GitLab

- no more FP16Trainer, we just have an FP16Optimizer wrapper
- most of the distributed code is moved to a new wrapper class called DistributedFairseqModel, which behaves like DistributedDataParallel and a FairseqModel at the same time
- Trainer now requires an extra dummy_batch argument at initialization, which we do fwd/bwd on when there's an uneven number of batches per worker. We hide the gradients from these dummy batches by multiplying the loss by 0
- Trainer.train_step now takes a list of samples, which will allow cleaner --update-freq

also don't crash if param does not recieve grads

adds -reset-optimizer, --reset-lr-scheduler, and --optimizer-overrides flags

This implements transformer based language model. It already obtains better perplexity on wikitext103 without any tuning. I will also train it on gbw where I also expect to get better ppl

Example training command:

python train.py /private/home/abaevski/data/wiki103 —save-dir /tmp —fp16 —max-epoch 80 —save-interval 1 —arch transformer_lm —task language_modeling —optimizer nag —lr 0.008 —lr-scheduler reduce_lr_on_plateau —lr-shrink 0.6 —dropout 0.2 —criterion adaptive_loss —adaptive-softmax-cutoff 10000,50000,200000 —max-tokens 512 —tokens-per-sample 512 —seed 1 —sample-break-mode none —log-format json —log-interval 50 —save-interval-updates 2500 —keep-interval-updates 25
small transformer got to 31.3 ppl on wiki text 103 (compared to 35 with fconv) while @myleott got a big transformer lm to 27 something ppl on wiki text 103

Two tiny changes to train/eval_lm. For train fix an off by one, while for eval_lm make it work when the task is translation'

A Task defines the data format, stores shared state (e.g., dictionaries) and provides helpers for building the model/criterion and calculating the loss.

Changes:
- Add TranslationTask and LanguageModelingTask. New tasks can be registered with @register_task decorator.
- Add EpochBatchIterator to encapsulate batching and saving/restoring dataloader position
- Remove LEFT_PAD_* constants and make them configurable per task

save best val loss in checkpoint and also print best so far

this way when training continues from an existing checkpoint, we dont immediately override checkpoint_best with a worse loss

This implements convolutional language model from https://arxiv.org/pdf/1612.08083.pdf

There are 3 modes for constructing batches:

- token block: fill each sample with a specified number of tokens without regard for sentence delimiters - this is what was used for training in the paper
- complete: fill each sample with a specified number of tokens but make sure it contains only complete sentences (i.e. if next sentence goes over token block limit, move it to the next sample) - this was used for evaluation in the paper
- eos: one sentence per sample (skip blank lines)

some results:

GCNN-13 - GBW - 37.46
GCNN-14B - GBW - 33.88
GCNN-8 - Wiki103 - 43.76
GCNN-14 - Wiki103 - 35.66

train:

python train.py /private/home/abaevski/data/wiki103 --save-dir /tmp --fp16 --max-epoch 35 --save-interval 1 --save-interval-updates 1000 --keep-interval-updates 25 --arch fconv_lm --optimizer nag --lr 1.0 --lr-scheduler reduce_lr_on_plateau --lr-shrink 0.5 --decoder-embed-dim 280 --decoder-layers '[(850, 6)] * 3 + [(850,1)] + [(850,5)] * 4 + [(850,1)] + [(850,4)] * 3 + [(1024,4)] + [(2048, 4)]' --clip-norm 0.1 --dropout 0.2 --weight-decay 5e-06 --criterion cross_entropy --max-tokens 1024 --max-target-positions 1024 --seed 1 --log-format json --log-interval 500

eval:

python eval_lm.py ~abaevski/data/wiki103 --path '/checkpoint02/abaevski/2018-04-27/lm_wiki.fp16.mxup300000.fconv.adam.lrs=reduce_lr_on_plateau.emb280.layers(850,6)*3+(850,1)+(850,5)*4+(850,1)+(850,4)*3+(1024,1)+(2048,4).lr0.0005.clp0.1.drp0.3.wd0.0.crt=cross_entropy.mxtk2048.smptk256.seed1.ngpu8/checkpoint_last.pt'

This PR includes breaking API changes to modularize fairseq-py and adds support for distributed training across multiple nodes.

Changes:
- c7033ef: add support for distributed training! See updated README for usage.
- e016299: modularize fairseq-py, adding support for register_model, register_criterion, register_optimizer, etc.
- 154e440: update LSTM implementation to use PackedSequence objects in the encoder, better following best practices and improving perf
- 90c2973 and 1da6265: improve unit test coverage