fix(client): make model serving of the examples work (#2356)

example/PennFudanPed/pfp/evaluator.py

@@ -108,7 +108,7 @@ def handler(file: str):
     with open(file, "rb") as f:
         data = f.read()
     img = Image(data, mime_type=MIMEType.PNG)
-    _, res = predict_mask_rcnn({"image": img}, 0)
+    _, res = predict_mask_rcnn({"image": img}, {"index": 0})
 
     bbox = res["bbox"]
     _img = PILImage.open(file)

example/cifar10/cifar/evaluator.py

@@ -22,7 +22,6 @@ def __init__(self) -> None:
         self.model = self._load_model(self.device)
 
     def predict(self, data, external):
-        print(f"index: {external['index']}")
         data_tensor = self._pre(data["image"])
         output = self.model(data_tensor)
         return self._post(output)
@@ -86,5 +85,5 @@ def online_eval(self, img: PILImage.Image):
             "ship",
             "truck",
         )
-        _, prob = self.predict(Image(fp=buf.getvalue()))
+        _, prob = self.predict({"image": Image(fp=buf.getvalue())}, {})
         return {classes[i]: p for i, p in enumerate(prob[0])}

example/nmt/nmt/evaluator.py

@@ -95,4 +95,4 @@ def _load_decoder_model(self, device):
         examples=["i m not afraid to die .", "i study mathematics ."],
     )
     def online_eval(self, content: str):
-        return self.ppl(Text(content))
+        return self.ppl({"english": Text(content)})

example/speech_command/sc/evaluator.py

@@ -80,7 +80,7 @@ def get_m5_model():
 @torch.no_grad()
 @evaluation.predict(resources={"nvidia.com/gpu": 1}, replicas=2)
 def predict_speech(data):
-    _audio = io.BytesIO(data.speech.to_bytes())
+    _audio = io.BytesIO(data["speech"].to_bytes())
     waveform, _ = torchaudio.load(_audio)
     waveform = torch.nn.utils.rnn.pad_sequence(
         [waveform.t()], batch_first=True, padding_value=0.0
@@ -112,14 +112,11 @@ def evaluate_speech(ppl_result):
 
 
 @api(
-    [
-        gradio.Audio(type="filepath"),
-        gradio.Audio(source="microphone", type="filepath"),
-    ],
+    gradio.Audio(type="filepath"),
     gradio.Label(),
 )
 def online_eval(file: str):
     with open(file, "rb") as f:
         data = f.read()
-    _, prob = predict_speech(Audio(fp=data))
+    _, prob = predict_speech({"speech": Audio(fp=data)})
     return {ALL_LABELS[i]: p for i, p in enumerate(prob)}

example/text_cls_AG_NEWS/tcan/evaluator.py

@@ -4,7 +4,7 @@
 import gradio
 from torchtext.data.utils import get_tokenizer, ngrams_iterator
 
-from starwhale import Text, PipelineHandler, multi_classification
+from starwhale import PipelineHandler, multi_classification
 from starwhale.api.service import api
 
 from .model import TextClassificationModel
@@ -24,7 +24,6 @@ def __init__(self) -> None:
 
     @torch.no_grad()
     def ppl(self, data: dict, **kw):
-        print(f"index: {kw['external']['index']}")
         ngrams = list(ngrams_iterator(self.tokenizer(data["text"]), 2))
         tensor = torch.tensor(self.vocab(ngrams)).to(self.device)
         output = self.model(tensor, torch.tensor([0]).to(self.device))
@@ -69,5 +68,5 @@ def _load_vocab(self):
         ],
     )
     def online_eval(self, content: str):
-        _, prob = self.ppl(Text(content))
+        _, prob = self.ppl({"text": content})
         return {_LABEL_NAMES[i]: p for i, p in enumerate(prob)}

example/ucf101/ucf101/evaluator.py

@@ -240,5 +240,5 @@ def cmp(self, ppl_result: t.Iterator) -> t.Any:
     def online_eval(self, file: str):
         with open(file, "rb") as f:
             data = f.read()
-        prob = self.ppl([Video(fp=data)])[0]
+        prob = self.ppl([{"video": Video(fp=data)}])[0]
         return {_LABELS[i]: p for i, p in enumerate(prob[1])}