str:
diff --git a/qai_hub_models/models/aotgan/export.py b/qai_hub_models/models/aotgan/export.py
index 1316347d..65f0b5c7 100644
--- a/qai_hub_models/models/aotgan/export.py
+++ b/qai_hub_models/models/aotgan/export.py
@@ -197,7 +197,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/common.py b/qai_hub_models/models/common.py
index bf635e0e..37020fd3 100644
--- a/qai_hub_models/models/common.py
+++ b/qai_hub_models/models/common.py
@@ -11,6 +11,7 @@
class TargetRuntime(Enum):
TFLITE = 0
QNN = 1
+ ORT = 2
def __str__(self):
return self.name.lower()
diff --git a/qai_hub_models/models/convnext_tiny/export.py b/qai_hub_models/models/convnext_tiny/export.py
index 4649e4ef..f8600751 100644
--- a/qai_hub_models/models/convnext_tiny/export.py
+++ b/qai_hub_models/models/convnext_tiny/export.py
@@ -189,7 +189,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model, supports_qnn=False)
+ parser = export_parser(model_cls=Model, supports_qnn=False, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/ddrnet23_slim/export.py b/qai_hub_models/models/ddrnet23_slim/export.py
index 62378aea..16a76e1f 100644
--- a/qai_hub_models/models/ddrnet23_slim/export.py
+++ b/qai_hub_models/models/ddrnet23_slim/export.py
@@ -197,7 +197,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model, supports_qnn=False)
+ parser = export_parser(model_cls=Model, supports_qnn=False, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/deeplabv3_resnet50/export.py b/qai_hub_models/models/deeplabv3_resnet50/export.py
index ff57a6fc..d2a2445e 100644
--- a/qai_hub_models/models/deeplabv3_resnet50/export.py
+++ b/qai_hub_models/models/deeplabv3_resnet50/export.py
@@ -193,7 +193,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/densenet121/export.py b/qai_hub_models/models/densenet121/export.py
index 5d02faa4..4268f1fe 100644
--- a/qai_hub_models/models/densenet121/export.py
+++ b/qai_hub_models/models/densenet121/export.py
@@ -189,7 +189,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/detr_resnet101/export.py b/qai_hub_models/models/detr_resnet101/export.py
index f46ca4e5..838626a5 100644
--- a/qai_hub_models/models/detr_resnet101/export.py
+++ b/qai_hub_models/models/detr_resnet101/export.py
@@ -189,7 +189,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model, supports_qnn=False)
+ parser = export_parser(model_cls=Model, supports_qnn=False, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/detr_resnet101_dc5/export.py b/qai_hub_models/models/detr_resnet101_dc5/export.py
index e8c61070..2f8b1653 100644
--- a/qai_hub_models/models/detr_resnet101_dc5/export.py
+++ b/qai_hub_models/models/detr_resnet101_dc5/export.py
@@ -189,7 +189,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model, supports_qnn=False)
+ parser = export_parser(model_cls=Model, supports_qnn=False, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/detr_resnet50/export.py b/qai_hub_models/models/detr_resnet50/export.py
index 70838101..f6e4697e 100644
--- a/qai_hub_models/models/detr_resnet50/export.py
+++ b/qai_hub_models/models/detr_resnet50/export.py
@@ -189,7 +189,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model, supports_qnn=False)
+ parser = export_parser(model_cls=Model, supports_qnn=False, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/detr_resnet50_dc5/export.py b/qai_hub_models/models/detr_resnet50_dc5/export.py
index 09e9b406..e3d443a3 100644
--- a/qai_hub_models/models/detr_resnet50_dc5/export.py
+++ b/qai_hub_models/models/detr_resnet50_dc5/export.py
@@ -189,7 +189,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model, supports_qnn=False)
+ parser = export_parser(model_cls=Model, supports_qnn=False, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/efficientnet_b0/export.py b/qai_hub_models/models/efficientnet_b0/export.py
index 44abf4a7..dfff029c 100644
--- a/qai_hub_models/models/efficientnet_b0/export.py
+++ b/qai_hub_models/models/efficientnet_b0/export.py
@@ -189,7 +189,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/esrgan/export.py b/qai_hub_models/models/esrgan/export.py
index 01c64d63..d03a9f09 100644
--- a/qai_hub_models/models/esrgan/export.py
+++ b/qai_hub_models/models/esrgan/export.py
@@ -193,7 +193,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/facebook_denoiser/export.py b/qai_hub_models/models/facebook_denoiser/export.py
index f536b473..997188b1 100644
--- a/qai_hub_models/models/facebook_denoiser/export.py
+++ b/qai_hub_models/models/facebook_denoiser/export.py
@@ -184,7 +184,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model, supports_qnn=False)
+ parser = export_parser(model_cls=Model, supports_qnn=False, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/fastsam_s/export.py b/qai_hub_models/models/fastsam_s/export.py
index 4902b3a5..1c56e30d 100644
--- a/qai_hub_models/models/fastsam_s/export.py
+++ b/qai_hub_models/models/fastsam_s/export.py
@@ -199,7 +199,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model, supports_qnn=False)
+ parser = export_parser(model_cls=Model, supports_qnn=False, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/fastsam_x/export.py b/qai_hub_models/models/fastsam_x/export.py
index 617e16c5..0bc1753c 100644
--- a/qai_hub_models/models/fastsam_x/export.py
+++ b/qai_hub_models/models/fastsam_x/export.py
@@ -199,7 +199,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model, supports_qnn=False)
+ parser = export_parser(model_cls=Model, supports_qnn=False, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/fcn_resnet50/export.py b/qai_hub_models/models/fcn_resnet50/export.py
index 3a5dd85a..78b7bd02 100644
--- a/qai_hub_models/models/fcn_resnet50/export.py
+++ b/qai_hub_models/models/fcn_resnet50/export.py
@@ -197,7 +197,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/ffnet_122ns_lowres/export.py b/qai_hub_models/models/ffnet_122ns_lowres/export.py
index 05689b50..cebe7622 100644
--- a/qai_hub_models/models/ffnet_122ns_lowres/export.py
+++ b/qai_hub_models/models/ffnet_122ns_lowres/export.py
@@ -193,7 +193,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/ffnet_40s/export.py b/qai_hub_models/models/ffnet_40s/export.py
index dac6aea6..40804e7e 100644
--- a/qai_hub_models/models/ffnet_40s/export.py
+++ b/qai_hub_models/models/ffnet_40s/export.py
@@ -193,7 +193,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/ffnet_40s_quantized/export.py b/qai_hub_models/models/ffnet_40s_quantized/export.py
index 08ed7624..bb562224 100644
--- a/qai_hub_models/models/ffnet_40s_quantized/export.py
+++ b/qai_hub_models/models/ffnet_40s_quantized/export.py
@@ -202,7 +202,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model, supports_qnn=False)
+ parser = export_parser(model_cls=Model, supports_qnn=False, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/ffnet_54s/export.py b/qai_hub_models/models/ffnet_54s/export.py
index 5a27ee33..ad77a2b3 100644
--- a/qai_hub_models/models/ffnet_54s/export.py
+++ b/qai_hub_models/models/ffnet_54s/export.py
@@ -193,7 +193,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/ffnet_54s_quantized/export.py b/qai_hub_models/models/ffnet_54s_quantized/export.py
index 83dc6e73..a1b046f8 100644
--- a/qai_hub_models/models/ffnet_54s_quantized/export.py
+++ b/qai_hub_models/models/ffnet_54s_quantized/export.py
@@ -202,7 +202,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model, supports_qnn=False)
+ parser = export_parser(model_cls=Model, supports_qnn=False, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/ffnet_78s/export.py b/qai_hub_models/models/ffnet_78s/export.py
index 09e78a83..f2fd96d5 100644
--- a/qai_hub_models/models/ffnet_78s/export.py
+++ b/qai_hub_models/models/ffnet_78s/export.py
@@ -193,7 +193,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/ffnet_78s_lowres/export.py b/qai_hub_models/models/ffnet_78s_lowres/export.py
index 08f9197f..b5d2fce0 100644
--- a/qai_hub_models/models/ffnet_78s_lowres/export.py
+++ b/qai_hub_models/models/ffnet_78s_lowres/export.py
@@ -193,7 +193,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/ffnet_78s_quantized/export.py b/qai_hub_models/models/ffnet_78s_quantized/export.py
index 49c83ae3..dd043de2 100644
--- a/qai_hub_models/models/ffnet_78s_quantized/export.py
+++ b/qai_hub_models/models/ffnet_78s_quantized/export.py
@@ -202,7 +202,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model, supports_qnn=False)
+ parser = export_parser(model_cls=Model, supports_qnn=False, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/googlenet/export.py b/qai_hub_models/models/googlenet/export.py
index 81262c38..d97de5a7 100644
--- a/qai_hub_models/models/googlenet/export.py
+++ b/qai_hub_models/models/googlenet/export.py
@@ -189,7 +189,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/googlenet_quantized/export.py b/qai_hub_models/models/googlenet_quantized/export.py
index 7ed01e3e..15aa2751 100644
--- a/qai_hub_models/models/googlenet_quantized/export.py
+++ b/qai_hub_models/models/googlenet_quantized/export.py
@@ -200,7 +200,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/googlenet_quantized/model.py b/qai_hub_models/models/googlenet_quantized/model.py
index e100c42a..0d486377 100644
--- a/qai_hub_models/models/googlenet_quantized/model.py
+++ b/qai_hub_models/models/googlenet_quantized/model.py
@@ -20,7 +20,6 @@
from qai_hub_models.models.googlenet.model import GoogLeNet
from qai_hub_models.utils.aimet.config_loader import get_default_aimet_config
from qai_hub_models.utils.asset_loaders import CachedWebModelAsset
-from qai_hub_models.utils.base_model import SourceModelFormat, TargetRuntime
from qai_hub_models.utils.quantization_aimet import tie_aimet_observer_groups
MODEL_ID = __name__.split(".")[-2]
@@ -44,11 +43,6 @@ def __init__(
sim_model,
)
- def preferred_hub_source_model_format(
- self, target_runtime: TargetRuntime
- ) -> SourceModelFormat:
- return SourceModelFormat.ONNX
-
@classmethod
def from_pretrained(
cls,
diff --git a/qai_hub_models/models/hrnet_pose/export.py b/qai_hub_models/models/hrnet_pose/export.py
index 338f35fd..0acf5119 100644
