stonesjtu / pytorch-nce

The Noise Contrastive Estimation for softmax output written in Pytorch

An NCE implementation in pytorch

About NCE

Noise Contrastive Estimation (NCE) is an approximation method that is used to work around the huge computational cost of large softmax layer. The basic idea is to convert the prediction problem into classification problem at training stage. It has been proved that these two criterions converges to the same minimal point as long as noise distribution is close enough to real one.

NCE bridges the gap between generative models and discriminative models, rather than simply speedup the softmax layer. With NCE, you can turn almost anything into posterior with less effort (I think).

Refs:

NCE:

http://www.cs.helsinki.fi/u/ahyvarin/papers/Gutmann10AISTATS.pdf

NCE on rnnlm:

https://pdfs.semanticscholar.org/144e/357b1339c27cce7a1e69f0899c21d8140c1f.pdf

Comparison with other methods

A review of softmax speedup methods:

http://ruder.io/word-embeddings-softmax/

NCE vs. IS (Importance Sampling): Nce is a binary classification while IS is sort of multi-class classification problem.

http://demo.clab.cs.cmu.edu/cdyer/nce_notes.pdf

NCE vs. GAN (Generative Adversarial Network):

https://arxiv.org/abs/1412.6515

On improving NCE

Sampling methods

In NCE, unigram distribution is usually used to approximate the noise distribution because it's fast to sample from. Sampling from a unigram is equal to multinomial sampling, which is of complexity $O(\log(N))$ via binary search tree. The cost of sampling becomes significant when noise ratio increases.

Since the unigram distribution can be obtained before training and remains unchanged across training, some works are proposed to make use of this property to speedup the sampling procedure. Alias method is one of them.

By constructing data structures, alias method can reduce the sampling complexity from $O(log(N))$ to $O(1)$, and it's easy to parallelize.

Refs:

alias method:

https://hips.seas.harvard.edu/blog/2013/03/03/the-alias-method-efficient-sampling-with-many-discrete-outcomes/

Generic NCE (full-NCE)

Conventional NCE only perform the contrasting on linear(softmax) layer, that is, given an input of a linear layer, the model outputs are $p(noise|input)$ and $p(target|input)$. In fact NCE can be applied to more general situations where models are capable to output likelihood values for both real data and noise data.

In this code base, I use a variant of generic NCE named full-NCE (f-NCE) to clarify. Unlike normal NCE, f-NCE samples the noises at input embedding.

Refs:

whole sentence language model by IBM (ICASSP2018)

Bi-LSTM language model by speechlab,SJTU (ICSLP2016?)

Batched NCE

Conventional NCE requires different noise samples per data token. Such computational pattern is not fully GPU-efficient because it needs batched matrix multiplication. A trick is to share the noise samples across the whole mini-batch, thus sparse batched matrix multiplication is converted to more efficient dense matrix multiplication. The batched NCE is already supported by Tensorflow.

A more aggressive approach is to called self contrasting (named by myself). Instead of sampling from noise distribution, the noises are simply the other training tokens the within the same mini-batch.

Ref:

batched NCE

https://arxiv.org/pdf/1708.05997.pdf

self contrasting:

https://www.isi.edu/natural-language/mt/simple-fast-noise.pdf

Run the word language model example

There's an example illustrating how to use the NCE module in example folder. This example is forked from the pytorch/examples repo.

Requirements

Please run pip install -r requirements first to see if you have the required python lib.

• tqdm is used for process bar during training
• dill is a more flexible replacement for pickle

NCE related Arguments

• --nce: whether to use NCE as approximation
• --noise-ratio <50>: numbers of noise samples per batch, the noise is shared among the tokens in a single batch, for training speed.
• --norm-term <9>: the constant normalization term Ln(z)
• --index-module <linear>: index module to use for NCE module (currently and available, does not support PPL calculating )
• --train: train or just evaluation existing model
• --vocab <None>: use vocabulary file if specified, otherwise use the words in train.txt
• --loss [full, nce, sampled, mix]: choose one of the loss type for training, the loss is converted to full for PPL evaluation automatically.

Examples

Run NCE criterion with linear module:

python main.py --cuda --noise-ratio 10 --norm-term 9 --nce --train

Run NCE criterion with gru module:

python main.py --cuda --noise-ratio 10 --norm-term 9 --nce --train --index-module gru

Run conventional CE criterion:

python main.py --cuda --train

A small benchmark in swbd+fisher dataset

It's a performance showcase. The dataset is not bundled in this repo however. The model is trained on concatenated sentences,but the hidden states are not passed across batches. An <s> is inserted between sentences. The model is evaluated on <s> padded sentences separately.

