#13477: Yolov7 ttnn implementation and demo

models/experimental/functional_yolov7/README.md

@@ -0,0 +1,29 @@
+## Yolov7 Model
+
+# Platforms:
+    WH N300, N150
+
+## command to run the model
+Use the following command to run the yolov7 model
+```
+pytest --disable-warnings tests/ttnn/integration_tests/yolov7/test_ttnn_yolov7.py
+```
+
+## Demo
+Batch size : 1
+Use the following command to run the demo
+```
+pytest --disable-warnings models/experimental/functional_yolov7/demo/yolov7_demo.py
+```
+
+## Outputs
+The Demo outputs are saved inside this directory: `models/experimental/functional_yolov7/demo/runs/detect`
+
+
+Torch ops: mul, pow, add, sub (Overall PCC of the model drops when converted to ttnn)
+
+Pending issues:
+[#12786](https://github.com/tenstorrent/tt-metal/issues/12786) - Failing Silu in Yolov7 model
+[#17370](https://github.com/tenstorrent/tt-metal/issues/17370) - Issue with repeat_interleave (Handled in torch in pre-processing of the model)
+[#17583](https://github.com/tenstorrent/tt-metal/issues/17583) - Yolov7 Trace+2cq fails with Out of Memoy issue
+[#17904](https://github.com/tenstorrent/tt-metal/issues/17904) - Detect sub-module fails with low PCC.

models/experimental/functional_yolov7/demo/yolov7_demo.py

@@ -0,0 +1,74 @@
+# SPDX-FileCopyrightText: © 2024 Tenstorrent Inc.
+
+# SPDX-License-Identifier: Apache-2.0
+
+import pytest
+import torch
+import ttnn
+import numpy as np
+from loguru import logger
+import sys
+from models.experimental.functional_yolov7.reference.model import Yolov7_model
+from models.experimental.functional_yolov7.tt.tt_yolov7 import ttnn_yolov7
+from ttnn.model_preprocessing import preprocess_model_parameters
+
+from models.experimental.functional_yolov7.ttnn_yolov7_detect import ttnn_detect, parse_opt
+
+from tests.ttnn.integration_tests.yolov7.test_ttnn_yolov7 import create_custom_preprocessor
+from models.experimental.functional_yolov7.reference.yolov7_utils import download_yolov7_weights
+
+
+@pytest.mark.parametrize("device_params", [{"l1_small_size": 16384}], indirect=True)
+def test_demo(device, reset_seeds):
+    sys.modules["models.yolo"] = sys.modules["models.experimental.functional_yolov7.reference.yolov7_model"]
+    sys.modules["models.common"] = sys.modules["models.experimental.functional_yolov7.reference.yolov7_utils"]
+
+    def load_weights(model, weights_path):
+        ckpt = torch.load(weights_path, map_location="cpu")
+        state_dict = ckpt["model"].float().state_dict()
+        model.load_state_dict(state_dict, strict=False)
+
+    args = [
+        "--weights",
+        "yolov7.pt",
+        "--source",
+        "models/experimental/functional_yolov7/demo/horses.jpg",
+        "--conf-thres",
+        "0.50",
+        "--img-size",
+        "640",
+    ]
+
+    opt = parse_opt(args)
+
+    torch_model = Yolov7_model()
+
+    new_state_dict = {}
+    keys = [name for name, parameter in torch_model.state_dict().items()]
+    ds_state_dict = {k: v for k, v in torch_model.state_dict().items()}
+    values = [parameter for name, parameter in ds_state_dict.items()]
+    for i in range(len(keys)):
+        new_state_dict[keys[i]] = values[i]
+    torch_model.load_state_dict(new_state_dict)
+    torch_model.eval()
+
+    weights_path = "tests/ttnn/integration_tests/yolov7/yolov7.pt"
+    weights_path = download_yolov7_weights(weights_path)
+    load_weights(torch_model, weights_path)
+
+    parameters = preprocess_model_parameters(
+        initialize_model=lambda: torch_model, custom_preprocessor=create_custom_preprocessor(None), device=None
+    )
+
+    grid = [torch.randn(1)] * 3
+    nx_ny = [80, 40, 20]
+    grid_tensors = []
+    for i in range(3):
+        yv, xv = torch.meshgrid([torch.arange(nx_ny[i]), torch.arange(nx_ny[i])])
+        grid_tensors.append(torch.stack((xv, yv), 2).view((1, 1, nx_ny[i], nx_ny[i], 2)).float())
+
+    ttnn_model = ttnn_yolov7(device, parameters, grid_tensors)
+
+    ttnn_output = ttnn_detect(opt, ttnn_model, device)
+
+    logger.info("Yolov7 Demo completed")