--- a/qai_hub_models/models/hrnet_pose/export.py
+++ b/qai_hub_models/models/hrnet_pose/export.py
@@ -116,7 +116,7 @@ def export_model(
model_compile_options = model.get_hub_compile_options(
target_runtime,
compile_options
- + " --force_channel_last_input image_tensor"
+ + " --force_channel_last_input image"
+ " --force_channel_last_output output_0",
)
print(f"Optimizing model {model_name} to run on-device")
@@ -156,7 +156,7 @@ def export_model(
sample_inputs = model.sample_inputs(input_spec)
# Convert inputs from channel first to channel last
hub_inputs = transpose_channel_first_to_last(
- "image_tensor", sample_inputs, target_runtime
+ "image", sample_inputs, target_runtime
)
submitted_inference_job = hub.submit_inference_job(
model=compile_job.get_target_model(),
@@ -197,7 +197,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/hrnet_pose_quantized/export.py b/qai_hub_models/models/hrnet_pose_quantized/export.py
index 57904574..ec61ebb0 100644
--- a/qai_hub_models/models/hrnet_pose_quantized/export.py
+++ b/qai_hub_models/models/hrnet_pose_quantized/export.py
@@ -121,7 +121,7 @@ def export_model(
model_compile_options = model.get_hub_compile_options(
target_runtime,
compile_options
- + " --force_channel_last_input image_tensor"
+ + " --force_channel_last_input image"
+ " --force_channel_last_output output_0",
)
print(f"Optimizing model {model_name} to run on-device")
@@ -165,7 +165,7 @@ def export_model(
hub_inputs = get_qnn_inputs(compile_job, sample_inputs)
# Convert inputs from channel first to channel last
hub_inputs = transpose_channel_first_to_last(
- "image_tensor", sample_inputs, target_runtime
+ "image", sample_inputs, target_runtime
)
submitted_inference_job = hub.submit_inference_job(
model=compile_job.get_target_model(),
@@ -206,7 +206,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model, supports_qnn=False)
+ parser = export_parser(model_cls=Model, supports_qnn=False, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/huggingface_wavlm_base_plus/export.py b/qai_hub_models/models/huggingface_wavlm_base_plus/export.py
index 5237cafc..59b3c4d5 100644
--- a/qai_hub_models/models/huggingface_wavlm_base_plus/export.py
+++ b/qai_hub_models/models/huggingface_wavlm_base_plus/export.py
@@ -180,7 +180,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model, supports_qnn=False)
+ parser = export_parser(model_cls=Model, supports_qnn=False, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/huggingface_wavlm_base_plus/perf.yaml b/qai_hub_models/models/huggingface_wavlm_base_plus/perf.yaml
index e930ee5d..324eb6fa 100644
--- a/qai_hub_models/models/huggingface_wavlm_base_plus/perf.yaml
+++ b/qai_hub_models/models/huggingface_wavlm_base_plus/perf.yaml
@@ -19,6 +19,7 @@ aggregated:
- Samsung Galaxy S23+
- Samsung Galaxy S24
- Samsung Galaxy S24 Ultra
+ - Samsung Galaxy S24+
- Samsung Galaxy Tab S8
- Xiaomi 12
- Xiaomi 12 Pro
@@ -31,19 +32,19 @@ models:
- name: HuggingFace-WavLM-Base-Plus
performance_metrics:
- torchscript_onnx_tflite:
- inference_time: 237767939.0
- throughput: 0.0042057815036197965
+ inference_time: 911873.0
+ throughput: 1.0966439405487387
estimated_peak_memory_range:
- min: 11886592
- max: 15703120
- primary_compute_unit: NPU
- precision: fp16
+ min: 149282816
+ max: 153276888
+ primary_compute_unit: CPU
+ precision: fp32
layer_info:
- layers_on_npu: 848
+ layers_on_npu: 0
layers_on_gpu: 0
- layers_on_cpu: 0
+ layers_on_cpu: 848
total_layers: 848
- job_id: jlpe928gr
+ job_id: jz57yq4q5
job_status: Passed
reference_device_info:
name: Samsung Galaxy S23
@@ -52,7 +53,7 @@ models:
os_name: Android
manufacturer: Samsung
chipset: Snapdragon® 8 Gen 2
- timestamp: '2024-03-15T00:13:01.265817Z'
+ timestamp: '2024-03-26T15:30:16.725161Z'
torchscript_onnx_qnn:
inference_time: 'null'
throughput: 'null'
@@ -69,19 +70,19 @@ models:
job_id: ''
job_status: Skipped
- torchscript_onnx_tflite:
- inference_time: 174470189.0
- throughput: 0.005731638199807303
+ inference_time: 712941.0
+ throughput: 1.4026406112146728
estimated_peak_memory_range:
- min: 11321344
- max: 711668304
- primary_compute_unit: NPU
- precision: fp16
+ min: 147787776
+ max: 179693888
+ primary_compute_unit: CPU
+ precision: fp32
layer_info:
- layers_on_npu: 848
+ layers_on_npu: 0
layers_on_gpu: 0
- layers_on_cpu: 0
+ layers_on_cpu: 848
total_layers: 848
- job_id: jygzew4g8
+ job_id: joprvzlvg
job_status: Passed
reference_device_info:
name: Samsung Galaxy S24
@@ -90,7 +91,7 @@ models:
os_name: Android
manufacturer: Samsung
chipset: Snapdragon® 8 Gen 3
- timestamp: '2024-03-15T00:13:01.265830Z'
+ timestamp: '2024-03-26T15:30:16.725174Z'
torchscript_onnx_qnn:
inference_time: 'null'
throughput: 'null'
diff --git a/qai_hub_models/models/inception_v3/export.py b/qai_hub_models/models/inception_v3/export.py
index 461d52ea..ff4ec962 100644
--- a/qai_hub_models/models/inception_v3/export.py
+++ b/qai_hub_models/models/inception_v3/export.py
@@ -189,7 +189,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/inception_v3_quantized/export.py b/qai_hub_models/models/inception_v3_quantized/export.py
index a5b83180..b723f39f 100644
--- a/qai_hub_models/models/inception_v3_quantized/export.py
+++ b/qai_hub_models/models/inception_v3_quantized/export.py
@@ -200,7 +200,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model, supports_qnn=False)
+ parser = export_parser(model_cls=Model, supports_qnn=False, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/inception_v3_quantized/model.py b/qai_hub_models/models/inception_v3_quantized/model.py
index 39dd2658..44c3fba5 100644
--- a/qai_hub_models/models/inception_v3_quantized/model.py
+++ b/qai_hub_models/models/inception_v3_quantized/model.py
@@ -20,7 +20,6 @@
from qai_hub_models.models.inception_v3.model import InceptionNetV3
from qai_hub_models.utils.aimet.config_loader import get_default_aimet_config
from qai_hub_models.utils.asset_loaders import CachedWebModelAsset
-from qai_hub_models.utils.base_model import SourceModelFormat, TargetRuntime
from qai_hub_models.utils.quantization_aimet import tie_aimet_observer_groups
MODEL_ID = __name__.split(".")[-2]
@@ -47,11 +46,6 @@ def __init__(
sim_model,
)
- def preferred_hub_source_model_format(
- self, target_runtime: TargetRuntime
- ) -> SourceModelFormat:
- return SourceModelFormat.ONNX
-
@classmethod
def from_pretrained(
cls,
@@ -196,11 +190,3 @@ def _tie_pre_concat_quantizers(cls, sim: QuantizationSimModel):
],
]
tie_aimet_observer_groups(groups)
-
- def get_hub_compile_options(
- self, target_runtime: TargetRuntime, other_compile_options: str = ""
- ) -> str:
- compile_options = super().get_hub_compile_options(
- target_runtime, other_compile_options
- )
- return compile_options + " --quantize_full_type int8 --quantize_io"
diff --git a/qai_hub_models/models/lama_dilated/export.py b/qai_hub_models/models/lama_dilated/export.py
index 02ae013e..7393373c 100644
--- a/qai_hub_models/models/lama_dilated/export.py
+++ b/qai_hub_models/models/lama_dilated/export.py
@@ -197,7 +197,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/litehrnet/export.py b/qai_hub_models/models/litehrnet/export.py
index 13f51b82..4c2df40a 100644
--- a/qai_hub_models/models/litehrnet/export.py
+++ b/qai_hub_models/models/litehrnet/export.py
@@ -184,7 +184,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model, supports_qnn=False)
+ parser = export_parser(model_cls=Model, supports_qnn=False, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/mediapipe_face/export.py b/qai_hub_models/models/mediapipe_face/export.py
index fe627869..007ffb4d 100644
--- a/qai_hub_models/models/mediapipe_face/export.py
+++ b/qai_hub_models/models/mediapipe_face/export.py
@@ -27,7 +27,6 @@
from qai_hub_models.utils.qai_hub_helpers import (
can_access_qualcomm_ai_hub,
export_without_hub_access,
- transpose_channel_first_to_last,
)
ALL_COMPONENTS = ["MediaPipeFaceDetector", "MediaPipeFaceLandmarkDetector"]
@@ -127,7 +126,7 @@ def export_model(
# 2. Compile the models to an on-device asset
model_compile_options = component.get_hub_compile_options(
- target_runtime, compile_options + " --force_channel_last_input image"
+ target_runtime, compile_options
)
print(f"Optimizing model {component_name} to run on-device")
submitted_compile_job = hub.submit_compile_job(
@@ -170,13 +169,9 @@ def export_model(
component_name
].get_hub_profile_options(target_runtime, profile_options)
sample_inputs = components_dict[component_name].sample_inputs()
- # Convert inputs from channel first to channel last
- hub_inputs = transpose_channel_first_to_last(
- "image", sample_inputs, target_runtime
- )
submitted_inference_job = hub.submit_inference_job(
model=compile_jobs[component_name].get_target_model(),
- inputs=hub_inputs,
+ inputs=sample_inputs,
device=hub.Device(device),
name=f"{model_name}_{component_name}",
options=profile_options_all,
@@ -223,7 +218,9 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model, components=ALL_COMPONENTS)
+ parser = export_parser(
+ model_cls=Model, components=ALL_COMPONENTS, supports_ort=False
+ )
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/mediapipe_hand/export.py b/qai_hub_models/models/mediapipe_hand/export.py
index 22ff6f03..2585cead 100644
--- a/qai_hub_models/models/mediapipe_hand/export.py
+++ b/qai_hub_models/models/mediapipe_hand/export.py
@@ -27,7 +27,6 @@
from qai_hub_models.utils.qai_hub_helpers import (
can_access_qualcomm_ai_hub,
export_without_hub_access,
- transpose_channel_first_to_last,
)
ALL_COMPONENTS = ["MediaPipeHandDetector", "MediaPipeHandLandmarkDetector"]
@@ -127,7 +126,7 @@ def export_model(
# 2. Compile the models to an on-device asset
model_compile_options = component.get_hub_compile_options(
- target_runtime, compile_options + " --force_channel_last_input image"
+ target_runtime, compile_options
)
print(f"Optimizing model {component_name} to run on-device")
submitted_compile_job = hub.submit_compile_job(
@@ -170,13 +169,9 @@ def export_model(
component_name
].get_hub_profile_options(target_runtime, profile_options)
