<think>
Let me analyze this
code and provide a helpful response. This appears to be a Python script for training a
convolutional neural network (CNN) on the MNIST handwritten digit dataset using PyTorch, and then exporting the trained model to ONNX format.
The
code consists of several parts:
Importing necessary libraries and setting hyperparameters
Preparing the MNIST dataset with transformations
Visualizing some examples from the dataset
Defining a CNN model architecture using PyTorch
Setting up loss function and optimizer
Defining training and testing functions
Training the model for 10 epochs and saving it
Exporting the trained model to ONNX format
At the end, the user mentions they want to convert their ONNX model to RKNN format but are having version discrepancies. They're using rknn-toolkit2 version 2.3.0 in their
code, but on their Ubuntu system they have rknn-toolkit2 version 1.6.0 with Python 3.10.
I'll focus on providing a solution for converting their ONNX model to RKNN format, considering the version discrepancy they're experiencing.
Let me first understand the model architecture better:
Input: MNIST images (1 channel, 28x28 pixels)
First
convolutional layer: 1→10 channels, 5x5 kernel, followed by ReLU and 2x2 max pooling
Second
convolutional layer: 10→20 channels, 5x5 kernel, followed by ReLU and 2x2 max pooling
Fully connected
layers: 320→50→10 (output classes for digits 0-9)
The exported ONNX model uses:
Input shape: [1, 1, 28, 28]
Output shape: [1, 10] (representing probabilities for digits 0-9)
ONNX opset version: 19
Now, the issue is the version discrepancy between the rknn-toolkit2 they're using in their
code (2.3.0) and what's installed on their Ubuntu system (1.6.0). I need to provide
code that will work with their installed version (1.6.0) to convert the ONNX model to RKNN format.
</think>
I see you've shared a PyTorch script for training a CNN model on the MNIST dataset, followed by ONNX conversion. You mentioned that you need to convert this ONNX model to RKNN format, but there's a version mismatch between the rknn-toolkit2 in your
code (2.3.0) and what's installed on your Ubuntu system (1.6.0).
Let me create a compatible RKNN conversion script for your system:
<artifact id="rknn_conversion_script" type="application/vnd.ant.
code" language="python" title="MNIST ONNX to RKNN Conversion Script"> from rknn.api import RKNN
Configuration
ONNX_MODEL = './model_Mnist_fixed.onnx'
RKNN_MODEL = './model_Mnist.rknn'
Initialize RKNN object
rknn = RKNN(verbose=True)
Pre-process config for input
print('--> Config model')
rknn.config(mean_values=[127.5], std_values=[127.5], target_platform='rk3588')
print('done')
Load ONNX model
print('--> Loading model')
ret = rknn.load_onnx(model=ONNX_MODEL)
if ret != 0:
print('Load ONNX model failed!')
exit(ret)
print('done')
Build model
print('--> Building model')
ret = rknn.build(do_quantization=True, dataset='./dataset.txt')
if ret != 0:
print('Build model failed!')
exit(ret)
print('done')
Export RKNN model
print('--> Export RKNN model')
ret = rknn.export_rknn(RKNN_MODEL)
if ret != 0:
print('Export RKNN model failed!')
exit(ret)
print('done')
Optional: Inference testing
print('--> Testing model on sample data')
Set inputs
img = cv2.imread('sample_digit.png', cv2.IMREAD_GRAYSCALE)
img = cv2.resize(img, (28, 28))
img = np.expand_dims(img, axis=0) # Add batch dimension
img = np.expand_dims(img, axis=0) # Add channel dimension
Init runtime environment
ret = rknn.init_runtime()
if ret != 0:
print('Init runtime environment failed')
exit(ret)
Inference
print('--> Running model')
outputs = rknn.inference(inputs=[img])
predicted = np.argmax(outputs[0][0])
print(f'Predicted digit: {predicted}')
Release RKNN context
rknn.release()
</artifact>
Additional Setup for RKNN Conversion
For the above script to work properly, you'll need:
Create a calibration dataset file: RKNN quantization requires a dataset file listing paths to sample images:
<artifact id="dataset_generator" type="application/vnd.ant.
code" language="python" title="Calibration Dataset Generator"> import os import numpy as np from torchvision import datasets, transforms import torch
Create dataset directory if it doesn't exist
os.makedirs('./calibration_images', exist_ok=True)
Load a subset of MNIST images for calibration
transform = transforms.Compose([transforms.ToTensor()])
dataset = datasets.MNIST(root='./data/mnist', train=True, transform=transform, download=True)
data_loader = torch.utils.data.DataLoader(dataset, batch_size=1, shuffle=True)
Save calibration images (using 100 images)
with open('./dataset.txt', 'w') as f:
for i, (img, ) in enumerate(data_loader):
if i >= 100: # Use 100 images for calibration
break
# Convert tensor to numpy and save as raw format
img_np = img.numpy().reshape(28, 28) * 255
img_np = img_np.astype(np.uint8)
img_path = f'./calibration_images/img{i:03d}.bin'
img_np.tofile(img_path)
f.write(f'{img_path}\n')
print(f"Created dataset.txt with {min(100, i+1)} image paths for calibration")
</artifact>
Version Compatibility
Notes
Since you're using rknn-toolkit2 version 1.6.0 on Ubuntu with Python 3.10, here are some important compatibility considerations:
API differences: The script I provided uses the API structure compatible with older versions. The newer 2.3.0 version might have slightly different parameter names or methods.
Target platform: I specified 'rk3588' as the target platform. Adjust this if you're targeting a different RK chip.
Quantization parameters: The script uses basic quantization settings. You may need to adjust these (particularly mean_values and std_values) based on your model's preprocessing.
Testing the model: The optional testing section requires you to have a sample image. You can generate one from your test dataset or skip this section if not needed.
Troubleshooting Common Issues
If you encounter errors:
ONNX compatibility: Older RKNN versions may not support ONNX opset 19. Try re-exporting your ONNX model with a lower opset version (e.g., 11 or 13).
Layer support: Some operations might not be supported. Check error messages for unsupported operations and consider simplifying your model if needed.
Quantization issues: If quantization fails, try do_quantization=False first to verify the conversion works without quantization.
Would you like me to explain any specific part of the conversion process in more detail?
按照这个分析来
本文介绍了卷积层的基本概念，包括卷积操作、步长(stride)及零填充(zerp-padding)的作用。其中，卷积操作通过将过滤器与图像进行滑动并计算点积来实现特征提取；步长决定了过滤器移动的步幅；零填充则有助于保留特征图的分辨率，并充分利用边界信息以提升模型性能。
扫一扫
779
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