Generally a model trained on concatenated sentences performs slightly worse than the one trained on separate sentences. But we saves 50% of training time by reducing the sentence padding operation.

dataset statistics

• training samples: 2200000 sentences, 22403872 words
• built vocabulary size: ~30K

testbed

• 1080 Ti
• i7 7700K
• pytorch-0.4.0
• cuda-8.0
• cudnn-6.0.1

how to run:

python main.py --train --batch-size 96 --cuda --loss nce --noise-ratio 500 --nhid 300 \
  --emsize 300 --log-interval 1000 --nlayers 1 --dropout 0 --weight-decay 1e-8 \
  --data data/swb --min-freq 3 --lr 2 --save nce-500-swb --concat

Running time

• crossentropy: 6.5 mins/epoch (56K tokens/sec)
• nce: 2 mins/epoch (187K tokens/sec)

performance

The rescore is performed on swbd 50-best, thanks to HexLee.

training loss type	evaluation type	PPL	WER
3gram	normed	??	19.4
CE(no concat)	normed(full)	53	13.1
CE	normed(full)	55	13.3
NCE	unnormed(NCE)	invalid	13.4
NCE	normed(full)	55	13.4
importance sample	normed(full)	55	13.4
importance sample	sampled(500)	invalid	19.0(worse than w/o rescore)

File structure

• example/log/: some log files of this scripts
• nce/: the NCE module wrapper
  • nce/nce_loss.py: the NCE loss
  • nce/alias_multinomial.py: alias method sampling
  • nce/index_linear.py: an index module used by NCE, as a replacement for normal Linear module
  • nce/index_gru.py: an index module used by NCE, as a replacement for the whole language model module
• sample.py: a simple script for NCE linear.
• example: a word langauge model sample to use NCE as loss.
  • example/vocab.py: a wrapper for vocabulary object
  • example/model.py: the wrapper of all nn.Modules.
  • example/generic_model.py: the model wrapper for index_gru NCE module
  • example/main.py: entry point
  • example/utils.py: some util functions for better code structure

---

Modified README from Pytorch/examples

This example trains a multi-layer LSTM on a language modeling task. By default, the training script uses the PTB dataset, provided.

python main.py --train --cuda --epochs 6        # Train a LSTM on PTB with CUDA

The model will automatically use the cuDNN backend if run on CUDA with cuDNN installed.

During training, if a keyboard interrupt (Ctrl-C) is received, training is stopped and the current model is evaluated against the test dataset.

The main.py script accepts the following arguments:

optional arguments:
  -h, --help         show this help message and exit
  --data DATA        location of the data corpus
  --emsize EMSIZE    size of word embeddings
  --nhid NHID        humber of hidden units per layer
  --nlayers NLAYERS  number of layers
  --lr LR            initial learning rate
  --lr-decay         learning rate decay when no progress is observed on validation set
  --weight-decay     weight decay(L2 normalization)
  --clip CLIP        gradient clipping
  --epochs EPOCHS    upper epoch limit
  --batch-size N     batch size
  --dropout DROPOUT  dropout applied to layers (0 = no dropout)
  --seed SEED        random seed
  --cuda             use CUDA
  --log-interval N   report interval
  --save SAVE        path to save the final model
  --bptt             max length of truncated bptt
  --concat           use concatenated sentence instead of individual sentence

CHANGELOG

• 2019.09.09: Improve numeric stability by directly calculation on logits

pytorch-nce's Issues

num_layers does't work

I choose different num_layers params, but it still 1, and I rewrite code, give me lots of error.

Why the target index is not removed from noise samples?

I read the source code, but didn't find such process which removed target from noise samples

Is the implementation batched NCE?

Hi, is the implementation batched NCE or the self-NCE or the normal NCE? Did you compare which one performs better? Thank you!

main.py does not run 'as is' on penn data

Hi there,

I'm trying out your code and couldn't run it 'as is' on penn data. I changed the import of data_sms to data in main.py.
Maybe you left this from some tryouts on another dataset?

Thanks for your implementation anyways.
F

Target Sample can be included in Noise sample

Hello. Thanks you for your NCE code in pytorch. It is very helpful.
I have some question about noise sampling.
In your code, target sample can be sampled as noise sample.
And "K" noise sample can be overlap. Is it OK ?
I think it is not valid in theory, but practically OK.
Do you have any idea for this ?

--nce

README.md refers to option --nce, for example python main.py --cuda --noise-ratio 10 --norm-term 9 --nce --train

example/utils.py does not have --nce in setup_parser()

Result:

think0 tonyr: python main.py --nce
usage: main.py [-h] [--data DATA] [--vocab VOCAB] [--min-freq MIN_FREQ]
               [--emsize EMSIZE] [--nhid NHID] [--nlayers NLAYERS] [--lr LR]
               [--bptt BPTT] [--concat] [--weight-decay WEIGHT_DECAY]
               [--lr-decay LR_DECAY] [--clip CLIP] [--epochs EPOCHS]
               [--batch-size N] [--dropout DROPOUT] [--seed SEED] [--cuda]
               [--log-interval N] [--save SAVE] [--loss LOSS]
               [--index-module INDEX_MODULE] [--noise-ratio NOISE_RATIO]
               [--norm-term NORM_TERM] [--train] [--tb-name TB_NAME] [--prof]
main.py: error: unrecognized arguments: --nce

It looks like --nce has been replaced by --loss nce and README.md should be updated.

