Pruned-Resnet

What is this repository for?

This repository consists of a 7-layered CNN architecture (named as pruned-ResNet) that can classify 32 x 32 color (RGB) images of CIFAR-10 dataset with more than 86 % accuracy.

Requirements

All the requirements to get the project running have been mentioned in the requirements.txt file. The Python version required is 3.8.15.

Here is a brief summary of each folder/file in this repository:

• Pruned_Resnet_CNN - This directory contains the input and output of each layer for the pruned-ResNet CNN architecture for CIFAR-10 dataset, the python code for pruned-ResNet (in both format - python file and jupyter notebook), and the required dataset (CIFAR-10). The input and output of each layer for pruned-ResNet is stored in a format which is suitable to carry out inference in Scale-Mamba. info_of_model_for_inference_txt_format contains all the input and output for each layer.

• /Pruned_Resnet_CNN/cifar10.tgz - It contains raw CIFAR-10 dataset.

• /Pruned_Resnet_CNN/Pruned_Resnet_CNN.ipynb - This file contains the code to create the CNN architecture, train it, and save the model in .onnx and .pth format. It also gives the information about the total number of parameters and total number of operations for the CNN model. This file contains the code in .ipynb format (jupyter notebook).

• /Pruned_Resnet_CNN/Pruned_Resnet_CNN.py - This file contains the code to create the CNN architecture, train it, and save the model in .onnx and .pth format. It also gives the information about the total number of parameters and total number of operations for the CNN model. This file contains the code in .py format (python file).

• /Pruned_Resnet_CNN/data - This directory contains the data after extracting it from raw data (cifar10.tgz). This data will be used for training.

• /Pruned_Resnet_CNN/diagram - This directory contains the schematic diagram of the CNN architecture.

• /Pruned_Resnet_CNN/inference - This directory contains the python code (in both format - python file and jupyter notebook) to perform inference using 3 test images and the saved model, stored in /Pruned_Resnet_CNN/inference/test_images and /Pruned_Resnet_CNN/saved_model respectively. It also contains the python code (in both format - jupyter notebook and python file) to extract the input and output of each layer which is obtained during inference and to extract the model's parameters.

• /Pruned_Resnet_CNN/saved_model - This directory contains the saved model as a result of training.

• requirements.txt - This file contains all the required dependencies to get the project running.

How to run the project?

Following are the steps to run the project (perform inference using PyTorch):

• Open the Jupyter notebook of inference tii-cnn-cifar10-pytorch/Pruned_Resnet_CNN/inference/inference_cifar10.ipynb.
• Make sure you have installed all the requirements from tii-cnn-cifar10-pytorch/requirements.txt and you have Python version 3.8.15.
• Run the cells of the Jupyter notebook one by one.

Results

Comparative analysis of ResNet-18 and 7-layered CNN for CIFAR-10

CNN architecture	No. of parameters (million)	No. of operations (GMAC)	Accuracy (%)
ResNet-18	11.69	1.82	90
CNN for CIFAR-10	1.26	0.13	86.97

Diagram of pruned-ResNet (7-layered CNN for CIFAR-10 dataset)

Parameters

Range of channels reserved: 100020 - 100049

Channels in use:

Bias

• 100020 → PrunedResnetData/all_parameters/bias/P0-conv_layer_1_bias.txt
• 100021 → PrunedResnetData/all_parameters/bias/P0-conv_layer_2_bias.txt
• 100022 → PrunedResnetData/all_parameters/bias/P0-conv_layer_3_bias.txt
• 100023 → PrunedResnetData/all_parameters/bias/P0-conv_layer_4_bias.txt
• 100024 → PrunedResnetData/all_parameters/bias/P0-conv_layer_5_bias.txt
• 100025 → PrunedResnetData/all_parameters/bias/P0-conv_layer_6_bias.txt
• 100026 → PrunedResnetData/all_parameters/bias/P0-linear_layer_1_bias.txt
• 100027 → PrunedResnetData/all_parameters/bias/P0-linear_layer_2_bias.txt

Beta

• 100028 → PrunedResnetData/all_parameters/new_beta/P0-bn_layer_1_new_beta.txt
• 100029 → PrunedResnetData/all_parameters/new_beta/P0-bn_layer_2_new_beta.txt
• 100030 → PrunedResnetData/all_parameters/new_beta/P0-bn_layer_3_new_beta.txt
• 100031 → PrunedResnetData/all_parameters/new_beta/P0-bn_layer_4_new_beta.txt
• 100032 → PrunedResnetData/all_parameters/new_beta/P0-bn_layer_5_new_beta.txt
• 100033 → PrunedResnetData/all_parameters/new_beta/P0-bn_layer_6_new_beta.txt

Gamma

• 100034 → PrunedResnetData/all_parameters/new_gamma/P0-bn_layer_1_new_gamma.txt
• 100035 → PrunedResnetData/all_parameters/new_gamma/P0-bn_layer_2_new_gamma.txt
• 100036 → PrunedResnetData/all_parameters/new_gamma/P0-bn_layer_3_new_gamma.txt
• 100037 → PrunedResnetData/all_parameters/new_gamma/P0-bn_layer_4_new_gamma.txt
• 100038 → PrunedResnetData/all_parameters/new_gamma/P0-bn_layer_5_new_gamma.txt
• 100039 → PrunedResnetData/all_parameters/new_gamma/P0-bn_layer_6_new_gamma.txt

Weights

• 100040 → PrunedResnetData/all_parameters/weights/P0-conv_layer_1_weight.txt
• 100041 → PrunedResnetData/all_parameters/weights/P0-conv_layer_2_weight.txt
• 100042 → PrunedResnetData/all_parameters/weights/P0-conv_layer_3_weight.txt
• 100043 → PrunedResnetData/all_parameters/weights/P0-conv_layer_4_weight.txt
• 100044 → PrunedResnetData/all_parameters/weights/P0-conv_layer_5_weight.txt
• 100045 → PrunedResnetData/all_parameters/weights/P0-conv_layer_6_weight.txt
• 100046 → PrunedResnetData/all_parameters/weights/P0-linear_layer_1_weight.txt
• 100047 → PrunedResnetData/all_parameters/weights/P0-linear_layer_2_weight.txt

First layer input

Range of channels reserved: 90050 - 90059

• 90050 → execution 1 ./PrunedResnetData/Inputs/Execution_0/P0-Conv_layer_1.input.txt

Who do I talk to?

Please contact Saurav Pawar for any further queries via Slack, Microsoft Teams, or Email ([email protected]).