DALL-E Mini - Generate image from text

Tentative Strategy of training (proposed by Luke and Suraj)

Data:

• Conceptual 12M Dataset (already loaded and preprocessed in TPU VM by Luke).
• YFCC100M Subset
• Coneptual Captions 3M

Architecture:

• Use the Taming Transformers VQ-GAN (with 16384 tokens)
• Use a seq2seq (language encoder --> image decoder) model with a pretrained non-autoregressive encoder (e.g. BERT) and an autoregressive decoder (like GPT).

Remaining Architecture Questions:

• Whether to freeze the text encoder?
• Whether to finetune the VQ-GAN?
• Which text encoder to use (e.g. BERT, RoBERTa, etc.)?
• Hyperparameter choices for the decoder (e.g. positional embedding, initialization, etc.)

TODO

• experiment with flax/jax and setup of the TPU instance that we should get shortly
• work on dataset loading - see suggested datasets
• Optionally create the OpenAI YFCC100M subset (see this post)
• work on text/image encoding
• concatenate inputs (not sure if we need fixed length for text or use a special token separating text & image)
• adapt training script
• create inference function
• integrate CLIP for better results (only if we have the time)
• work on a demo (streamlit or colab or maybe just HF widget)
• document (set up repo on model hub per instructions, start on README writeup…)
• help with coordinating activities & progress

Dependencies Installation

You should create a new python virtual environment and install the project dependencies inside the virtual env. You need to use the -f (--find-links) option for pip to be able to find the appropriate libtpu required for the TPU hardware:

$ pip install -r requirements.txt -f https://storage.googleapis.com/jax-releases/libtpu_releases.html

If you use conda, you can create the virtual env and install everything using: conda env update -f environments.yaml