SLiM-Performer (PyTorch)

Evaluated under Python 3.6.9 and PyTorch 1.7.0.

To reduce repository memory, the following data can be downloaded:

• Penn Treebank
• Enwik8

The copy task data is generated randomly.

The following are the commands used to produce the experimental results in the paper.

Penn Treebank commands:

• python3 train.py --arg_code=0 for the full setup.
• python3 train.py --arg_code=1 for the C = 512 setup.
• python3 train.py --arg_code=2 for the C = 256 setup.
• python3 train.py --arg_code=3 for the C = 512 fine-tuning setup.
• python3 train.py --arg_code=4 for the C = 256 fine-tuning setup.

Enwik8 commands:

• python3 train.py --arg_code=5 for the full setup.
• python3 train.py --arg_code=6 for the C = 2048 setup.
• python3 train.py --arg_code=7 for the C = 1366 setup.
• python3 train.py --arg_code=8 for the C = 2048 fine-tuning setup.
• python3 train.py --arg_code=9 for the C = 1366 fine-tuning setup.

Copying task commands:

• python3 train.py --arg_code 10 for the full setup.
• python3 train.py --arg_code 11 for the C = 128 setup.
• python3 train.py --arg_code 12 for the C = 64 setup.
• python3 train.py --arg_code 13 for the C = 128 fine-tuning setup.
• python3 train.py --arg_code 14 for the C = 64 fine-tuning setup.