simplified

README.md

@@ -1,4 +1,4 @@
-# min DALL·E
+# min(DALL·E)
 
 This is a minimal implementation of [DALL·E Mini](https://github.com/borisdayma/dalle-mini) in both Flax and PyTorch
 
@@ -11,21 +11,19 @@ Run `sh setup.sh` to install dependencies and download pretrained models.  The o
 Here are some examples
 
 ```
-python3 image_from_text_flax.py \
-  --dalle_bart_path='./pretrained/dalle_bart_mega' \
-  --vqgan_path='./pretrained/vqgan' \
-  --image_path='./generated/avacado_armchair_flax.png' \
-  --seed=4 \
+python3 image_from_text.py --text='alien life' --seed=7
+```
+![Alien](examples/alien.png)
+
+```
+python3 image_from_text.py --mega --seed=4 \
   --text='a comfy chair that looks like an avocado'
 ```
 ![Avocado Armchair](examples/avocado_armchair.png)
 
 
 ```
-python3 image_from_text_flax.py \
-  --dalle_path='./pretrained/dalle-mega' \
-  --seed=100 \
-  --image_path='./generated/godzilla_trial.png' \
+python3 image_from_text.py --mega --seed=100 \
   --text='court sketch of godzilla on trial'
 ```
 

image_from_text.py

@@ -0,0 +1,129 @@
+import argparse
+import os
+import json
+import numpy
+from PIL import Image
+from typing import Tuple, List
+
+from min_dalle.load_params import load_dalle_bart_flax_params
+from min_dalle.text_tokenizer import TextTokenizer
+from min_dalle.min_dalle_flax import generate_image_tokens_flax
+from min_dalle.min_dalle_torch import (
+    generate_image_tokens_torch,
+    detokenize_torch
+)
+
+
+parser = argparse.ArgumentParser()
+parser.add_argument(
+    '--text',
+    help='text to generate image from',
+    type=str
+)
+parser.add_argument(
+    '--seed',
+    help='random seed',
+    type=int,
+    default=0
+)
+parser.add_argument(
+    '--mega',
+    help='use larger dalle mega model',
+    action=argparse.BooleanOptionalAction
+)
+parser.add_argument(
+    '--torch',
+    help='use torch transformers',
+    action=argparse.BooleanOptionalAction
+)
+parser.add_argument(
+    '--image_path',
+    help='path to save generated image',
+    type=str,
+    default='generated.png'
+)
+parser.add_argument(
+    '--image_token_count',
+    help='number of image tokens to generate (for debugging)',
+    type=int,
+    default=256
+)
+
+
+def load_dalle_bart_metadata(path: str) -> Tuple[dict, dict, List[str]]:
+    print("loading model")
+    for f in ['config.json', 'flax_model.msgpack', 'vocab.json', 'merges.txt']:
+        assert(os.path.exists(os.path.join(path, f)))
+    with open(path + '/config.json', 'r') as f: 
+        config = json.load(f)
+    with open(path + '/vocab.json') as f:
+        vocab = json.load(f)
+    with open(path + '/merges.txt') as f:
+        merges = f.read().split("\n")[1:-1]
+    return config, vocab, merges
+
+
+def ascii_from_image(image: Image.Image, size: int) -> str:
+    rgb_pixels = image.resize((size, int(0.55 * size))).convert('L').getdata()
+    chars = list('.,;/IOX')
+    chars = [chars[i * len(chars) // 256] for i in rgb_pixels]
+    chars = [chars[i * size: (i + 1) * size] for i in range(size // 2)]
+    return '\n'.join(''.join(row) for row in chars)
+
+
+def save_image(image: numpy.ndarray, path: str) -> Image.Image:
+    if os.path.isdir(path):
+        path = os.path.join(path, 'generated.png')
+    elif not path.endswith('.png'):
+        path += '.png'
+    print("saving image to", path)
+    image: Image.Image = Image.fromarray(numpy.asarray(image))
+    image.save(path)
+    return image
+
+
+def tokenize_text(
+    text: str, 
+    config: dict,
+    vocab: dict,
+    merges: List[str]
+) -> numpy.ndarray:
+    print("tokenizing text")
+    tokens = TextTokenizer(vocab, merges)(text)
+    print("text tokens", tokens)
+    text_tokens = numpy.ones((2, config['max_text_length']), dtype=numpy.int32)
+    text_tokens[0, :len(tokens)] = tokens
+    text_tokens[1, :2] = [tokens[0], tokens[-1]]
+    return text_tokens
+
+
+if __name__ == '__main__':
+    args = parser.parse_args()
+
+    model_name = 'mega' if args.mega == True else 'mini'
+    model_path = './pretrained/dalle_bart_{}'.format(model_name)
+    config, vocab, merges = load_dalle_bart_metadata(model_path)
+    text_tokens = tokenize_text(args.text, config, vocab, merges)
+    params_dalle_bart = load_dalle_bart_flax_params(model_path)
+
+    image_tokens = numpy.zeros(config['image_length'])
+    if args.torch == True:
+        image_tokens[:args.image_token_count] = generate_image_tokens_torch(
+            text_tokens = text_tokens,
+            seed = args.seed,
+            config = config,
+            params = params_dalle_bart,
+            image_token_count = args.image_token_count
+        )
+    else:
+        image_tokens[...] = generate_image_tokens_flax(
+            text_tokens = text_tokens, 
+            seed = args.seed,
+            config = config,
+            params = params_dalle_bart,
+        )
+
+    if args.image_token_count == config['image_length']:
+        image = detokenize_torch(image_tokens)
+        image = save_image(image, args.image_path)
+        print(ascii_from_image(image, size=128))

