Trying coremltools

snippets/coreml/mobilenet.py

@@ -0,0 +1,42 @@
+import time
+
+import coremltools
+import torch
+import torchvision
+
+from PIL import Image
+import numpy as np
+from torchvision import transforms
+
+
+labels = open("../../models/imagenet_classes.txt").readlines()
+image = Image.open("../../models/cat.png")
+
+model = torchvision.models.mobilenet_v3_large(pretrained=True)
+model.eval()
+model = torch.jit.trace(model, torch.zeros(1, 3, 224, 224))
+
+coreml_model = coremltools.convert(
+    model,
+    inputs=[coremltools.TensorType(name="input_1", shape=(1, 3, 224, 224))],
+    outputs=[coremltools.TensorType(name="output_1")],
+)
+
+preprocess = transforms.Compose(
+    [
+        transforms.Resize(256),
+        transforms.CenterCrop(224),
+        transforms.ToTensor(),
+        transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
+    ]
+)
+input = preprocess(image)
+input = input.unsqueeze(0).numpy()
+
+for i in range(100):
+    start = time.time()
+    pred = coreml_model.predict({"input_1": input})["output_1"][0]
+    print(f"elapsed: {(time.time() - start) * 1000:.2f}ms")
+output = np.argsort(pred)[::-1][:5]
+output = [labels[i].strip() for i in output]
+print(f"top5: {output}")

snippets/coreml/requirements.txt

@@ -0,0 +1,3 @@
+coremltools==8.0b2
+torch==2.3.0
+torchvision