It's not clear that the rest of the code still works. This has two issues:

think7 tonyr: python main.py --cuda --noise-ratio 10 --norm-term 9 --train --index-module gru --dropout 0
...
Falling into one layer GRU due to Index_GRU supporting
/usr/lib/python3/dist-packages/torch/nn/modules/rnn.py:71: UserWarning: dropout option adds dropout after all but last recurrent layer, so non-zero dropout expects num_layers greater than 1, but got dropout=0.2 and num_layers=1
  warnings.warn("dropout option adds dropout after all but last "
Training PPL 0.0:  11%|██▎                  | 227/2104 [00:02<00:17, 108.01it/s]

firstly is the warning from rnn.py, secondly the perplexities are all zero.

Moving on to NCE, the reported train PPL is very low, the valid PPL very high.

...
Training PPL 34.3: 100%|████████████████████| 2104/2104 [00:32<00:00, 64.85it/s]
| end of epoch   1 | time: 33.14s |valid ppl   895.19
...
Training PPL 9.8: 100%|█████████████████████| 2104/2104 [00:32<00:00, 64.55it/s]
| end of epoch  40 | time: 33.25s |valid ppl   187.99
| End of training | test ppl   183.84

truncated bptt without padding?

Hi,

Thanks for the great example. I noticed that you pad sentences to the max length per mini-batch, which is a bit different from the truncated bptt approach of the original word_language_model without NCE on pytorch. I wonder if you have compared the two approach and investigated if it makes a difference in terms of the final ppl?

I'm also interested to know how much better this model can be on the ptb dataset. I'm also reading the torch blog post on NCE

Why need to sub math.log(self.noise_ratio)

Hi Stonesjtu,

Thanks for the sharing this NCE implement. I have a question about details. I'd like to know why we need to sub math.log(self.noise_ratio) here:

Pytorch-NCE/nce/nce_loss.py

Line 228 in 1fae107

logit_true = logit_model - logit_noise - math.log(self.noise_ratio)

In this tutorial https://www.tensorflow.org/extras/candidate_sampling.pdf, see the Table of Candidate Sampling Algorithms. The input to training loss of NCE is G(x, y) = F(x, y) - log(Q(y|x)).

Thanks,
Bin

why the labels in sampled_softmax_loss func are all zero?

I read the code in nce_loss.py:

    def sampled_softmax_loss(self, logit_target_in_model, logit_noise_in_model, logit_noise_in_noise, logit_target_in_noise):
        """Compute the sampled softmax loss based on the tensorflow's impl"""
        logits = torch.cat([logit_target_in_model.unsqueeze(2), logit_noise_in_model], dim=2)
        q_logits = torch.cat([logit_target_in_noise.unsqueeze(2), logit_noise_in_noise], dim=2)
        # subtract Q for correction of biased sampling
        logits = logits - q_logits
        labels = torch.zeros_like(logits.narrow(2, 0, 1)).squeeze(2).long()
        loss = self.ce(
            logits.view(-1, logits.size(-1)),
            labels.view(-1),
        ).view_as(labels)

        return loss

The labels are created by 'torch.zeros_like' function, so they are all zeros.
Is this a bug? Because the target label should be one?

Why the nec_linear output loss while output prob for testing?

i don't understand the code below in sample.py:

# training mode
loss = nce_linear(target, input).mean()
print(loss.item())

# evaluation mode for fast probability computation
nce_linear.eval()
prob = nce_linear(target, input).mean()
print(prob.item())

Besides, why need target input parameter for inference?

Error in NCE expression?

In line 227 of nce_loss.py: logit_true = logit_model - logit_noise - math.log(self.noise_ratio), but this is not the same as the log of line 223:# p_true = logit_model.exp() / (logit_model.exp() + self.noise_ratio * logit_noise.exp())
where is the logit_model in the denominator? shouldn't be logit_true = logit_model - log( logit_model.exp() + self.noise_ratio * logit_noise.exp())?

How to select negative samples for NCE loss

Thanks for the efforts. A quick question: How does the model select negative samples for NCE loss? Randomly?

why squeeze here?

Hi, I think there is a bug here:

Pytorch-NCE/nce.py

Line 198 in 862afc6

).squeeze()

For RNN model which the last layer before softmax has shape [B * N * D] where time steps N>1, I believe the squeeze do not have any effect. Maybe for batch size B=1? If that is the case, squeeze(0) might be a better choice.

I am using your code for predicting the last state (in other words, N=1). The squeeze here will give a model_loss.shape = (B , 1) and noise_loss.shape = (B,) and then the total loss.shape = (B, B), which should be (B,1) I think.

Can I use this loss on my customization model ?

Hi i try to use this loss on my model ,
can u provide me more guide ?