sample_inputs = components_dict[component_name].sample_inputs()
- # Convert inputs from channel first to channel last
- hub_inputs = transpose_channel_first_to_last(
- "image", sample_inputs, target_runtime
- )
submitted_inference_job = hub.submit_inference_job(
model=compile_jobs[component_name].get_target_model(),
- inputs=hub_inputs,
+ inputs=sample_inputs,
device=hub.Device(device),
name=f"{model_name}_{component_name}",
options=profile_options_all,
@@ -223,7 +218,9 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model, components=ALL_COMPONENTS)
+ parser = export_parser(
+ model_cls=Model, components=ALL_COMPONENTS, supports_ort=False
+ )
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/mediapipe_pose/export.py b/qai_hub_models/models/mediapipe_pose/export.py
index ec0c57a3..3ed5d886 100644
--- a/qai_hub_models/models/mediapipe_pose/export.py
+++ b/qai_hub_models/models/mediapipe_pose/export.py
@@ -27,7 +27,6 @@
from qai_hub_models.utils.qai_hub_helpers import (
can_access_qualcomm_ai_hub,
export_without_hub_access,
- transpose_channel_first_to_last,
)
ALL_COMPONENTS = ["MediaPipePoseDetector", "MediaPipePoseLandmarkDetector"]
@@ -127,7 +126,7 @@ def export_model(
# 2. Compile the models to an on-device asset
model_compile_options = component.get_hub_compile_options(
- target_runtime, compile_options + " --force_channel_last_input image"
+ target_runtime, compile_options
)
print(f"Optimizing model {component_name} to run on-device")
submitted_compile_job = hub.submit_compile_job(
@@ -170,13 +169,9 @@ def export_model(
component_name
].get_hub_profile_options(target_runtime, profile_options)
sample_inputs = components_dict[component_name].sample_inputs()
- # Convert inputs from channel first to channel last
- hub_inputs = transpose_channel_first_to_last(
- "image", sample_inputs, target_runtime
- )
submitted_inference_job = hub.submit_inference_job(
model=compile_jobs[component_name].get_target_model(),
- inputs=hub_inputs,
+ inputs=sample_inputs,
device=hub.Device(device),
name=f"{model_name}_{component_name}",
options=profile_options_all,
@@ -223,7 +218,9 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model, components=ALL_COMPONENTS)
+ parser = export_parser(
+ model_cls=Model, components=ALL_COMPONENTS, supports_ort=False
+ )
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/mediapipe_selfie/export.py b/qai_hub_models/models/mediapipe_selfie/export.py
index ce654b17..9ef6cf58 100644
--- a/qai_hub_models/models/mediapipe_selfie/export.py
+++ b/qai_hub_models/models/mediapipe_selfie/export.py
@@ -197,7 +197,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/mnasnet05/export.py b/qai_hub_models/models/mnasnet05/export.py
index c500d613..794bb948 100644
--- a/qai_hub_models/models/mnasnet05/export.py
+++ b/qai_hub_models/models/mnasnet05/export.py
@@ -189,7 +189,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/mobilenet_v2/export.py b/qai_hub_models/models/mobilenet_v2/export.py
index d5d83eb6..67c9c88e 100644
--- a/qai_hub_models/models/mobilenet_v2/export.py
+++ b/qai_hub_models/models/mobilenet_v2/export.py
@@ -189,7 +189,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/mobilenet_v2_quantized/export.py b/qai_hub_models/models/mobilenet_v2_quantized/export.py
index f73c0796..e01ee5a7 100644
--- a/qai_hub_models/models/mobilenet_v2_quantized/export.py
+++ b/qai_hub_models/models/mobilenet_v2_quantized/export.py
@@ -200,7 +200,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/mobilenet_v2_quantized/model.py b/qai_hub_models/models/mobilenet_v2_quantized/model.py
index d72efcb4..b8849fb8 100644
--- a/qai_hub_models/models/mobilenet_v2_quantized/model.py
+++ b/qai_hub_models/models/mobilenet_v2_quantized/model.py
@@ -23,7 +23,6 @@
)
from qai_hub_models.utils.aimet.config_loader import get_default_aimet_config
from qai_hub_models.utils.asset_loaders import CachedWebModelAsset
-from qai_hub_models.utils.base_model import SourceModelFormat, TargetRuntime
from qai_hub_models.utils.quantization_aimet import convert_all_depthwise_to_per_tensor
MODEL_ID = __name__.split(".")[-2]
@@ -47,11 +46,6 @@ def __init__(
quant_sim_model,
)
- def preferred_hub_source_model_format(
- self, target_runtime: TargetRuntime
- ) -> SourceModelFormat:
- return SourceModelFormat.ONNX
-
@classmethod
def from_pretrained(
cls,
@@ -104,11 +98,3 @@ def from_pretrained(
sim.model.eval()
return cls(sim)
-
- def get_hub_compile_options(
- self, target_runtime: TargetRuntime, other_compile_options: str = ""
- ) -> str:
- compile_options = super().get_hub_compile_options(
- target_runtime, other_compile_options
- )
- return compile_options + " --quantize_full_type int8 --quantize_io"
diff --git a/qai_hub_models/models/mobilenet_v3_large/export.py b/qai_hub_models/models/mobilenet_v3_large/export.py
index 5bb2fffd..2a013f28 100644
--- a/qai_hub_models/models/mobilenet_v3_large/export.py
+++ b/qai_hub_models/models/mobilenet_v3_large/export.py
@@ -189,7 +189,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model, supports_qnn=False)
+ parser = export_parser(model_cls=Model, supports_qnn=False, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/mobilenet_v3_large_quantized/export.py b/qai_hub_models/models/mobilenet_v3_large_quantized/export.py
index 5305d6b9..ce0375ef 100644
--- a/qai_hub_models/models/mobilenet_v3_large_quantized/export.py
+++ b/qai_hub_models/models/mobilenet_v3_large_quantized/export.py
@@ -193,7 +193,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model, supports_qnn=False)
+ parser = export_parser(model_cls=Model, supports_qnn=False, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/mobilenet_v3_large_quantized/model.py b/qai_hub_models/models/mobilenet_v3_large_quantized/model.py
index 55b92db4..c26dbf43 100644
--- a/qai_hub_models/models/mobilenet_v3_large_quantized/model.py
+++ b/qai_hub_models/models/mobilenet_v3_large_quantized/model.py
@@ -20,7 +20,6 @@
from qai_hub_models.models.mobilenet_v3_large.model import MobileNetV3Large
from qai_hub_models.utils.aimet.config_loader import get_default_aimet_config
from qai_hub_models.utils.asset_loaders import CachedWebModelAsset
-from qai_hub_models.utils.base_model import SourceModelFormat, TargetRuntime
MODEL_ID = __name__.split(".")[-2]
MODEL_ASSET_VERSION = 1
@@ -43,11 +42,6 @@ def __init__(
sim_model,
)
- def preferred_hub_source_model_format(
- self, target_runtime: TargetRuntime
- ) -> SourceModelFormat:
- return SourceModelFormat.ONNX
-
@classmethod
def from_pretrained(
cls,
diff --git a/qai_hub_models/models/mobilenet_v3_small/export.py b/qai_hub_models/models/mobilenet_v3_small/export.py
index 92037cb9..7bde62c9 100644
--- a/qai_hub_models/models/mobilenet_v3_small/export.py
+++ b/qai_hub_models/models/mobilenet_v3_small/export.py
@@ -189,7 +189,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model, supports_qnn=False)
+ parser = export_parser(model_cls=Model, supports_qnn=False, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/openai_clip/export.py b/qai_hub_models/models/openai_clip/export.py
index 780dff90..3fa8dac3 100644
--- a/qai_hub_models/models/openai_clip/export.py
+++ b/qai_hub_models/models/openai_clip/export.py
@@ -27,7 +27,6 @@
from qai_hub_models.utils.qai_hub_helpers import (
can_access_qualcomm_ai_hub,
export_without_hub_access,
- transpose_channel_first_to_last,
)
ALL_COMPONENTS = ["CLIPTextEncoder", "CLIPImageEncoder"]
@@ -127,7 +126,7 @@ def export_model(
# 2. Compile the models to an on-device asset
model_compile_options = component.get_hub_compile_options(
- target_runtime, compile_options + " --force_channel_last_input image"
+ target_runtime, compile_options
)
print(f"Optimizing model {component_name} to run on-device")
submitted_compile_job = hub.submit_compile_job(
@@ -170,13 +169,9 @@ def export_model(
component_name
].get_hub_profile_options(target_runtime, profile_options)
sample_inputs = components_dict[component_name].sample_inputs()
- # Convert inputs from channel first to channel last
- hub_inputs = transpose_channel_first_to_last(
- "image", sample_inputs, target_runtime
- )
submitted_inference_job = hub.submit_inference_job(
model=compile_jobs[component_name].get_target_model(),
- inputs=hub_inputs,
+ inputs=sample_inputs,
device=hub.Device(device),
name=f"{model_name}_{component_name}",
options=profile_options_all,
@@ -223,7 +218,9 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model, components=ALL_COMPONENTS)
+ parser = export_parser(
+ model_cls=Model, components=ALL_COMPONENTS, supports_ort=False
+ )
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/openai_clip/perf.yaml b/qai_hub_models/models/openai_clip/perf.yaml
index 0989352c..2a6ddaed 100644
--- a/qai_hub_models/models/openai_clip/perf.yaml
+++ b/qai_hub_models/models/openai_clip/perf.yaml
@@ -17,63 +17,22 @@ aggregated:
- Samsung Galaxy S23
- Samsung Galaxy S23 Ultra
- Samsung Galaxy S23+
- - Samsung Galaxy S24
- - Samsung Galaxy S24 Ultra
- Samsung Galaxy Tab S8
- Xiaomi 12
- Xiaomi 12 Pro
supported_chipsets:
- Snapdragon® 8 Gen 1
- Snapdragon® 8 Gen 2
- - Snapdragon® 8 Gen 3
- Snapdragon® 888
models:
- name: CLIPTextEncoder
performance_metrics:
- torchscript_onnx_tflite:
- inference_time: 15516.0
- throughput: 64.44960041247744
+ inference_time: 15528.0
+ throughput: 64.39979392065945
estimated_peak_memory_range:
- min: 49152
- max: 3267008
- primary_compute_unit: NPU
- precision: fp16
- layer_info:
- layers_on_npu: 574
- layers_on_gpu: 0
- layers_on_cpu: 2
- total_layers: 576
- job_id: jz5worjp1
- job_status: Passed
- reference_device_info:
- name: Samsung Galaxy S23
- os: '13'
- form_factor: Phone
- os_name: Android
- manufacturer: Samsung
- chipset: Snapdragon® 8 Gen 2
- timestamp: '2024-03-14T23:47:17.422656Z'
- torchscript_onnx_qnn:
- inference_time: 15586.0
- throughput: 64.16014371872193
- estimated_peak_memory_range:
- min: 45056
- max: 2975720
- primary_compute_unit: NPU
- precision: fp16
- layer_info:
- layers_on_npu: 574
- layers_on_gpu: 0
- layers_on_cpu: 2
- total_layers: 576
