Minimal text diffusion

A minimal implementation of diffusion models of text: learns a diffusion model of a given text corpus, allowing to generate text samples from the learned model.

Diffusion in action: a DDPM model gradually denoising random text hotnutggy pi greentedsty rawyaented to the white eggplant is dried and mac clement star fe honey spin theapple purpleip to the brown radicchio is sour

This repo has been refactored by taking a large amount of code from https://github.com/XiangLi1999/Diffusion-LM (which includes some code from: https://github.com/openai/glide-text2im), thanks to the authors for their work!

The main idea was to retain just enough code to allow training a simple diffusion model and generating samples, remove image-related terms, and make it easier to use.

I've included an extremely simple corpus (data/simple-{train,test}.txt) I used for quick iterations and testing.

Table of Contents

• Minimal text diffusion
  • Table of Contents
  • Getting started
  • Setup
  • Preparing dataset
  • Training
  • Inference
  • Training from scratch on the greetings dataset
  • Experiments with using pre-trained models and embeddings
  • Controllable Generation
  • Gory details
  • Training
  • Evolving input
  • Sampling
  • TODO
  • Opportunities for further minimization
  • Acknowledgements
  • License

Getting started

Setup

• Install the requirements: pip install -r requirements.txt

• Some of the dependencies might be easier to install via conda:

  conda install mpi4py
  conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch

Preparing dataset

• We will use data/simple.txt as a running example. To begin, we need to create a tokenizer over the dataset. I found that word-level tokenization works best, but the implementation in src/utils/custom_tokenizer includes options to create BPE tokenizer.

python src/utils/custom_tokenizer.py train-word-level data/simple/simple.txt 

Training

• To train a model, run scripts/train.sh. By default, this will train a model on the simple corpus. However, you can change this to any text file using --train_data argument. Note that you may have to increase the sequence length (--seq_len) if your corpus is longer than the simple corpus. The other default arguments are set to match the best setting I found for the simple corpus (see discussion below).

• Once training finishes, the model will be saved in ckpts/simple. You can then use this model to generate samples.

• The checkpoint can also be downloaded from here.

Inference

• To generate samples, run:

bash scripts/text_sample.sh ckpts/simple/ema_0.9999_025000.pt 2000 10

• Here:

  • ckpts/simple/ema_0.9999_025000.pt is the path to the checkpoint
  • 2000 is the number of diffusion steps.
  • 10 is the number of samples to generate.

• By default, this will generate 10 samples from the model trained on the simple corpus. Changing SEED in scripts/text_sample.sh will generate different samples. You can also change the number of samples generated by changing the NUM_SAMPLES argument.

• During inference (denoising), the intermediate sentences will be printed to the console.

• The generated samples will be saved in ckpt/simple/.

• Complete set of outputs are available here.

Training from scratch on the greetings dataset

• I've added another trainign from scratch tutorial here: greetings.

Experiments with using pre-trained models and embeddings

• Update 10/24: The most fluent/realistic outputs are obtained using i) word-level tokenization, ii) initializing a model from scratch, and iii) fine-tuning the embeddings. This is the default in run_train.sh now. Please see docs/old_experiments.md for details on the experiments I ran before this update.

Controllable Generation

• The diffusion model can be combined with a classifier to perform classifier-guided diffusion. Please see details in docs/controllable.md.

Gory details

• Below are my rough notes on how the code works. [TODO] Clean this up and add more details.

Training

• Input text is embedded. This is the mean of x_start_mean. Some noise is added to x_start_mean to get x_start.

• Using random t, a noisy version of the input is created from q(x_t | x_0). This is simply x_t = x_0 sqrt(1 - \beta_t) + \epsilon_t sqrt(\beta_t). The function used for this is q_sample. Any operation that involves going ahead in the diffusion process is carried out by functions that start with q_.

• x_t is fed to the transformer model. Then, the transformer model is trained to generate an approximation of x_start given x_t and t (the timestep). Specifically, the embedded text is passed through a BERT encoder and downsampled. The size of the output embeddings and input embeddings is the same for this reason. Maybe this is the trick mentioned in the paper where they want to tie each weight with the x_start term, but I'm not sure how it's different from DDIM.

• The loss has several terms: 1) Difference between the actual x_start and the output of the transformer model. This is the MSE loss. 2) Mean of the xT should be close to zero. This is the tT_loss term. It is obtained by calling q_mean_variance for the t=T. q_mean_variance is like q_sample, but it returns the mean and variance of the distribution x_t | x0 instead of a sample.

3) Decoder NLL loss. This is the decoder_nll term. It is obtained by calling token_discrete_loss. token_discrete_loss calls get_logits, which in turns uses the embeddings to convert to logits. The logits are then used to calculate the NLL loss. Essentially this is how the embeddings are trained.

    def get_logits(self, hidden_repr):
        return self.lm_head(hidden_repr)

• One thing to note is that:

    print(model.lm_head.weight == model.word_embedding.weight)
    print(model.lm_head.weight.shape, model.word_embedding.weight.shape)

They are identical! Intuitively, the model is trained to predict the embedded input. Thus, having a linear layer with the weights from word_embedding is like doing a nearest neighbor search. While initializing, the weights are assigned to lm_head from word_embedding under torch.no_grad(), so that the gradients are not computed for lm_head.

Evolving input

• Note that the embeddings are trained. Although initial embeddings are passed in training losses, they are not used. Instead, the get_embeds method is used to get the embeddings. This is because the embeddings are trained to predict the input text. Thus, the embeddings are not the same as the input embeddings.

Sampling

• p_mean_variance: returns the distribution p(x_{t-1} | x_t) (the mean and variance). In addition, returns a prediction for the initial x_0.

• q_posterior_mean_variance: returns the distribution q(x_{t-1} | x_t, x_0).

• Additionally, recall that our model is trained to predict x_start given x_t and t.

• Putting these together, we can sample from the model. The sampling is done in the following way:

1. Starting with noise xT, a noisy x_start is first generated using the model.

2. The xT and x_start are used to generate x_{T-1} using q_posterior_mean_variance (x_{T-1} ~ q(x_{T-1} | x_T, x_start)).

The process is repeated until x_0 is generated.

TODO

• [ ] Add more details to the inner workings section.
• [ ] Add classifier-guided sampling.
• [ ] Add more experiments.

Opportunities for further minimization

• [ ] logger.py can be completely deleted.
• [ ] args.py and factory_methods.py can be combined.

Acknowledgements

• Thanks to the team behind Diffusion-LM Improves Controllable Text Generation for releasing their code, which I used as a starting point.
• Thanks to the authors of several open-source implementations of DDPM/DDIM, helpful blogs, and videos. Some of the ones I bookmarked are:

License

• MIT License