image_from_text_flax.py

@@ -1,126 +0,0 @@
-import jax
-from jax import numpy as jnp
-import numpy
-import argparse
-
-from min_dalle.load_params import load_dalle_bart_flax_params
-from min_dalle.image_from_text import (
-    load_dalle_bart_metadata, 
-    tokenize, 
-    detokenize_torch,
-    save_image, 
-    ascii_from_image
-)
-from min_dalle.models.dalle_bart_encoder_flax import DalleBartEncoderFlax
-from min_dalle.models.dalle_bart_decoder_flax import DalleBartDecoderFlax
-
-
-parser = argparse.ArgumentParser()
-parser.add_argument(
-    '--text',
-    help='text to generate image from',
-    type=str
-)
-parser.add_argument(
-    '--seed',
-    help='random seed',
-    type=int,
-    default=0
-)
-parser.add_argument(
-    '--image_path',
-    help='generated image path',
-    type=str,
-    default='generated.png'
-)
-parser.add_argument(
-    '--dalle_bart_path',
-    help='pretraied dalle bart path',
-    type=str,
-    default='./pretrained/dalle_bart_mini'
-)
-parser.add_argument(
-    '--vqgan_path',
-    help='pretraied vqgan path',
-    type=str,
-    default='./pretrained/vqgan'
-)
-
-
-def encode_flax(
-    text_tokens: numpy.ndarray,
-    config: dict, 
-    params: dict
-) -> jnp.ndarray:
-    print("loading flax encoder")
-    encoder: DalleBartEncoderFlax = DalleBartEncoderFlax(
-        attention_head_count = config['encoder_attention_heads'],
-        embed_count = config['d_model'],
-        glu_embed_count = config['encoder_ffn_dim'],
-        text_token_count = config['max_text_length'],
-        text_vocab_count = config['encoder_vocab_size'],
-        layer_count = config['encoder_layers']
-    ).bind({'params': params.pop('encoder')})
-
-    print("encoding text tokens")
-    encoder_state = encoder(text_tokens)
-    del encoder
-    return encoder_state
-
-def decode_flax(
-    text_tokens: jnp.ndarray,
-    encoder_state: jnp.ndarray,
-    config: dict,
-    seed: int,
-    params: dict
-) -> jnp.ndarray:
-    print("loading flax decoder")
-    decoder = DalleBartDecoderFlax(
-        image_token_count = config['image_length'],
-        text_token_count = config['max_text_length'],
-        image_vocab_count = config['image_vocab_size'],
-        attention_head_count = config['decoder_attention_heads'],
-        embed_count = config['d_model'],
-        glu_embed_count = config['decoder_ffn_dim'],
-        layer_count = config['decoder_layers'],
-        start_token = config['decoder_start_token_id']
-    )
-    print("sampling image tokens")
-    image_tokens = decoder.sample_image_tokens(
-        text_tokens,
-        encoder_state,
-        jax.random.PRNGKey(seed),
-        params.pop('decoder')
-    )
-    del decoder
-    return image_tokens
-
-def generate_image_tokens_flax(
-    text: str,
-    seed: int, 
-    dalle_bart_path: str
-) -> numpy.ndarray:
-    config, vocab, merges = load_dalle_bart_metadata(dalle_bart_path)
-    text_tokens = tokenize(text, config, vocab, merges)
-    params_dalle_bart = load_dalle_bart_flax_params(dalle_bart_path)
-    encoder_state = encode_flax(text_tokens, config, params_dalle_bart)
-    image_tokens = decode_flax(
-        text_tokens, 
-        encoder_state, 
-        config, seed, 
-        params_dalle_bart
-    )
-    return numpy.array(image_tokens)
-
-if __name__ == '__main__':
-    args = parser.parse_args()
-
-    image_tokens = generate_image_tokens_flax(
-        args.text, 
-        args.seed, 
-        args.dalle_bart_path
-    )
-    print("image tokens", list(image_tokens))
-    image = detokenize_torch(image_tokens, args.vqgan_path)
-    image = save_image(image, args.image_path)
-    print(ascii_from_image(image, size=128))

min_dalle/image_from_text.py

@@ -1,70 +0,0 @@
-import os
-import json
-import numpy
-import torch
-from PIL import Image
-from typing import Tuple, List
-
-from .text_tokenizer import TextTokenizer
-from .models.vqgan_detokenizer import VQGanDetokenizer
-from .load_params import load_vqgan_torch_params
-
-
-def load_dalle_bart_metadata(path: str) -> Tuple[dict, dict, List[str]]:
-    print("loading model")
-    for f in ['config.json', 'flax_model.msgpack', 'vocab.json', 'merges.txt']:
-        assert(os.path.exists(os.path.join(path, f)))
-    with open(path + '/config.json', 'r') as f: 
-        config = json.load(f)
-    with open(path + '/vocab.json') as f:
-        vocab = json.load(f)
-    with open(path + '/merges.txt') as f:
-        merges = f.read().split("\n")[1:-1]
-    return config, vocab, merges
-
-
-def ascii_from_image(image: Image.Image, size: int) -> str:
-    rgb_pixels = image.resize((size, int(0.55 * size))).convert('L').getdata()
-    chars = list('.,;/IOX')
-    chars = [chars[i * len(chars) // 256] for i in rgb_pixels]
-    chars = [chars[i * size: (i + 1) * size] for i in range(size // 2)]
-    return '\n'.join(''.join(row) for row in chars)
-
-
-def save_image(image: numpy.ndarray, path: str) -> Image.Image:
-    if os.path.isdir(path):
-        path = os.path.join(path, 'generated.png')
-    elif not path.endswith('.png'):
-        path += '.png'
-    print("saving image to", path)
-    image: Image.Image = Image.fromarray(numpy.asarray(image))
-    image.save(path)
-    return image
-
-
-def tokenize(
-    text: str, 
-    config: dict,
-    vocab: dict,
-    merges: List[str]
-) -> numpy.ndarray:
-    print("tokenizing text")
-    tokens = TextTokenizer(vocab, merges)(text)
-    print("text tokens", tokens)
-    text_tokens = numpy.ones((2, config['max_text_length']), dtype=numpy.int32)
-    text_tokens[0, :len(tokens)] = tokens
-    text_tokens[1, :2] = [tokens[0], tokens[-1]]
-    return text_tokens
-
-
-def detokenize_torch(
-    image_tokens: numpy.ndarray, 
-    model_path: str
-) -> numpy.ndarray:
-    print("detokenizing image")
-    params = load_vqgan_torch_params(model_path)
-    detokenizer = VQGanDetokenizer()
-    detokenizer.load_state_dict(params)
-    image_tokens = torch.tensor(image_tokens).to(torch.long)
-    image = detokenizer.forward(image_tokens).to(torch.uint8)
-    return image.detach().numpy()

