README.md

data preparation

In data.py, there are some tools to transform TFRecord into HDF5. As our models are built on Pytorch, it is not beneficial to use TF-Record. To run the script in data.py, please change the csv_path/train_path/valid_path to the paths of the categories csv file, the training data and the valid data. We don't use test data in our experiment. The function TFRecord2hdf5(mode='train' or 'valid') will save the hdf5 file in ./data. Each hdf5 file has the following domains:

• 'features' -- the input feature of data;
• 'labels' -- the ground-true label of data;
• .attrs['num_classes'] -- the number of classes in this task.

Pytorch Data Loader

The Pytorch data loader is implemented in torch_loader.py as MultiTask_Dataloader. If you change the saveto path in data.py in pre-processing, please also change the data_path in torch_loader.py.

Examples command lines to train models

Note: all the models will have three layers.

• Single-Task:

python  Trainer.py --batch_size 256 --max_epoch 10 --hidden_feature 1024 --lr 0.0001 \
--model STL --saveto results/STL

• Hard-Sharing Model (sharing all the hidden layers):

python  Trainer.py --batch_size 256 --max_epoch 10 --hidden_feature 1024 --lr 0.0001 \
--model Hard3 --saveto results/Hard3

• Soft-Order:

python  Trainer.py --batch_size 256 --max_epoch 10 --hidden_feature 1024 --lr 0.0001 \
--model SoftOrder --saveto results/SoftOrder

• Cross-Stitch:

python  Trainer.py --batch_size 256 --max_epoch 10 --hidden_feature 1024 --lr 0.0001 \
--model CrossStitch --saveto results/CrossStitch

• MMoE:

python  Trainer.py --batch_size 256 --max_epoch 10 --hidden_feature 1024 --lr 0.0001 \
--model MMoE --saveto results/MMoE

• MRN:

python  Trainer.py --batch_size 256 --max_epoch 10 --hidden_feature 1024 --lr 0.0001 \
--model MRN --regularization_task True --regularization_feature False --regularization_input False\
--mrn_constant 1e-3 --saveto results/MRN

• DMTRL:

  • Tucker

  python  Trainer.py --batch_size 256 --max_epoch 10 --hidden_feature 1024 --lr 0.0001 \
  --model DMTRL --method Tucker --saveto results/DMTRL_Tucker

  • TT

  python  Trainer.py --batch_size 256 --max_epoch 10 --hidden_feature 1024 --lr 0.0001 \
  --model DMTRL --method TT --saveto results/DMTRL_TT

  • LAF

  python  Trainer.py --batch_size 256 --max_epoch 10 --hidden_feature 1024 --lr 0.0001 \
  --model DMTRL --method LAF --saveto results/DMTRL_LAF

• TAAN:

  • TAAN w/o regularization

  python  Trainer.py --batch_size 256 --max_epoch 10 --hidden_feature 1024 --lr 0.0001 \
  --model TAAN --basis 64 --saveto results/TAAN

  • TAAN + TraceNorm

  python  Trainer.py --batch_size 256 --max_epoch 10 --hidden_feature 1024 --lr 0.0001 \
  --model TAAN --basis 64 --regularize tracenorm --taan_constant 0.1 --saveto results/TAAN_TN

  • TAAN + Cosine

  python  Trainer.py --batch_size 256 --max_epoch 10 --hidden_feature 1024 --lr 0.0001 \
  --model TAAN --basis 64 --regularize cosine --taan_constant 0.1 --saveto results/TAAN_TN

  • TAAN + Distance

  python  Trainer.py --batch_size 256 --max_epoch 10 --hidden_feature 1024 --lr 0.0001 \
  --model TAAN --basis 64 --regularize distance --taan_constant 0.1 --saveto results/TAAN_TN

Visualization

After training TAAN, you can visualize the distance between task-specific activation functions on each hidden layers by Visualizer.py. A example command line is given as

    python  Visualizer.py --hidden_feature 1024 --basis 64 --path results/TAAN