DALL-E in Mesh-Tensorflow [WIP]

Open-AI's DALL-E in Mesh-Tensorflow.

If this is similarly efficient to GPT-Neo, this repo should be able to train models up to, and larger than, the size of Open-AI's DALL-E (12B params).

No pretrained models... Yet.

Thanks to Ben Wang for the tf vae implementation as well as getting the mtf version working, and Aran Komatsuzaki for help building the mtf VAE and input pipeline.

Setup

git clone https://github.com/EleutherAI/GPTNeo
cd GPTNeo
pip3 install -r requirements.txt

Training Setup

Runs on TPUs, untested on GPUs but should work in theory. The example configs are designed to run on a TPU v3-32 pod.

To set up TPUs, sign up for Google Cloud Platform, and create a storage bucket.

Create your VM through a google shell (https://ssh.cloud.google.com/) with ctpu up --vm-only so that it can connect to your Google bucket and TPUs and setup the repo as above.

VAE pretraining

DALLE needs a pretrained VAE to compress images to tokens. To run the VAE pretraining, adjust the params in configs/vae_example.json to a glob path pointing to a dataset of jpgs, and adjust image size to the appropriate size.

  "dataset": {
    "train_path": "gs://neo-datasets/CIFAR-10-images/train/**/*.jpg",
    "eval_path": "gs://neo-datasets/CIFAR-10-images/test/**/*.jpg",
    "image_size": 32
  }

Once this is all set up, create your TPU, then run:

python train_vae_tf.py --tpu your_tpu_name --model vae_example

The training logs image tensors and loss values, to check progress, you can run:

tensorboard --logdir your_model_dir

Dataset Creation [DALL-E]

Once the VAE is pretrained, you can move on to DALL-E.

Currently we are training on a dummy dataset. A public, large-scale dataset for DALL-E is in the works. In the meantime, to generate some dummy data, run:

python src/data/create_tfrecords.py

This should download CIFAR-10, and generate some random captions to act as text inputs.

Custom datasets should be formatted in a folder, with a jsonl file in the root folder containing caption data and paths to the respective images, as follows:

Folder structure:

        data_folder
            jsonl_file
            folder_1
                img1
                img2
                ...
            folder_2
                img1
                img2
                ...
            ...

jsonl structure:
    {"image_path": folder_1/img1, "caption": "some words"}
    {"image_path": folder_2/img2, "caption": "more words"}
    ...

you can then use the create_paired_dataset function in src/data/create_tfrecords.py to encode the dataset into tfrecords for use in training.

Once the dataset is created, copy it over to your bucket with gsutil:

gsutil cp -r DALLE-tfrecords gs://neo-datasets/

And finally, run training with

python train_dalle.py --tpu your_tpu_name --model dalle_example

Config Guide

VAE:

{
  "model_type": "vae",
  "dataset": {
    "train_path": "gs://neo-datasets/CIFAR-10-images/train/**/*.jpg", # glob path to training images
    "eval_path": "gs://neo-datasets/CIFAR-10-images/test/**/*.jpg", # glob path to eval images
    "image_size": 32 # size of images (all images will be cropped / padded to this size)
  },
  "train_batch_size": 32, 
  "eval_batch_size": 32,
  "predict_batch_size": 32,
  "steps_per_checkpoint": 1000, # how often to save a checkpoint
  "iterations": 500, # number of batches to infeed to the tpu at a time. Must be < steps_per_checkpoint
  "train_steps": 100000, # total training steps
  "eval_steps": 0, # run evaluation for this many steps every steps_per_checkpoint
  "model_path": "gs://neo-models/vae_test2/", # directory in which to save the model
  "mesh_shape": "data:16,model:2", # mapping of processors to named dimensions - see mesh-tensorflow repo for more info
  "layout": "batch_dim:data", # which named dimensions of the model to split across the mesh - see mesh-tensorflow repo for more info
  "num_tokens": 512, # vocab size
  "dim": 512, 
  "hidden_dim": 64, # size of hidden dim
  "n_channels": 3, # number of input channels
  "bf_16": false, # if true, the model is trained with bfloat16 precision
  "lr": 0.001, # learning rate [by default learning rate starts at this value, then decays to 10% of this value over the course of the training]
  "num_layers": 3, # number of blocks in the encoder / decoder
  "train_gumbel_hard": true, # whether to use hard or soft gumbel_softmax
  "eval_gumbel_hard": true
}

DALL-E: