This is a code for training Transformers on sequential tasks such as language modeling. Unlike the original Transformer architecture, it uses caching of previous representations and relative position embeddings to better adapt to sequential tasks. In addition, the code also implements the following projects as described below and in this blog post:
You need PyTorch 0.4.1 or above and a cuda-enabled GPU to run the code. If there are multiple GPUs available, the code uses nn.DataParallel to utilize them. For better efficiency, enable distributed training by --distributed argument, which can run on multiple nodes.
This code can be used for running experiments in Adaptive Attention Span for Transformers paper. The adaptive span allows a model to learn an optimal context size for each self-attention head from training data. As shown in the below figure, only few heads require long attention span, thus making it possible to increase the context size to 8k tokens without increasing computation time and memory footprint significantly.
An argument --adapt-span enables adaptive span. Otherwise a model will have a fixed attention span. The adaptive-span is implemented as a nn.Module to make it easier to plug it into other models.
Scripts for running experiments in the paper are located in ./experiments/ directory. For example, a smaller 8-layer version of our model can be trained on a single GPU by running:
bash experiments/enwik8_small.shIt should reach about 1.3bpc on dev after 150k steps.
For training larger models, multiple GPUs are recommended. In the script files, you can configure the number of available GPUs. Increase the --batch-split argument if you run out of GPU memory (it splits batches into smaller pieces without changing the final result).
We obtained the following results in our experiments:
| Experiment | #params | dev | test |
|---|---|---|---|
| enwik8 | 38M | 1.04 bpb | 1.02 bpb |
| enwik8_large | 209M | 1.00 bpb | 0.98 bpb |
| text8 | 39M | 1.05 bpc | 1.11 bpc |
| text8_large | 209M | 1.01 bpc | 1.07 bpc |
A large model training takes about 1.2sec/batch near the end (initially it's faster because the attention spans are smaller) on 8 V100 GPUs. So, for example, the whole enwik8_large training of 170k steps should take less than 2.4 days.
You can download pre-trained models by running the get_pretrained.sh script. Then the same scripts in ./experiments/ can be used to evaluate those models. Since the download script puts models in ./checkpoints/, make sure there is no file with the same name. Note that these pre-trained models are obtained by rerunning the training scripts after the code cleanup, so there are small differences from the above results due to the randomness of the training.
The code also can be used for training All-attention Networks introduced in Augmenting Self-attention with Persistent Memory. If --pers-mem-size argument is set to N, all FF sublayers will be removed from the model and N persistent memory vectors will be added to every self-attention sublayer. The following experiments can be found in ./experiments/ directory.
| Experiment | #params | dev | test |
|---|---|---|---|
| enwik8_pers_small.sh | 39M | 1.03 bpb | 1.01 bpb |
| enwik8_pers.sh | 114M | 1.00 bpb | 0.98 bpb |
| wiki103_pers.sh | 133M | 18.8 ppl * | 19.7 ppl * |
(*This number is slightly better than the paper because it includes end-of-line as a token.)
The code is licensed under CC-BY-NC license. See the LICENSE file for more details.
We thank Xavier Martinet for helping with cleaning the code. The data preprocessing scripts are downloaded from awd-lstm and transformer-XL repos. The adagrad_with_grad_clip.py is mostly adapted from PyTorch.