ak9250 / DALLE-pytorch

Implementation / replication of DALL-E, OpenAI's Text to Image Transformer, in Pytorch

DALL-E in Pytorch (wip)

Implementation / replication of DALL-E, OpenAI's Text to Image Transformer, in Pytorch. It will also contain CLIP for ranking the generations.

Yannic Kilcher's video

Install

$ pip install dalle-pytorch

Usage

Train VAE

import torch
from dalle_pytorch import DiscreteVAE

vae = DiscreteVAE(
    num_layers = 3,
    num_tokens = 2000,
    dim = 512,
    hidden_dim = 64
)

images = torch.randn(8, 3, 256, 256)

loss = vae(images, return_recon_loss = True)
loss.backward()

Train CLIP

import torch
from dalle_pytorch import CLIP

clip = CLIP(
    dim_text = 512,
    dim_image = 512,
    dim_latent = 512,
    num_text_tokens = 10000,
    num_visual_tokens = 512,
    text_enc_depth = 6,
    visual_enc_depth = 6,
    text_seq_len = 256,
    visual_seq_len = 1024,
    text_heads = 8,
    visual_heads = 8
)

text = torch.randint(0, 10000, (2, 256))
images = torch.randint(0, 512, (2, 1024))
mask = torch.ones_like(text).bool()

loss = clip(text, images, text_mask = mask, return_loss = True)
loss.backward()

Combine pretrained VAE with CLIP, and train off raw images

import torch
from dalle_pytorch import DiscreteVAE, CLIP

vae = DiscreteVAE(
    num_layers = 3,
    num_tokens = 2000,
    dim = 512,
    hidden_dim = 64
)

clip = CLIP(
    dim_text = 512,
    dim_image = 512,
    dim_latent = 512,
    num_text_tokens = 10000,
    num_visual_tokens = 512,
    text_enc_depth = 6,
    visual_enc_depth = 6,
    text_seq_len = 256,
    visual_seq_len = 1024,
    text_heads = 8,
    visual_heads = 8,
    vae = vae
)

text = torch.randint(0, 10000, (2, 256))
images = torch.randn(2, 3, 256, 256) # raw images
mask = torch.ones_like(text).bool()

loss = clip(text, images, text_mask = mask, return_loss = True)
loss.backward()

Train DALL-E

import torch
from dalle_pytorch import DALLE

dalle = DALLE(
    dim = 512,
    num_text_tokens = 10000,
    num_image_tokens = 512,
    text_seq_len = 256,
    image_seq_len = 1024,
    depth = 6, # should be 64
    heads = 8
)

text = torch.randint(0, 10000, (2, 256))
images = torch.randint(0, 512, (2, 1024))
mask = torch.ones_like(text).bool()

loss = dalle(text, images, mask = mask, return_loss = True)
loss.backward()

Combine pretrained VAE with DALL-E, and train off raw images

import torch
from dalle_pytorch import DiscreteVAE, DALLE

vae = DiscreteVAE(
    num_tokens = 512,
    dim = 512
)

dalle = DALLE(
    dim = 512,
    vae = vae,
    num_text_tokens = 10000,
    num_image_tokens = 512,
    text_seq_len = 256,
    image_seq_len = 1024,
    depth = 6, # should be 64
    heads = 8
)

text = torch.randint(0, 10000, (2, 256))
images = torch.randn(2, 3, 256, 256) # train directly on raw images, VAE converts to proper embeddings
mask = torch.ones_like(text).bool()

loss = dalle(text, images, mask = mask, return_loss = True)
loss.backward()

Finally, to generate images

from dalle_pytorch import generate_images

images = generate_images(
    dalle,
    vae = vae,
    text = text,
    mask = mask
)

images.shape # (2, 3, 256, 256)

To get the similarity scores from your trained Clipper, just do

from dalle_pytorch import generate_images

images, scores = generate_images(
    dalle,
    vae = vae,
    text = text,
    mask = mask,
    clipper = clip
)

scores.shape # (2,)
images.shape # (2, 3, 256, 256)

# do your topk here, in paper they sampled 512 and chose top 32

Citations

@misc{unpublished2021dalle,
    title   = {DALL·E: Creating Images from Text},
    author  = {Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray},
    year    = {2021}
}

@misc{unpublished2021clip,
    title  = {CLIP: Connecting Text and Images},
    author = {Alec Radford, Ilya Sutskever, Jong Wook Kim, Gretchen Krueger, Sandhini Agarwal},
    year   = {2021}
}