- job_id: jz57z1rp3
- job_status: Passed
- - torchscript_onnx_tflite:
- inference_time: 11115.0
- throughput: 89.9685110211426
- estimated_peak_memory_range:
- min: 16384
- max: 204316144
+ min: 40960
+ max: 3106072
primary_compute_unit: NPU
precision: fp16
layer_info:
@@ -81,62 +40,39 @@ models:
layers_on_gpu: 0
layers_on_cpu: 2
total_layers: 576
- job_id: jnp10ml5q
+ job_id: j2p0m2veg
job_status: Passed
- reference_device_info:
- name: Samsung Galaxy S24
- os: '14'
- form_factor: Phone
- os_name: Android
- manufacturer: Samsung
- chipset: Snapdragon® 8 Gen 3
- timestamp: '2024-03-14T23:49:22.781059Z'
torchscript_onnx_qnn:
- inference_time: 11246.0
- throughput: 88.92050506846878
+ inference_time: 8149.0
+ throughput: 122.71444348999877
estimated_peak_memory_range:
min: 40960
- max: 205502128
+ max: 23728064
primary_compute_unit: NPU
precision: fp16
layer_info:
- layers_on_npu: 574
- layers_on_gpu: 0
- layers_on_cpu: 2
- total_layers: 576
- job_id: j0pxv89g7
- job_status: Passed
-- name: CLIPImageEncoder
- performance_metrics:
- - torchscript_onnx_tflite:
- inference_time: 128196.0
- throughput: 7.800555399544447
- estimated_peak_memory_range:
- min: 143360
- max: 3847064
- primary_compute_unit: NPU
- precision: fp16
- layer_info:
- layers_on_npu: 575
+ layers_on_npu: 377
layers_on_gpu: 0
layers_on_cpu: 0
- total_layers: 575
- job_id: jmg9vqv57
+ total_layers: 377
+ job_id: jogk2q9og
job_status: Passed
reference_device_info:
- name: Samsung Galaxy S23
+ name: Samsung Galaxy S23 Ultra
os: '13'
form_factor: Phone
os_name: Android
manufacturer: Samsung
chipset: Snapdragon® 8 Gen 2
- timestamp: '2024-03-14T23:59:18.769511Z'
- torchscript_onnx_qnn:
- inference_time: 127795.0
- throughput: 7.825032278258147
+ timestamp: '2024-02-21T16:25:08.294036Z'
+- name: CLIPImageEncoder
+ performance_metrics:
+ - torchscript_onnx_tflite:
+ inference_time: 127729.0
+ throughput: 7.829075621041424
estimated_peak_memory_range:
- min: 180224
- max: 4074336
+ min: 159744
+ max: 3867320
primary_compute_unit: NPU
precision: fp16
layer_info:
@@ -144,43 +80,28 @@ models:
layers_on_gpu: 0
layers_on_cpu: 0
total_layers: 575
- job_id: jqp4q6lgo
+ job_id: j1p8em48p
job_status: Passed
- - torchscript_onnx_tflite:
- inference_time: 98556.0
- throughput: 10.14651568651325
+ torchscript_onnx_qnn:
+ inference_time: 50903.0
+ throughput: 19.645207551617784
estimated_peak_memory_range:
- min: 163840
- max: 781391856
+ min: 86016
+ max: 59741752
primary_compute_unit: NPU
precision: fp16
layer_info:
- layers_on_npu: 575
+ layers_on_npu: 370
layers_on_gpu: 0
layers_on_cpu: 0
- total_layers: 575
- job_id: jvgdwml5j
+ total_layers: 370
+ job_id: jn5qlrmmp
job_status: Passed
reference_device_info:
- name: Samsung Galaxy S24
- os: '14'
+ name: Samsung Galaxy S23 Ultra
+ os: '13'
form_factor: Phone
os_name: Android
manufacturer: Samsung
- chipset: Snapdragon® 8 Gen 3
- timestamp: '2024-03-15T00:01:23.890974Z'
- torchscript_onnx_qnn:
- inference_time: 97281.0
- throughput: 10.279499593959766
- estimated_peak_memory_range:
- min: 237568
- max: 783870384
- primary_compute_unit: NPU
- precision: fp16
- layer_info:
- layers_on_npu: 575
- layers_on_gpu: 0
- layers_on_cpu: 0
- total_layers: 575
- job_id: jo5mr1qgk
- job_status: Passed
+ chipset: Snapdragon® 8 Gen 2
+ timestamp: '2024-02-21T16:30:00.084732Z'
diff --git a/qai_hub_models/models/openpose/export.py b/qai_hub_models/models/openpose/export.py
index 98bbb750..c432a301 100644
--- a/qai_hub_models/models/openpose/export.py
+++ b/qai_hub_models/models/openpose/export.py
@@ -197,7 +197,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/protocols.py b/qai_hub_models/models/protocols.py
index e263482e..86b53dc9 100644
--- a/qai_hub_models/models/protocols.py
+++ b/qai_hub_models/models/protocols.py
@@ -19,12 +19,17 @@
from __future__ import annotations
from abc import abstractmethod
-from typing import Protocol, Type, TypeVar, runtime_checkable
+from pathlib import Path
+from typing import Any, Protocol, Type, TypeVar, runtime_checkable
-from qai_hub.client import DatasetEntries
+from qai_hub.client import DatasetEntries, SourceModel
from qai_hub_models.evaluators.base_evaluators import BaseEvaluator, _DataLoader
-from qai_hub_models.models.common import SampleInputsType, TargetRuntime
+from qai_hub_models.models.common import (
+ SampleInputsType,
+ SourceModelFormat,
+ TargetRuntime,
+)
from qai_hub_models.utils.input_spec import InputSpec
FromPretrainedTypeVar = TypeVar("FromPretrainedTypeVar", bound="FromPretrainedProtocol")
@@ -175,6 +180,51 @@ def from_pretrained(
...
+class PretrainedHubModelProtocol(HubModelProtocol, FromPretrainedProtocol):
+ """
+ All pretrained AI Hub Models must, at minimum, implement this interface.
+ """
+
+ @abstractmethod
+ def convert_to_torchscript(
+ self, input_spec: InputSpec | None = None, check_trace: bool = True
+ ) -> Any:
+ """
+ Converts the torch module to a torchscript trace, which
+ is the format expected by qai hub.
+
+ This is a default implementation that may be overriden by a subclass.
+ """
+ ...
+
+ def convert_to_hub_source_model(
+ self,
+ target_runtime: TargetRuntime,
+ output_path: str | Path,
+ input_spec: InputSpec | None = None,
+ check_trace: bool = True,
+ ) -> SourceModel:
+ ...
+
+ def get_hub_compile_options(
+ self,
+ target_runtime: TargetRuntime,
+ other_compile_options: str = "",
+ ) -> str:
+ """
+ AI Hub compile options recommended for the model.
+ """
+ ...
+
+ def preferred_hub_source_model_format(
+ self, target_runtime: TargetRuntime
+ ) -> SourceModelFormat:
+ """
+ Source model format preferred for conversion on AI Hub.
+ """
+ ...
+
+
class FromPrecompiledProtocol(Protocol):
"""
Models follow this protocol if they can be initiated from a precompiled torch model.
diff --git a/qai_hub_models/models/quicksrnetlarge/export.py b/qai_hub_models/models/quicksrnetlarge/export.py
index 9e1976a4..ed3efe30 100644
--- a/qai_hub_models/models/quicksrnetlarge/export.py
+++ b/qai_hub_models/models/quicksrnetlarge/export.py
@@ -197,7 +197,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/quicksrnetlarge_quantized/export.py b/qai_hub_models/models/quicksrnetlarge_quantized/export.py
index ea5568eb..27776bd4 100644
--- a/qai_hub_models/models/quicksrnetlarge_quantized/export.py
+++ b/qai_hub_models/models/quicksrnetlarge_quantized/export.py
@@ -206,7 +206,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model, supports_qnn=False)
+ parser = export_parser(model_cls=Model, supports_qnn=False, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/quicksrnetlarge_quantized/model.py b/qai_hub_models/models/quicksrnetlarge_quantized/model.py
index 9bca792d..e9185b68 100644
--- a/qai_hub_models/models/quicksrnetlarge_quantized/model.py
+++ b/qai_hub_models/models/quicksrnetlarge_quantized/model.py
@@ -16,6 +16,7 @@
from aimet_torch.cross_layer_equalization import equalize_model
from aimet_torch.quantsim import QuantizationSimModel, load_encodings_to_sim
+from qai_hub_models.models.common import SourceModelFormat, TargetRuntime
from qai_hub_models.models.quicksrnetlarge.model import QuickSRNetLarge
from qai_hub_models.utils.aimet.config_loader import get_default_aimet_config_legacy_v2
from qai_hub_models.utils.asset_loaders import CachedWebModelAsset
@@ -97,3 +98,11 @@ def from_pretrained(
sim.model.eval()
return cls(sim)
+
+ def preferred_hub_source_model_format(
+ self, target_runtime: TargetRuntime
+ ) -> SourceModelFormat:
+ if target_runtime == TargetRuntime.QNN:
+ return SourceModelFormat.ONNX
+ else:
+ return SourceModelFormat.TORCHSCRIPT
diff --git a/qai_hub_models/models/quicksrnetmedium/export.py b/qai_hub_models/models/quicksrnetmedium/export.py
index ad1016dc..d98eed32 100644
--- a/qai_hub_models/models/quicksrnetmedium/export.py
+++ b/qai_hub_models/models/quicksrnetmedium/export.py
@@ -197,7 +197,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/quicksrnetmedium_quantized/export.py b/qai_hub_models/models/quicksrnetmedium_quantized/export.py
index da516367..bc1ebf42 100644
--- a/qai_hub_models/models/quicksrnetmedium_quantized/export.py
+++ b/qai_hub_models/models/quicksrnetmedium_quantized/export.py
@@ -206,7 +206,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model, supports_qnn=False)
+ parser = export_parser(model_cls=Model, supports_qnn=False, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/quicksrnetmedium_quantized/model.py b/qai_hub_models/models/quicksrnetmedium_quantized/model.py
index 3b6cb7b6..939d8e67 100644
--- a/qai_hub_models/models/quicksrnetmedium_quantized/model.py
+++ b/qai_hub_models/models/quicksrnetmedium_quantized/model.py
@@ -16,6 +16,7 @@
from aimet_torch.cross_layer_equalization import equalize_model
from aimet_torch.quantsim import QuantizationSimModel, load_encodings_to_sim
+from qai_hub_models.models.common import SourceModelFormat, TargetRuntime
from qai_hub_models.models.quicksrnetmedium.model import QuickSRNetMedium
from qai_hub_models.utils.aimet.config_loader import get_default_aimet_config_legacy_v2
from qai_hub_models.utils.asset_loaders import CachedWebModelAsset
@@ -96,3 +97,11 @@ def from_pretrained(
sim.model.eval()
return cls(sim)
+
+ def preferred_hub_source_model_format(
+ self, target_runtime: TargetRuntime
+ ) -> SourceModelFormat:
+ if target_runtime == TargetRuntime.QNN:
+ return SourceModelFormat.ONNX
+ else:
+ return SourceModelFormat.TORCHSCRIPT
diff --git a/qai_hub_models/models/quicksrnetsmall/export.py b/qai_hub_models/models/quicksrnetsmall/export.py
index f4ecbcca..ec138a32 100644
--- a/qai_hub_models/models/quicksrnetsmall/export.py
+++ b/qai_hub_models/models/quicksrnetsmall/export.py
@@ -197,7 +197,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/quicksrnetsmall_quantized/export.py b/qai_hub_models/models/quicksrnetsmall_quantized/export.py
index 4944af8c..d5ad9ec4 100644
--- a/qai_hub_models/models/quicksrnetsmall_quantized/export.py
+++ b/qai_hub_models/models/quicksrnetsmall_quantized/export.py
@@ -206,7 +206,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model, supports_qnn=False)
