The COCO dataset is a high-quality dataset both in terms of images and text. Each image has multiple captions and it consists of around 100 000 images. I have created a Google Colab which does the following:

1. Fetches the COCO images and captions
2. Allows the user to specify image dimensions which the images will be resized to
3. Stores information in two files for easy access in the following format:

od-captions.txt (<image_path> : <image_caption>)

train2017/000000203564.jpg : A bicycle replica with a clock as the front wheel. train2017/000000322141.jpg : A room with blue walls and a white sink and door. train2017/00000016977.jpg : A car that seems to be parked illegally behind a legally parked car train2017/000000106140.jpg : A large passenger airplane flying through the air. train2017/000000571635.jpg : A bathroom with a toilet, sink, and shower.

od-captionsonly.txt (<image_caption>)

A bicycle replica with a clock as the front wheel. A room with blue walls and a white sink and door. A car that seems to be parked illegally behind a legally parked car A large passenger airplane flying through the air. A bathroom with a toilet, sink, and shower.

Here is an example image and caption: The man at bat readies to swing at the pitch while the umpire looks on.

I have written this hoping it will be somewhat compatible with @htoyryla's fork of a similar project. Might be of interest to you. Feel free to use this to generate a dataset for DALL-E!

https://github.com/lucidrains/DALLE-pytorch/issues/10#issuecomment-757132197

"wow! temperature feature is awesome! Gradually decreasing it from 5 to 0.05 over 5 epochs and convergence is really fast as well as results look much better!!!"

You can pretty much copy the following code with appropriate modifications. To be more specific, the following is important ingredients:

• Each resolution consists of down/up-sampler and blocks(), consisting of a certain number of Res blocks.
• Each Res block looks as in block() (no modification preferred)
• For the specifics of up/down-sampler, please refer to the comments.
• The hidden dimension of each layer depends on the current spatial scale, and the order of blocks() and up/down-sampler() matters.

def block(self, x, dim, name=None):
    with tf.variable_scope(name):
        hidden_dim = dim // 4
        res = x
        x = self.activation(x, name="activ1")   
        x = self.conv2d(x, hidden_dim, (1, 1), (1, 1), padding="SAME", name="conv1",
                        variable_dtype=self.variable_dtype)
        x = self.activation(x, name="activ2")
        x = self.conv2d(x, hidden_dim, (3, 3), (1, 1), padding="SAME", name="conv2",
                        variable_dtype=self.variable_dtype)
        x = self.activation(x, name="activ3")
        x = self.conv2d(x, dim, (1, 1), (1, 1), padding="SAME", name="conv3",
                        variable_dtype=self.variable_dtype)
        x += res     
        return x


def blocks(self, x, dim, num_blocks, name=None):
    with tf.variable_scope(name):
        for idx in range(num_blocks):
            x = block(x, dim, name="block"+str(idx))    


def decoder(self, num_res, num_blocks):
    with tf.variable_scope("decoder"):
        dim = tf.shape(x)[1] # dim = c
        # num_res is the number of resolutions. e.g. if 32 -> 64 -> 128 -> 256, then = 4
        for idx in range(num_res):
            x = self.blocks(x, dim//(2 ** idx), num_blocks, name='blocks'+str(idx))
            if idx != num_res - 1: # not the last layer
                x = upsample(x) # nearest neighbor interpolation upsampling by the factor of 2
        x = self.conv2d(x, self.channels_dim, (1, 1), (1, 1), variable_dtype=self.variable_dtype)
        return x


def encoder(self, x, num_res, dim, num_blocks):
    with tf.variable_scope("encoder"):
        # dim is the embedding dimension of the first layer of encoder
        x = self.conv2d(x, dim, (3, 3), (1, 1), padding="SAME", name="conv_in",
                        variable_dtype=self.variable_dtype)        
        dim = tf.shape(x)[1] # dim = c
        for idx in range(num_res):
            x = self.blocks(x, dim * (2 ** idx), num_blocks, name='blocks'+str(idx))
            if idx != num_res - 1: # not the last layer
                x = downsample(x) # average pooling with stride = 2
        return x

I got error when run the command: python3 src/data/create_tfrecords.py

Error: Downloading: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1.04M/1.04M [00:00<00:00, 6.36MB/s] Downloading: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 456k/456k [00:00<00:00, 3.31MB/s] Downloading: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1.36M/1.36M [00:00<00:00, 7.19MB/s] 0it [00:00, ?it/s]Traceback (most recent call last): File "src/data/create_tfrecords.py", line 184, in <module> tokenizer=None) File "src/data/create_tfrecords.py", line 158, in create_paired_dataset data = load_jsonl(path) File "src/data/create_tfrecords.py", line 32, in load_jsonl with open(input_path, 'r', encoding='utf-8') as f: FileNotFoundError: [Errno 2] No such file or directory: '/home/data/coco/coco_captions.jsonl' 0it [00:00, ?it/s]

The image generation takes a good amount of time because of the training.

When the pretrained models are released, how big is the size of the pretrained model, and how long will image generation take then? And how much computing power?

And around what time to you plan to release pretrained models?

Best regards