This repository contains the offical implementation for our paper
Deep Model Reassembly (NeurIPS2022)
[arxiv] [project page] [code]
Xingyi Yang, Zhou Daquan, Songhua Liu, Jingwen Ye, Xinchao Wang
In this work, we explore a novel knowledge-transfer task, termed as Deep Model Reassembly (DeRy), for general-purpose model reuse. DeRy first dissect each model into distinctive building blocks, and then selectively reassemble the derived blocks to produce customized networks under both the hardware resource and performance constraints.
- 2022/12/07 Code Updated for better usage.
DeRy/
โโโ blocklize/block_meta.py [Meta Information & Node Defnition]
โโโ similarity/
โ โโโ get_rep.py [Compute and save the feature embeddings]
โ โโโ get_sim.py [Compute representation similarity given the saved features]
| โโโ partition.py [Network partition by cover set problem]
| โโโ zeroshot_reassembly.py [Network reassembly by solving integer program]
โโโ configs/
| โโโ compute_sim/ [Model configs in the model zoo to compute the feature similarity]
| โโโ dery/XXX/$ModelSize_$DataSet_$BatchSize_$TrainTime_dery_$Optimizor.py [Config files for transfer experiments]
โโโ mmcls_addon/
| โโโ datasets/ [Dataset definitions]
| โโโ models/backbones/dery.py [DeRy backbone definition]
โโโ third_package/timm [Modified timm package]
The model training part is based on mmclassification. Some of the pre-trained weights are from timm.
# Create python env
conda create -n open-mmlab python=3.8 pytorch=1.10 cudatoolkit=11.3 torchvision -c pytorch -y
conda activate open-mmlab
# Install mmcv and mmcls
pip3 install openmim
mim install mmcv-full==1.4.8
mim install mmcls==0.18.0
# Install timm
pip3 install timm
Note: Our code needs torch.fx
to support the computational graph extraction from the torch model. Therefore, please install the torch > 1.10
.
To run the code for *DeRy, we need to go through 4 steps
-
[Model Zoo Preparation] Compute the model feature embeddings and representation similarity. We first write model configuration and its weight path, and run the configs in
configs/compute_sim
PYTHONPATH="$PWD" python simlarity/get_rep.py \ $Config_file \ # configs in `configs/compute_sim` --out $Feature_path \ # Save feature embeddings in *.pth* files [--checkpoint $Checkpoint] # download checkpoint if any
All feature embeddings need to be saved in .pth files in the same $Feat_dictionary. We then load them and compute the feature similarity. Similarity will be saved as net1.net2.pkl files.
PYTHONPATH="$PWD" python simlarity/compute_sim.py / --feat_path $Feat_dictionary / --sim_func $Similarity_function [cka, rbf_cka, lr]
Pre-computed similarity on ImageNet for Linear CKA and Linear Regression.
We also need to compute the feature size (input-output feature dimensions). It can be done by running
PYTHONPATH="$PWD" python simlarity/count_inout_size.py / --root $Feat_dictionary
The results is a json file containing the input-output shape for all network layers, like MODEL_INOUT_SHAPE.json.
-
[Network Partition] Solve the cover set optimization to get the network partition. The results is an assignment file in .pkl.
PYTHONPATH="$PWD" python simlarity/partition.py / --sim_path $Feat_similarity_path / --K $Num_partition / # default=4 --trial $Num_repeat_runs / # default=200 --eps $Size_ratio_each_block / # default=0.2 --num_iter $Maximum_num_iter_eachrun # default=200
-
[Reassemby] Reassemble the partitioned building blocks into a full model, by solving a integer program with training-free proxy. The results are a series of model configs in .py.
PYTHONPATH="$PWD" python simlarity/zeroshot_reassembly.py \ --path $Block_partition_file [Saved in the partition step] \ --C $Maximum_parameter_num \ --minC $Minimum_parameter_num \ --flop_C $Maximum_FLOPs_num \ --minflop_C $Minimum_FLOPs_num \ --num_batch $Number_batch_average_to_compute_score \ --batch_size $Number_sample_each_batch \ --trial $Search_time \ --zero_proxy $Type_train_free_proxy [Default NASWOT] \ --data_config $Config_target_data
-
[Fune-tuning] Train the reassembled model on target data. You may refers to mmclassification for the model training.
Note: Partitioning and reassembling results may not be identical because of the algorithmatic stochaticity. It may slightly affect the performance.
- We use several pre-trained models not included in timm and mmcls, listed in Pre-trained.
@article{yang2022dery,
author = {Xingyi Yang, Daquan Zhou, Songhua Liu, Jingwen Ye, Xinchao Wang},
title = {Deep Model Reassembly},
journal = {NeurIPS},
year = {2022},
}