+ parser = export_parser(model_cls=Model, supports_qnn=False, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/quicksrnetsmall_quantized/model.py b/qai_hub_models/models/quicksrnetsmall_quantized/model.py
index 5ba7fb5e..9102f5f9 100644
--- a/qai_hub_models/models/quicksrnetsmall_quantized/model.py
+++ b/qai_hub_models/models/quicksrnetsmall_quantized/model.py
@@ -16,6 +16,7 @@
from aimet_torch.cross_layer_equalization import equalize_model
from aimet_torch.quantsim import QuantizationSimModel, load_encodings_to_sim
+from qai_hub_models.models.common import SourceModelFormat, TargetRuntime
from qai_hub_models.models.quicksrnetsmall.model import QuickSRNetSmall
from qai_hub_models.utils.aimet.config_loader import get_default_aimet_config_legacy_v2
from qai_hub_models.utils.asset_loaders import CachedWebModelAsset
@@ -95,3 +96,11 @@ def from_pretrained(
sim.model.eval()
return cls(sim)
+
+ def preferred_hub_source_model_format(
+ self, target_runtime: TargetRuntime
+ ) -> SourceModelFormat:
+ if target_runtime == TargetRuntime.QNN:
+ return SourceModelFormat.ONNX
+ else:
+ return SourceModelFormat.TORCHSCRIPT
diff --git a/qai_hub_models/models/real_esrgan_general_x4v3/export.py b/qai_hub_models/models/real_esrgan_general_x4v3/export.py
index 7f5ce2b6..b91d81b8 100644
--- a/qai_hub_models/models/real_esrgan_general_x4v3/export.py
+++ b/qai_hub_models/models/real_esrgan_general_x4v3/export.py
@@ -197,7 +197,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/real_esrgan_x4plus/export.py b/qai_hub_models/models/real_esrgan_x4plus/export.py
index 73ab228c..a53fe803 100644
--- a/qai_hub_models/models/real_esrgan_x4plus/export.py
+++ b/qai_hub_models/models/real_esrgan_x4plus/export.py
@@ -184,7 +184,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/regnet/export.py b/qai_hub_models/models/regnet/export.py
index 2d4e54ce..911f9a7b 100644
--- a/qai_hub_models/models/regnet/export.py
+++ b/qai_hub_models/models/regnet/export.py
@@ -189,7 +189,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/resnet101/export.py b/qai_hub_models/models/resnet101/export.py
index 92c2ea4b..44b0bf51 100644
--- a/qai_hub_models/models/resnet101/export.py
+++ b/qai_hub_models/models/resnet101/export.py
@@ -189,7 +189,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/resnet101_quantized/export.py b/qai_hub_models/models/resnet101_quantized/export.py
index 774ea807..2e1658f6 100644
--- a/qai_hub_models/models/resnet101_quantized/export.py
+++ b/qai_hub_models/models/resnet101_quantized/export.py
@@ -200,7 +200,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/resnet101_quantized/model.py b/qai_hub_models/models/resnet101_quantized/model.py
index bcfb3730..c0a803ef 100644
--- a/qai_hub_models/models/resnet101_quantized/model.py
+++ b/qai_hub_models/models/resnet101_quantized/model.py
@@ -23,7 +23,6 @@
from qai_hub_models.models.resnet101.model import ResNet101
from qai_hub_models.utils.aimet.config_loader import get_default_aimet_config
from qai_hub_models.utils.asset_loaders import CachedWebModelAsset
-from qai_hub_models.utils.base_model import SourceModelFormat, TargetRuntime
MODEL_ID = __name__.split(".")[-2]
MODEL_ASSET_VERSION = 4
@@ -49,11 +48,6 @@ def __init__(
sim_model,
)
- def preferred_hub_source_model_format(
- self, target_runtime: TargetRuntime
- ) -> SourceModelFormat:
- return SourceModelFormat.ONNX
-
@classmethod
def from_pretrained(
cls,
@@ -91,11 +85,3 @@ def from_pretrained(
sim.model.eval()
return cls(sim)
-
- def get_hub_compile_options(
- self, target_runtime: TargetRuntime, other_compile_options: str = ""
- ) -> str:
- compile_options = super().get_hub_compile_options(
- target_runtime, other_compile_options
- )
- return compile_options + " --quantize_full_type int8 --quantize_io"
diff --git a/qai_hub_models/models/resnet18/export.py b/qai_hub_models/models/resnet18/export.py
index 6dc6e2b2..05728934 100644
--- a/qai_hub_models/models/resnet18/export.py
+++ b/qai_hub_models/models/resnet18/export.py
@@ -189,7 +189,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/resnet18_quantized/export.py b/qai_hub_models/models/resnet18_quantized/export.py
index d8d39572..f0c0c9ab 100644
--- a/qai_hub_models/models/resnet18_quantized/export.py
+++ b/qai_hub_models/models/resnet18_quantized/export.py
@@ -200,7 +200,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/resnet18_quantized/model.py b/qai_hub_models/models/resnet18_quantized/model.py
index e000d973..4a147fcb 100644
--- a/qai_hub_models/models/resnet18_quantized/model.py
+++ b/qai_hub_models/models/resnet18_quantized/model.py
@@ -20,7 +20,6 @@
from qai_hub_models.models.resnet18.model import ResNet18
from qai_hub_models.utils.aimet.config_loader import get_default_aimet_config
from qai_hub_models.utils.asset_loaders import CachedWebModelAsset
-from qai_hub_models.utils.base_model import SourceModelFormat, TargetRuntime
MODEL_ID = __name__.split(".")[-2]
MODEL_ASSET_VERSION = 7
@@ -43,11 +42,6 @@ def __init__(
resnet18_model,
)
- def preferred_hub_source_model_format(
- self, target_runtime: TargetRuntime
- ) -> SourceModelFormat:
- return SourceModelFormat.ONNX
-
@classmethod
def from_pretrained(
cls,
@@ -83,11 +77,3 @@ def from_pretrained(
sim.model.eval()
return cls(sim)
-
- def get_hub_compile_options(
- self, target_runtime: TargetRuntime, other_compile_options: str = ""
- ) -> str:
- compile_options = super().get_hub_compile_options(
- target_runtime, other_compile_options
- )
- return compile_options + " --quantize_full_type int8 --quantize_io"
diff --git a/qai_hub_models/models/resnet50/export.py b/qai_hub_models/models/resnet50/export.py
index e7835d92..ec590470 100644
--- a/qai_hub_models/models/resnet50/export.py
+++ b/qai_hub_models/models/resnet50/export.py
@@ -189,7 +189,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/resnext101/export.py b/qai_hub_models/models/resnext101/export.py
index e1b1fcd6..1d43fc58 100644
--- a/qai_hub_models/models/resnext101/export.py
+++ b/qai_hub_models/models/resnext101/export.py
@@ -189,7 +189,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/resnext101_quantized/export.py b/qai_hub_models/models/resnext101_quantized/export.py
index 865e3205..d59544be 100644
--- a/qai_hub_models/models/resnext101_quantized/export.py
+++ b/qai_hub_models/models/resnext101_quantized/export.py
@@ -200,7 +200,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model, supports_qnn=False)
+ parser = export_parser(model_cls=Model, supports_qnn=False, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/resnext101_quantized/model.py b/qai_hub_models/models/resnext101_quantized/model.py
index 98ea0d76..ce13db7a 100644
--- a/qai_hub_models/models/resnext101_quantized/model.py
+++ b/qai_hub_models/models/resnext101_quantized/model.py
@@ -20,7 +20,6 @@
from qai_hub_models.models.resnext101.model import ResNeXt101
from qai_hub_models.utils.aimet.config_loader import get_default_aimet_config
from qai_hub_models.utils.asset_loaders import CachedWebModelAsset
-from qai_hub_models.utils.base_model import SourceModelFormat, TargetRuntime
MODEL_ID = __name__.split(".")[-2]
MODEL_ASSET_VERSION = 4
@@ -43,11 +42,6 @@ def __init__(
sim_model,
)
- def preferred_hub_source_model_format(
- self, target_runtime: TargetRuntime
- ) -> SourceModelFormat:
- return SourceModelFormat.ONNX
-
@classmethod
def from_pretrained(
cls,
@@ -83,11 +77,3 @@ def from_pretrained(
sim.model.eval()
return cls(sim)
-
- def get_hub_compile_options(
- self, target_runtime: TargetRuntime, other_compile_options: str = ""
- ) -> str:
- compile_options = super().get_hub_compile_options(
- target_runtime, other_compile_options
- )
- return compile_options + " --quantize_full_type int8 --quantize_io"
diff --git a/qai_hub_models/models/resnext50/export.py b/qai_hub_models/models/resnext50/export.py
index 7baf5fa2..3f66523c 100644
--- a/qai_hub_models/models/resnext50/export.py
+++ b/qai_hub_models/models/resnext50/export.py
@@ -189,7 +189,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/resnext50_quantized/export.py b/qai_hub_models/models/resnext50_quantized/export.py
index 7f4a6bcc..5977503b 100644
--- a/qai_hub_models/models/resnext50_quantized/export.py
+++ b/qai_hub_models/models/resnext50_quantized/export.py
@@ -200,7 +200,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model, supports_qnn=False)
+ parser = export_parser(model_cls=Model, supports_qnn=False, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/resnext50_quantized/model.py b/qai_hub_models/models/resnext50_quantized/model.py
index dca50076..1e1a4930 100644
--- a/qai_hub_models/models/resnext50_quantized/model.py
+++ b/qai_hub_models/models/resnext50_quantized/model.py
@@ -20,7 +20,6 @@
from qai_hub_models.models.resnext50.model import ResNeXt50
from qai_hub_models.utils.aimet.config_loader import get_default_aimet_config
from qai_hub_models.utils.asset_loaders import CachedWebModelAsset
-from qai_hub_models.utils.base_model import SourceModelFormat, TargetRuntime
MODEL_ID = __name__.split(".")[-2]
MODEL_ASSET_VERSION = 1
@@ -43,11 +42,6 @@ def __init__(
sim_model,
)
- def preferred_hub_source_model_format(
- self, target_runtime: TargetRuntime
- ) -> SourceModelFormat:
- return SourceModelFormat.ONNX
-
@classmethod
def from_pretrained(
cls,
@@ -83,11 +77,3 @@ def from_pretrained(
sim.model.eval()
return cls(sim)
-
- def get_hub_compile_options(
- self, target_runtime: TargetRuntime, other_compile_options: str = ""
- ) -> str:
- compile_options = super().get_hub_compile_options(
- target_runtime, other_compile_options
- )
- return compile_options + " --quantize_full_type int8 --quantize_io"
diff --git a/qai_hub_models/models/sam/export.py b/qai_hub_models/models/sam/export.py
index a5ed59d6..9fcbb110 100644
--- a/qai_hub_models/models/sam/export.py
+++ b/qai_hub_models/models/sam/export.py
@@ -230,7 +230,10 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
parser = export_parser(
- model_cls=Model, components=ALL_COMPONENTS, supports_qnn=False
