u-keisuke / DL_Project_LTH_Fusing

The GitRepo for the project of the Deep Learning class at ETH Zurich.

Our team decided to explore the performance of a network that is initialized by a fusion of winning lottery tickets of trained networks N1, N2. How will this perform when compared to the fusion of non-pruned N1, N2?

File Explanations

1. base_convNN.py: Call it to train two MLPs or two CNNs. Networks are saved in the models folder. See Sample Commands below.
2. main.py: Call it to fuse two models, whose networks are saved in the models folder. The accuracy will be printed in the terminal. See Sample Commands below.
3. fusion.py: Takes two models as input and outputs the fused model.

Sample Commands

Training two MLPs

Results will be saved in the models folder.

python base_convNN.py --num-models 2 --gpu-id 0 --model-name mlp

Training two CNNs

Results will be saved in the models folder.

python base_convNN.py --num-models 2 --gpu-id 0 --model-name cnn

Fusing two MLPs

Trained networks have to be in the models folder.

python main.py --num-models 2 --gpu-id 0 --model-name mlp

Fusing two CNNs

Trained networks have to be in the models folder.

python main.py --num-models 2 --gpu-id 0 --model-name cnn

Sample workflow for fusing two MLPs

1. Train two MLPs. The result will be saved in the models folder.

python base_convNN.py --num-models 2 --gpu-id 0 --model-name mlp

2. Fuse two MLPs. The accuracy of the individual models and the fused model will be printed to the terminal.

python main.py --num-models 2 --gpu-id 0 --model-name mlp