Dynamic evaluation for pytorch language models as implemented in Dynamic Evaluation of Neural Sequence Models.

Requirements: python 3 (tested in 3.5, 3.6), pytorch (tested in 0.1.12, 0.2)

Instructions for use:

1. Train a language model using an existing repository, such as the pytorch language modeling tutorial . This should save a .pt file with the trained model

2. Copy the file dynamiceval.py into the repository

3. Run dynamic evaluation with: python dynamiceval.py --model modelname.pt

AWD-LSTM

To replicate results in paper for AWD-LSTM + dynamic eval, train the language model using the Salesforce AWD-LSTM repository. We used the original codebase from this repository, with the goal of exact replication of results from their paper (which we failed to achieve). The default settings and hyper-parameters for dynamiceval.py are tuned for for AWD-LSTM + dynamic eval on PTB.

AWD-QRNN

This code also supports the pytorch QRNN with the --QRNN option. AWD-QRNN + dynamic eval obtains very similar results to AWD-LSTM + dynamic eval, and is much faster to train and evaluate. Training an AWD-QRNN on PTB using the Salesforce AWD-LSTM repository, and running dynamic eval with the default settings gives a test perplexity of 50.5. Increasing the sequence segment length from 5 to 20 runs 3x faster (1 minute vs. 3 minutes on PTB), and gives validation (use --val flag) and test perplexities of 51.4/50.5 with the following arguments:

python dynamiceval.py --model PTB.pt --QRNN --lr 0.00012 --lamb 0.02 --bptt 20

AWD-QRNN trained on wikitext-2 gives validation (use --val flag) and test perplexities of 45.9/44.0 with the following arguments:

python dynamiceval.py --model WT2.pt --QRNN --lr 0.00012 --lamb 0.008 --bptt 20 --data data/wikitext-2

Hyper-parameter search

To get stronger results with any other model or dataset, you can run with:

python dynamiceval.py --model modelname.pt --grid

This will do a hyper-parameter search on the validation set, takes a few hours on PTB with LSTM and default settings, can be much faster with QRNN and/or larger --bptt. If the default model size/settings are changed too much, the hyper-parameters in lrlist and lamblist may need to be changed for best results. If you want to do a faster search, you can try running with --gridfast to use a subset of the validation set, or you can reduce the number of elements in lamblist (tuning lr is more important).

Command line arguments:

--model (required) -filename of the trained model to be evaluated

--data -location of the data corpus

--grid -hyper-parameter grid search over lambda and eta, gives both valid and test error

--gridfast -same as grid, but only uses first 30k validation tokens for search

--val -measure validation error instead of test error

--gpu -specify a gpu device, uses device 0 by default (set negative for cpu)

--QRNN -apply dynamic eval to a QRNN

--bptt -sequence segment length for dynamic eval, also used for gradient statistics on training data

--batch_size -batch size for gradient statistics on training data

--lr -learning rate eta (ignored if --grid is set)

--lamb -decay rate lambda (ignored if --grid is set)

--epsilon -stabilization parameter epsilon

--max_batches -max number of batches for training gradient statistics (-1 uses full training set)

--oldhyper

-The original version code inadvertently scaled a couple of terms differently from the equations described in the paper, which affects the hyper-parameters. The code was changed to reflect the paper equations, and this flag applies a hyper-parameter transformation in a way that accounts for this change. If you run this version of the code with the --oldhyper flag, it is equivalent to running the old version of the code. This will also print out hyper-parameter values that can be used with the new version of the code (without this flag) to obtain the same results. Some previous results that report hyper-parameters with the old code would require applying this hyper-parameter scaling to achieve the same results with this code. The following replicates the exact settings used to obtain results for AWD-LSTM in the paper for PTB:

python dynamiceval.py --model PTB.pt --lr 0.002 --lamb 0.02 --epsilon 0.001 --oldhyper

and for wikitext-2:

python dynamiceval.py --model WT2.pt --lr 0.002 --lamb 0.02 --epsilon 0.001 --oldhyper --data data/wikitext-2

Note that while the original hyper-parameters are the same for Wikitext-2 and PTB, the new/scaled hyper-parameters are a little different.