+ model_cls=Model,
+ components=ALL_COMPONENTS,
+ supports_qnn=False,
+ supports_ort=False,
)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/sesr_m5/export.py b/qai_hub_models/models/sesr_m5/export.py
index 51517850..ed685d85 100644
--- a/qai_hub_models/models/sesr_m5/export.py
+++ b/qai_hub_models/models/sesr_m5/export.py
@@ -193,7 +193,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/sesr_m5_quantized/export.py b/qai_hub_models/models/sesr_m5_quantized/export.py
index 180d06e3..5c6e95d8 100644
--- a/qai_hub_models/models/sesr_m5_quantized/export.py
+++ b/qai_hub_models/models/sesr_m5_quantized/export.py
@@ -194,7 +194,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model, supports_qnn=False)
+ parser = export_parser(model_cls=Model, supports_qnn=False, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/sesr_m5_quantized/model.py b/qai_hub_models/models/sesr_m5_quantized/model.py
index 7e2bac70..8782ceaf 100644
--- a/qai_hub_models/models/sesr_m5_quantized/model.py
+++ b/qai_hub_models/models/sesr_m5_quantized/model.py
@@ -17,6 +17,7 @@
from aimet_torch.quantsim import QuantizationSimModel, load_encodings_to_sim
from qai_hub_models.models._shared.sesr.common import _load_sesr_source_model
+from qai_hub_models.models.common import SourceModelFormat, TargetRuntime
from qai_hub_models.models.sesr_m5.model import (
NUM_CHANNELS,
NUM_LBLOCKS,
@@ -97,3 +98,11 @@ def from_pretrained(
sim.model.eval()
return cls(sim)
+
+ def preferred_hub_source_model_format(
+ self, target_runtime: TargetRuntime
+ ) -> SourceModelFormat:
+ if target_runtime == TargetRuntime.QNN:
+ return SourceModelFormat.ONNX
+ else:
+ return SourceModelFormat.TORCHSCRIPT
diff --git a/qai_hub_models/models/shufflenet_v2/export.py b/qai_hub_models/models/shufflenet_v2/export.py
index cdb14643..1d84398e 100644
--- a/qai_hub_models/models/shufflenet_v2/export.py
+++ b/qai_hub_models/models/shufflenet_v2/export.py
@@ -189,7 +189,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/shufflenet_v2_quantized/export.py b/qai_hub_models/models/shufflenet_v2_quantized/export.py
index d4cd288c..bbc810cc 100644
--- a/qai_hub_models/models/shufflenet_v2_quantized/export.py
+++ b/qai_hub_models/models/shufflenet_v2_quantized/export.py
@@ -200,7 +200,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/shufflenet_v2_quantized/model.py b/qai_hub_models/models/shufflenet_v2_quantized/model.py
index ba13c2c4..783cc2b5 100644
--- a/qai_hub_models/models/shufflenet_v2_quantized/model.py
+++ b/qai_hub_models/models/shufflenet_v2_quantized/model.py
@@ -23,7 +23,6 @@
from qai_hub_models.models.shufflenet_v2.model import ShufflenetV2
from qai_hub_models.utils.aimet.config_loader import get_default_aimet_config
from qai_hub_models.utils.asset_loaders import CachedWebModelAsset
-from qai_hub_models.utils.base_model import SourceModelFormat, TargetRuntime
from qai_hub_models.utils.quantization_aimet import (
convert_all_depthwise_to_per_tensor,
tie_aimet_observer_groups,
@@ -53,11 +52,6 @@ def __init__(
sim_model,
)
- def preferred_hub_source_model_format(
- self, target_runtime: TargetRuntime
- ) -> SourceModelFormat:
- return SourceModelFormat.ONNX
-
@classmethod
def from_pretrained(
cls,
@@ -98,14 +92,6 @@ def from_pretrained(
sim.model.eval()
return cls(sim)
- def get_hub_compile_options(
- self, target_runtime: TargetRuntime, other_compile_options: str = ""
- ) -> str:
- compile_options = super().get_hub_compile_options(
- target_runtime, other_compile_options
- )
- return compile_options + " --quantize_full_type int8 --quantize_io"
-
@classmethod
def _tie_pre_concat_quantizers(cls, sim: QuantizationSimModel):
"""
diff --git a/qai_hub_models/models/sinet/export.py b/qai_hub_models/models/sinet/export.py
index 780ac793..81eabff3 100644
--- a/qai_hub_models/models/sinet/export.py
+++ b/qai_hub_models/models/sinet/export.py
@@ -197,7 +197,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/squeezenet1_1/export.py b/qai_hub_models/models/squeezenet1_1/export.py
index 1586bb36..917ebdfa 100644
--- a/qai_hub_models/models/squeezenet1_1/export.py
+++ b/qai_hub_models/models/squeezenet1_1/export.py
@@ -189,7 +189,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/squeezenet1_1_quantized/export.py b/qai_hub_models/models/squeezenet1_1_quantized/export.py
index e4256985..b97b3e7f 100644
--- a/qai_hub_models/models/squeezenet1_1_quantized/export.py
+++ b/qai_hub_models/models/squeezenet1_1_quantized/export.py
@@ -198,7 +198,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/squeezenet1_1_quantized/model.py b/qai_hub_models/models/squeezenet1_1_quantized/model.py
index 67a3f532..70ea93b9 100644
--- a/qai_hub_models/models/squeezenet1_1_quantized/model.py
+++ b/qai_hub_models/models/squeezenet1_1_quantized/model.py
@@ -20,7 +20,6 @@
from qai_hub_models.models.squeezenet1_1.model import SqueezeNet
from qai_hub_models.utils.aimet.config_loader import get_default_aimet_config
from qai_hub_models.utils.asset_loaders import CachedWebModelAsset
-from qai_hub_models.utils.base_model import SourceModelFormat, TargetRuntime
MODEL_ID = __name__.split(".")[-2]
MODEL_ASSET_VERSION = 2
@@ -43,11 +42,6 @@ def __init__(
sim_model,
)
- def preferred_hub_source_model_format(
- self, target_runtime: TargetRuntime
- ) -> SourceModelFormat:
- return SourceModelFormat.ONNX
-
@classmethod
def from_pretrained(
cls,
@@ -83,11 +77,3 @@ def from_pretrained(
sim.model.eval()
return cls(sim)
-
- def get_hub_compile_options(
- self, target_runtime: TargetRuntime, other_compile_options: str = ""
- ) -> str:
- compile_options = super().get_hub_compile_options(
- target_runtime, other_compile_options
- )
- return compile_options + " --quantize_full_type int8 --quantize_io"
diff --git a/qai_hub_models/models/stylegan2/export.py b/qai_hub_models/models/stylegan2/export.py
index 520ed574..a430737a 100644
--- a/qai_hub_models/models/stylegan2/export.py
+++ b/qai_hub_models/models/stylegan2/export.py
@@ -185,7 +185,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model, supports_qnn=False)
+ parser = export_parser(model_cls=Model, supports_qnn=False, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/swin_base/export.py b/qai_hub_models/models/swin_base/export.py
index b9e657bd..7b7b0968 100644
--- a/qai_hub_models/models/swin_base/export.py
+++ b/qai_hub_models/models/swin_base/export.py
@@ -189,7 +189,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model, supports_qnn=False)
+ parser = export_parser(model_cls=Model, supports_qnn=False, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/swin_small/export.py b/qai_hub_models/models/swin_small/export.py
index ed29fb83..75769cb2 100644
--- a/qai_hub_models/models/swin_small/export.py
+++ b/qai_hub_models/models/swin_small/export.py
@@ -189,7 +189,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model, supports_qnn=False)
+ parser = export_parser(model_cls=Model, supports_qnn=False, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/swin_tiny/export.py b/qai_hub_models/models/swin_tiny/export.py
index 3f43d39b..12d81836 100644
--- a/qai_hub_models/models/swin_tiny/export.py
+++ b/qai_hub_models/models/swin_tiny/export.py
@@ -189,7 +189,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model, supports_qnn=False)
+ parser = export_parser(model_cls=Model, supports_qnn=False, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/trocr/export.py b/qai_hub_models/models/trocr/export.py
index 37af35b9..3e336783 100644
--- a/qai_hub_models/models/trocr/export.py
+++ b/qai_hub_models/models/trocr/export.py
@@ -27,7 +27,6 @@
from qai_hub_models.utils.qai_hub_helpers import (
can_access_qualcomm_ai_hub,
export_without_hub_access,
- transpose_channel_first_to_last,
)
ALL_COMPONENTS = ["TrOCREncoder", "TrOCRDecoder"]
@@ -127,7 +126,7 @@ def export_model(
# 2. Compile the models to an on-device asset
model_compile_options = component.get_hub_compile_options(
- target_runtime, compile_options + " --force_channel_last_input pixel_values"
+ target_runtime, compile_options
)
print(f"Optimizing model {component_name} to run on-device")
submitted_compile_job = hub.submit_compile_job(
@@ -170,13 +169,9 @@ def export_model(
component_name
].get_hub_profile_options(target_runtime, profile_options)
sample_inputs = components_dict[component_name].sample_inputs()
- # Convert inputs from channel first to channel last
- hub_inputs = transpose_channel_first_to_last(
- "pixel_values", sample_inputs, target_runtime
- )
submitted_inference_job = hub.submit_inference_job(
model=compile_jobs[component_name].get_target_model(),
- inputs=hub_inputs,
+ inputs=sample_inputs,
device=hub.Device(device),
name=f"{model_name}_{component_name}",
options=profile_options_all,
@@ -224,7 +219,10 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
parser = export_parser(
- model_cls=Model, components=ALL_COMPONENTS, supports_qnn=False
+ model_cls=Model,
+ components=ALL_COMPONENTS,
+ supports_qnn=False,
+ supports_ort=False,
)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/unet_segmentation/export.py b/qai_hub_models/models/unet_segmentation/export.py
index 11489e1b..acc1677d 100644
--- a/qai_hub_models/models/unet_segmentation/export.py
+++ b/qai_hub_models/models/unet_segmentation/export.py
@@ -197,7 +197,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/vit/export.py b/qai_hub_models/models/vit/export.py
index bc43f4ae..e88a1575 100644
--- a/qai_hub_models/models/vit/export.py
+++ b/qai_hub_models/models/vit/export.py
@@ -191,7 +191,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model, supports_qnn=False)
+ parser = export_parser(model_cls=Model, supports_qnn=False, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/whisper_base_en/export.py b/qai_hub_models/models/whisper_base_en/export.py
index 5eaddcb4..019486f5 100644
--- a/qai_hub_models/models/whisper_base_en/export.py
+++ b/qai_hub_models/models/whisper_base_en/export.py
@@ -219,7 +219,10 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
parser = export_parser(