min_dalle/min_dalle_flax.py

@@ -1,50 +1,9 @@
 import jax
 from jax import numpy as jnp
 import numpy
-import argparse
 
-from load_params import load_dalle_bart_flax_params
-from image_from_text import (
-    load_dalle_bart_metadata, 
-    tokenize, 
-    detokenize_torch,
-    save_image, 
-    ascii_from_image
-)
-from models.dalle_bart_encoder_flax import DalleBartEncoderFlax
-from models.dalle_bart_decoder_flax import DalleBartDecoderFlax
-
-
-parser = argparse.ArgumentParser()
-parser.add_argument(
-    '--text',
-    help='text to generate image from',
-    type=str
-)
-parser.add_argument(
-    '--seed',
-    help='random seed',
-    type=int,
-    default=0
-)
-parser.add_argument(
-    '--image_path',
-    help='generated image path',
-    type=str,
-    default='generated.png'
-)
-parser.add_argument(
-    '--dalle_bart_path',
-    help='pretraied dalle bart path',
-    type=str,
-    default='./pretrained/dalle_bart_mini'
-)
-parser.add_argument(
-    '--vqgan_path',
-    help='pretraied vqgan path',
-    type=str,
-    default='./pretrained/vqgan'
-)
+from .models.dalle_bart_encoder_flax import DalleBartEncoderFlax
+from .models.dalle_bart_decoder_flax import DalleBartDecoderFlax
 
 
 def encode_flax(
@@ -67,6 +26,7 @@ def encode_flax(
     del encoder
     return encoder_state
 
+
 def decode_flax(
     text_tokens: jnp.ndarray,
     encoder_state: jnp.ndarray,
@@ -95,32 +55,25 @@ def decode_flax(
     del decoder
     return image_tokens
 
+
 def generate_image_tokens_flax(
-    text: str,
+    text_tokens: numpy.ndarray,
     seed: int,
-    dalle_bart_path: str
+    config: dict,
+    params: dict
 ) -> numpy.ndarray:
-    config, vocab, merges = load_dalle_bart_metadata(dalle_bart_path)
-    text_tokens = tokenize(text, config, vocab, merges)
-    params_dalle_bart = load_dalle_bart_flax_params(dalle_bart_path)
-    encoder_state = encode_flax(text_tokens, config, params_dalle_bart)
+    encoder_state = encode_flax(
+        text_tokens, 
+        config, 
+        params
+    )
     image_tokens = decode_flax(
         text_tokens, 
         encoder_state, 
-        config, seed, 
-        params_dalle_bart
-    )
-    return numpy.array(image_tokens)
-
-if __name__ == '__main__':
-    args = parser.parse_args()
-
-    image_tokens = generate_image_tokens_flax(
-        args.text, 
-        args.seed, 
-        args.dalle_bart_path
+        config, 
+        seed, 
+        params
     )
+    image_tokens = numpy.array(image_tokens)
     print("image tokens", list(image_tokens))
-    image = detokenize_torch(image_tokens, args.vqgan_path)
-    image = save_image(image, args.image_path)
-    print(ascii_from_image(image, size=128))
+    return image_tokens

min_dalle/min_dalle_torch.py

@@ -1,61 +1,17 @@
 import numpy
 import torch
 from torch import Tensor
-import argparse
 from typing import Dict
 
