[GSoC] Little help on GSoC 2023 Pre-requisite task

[adrianboguszewski]

Discussed in https://github.com/openvinotoolkit/openvino/discussions/16294

^{Originally posted by **Matrixmang0** March 14, 2023} Hello world! My name is Santhosh, and I am an undergrad junior Electrical and Electronics Engineering from India. I would really rooting to work on OpenVINO projet, this summer. So, I am currently working on my GSOC pre-requisite task on `Optimizing YOWOv2 model using OpenVINO`. I am actually stuck at Converting the model to ONNX format, dealing with an error. Below is the code snippet, where I am facing the follwing error. ``` import os # Define the path to the weights weights_path = Path('model/yowo_v2_tiny_ava.pth') # Define the path where the ONNX model will be saved onnx_path = weights_path.with_suffix('.onnx') # Define the width and height of the input image and the number of frames in the input video IMAGE_WIDTH = 208 IMAGE_HEIGHT = 208 NUM_FRAMES = 16 # If an ONNX model with the same name already exists, remove it if os.path.exists(onnx_path): os.remove(onnx_path) # Create a dummy input tensor for the PyTorch model, with random values dummy_input = torch.randn(1, 3, NUM_FRAMES, IMAGE_HEIGHT, IMAGE_WIDTH) # Export the PyTorch model to ONNX format using the `torch.onnx.export()` function torch.onnx.export( YOWO_model, # The PyTorch model to export dummy_input, # The dummy input tensor to use for tracing onnx_path, # The path to save the ONNX model verbose=False, opset_version=12, input_names=['input'], output_names=['output'], dynamic_axes={ 'input': {0: 'batch_size', 2: 'seq_length'}, 'output': {0: 'batch_size', 2: 'seq_length'} } ) # Load the exported ONNX model onnx_model = onnx.load(onnx_path) # Save the ONNX model onnx.save(onnx_model, onnx_path) print(f"ONNX model exported to {onnx_path}.") ``` Error for the code! ``` File ~/Openvino-projects/openvino_notebooks/notebooks/YOWOv2-optimization/YOWOv2/models/backbone/backbone_2d/cnn_2d/yolo_free/yolo_free_fpn.py:204, in PaFPNELAN.forward(self, features) 202 c6 = self.cv1(c5) 203 c7 = F.interpolate(c6, scale_factor=2.0) --> 204 c8 = torch.cat([c7, self.cv2(c4)], dim=1) 205 c9 = self.head_elan_1(c8) 206 ## P4 -> P3 RuntimeError: Sizes of tensors must match except in dimension 1. Expected size 12 but got size 13 for tensor number 1 in the list. ``` This is just the final trace of the error, as the entire error log is real huge Any help on this would really be appreciated!

[adrianboguszewski]

@Matrixmang0, transferred here as related to the prerequisite task.

[blaz-r]

@Matrixmang0 hello, just to make sure, does inference in PyTorch format work for you? I don't see any code on your repo regarding yowov2 and from above I'm not sure if that works. I'd say that if you didn't try that, it'd be great to verify the inference of the original model first, and then possibly check the dimensions with debugger, because it might be something with that.

[adrianboguszewski]

Closing because of the lack of activity. Please reopen if needed.