- model_cls=Model, components=ALL_COMPONENTS, supports_qnn=False
+ model_cls=Model,
+ components=ALL_COMPONENTS,
+ supports_qnn=False,
+ supports_ort=False,
)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/whisper_small_en/export.py b/qai_hub_models/models/whisper_small_en/export.py
index 348716d4..4ab0b290 100644
--- a/qai_hub_models/models/whisper_small_en/export.py
+++ b/qai_hub_models/models/whisper_small_en/export.py
@@ -219,7 +219,10 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
parser = export_parser(
- model_cls=Model, components=ALL_COMPONENTS, supports_qnn=False
+ model_cls=Model,
+ components=ALL_COMPONENTS,
+ supports_qnn=False,
+ supports_ort=False,
)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/whisper_tiny_en/export.py b/qai_hub_models/models/whisper_tiny_en/export.py
index 57b7c76b..ecfa1e54 100644
--- a/qai_hub_models/models/whisper_tiny_en/export.py
+++ b/qai_hub_models/models/whisper_tiny_en/export.py
@@ -219,7 +219,10 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
parser = export_parser(
- model_cls=Model, components=ALL_COMPONENTS, supports_qnn=False
+ model_cls=Model,
+ components=ALL_COMPONENTS,
+ supports_qnn=False,
+ supports_ort=False,
)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/wideresnet50/export.py b/qai_hub_models/models/wideresnet50/export.py
index c86de65a..21d3369d 100644
--- a/qai_hub_models/models/wideresnet50/export.py
+++ b/qai_hub_models/models/wideresnet50/export.py
@@ -189,7 +189,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/wideresnet50_quantized/export.py b/qai_hub_models/models/wideresnet50_quantized/export.py
index ad05928c..a8304a80 100644
--- a/qai_hub_models/models/wideresnet50_quantized/export.py
+++ b/qai_hub_models/models/wideresnet50_quantized/export.py
@@ -198,7 +198,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/wideresnet50_quantized/model.py b/qai_hub_models/models/wideresnet50_quantized/model.py
index 9b1086a0..f1bd3f00 100644
--- a/qai_hub_models/models/wideresnet50_quantized/model.py
+++ b/qai_hub_models/models/wideresnet50_quantized/model.py
@@ -23,7 +23,6 @@
from qai_hub_models.models.wideresnet50.model import WideResNet50
from qai_hub_models.utils.aimet.config_loader import get_default_aimet_config
from qai_hub_models.utils.asset_loaders import CachedWebModelAsset
-from qai_hub_models.utils.base_model import SourceModelFormat, TargetRuntime
MODEL_ID = __name__.split(".")[-2]
MODEL_ASSET_VERSION = 2
@@ -46,11 +45,6 @@ def __init__(
sim_model,
)
- def preferred_hub_source_model_format(
- self, target_runtime: TargetRuntime
- ) -> SourceModelFormat:
- return SourceModelFormat.ONNX
-
@classmethod
def from_pretrained(
cls,
@@ -88,11 +82,3 @@ def from_pretrained(
sim.model.eval()
return cls(sim)
-
- def get_hub_compile_options(
- self, target_runtime: TargetRuntime, other_compile_options: str = ""
- ) -> str:
- compile_options = super().get_hub_compile_options(
- target_runtime, other_compile_options
- )
- return compile_options + " --quantize_full_type int8 --quantize_io"
diff --git a/qai_hub_models/models/xlsr/export.py b/qai_hub_models/models/xlsr/export.py
index c45d8d65..ae395081 100644
--- a/qai_hub_models/models/xlsr/export.py
+++ b/qai_hub_models/models/xlsr/export.py
@@ -193,7 +193,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/xlsr_quantized/export.py b/qai_hub_models/models/xlsr_quantized/export.py
index d8b04a25..e5209fb7 100644
--- a/qai_hub_models/models/xlsr_quantized/export.py
+++ b/qai_hub_models/models/xlsr_quantized/export.py
@@ -202,7 +202,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model, supports_qnn=False)
+ parser = export_parser(model_cls=Model, supports_qnn=False, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/xlsr_quantized/model.py b/qai_hub_models/models/xlsr_quantized/model.py
index 9747be53..cbf2ec5a 100644
--- a/qai_hub_models/models/xlsr_quantized/model.py
+++ b/qai_hub_models/models/xlsr_quantized/model.py
@@ -15,6 +15,7 @@
import torch
from aimet_torch.quantsim import QuantizationSimModel, load_encodings_to_sim
+from qai_hub_models.models.common import SourceModelFormat, TargetRuntime
from qai_hub_models.models.xlsr.model import XLSR, _load_xlsr_source_model
from qai_hub_models.utils.asset_loaders import CachedWebModelAsset
@@ -88,3 +89,11 @@ def from_pretrained(
load_encodings_to_sim(sim, aimet_encodings)
return cls(sim)
+
+ def preferred_hub_source_model_format(
+ self, target_runtime: TargetRuntime
+ ) -> SourceModelFormat:
+ if target_runtime == TargetRuntime.QNN:
+ return SourceModelFormat.ONNX
+ else:
+ return SourceModelFormat.TORCHSCRIPT
diff --git a/qai_hub_models/models/yolov6/export.py b/qai_hub_models/models/yolov6/export.py
index 1fd64724..9b974e35 100644
--- a/qai_hub_models/models/yolov6/export.py
+++ b/qai_hub_models/models/yolov6/export.py
@@ -191,7 +191,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/yolov7/export.py b/qai_hub_models/models/yolov7/export.py
index 4a68cc79..77c632f9 100644
--- a/qai_hub_models/models/yolov7/export.py
+++ b/qai_hub_models/models/yolov7/export.py
@@ -191,7 +191,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model, supports_qnn=False)
+ parser = export_parser(model_cls=Model, supports_qnn=False, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/yolov8_det/export.py b/qai_hub_models/models/yolov8_det/export.py
index 2b207ddb..43880658 100644
--- a/qai_hub_models/models/yolov8_det/export.py
+++ b/qai_hub_models/models/yolov8_det/export.py
@@ -193,7 +193,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model)
+ parser = export_parser(model_cls=Model, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/models/yolov8_seg/export.py b/qai_hub_models/models/yolov8_seg/export.py
index 227c0a81..a39ac75e 100644
--- a/qai_hub_models/models/yolov8_seg/export.py
+++ b/qai_hub_models/models/yolov8_seg/export.py
@@ -193,7 +193,7 @@ def export_model(
def main():
warnings.filterwarnings("ignore")
- parser = export_parser(model_cls=Model, supports_qnn=False)
+ parser = export_parser(model_cls=Model, supports_qnn=False, supports_ort=False)
args = parser.parse_args()
export_model(**vars(args))
diff --git a/qai_hub_models/utils/args.py b/qai_hub_models/utils/args.py
index 8b60a63d..688961f4 100644
--- a/qai_hub_models/utils/args.py
+++ b/qai_hub_models/utils/args.py
@@ -334,6 +334,7 @@ def export_parser(
model_cls: Type[FromPretrainedTypeVar] | Type[FromPrecompiledTypeVar],
components: Optional[List[str]] = None,
supports_qnn=True,
+ supports_ort=True,
exporting_compiled_model=False,
) -> argparse.ArgumentParser:
"""
@@ -348,6 +349,9 @@ def export_parser(
supports_qnn:
Whether QNN export is supported.
Default=True.
+ supports_ort:
+ Whether ORT export is supported.
+ Default=True.
exporting_compiled_model:
True when exporting compiled model.
If set, removing skip_profiling flag from export arguments.
diff --git a/qai_hub_models/utils/base_model.py b/qai_hub_models/utils/base_model.py
index b0763a06..353db64c 100644
--- a/qai_hub_models/utils/base_model.py
+++ b/qai_hub_models/utils/base_model.py
@@ -18,8 +18,8 @@
from qai_hub_models.models.protocols import (
ExecutableModelProtocol,
FromPrecompiledProtocol,
- FromPretrainedProtocol,
HubModelProtocol,
+ PretrainedHubModelProtocol,
)
from qai_hub_models.utils.input_spec import InputSpec, make_torch_inputs
@@ -93,7 +93,7 @@ def get_hub_profile_options(
class BaseModel(
torch.nn.Module,
HubModel,
- FromPretrainedProtocol,
+ PretrainedHubModelProtocol,
ExecutableModelProtocol,
):
"""
@@ -154,6 +154,8 @@ def get_hub_compile_options(
compile_options = ""
if target_runtime == TargetRuntime.QNN:
compile_options = "--target_runtime qnn_lib_aarch64_android"
+ if target_runtime == TargetRuntime.ORT:
+ compile_options = "--target_runtime onnx"
if other_compile_options != "":
return compile_options + " " + other_compile_options
return compile_options
diff --git a/qai_hub_models/utils/config_loaders.py b/qai_hub_models/utils/config_loaders.py
index ab79e346..14b4136c 100644
--- a/qai_hub_models/utils/config_loaders.py
+++ b/qai_hub_models/utils/config_loaders.py
@@ -8,7 +8,7 @@
from dataclasses import dataclass
from enum import Enum
from pathlib import Path
-from typing import Any, Dict, List, Optional, Tuple, Union
+from typing import Any, Dict, List, Optional, Tuple, Type, Union
import requests
from qai_hub.util.session import create_session
@@ -472,6 +472,129 @@ def get_perf_details(
return perf_details
+class QAIHMModelCodeGen:
+ def __init__(
+ self,
+ is_aimet: bool,
+ has_on_target_demo: bool,
+ qnn_export_failure_reason: str,
+ tflite_export_failure_reason: str,
+ has_demo: bool,
+ check_trace: bool,
+ channel_last_input: List[str],
+ channel_last_output: List[str],
+ outputs_to_skip_validation: List[str],
+ export_test_model_kwargs: Dict[str, str],
+ components: Dict[str, Any],
+ default_components: List[str],
+ skip_tests: bool,
+ is_precompiled: bool,
+ no_assets: bool,
+ global_requirements_incompatible: bool,
+ torchscript_opt: List[str],
+ inference_metrics: str,
+ supports_ort: bool,
+ ) -> None:
+ self.is_aimet = is_aimet
+ self.has_on_target_demo = has_on_target_demo
+ self.qnn_export_failure_reason = qnn_export_failure_reason
+ self.tflite_export_failure_reason = tflite_export_failure_reason
+ self.has_demo = has_demo
+ self.check_trace = check_trace
+ self.channel_last_input = channel_last_input
+ self.channel_last_output = channel_last_output
+ self.outputs_to_skip_validation = outputs_to_skip_validation
+ self.export_test_model_kwargs = export_test_model_kwargs
+ self.components = components
+ self.default_components = default_components
+ self.skip_tests = skip_tests
+ self.is_precompiled = is_precompiled
+ self.no_assets = no_assets
+ self.global_requirements_incompatible = global_requirements_incompatible
+ self.torchscript_opt = torchscript_opt
+ self.inference_metrics = inference_metrics
+ self.supports_ort = supports_ort
+
+ def validate(self) -> Tuple[bool, Optional[str]]:
+ """Returns false with a reason if the info spec for this model is not valid."""
