CNN LEGO: Disassembling and Assembling Convolutional Neural Networks

Requirements

• Python Version: 3.9
• PyTorch Version: 2.0.1
• GPU: NVIDIA RTX A6000 / NVIDIA A40

Quick Start

Prepare the Source Models

• Train a Pre-trained Model:

python engines/train.py \
  --model_name 'vgg16' \
  --data_name 'cifar10' \
  --num_classes 10 \
  --num_epochs 200 \
  --model_dir ${model_dir} \
  --data_train_dir ${data_train_dir} \
  --data_test_dir ${data_test_dir} \
  --log_dir ${log_dir}

Model Disassembling

• Select the Top 1% of Samples with High Confidence:

python core/sample_select.py \
  --model_name 'vgg16' \
  --data_name 'cifar10' \
  --num_classes 10 \
  --model_path ${model_path} \
  --data_dir ${data_dir} \
  --save_dir ${save_dir} \
  --num_samples 50

• Relevant Features Identifying (\alpha and \beta can be configured in core/relevant_feature_identifying.py):

python core/relevant_feature_identifying.py \
  --model_name 'vgg16' \
  --data_name cifar10 \
  --num_classes 10 \
  --model_path ${model_path} \
  --data_dir ${data_dir} \
  --save_dir ${save_dir}

• Parameter Linking and Model Assembling (output the disassembled task-aware component):

python core/model_disassemble.py \
  --model_name 'vgg16' \
  --num_classes 10 \
  --model_path ${model_path} \
  --mask_dir ${mask_dir} \
  --save_dir ${save_dir} \
  --disa_layers ${disa_layers} \
  --disa_labels ${disa_labels}

Model Assembling

• Parameter Scaling (optional):

python core/parameter_scaling.py \
  --model_name 'vgg16' \
  --data_name 'cifar10' \
  --num_classes 10 \
  --model_path ${model_path} \
  --data_dir ${data_dir}

• Alignment Padding and Model Assembling (output the assembled model):

python core/model_assemble.py \
  --model1_path ${model1_path} \
  --model2_path ${model2_path} \
  --asse_path ${asse_path}

Others

• Evaluate the Accuracy of the Model or Task-aware Component:

python engines/test.py \
  --model_name 'vgg16' \
  --data_name cifar10 \
  --num_classes 10 \
  --model_path ${model_path} \
  --data_dir ${data_dir}

• Visualize Model Decision Routes:

python core/model_decision_route_visualizing.py \
  --mask_dir ${mask_dir} \
  --layers ${layers} \
  --labels ${labels}