F-LM

Language modeling. This codebase contains implementation of G-LSTM and F-LSTM cells from [1]. It also might contain some ongoing experiments.

This code was forked from https://github.com/rafaljozefowicz/lm and contains "BIGLSTM" language model baseline from [2].

Current code runs on Tensorflow r1.0 and supports multi-GPU data parallelism using synchronized gradient updates.

Best perplexity

On One Billion Words benchmark using 8 GPUs in one DGX-1, BIG G-LSTM G4 was able to achieve 24.29 after 2 weeks of training and 23.36 after 3 weeks.

Performance

Not using XLA optimization for now. To be tested. (In all experiments minibatch of 128 per GPU is used)

• SMALLLSTM model on 1xGP100 is getting about ~34K wps.
• SMALLLSTM model on 2xGP100 is getting about ~54.9K wps.
• BIGLSTM model on 1xGP100 is getting about ~4.8K wps
• BIGLSTM model on 2xGP100 is getting about ~8.5K wps
• BIG G-LSTM G4 model on 2xGP100 is getting about ~17.4K wps
• BIG F-LSTM F512 model on 2xGP100 is getting about ~18.5K wps

On DGX-1, from [1], after 1 week of training on DGX-1 using all 8 GPUs. (newer code should be faster).

Model	Perplexity	Steps	WPS
BIGLSTM	31.001	584.6K	20.3K
BIG F-LSTM F512	28.11	1.217M	42.9K
BIG G-LSTM G4	28.17	1.128M	41.7K
BIG G-LSTM G16	34.789	850.4K	41.1K

Exact commit used to produce results from [1]: d98fb110053c187354caf68ff56f5a8535926b5d (should work with TF r1.0)

Dependencies

• TensorFlow r1.0
• Python 2.7 (should work with Python 3 too)
• 1B Word Benchmark Dataset (https://github.com/ciprian-chelba/1-billion-word-language-modeling-benchmark to get data)

To run

Assuming the data directory is in: /raid/okuchaiev/Data/LM1B/1-billion-word-language-modeling-benchmark-r13output/, execute:

export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7

SECONDS=604800
LOGSUFFIX=BIGLSTM

#train
python /home/okuchaiev/repos/f-lm/single_lm_train.py --logdir=/raid/okuchaiev/Workspace/LM/FGLSTM/$LOGSUFFIX --num_gpus=8 --datadir=/raid/okuchaiev/Data/LM1B/1-billion-word-language-modeling-benchmark-r13output/ --hpconfig run_profiler=False,float16_rnn=False,max_time=$SECONDS,num_steps=20,num_shards=8,num_layers=2,learning_rate=0.2,max_grad_norm=1,keep_prob=0.9,emb_size=1024,projected_size=1024,state_size=8192,num_sampled=8192,batch_size=128  > train_$LOGSUFFIX.log 2>&1

#eval
python /home/okuchaiev/repos/f-lm/single_lm_train.py --logdir=/raid/okuchaiev/Workspace/LM/FGLSTM/$LOGSUFFIX --num_gpus=8 --datadir=/raid/okuchaiev/Data/LM1B/1-billion-word-language-modeling-benchmark-r13output/ --mode=eval_full --hpconfig run_profiler=False,float16_rnn=False,max_time=$SECONDS,num_steps=20,num_shards=8,num_layers=2,learning_rate=0.2,max_grad_norm=1,keep_prob=0.9,emb_size=1024,projected_size=1024,state_size=8192,num_sampled=8192,batch_size=16 > eval_full_$LOGSUFFIX.log 2>&1

• To use G-LSTM cell specify num_of_groups parameter.
• To use F-LSTM cell specify fact_size parameter.

To change hyper-parameters

The command accepts and additional argument --hpconfig which allows to override various hyper-parameters, including:

• batch_size=128 - batch size per GPU. Global batch size = batch_size*num_gpus
• num_steps=20 - number of LSTM cell timesteps
• num_shards=8 - embedding and softmax matrices are split into this many shards
• num_layers=1 - numer of LSTM layers
• learning_rate=0.2 - learning rate for optimizer
• max_grad_norm=10.0 - maximum acceptable gradient norm for LSTM layers
• keep_prob=0.9 - dropout keep probability
• optimizer=0 - which optimizer to use: Adagrad(0), Momentum(1), Adam(2), RMSProp(3), SGD(4)
• vocab_size=793470 - vocabluary size
• emb_size=512 - size of the embedding (should be same as projected_size)
• state_size=2048 - LSTM cell size
• projected_size=512 - LSTM projection size
• num_sampled=8192 - training uses sampled softmax, number of samples)
• do_summaries=False - generate weight and grad stats for Tensorboard
• max_time=180 - max time (in seconds) to run
• fact_size - to use F-LSTM cell, this should be set to factor size
• num_of_groups=0 - to use G-LSTM cell, this should be set to number of groups
• save_model_every_min=30 - how often to checkpoint
• save_summary_every_min=16 - how often to save summaries
• use_residual=False - whether to use LSTM residual connections

Feedback

Forked code and GLSTM/FLSTM cells: okuchaiev@nvidia.com.

References

• [1] Factorization tricks for LSTM networks, ICLR 2017 workshop.
• [2] Exploring the Limits of Language Modeling