Update inception v3 readme

examples/inception_v3/README.md

@@ -0,0 +1,43 @@
+# Inception v3
+
+Inception v3 Inference on CUDANet
+
+## Usage
+
+1. Export pytorch Inception v3 weights pretrained on ImageNet (requires pytorch and torchvision):
+
+```sh
+python inception_v3.py
+```
+
+2. Follow the instructions from the repository root to build the CUDANet library.
+
+3. Build Inception v3 (requires [OpenCV](https://opencv.org/) for image loading and preprocessing):
+
+```sh
+mkdir build
+cd build
+cmake -S ..
+make
+```
+
+4. (Optional) Run tests
+
+Generate test input/output and resources by running `inception_blocks.py` in the `test` folder
+
+Build and run tests (requires [Google Test](https://github.com/google/googletest))
+
+```sh
+cd build
+make test_inception_v3
+./tests/test_inception_v3
+```
+
+5. Run Inception v3 inference:
+```sh
+inception_v3 ../inception_v3_weights.bin ../image.jpg
+```
+
+## Note on Preprocessing
+
+The image preprocessing in this implementation uses OpenCV, which may produce slightly different results compared to PyTorch's Pillow-based preprocessing due to differences in interpolation methods during resizing.

examples/inception_v3/inception_v3.py

@@ -4,14 +4,21 @@ import sys
 sys.path.append("../../tools")  # Ugly hack
 from utils import export_model_weights, print_model_parameters, predict
 
+import torch
 
 if __name__ == "__main__":
-    inception = torchvision.models.inception_v3(
+
     weights = torchvision.models.Inception_V3_Weights.DEFAULT
+    inception = torchvision.models.inception_v3(
+        weights=weights,
+        transform_input=False
     )
+
+    inception.transform_input = False
     inception.eval()
 
-    # print_model_parameters(inception)  # print layer names and number of parameters
     export_model_weights(inception, "inception_v3_weights.bin")
 
-    # print(predict(inception, "bird.jpg"))
+    # class_labels = weights.meta["categories"]
+    # prediction = predict(inception, "bird.jpg")
+    # print(prediction, class_labels[prediction])

tools/utils.py

@@ -35,8 +35,6 @@ def export_model_weights(model: torch.nn.Module, filename):
 
             tensor_data += tensor_bytes
 
-        # print(model.named_buffers)
-
         # Add buffers (for running_mean and running_var)
         for name, buf in model.named_buffers():
             if "running_mean" not in name and "running_var" not in name:
@@ -76,9 +74,7 @@ def predict(model, image_path, resize=299, crop=299, preprocess=None):
         )
 
     input_tensor = preprocess(input_image)
-    input_batch = input_tensor.unsqueeze(
+    input_batch = input_tensor.unsqueeze(0)  
-        0
-    )  # create a mini-batch as expected by the model
 
     # move the input and model to GPU for speed if available
     if torch.cuda.is_available():
@@ -87,4 +83,4 @@ def predict(model, image_path, resize=299, crop=299, preprocess=None):
 
     with torch.no_grad():
         output = model(input_batch)  
-        return torch.argmax(output)
+        return torch.argmax(output).item()