-from min_dalle.image_from_text import (
-    load_dalle_bart_metadata, 
-    tokenize,
-    detokenize_torch,
-    save_image, 
-    ascii_from_image
-)
-from min_dalle.models.dalle_bart_encoder_torch import DalleBartEncoderTorch
-from min_dalle.models.dalle_bart_decoder_torch import DalleBartDecoderTorch
+from .models.vqgan_detokenizer import VQGanDetokenizer
+from .models.dalle_bart_encoder_torch import DalleBartEncoderTorch
+from .models.dalle_bart_decoder_torch import DalleBartDecoderTorch
 
-from min_dalle.load_params import (
-    load_dalle_bart_flax_params, 
+from .load_params import (
+    load_vqgan_torch_params,
     convert_dalle_bart_torch_from_flax_params
 )
 
-parser = argparse.ArgumentParser()
-parser.add_argument(
-    '--text',
-    help='text to generate image from',
-    type=str
-)
-parser.add_argument(
-    '--seed',
-    help='random seed',
-    type=int,
-    default=0
-)
-parser.add_argument(
-    '--image_token_count',
-    help='image tokens to sample',
-    type=int,
-    default=256
-)
-parser.add_argument(
-    '--image_path',
-    help='generated image path',
-    type=str,
-    default='generated.png'
-)
-parser.add_argument(
-    '--dalle_bart_path',
-    help='pretraied dalle bart path',
-    type=str,
-    default='./pretrained/dalle_bart_mini'
-)
-parser.add_argument(
-    '--vqgan_path',
-    help='pretraied vqgan path',
-    type=str,
-    default='./pretrained/vqgan'
-)
-
 
 def encode_torch(
     text_tokens: numpy.ndarray,
@@ -123,37 +79,35 @@ def decode_torch(
 
 
 def generate_image_tokens_torch(
-    text: str, 
+    text_tokens: numpy.ndarray,
     seed: int,
-    image_token_count: int,
-    dalle_bart_path: str
+    config: dict,
+    params: dict,
+    image_token_count: int
 ) -> numpy.ndarray:
-    config, vocab, merges = load_dalle_bart_metadata(dalle_bart_path)
-    text_tokens = tokenize(text, config, vocab, merges)
-    params_dalle_bart = load_dalle_bart_flax_params(dalle_bart_path)
-    encoder_state = encode_torch(text_tokens, config, params_dalle_bart)
+    encoder_state = encode_torch(
+        text_tokens, 
+        config, 
+        params
+    )
     image_tokens = decode_torch(
         text_tokens, 
         encoder_state, 
-        config, seed, params_dalle_bart,
+        config, 
+        seed, 
+        params,
         image_token_count
     )
     return image_tokens.detach().numpy()
 
 
-if __name__ == '__main__':
-    args = parser.parse_args()
-    image_tokens = generate_image_tokens_torch(
-        args.text, 
-        args.seed, 
-        args.image_token_count, 
-        args.dalle_bart_path
-    )
-    if args.image_token_count < 256:
-        print("image tokens", list(image_tokens, ))
-    else:
-        image = detokenize_torch(image_tokens, args.vqgan_path)
-        image = save_image(image, args.image_path)
-        print(ascii_from_image(image, size=128))
-
+def detokenize_torch(image_tokens: numpy.ndarray) -> numpy.ndarray:
+    print("detokenizing image")
+    model_path = './pretrained/vqgan'
+    params = load_vqgan_torch_params(model_path)
+    detokenizer = VQGanDetokenizer()
+    detokenizer.load_state_dict(params)
+    image_tokens = torch.tensor(image_tokens).to(torch.long)
+    image = detokenizer.forward(image_tokens).to(torch.uint8)
+    return image.detach().numpy()