+ return True, None
+
+ @classmethod
+ def from_model(cls: Type[QAIHMModelCodeGen], model_id: str) -> QAIHMModelCodeGen:
+ code_gen_path = QAIHM_MODELS_ROOT / model_id / "code-gen.yaml"
+ if not os.path.exists(code_gen_path):
+ raise ValueError(f"{model_id} does not exist")
+ return cls.from_yaml(code_gen_path)
+
+ @classmethod
+ def from_yaml(
+ cls: Type[QAIHMModelCodeGen], code_gen_path: str | Path | None = None
+ ) -> QAIHMModelCodeGen:
+ # Load CFG and params
+ code_gen_config = QAIHMModelCodeGen.load_code_gen_yaml(code_gen_path)
+ return cls(
+ code_gen_config["is_aimet"],
+ code_gen_config["has_on_target_demo"],
+ code_gen_config["qnn_export_failure_reason"],
+ code_gen_config["tflite_export_failure_reason"],
+ code_gen_config["has_demo"],
+ code_gen_config["check_trace"],
+ code_gen_config["channel_last_input"],
+ code_gen_config["channel_last_output"],
+ code_gen_config["outputs_to_skip_validation"],
+ code_gen_config["export_test_model_kwargs"],
+ code_gen_config["components"],
+ code_gen_config["default_components"],
+ code_gen_config["skip_tests"],
+ code_gen_config["is_precompiled"],
+ code_gen_config["no_assets"],
+ code_gen_config["global_requirements_incompatible"],
+ code_gen_config["torchscript_opt"],
+ code_gen_config["inference_metrics"],
+ code_gen_config["supports_ort"],
+ )
+
+ # Schema for code-gen.yaml
+ CODE_GEN_YAML_SCHEMA = Schema(
+ And(
+ {
+ OptionalSchema("has_components", default=""): str,
+ OptionalSchema("is_aimet", default=False): bool,
+ OptionalSchema("has_on_target_demo", default=False): bool,
+ OptionalSchema("qnn_export_failure_reason", default=""): str,
+ OptionalSchema("tflite_export_failure_reason", default=""): str,
+ OptionalSchema("has_demo", default=True): bool,
+ OptionalSchema("check_trace", default=True): bool,
+ OptionalSchema("channel_last_input", default=[]): list,
+ OptionalSchema("channel_last_output", default=[]): list,
+ OptionalSchema("outputs_to_skip_validation", default=[]): list,
+ OptionalSchema("export_test_model_kwargs", default={}): dict,
+ OptionalSchema("components", default={}): dict,
+ OptionalSchema("default_components", default=[]): list,
+ OptionalSchema("skip_tests", default=False): bool,
+ OptionalSchema("is_precompiled", default=False): bool,
+ OptionalSchema("no_assets", default=False): bool,
+ OptionalSchema("global_requirements_incompatible", default=False): bool,
+ OptionalSchema("torchscript_opt", default=[]): list,
+ OptionalSchema("inference_metrics", default="psnr"): str,
+ OptionalSchema("supports_ort", default=False): bool,
+ }
+ )
+ )
+
+ @staticmethod
+ def load_code_gen_yaml(path: str | Path | None = None):
+ if not path or not os.path.exists(path):
+ return QAIHMModelCodeGen.CODE_GEN_YAML_SCHEMA.validate({}) # Default Schema
+ data = load_yaml(path)
+ try:
+ # Validate high level-schema
+ data = QAIHMModelCodeGen.CODE_GEN_YAML_SCHEMA.validate(data)
+ except SchemaError as e:
+ assert 0, f"{e.code} in {path}"
+ return data
+
+
class QAIHMModelInfo:
def __init__(
self,
@@ -494,7 +617,7 @@ def __init__(
form_factors: List[FORM_FACTOR],
has_static_banner: bool,
has_animated_banner: bool,
- code_gen_config: Dict[str, str | bool],
+ code_gen_config: QAIHMModelCodeGen,
license_type: str,
deploy_license_type: str,
dataset: List[str],
@@ -593,9 +716,10 @@ def validate(self) -> Tuple[bool, Optional[str]]:
if not os.path.exists(self.get_package_path() / "info.yaml"):
return False, "All public models must have an info.yaml"
- if self.code_gen_config.get(
- "tflite_export_failure_reason", False
- ) and self.code_gen_config.get("qnn_export_failure_reason", False):
+ if (
+ self.code_gen_config.tflite_export_failure_reason
+ and self.code_gen_config.qnn_export_failure_reason
+ ):
return False, "Public models must support at least one export path"
session = create_session()
@@ -684,20 +808,23 @@ def get_requirements_path(self, root: Path = QAIHM_PACKAGE_ROOT):
def has_model_requirements(self, root: Path = QAIHM_PACKAGE_ROOT):
return os.path.exists(self.get_requirements_path(root))
- @staticmethod
- def from_model(model_id: str):
+ @classmethod
+ def from_model(cls: Type[QAIHMModelInfo], model_id: str) -> QAIHMModelInfo:
schema_path = QAIHM_MODELS_ROOT / model_id / "info.yaml"
code_gen_path = QAIHM_MODELS_ROOT / model_id / "code-gen.yaml"
if not os.path.exists(schema_path):
raise ValueError(f"{model_id} does not exist")
- return QAIHMModelInfo.from_yaml(schema_path, code_gen_path)
-
- @staticmethod
- def from_yaml(info_path: str | Path, code_gen_path: str | Path | None = None):
+ return cls.from_yaml(schema_path, code_gen_path)
+
+ @classmethod
+ def from_yaml(
+ cls: Type[QAIHMModelInfo],
+ info_path: str | Path,
+ code_gen_path: str | Path | None = None,
+ ) -> QAIHMModelInfo:
# Load CFG and params
info_yaml = QAIHMModelInfo.load_info_yaml(info_path)
- code_gen_config = QAIHMModelInfo.load_code_gen_yaml(code_gen_path)
- return QAIHMModelInfo(
+ return cls(
info_yaml["name"],
info_yaml["id"],
MODEL_STATUS.from_string(info_yaml["status"]),
@@ -717,7 +844,7 @@ def from_yaml(info_path: str | Path, code_gen_path: str | Path | None = None):
[FORM_FACTOR.from_string(ff) for ff in info_yaml["form_factors"]],
info_yaml["has_static_banner"],
info_yaml["has_animated_banner"],
- code_gen_config,
+ QAIHMModelCodeGen.from_yaml(code_gen_path),
info_yaml["license_type"],
info_yaml["deploy_license_type"],
info_yaml["dataset"],
@@ -755,33 +882,6 @@ def from_yaml(info_path: str | Path, code_gen_path: str | Path | None = None):
)
)
- # Schema for code-gen.yaml
- CODE_GEN_YAML_SCHEMA = Schema(
- And(
- {
- OptionalSchema("has_components", default=""): str,
- OptionalSchema("is_aimet", default=False): bool,
- OptionalSchema("has_on_target_demo", default=False): bool,
- OptionalSchema("qnn_export_failure_reason", default=""): str,
- OptionalSchema("tflite_export_failure_reason", default=""): str,
- OptionalSchema("has_demo", default=True): bool,
- OptionalSchema("check_trace", default=True): bool,
- OptionalSchema("channel_last_input", default=""): str,
- OptionalSchema("channel_last_output", default=""): str,
- OptionalSchema("outputs_to_skip_validation", default=[]): list,
- OptionalSchema("export_test_model_kwargs", default={}): dict,
- OptionalSchema("components", default={}): dict,
- OptionalSchema("default_components", default=[]): list,
- OptionalSchema("skip_tests", default=False): bool,
- OptionalSchema("is_precompiled", default=False): bool,
- OptionalSchema("no_assets", default=False): bool,
- OptionalSchema("global_requirements_incompatible", default=False): bool,
- OptionalSchema("torchscript_opt", default=[]): list,
- OptionalSchema("inference_metrics", default="psnr"): str,
- }
- )
- )
-
@staticmethod
def load_info_yaml(path: str | Path) -> Dict[str, Any]:
data = load_yaml(path)
@@ -791,15 +891,3 @@ def load_info_yaml(path: str | Path) -> Dict[str, Any]:
except SchemaError as e:
assert 0, f"{e.code} in {path}"
return data
-
- @staticmethod
- def load_code_gen_yaml(path: str | Path | None):
- if not path or not os.path.exists(path):
- return QAIHMModelInfo.CODE_GEN_YAML_SCHEMA.validate({}) # Default Schema
- data = load_yaml(path)
- try:
- # Validate high level-schema
- data = QAIHMModelInfo.CODE_GEN_YAML_SCHEMA.validate(data)
- except SchemaError as e:
- assert 0, f"{e.code} in {path}"
- return data
diff --git a/qai_hub_models/utils/printing.py b/qai_hub_models/utils/printing.py
index 5efd51ec..f47f17e2 100644
--- a/qai_hub_models/utils/printing.py
+++ b/qai_hub_models/utils/printing.py
@@ -75,6 +75,8 @@ def print_profile_metrics_from_job(
runtime = TargetRuntime.TFLITE
elif is_qnn_hub_model(profile_job.model):
runtime = TargetRuntime.QNN
+ elif profile_job.model.model_type == SourceModelType.ORT:
+ runtime = TargetRuntime.ORT
else:
raise NotImplementedError()
diff --git a/qai_hub_models/utils/quantization_aimet.py b/qai_hub_models/utils/quantization_aimet.py
index 56517e76..665b3261 100644
--- a/qai_hub_models/utils/quantization_aimet.py
+++ b/qai_hub_models/utils/quantization_aimet.py
@@ -46,7 +46,7 @@
from qai_hub_models.models.common import SourceModelFormat, TargetRuntime
from qai_hub_models.models.protocols import (
EvalModelProtocol,
- HubModelProtocol,
+ PretrainedHubModelProtocol,
QuantizableModelProtocol,
)
from qai_hub_models.utils.input_spec import InputSpec, make_torch_inputs
@@ -103,7 +103,7 @@ def convert_depthwise_to_per_tensor(op, parent_module, name):
)
-class AIMETQuantizableMixin(HubModelProtocol, QuantizableModelProtocol):
+class AIMETQuantizableMixin(PretrainedHubModelProtocol, QuantizableModelProtocol):
"""
Mixin that allows a model to be quantized & exported to disk using AIMET.
@@ -118,14 +118,6 @@ def __init__(
self.quant_sim = quant_sim
self.needs_onnx_direct_aimet_export = needs_onnx_direct_aimet_export
- def preferred_hub_source_model_format(
- self, target_runtime: TargetRuntime
- ) -> SourceModelFormat:
- if target_runtime == TargetRuntime.QNN:
- return SourceModelFormat.ONNX
- else:
- return SourceModelFormat.TORCHSCRIPT
-
def quantize(
self,
data: _DataLoader,
@@ -320,3 +312,16 @@ def get_calibration_data(
input_spec = self.get_input_spec()
inputs = make_torch_inputs(input_spec)
return {k: v.numpy() for k, v in zip(input_spec.keys(), inputs)}
+
+ def get_hub_compile_options(
+ self, target_runtime: TargetRuntime, other_compile_options: str = ""
+ ) -> str:
+ compile_options = super().get_hub_compile_options(
+ target_runtime, other_compile_options
+ )
+ return compile_options + " --quantize_full_type int8 --quantize_io"
+
+ def preferred_hub_source_model_format(
+ self, target_runtime: TargetRuntime
+ ) -> SourceModelFormat:
+ return SourceModelFormat.ONNX