|
--- |
|
dataset_info: |
|
features: |
|
- name: __key__ |
|
dtype: string |
|
- name: image_tokens |
|
sequence: int64 |
|
- name: text_tokens |
|
sequence: int64 |
|
- name: text |
|
dtype: string |
|
- name: data |
|
dtype: string |
|
splits: |
|
- name: train |
|
num_bytes: 2727128395 |
|
num_examples: 2905954 |
|
- name: validation |
|
num_bytes: 12618157 |
|
num_examples: 13443 |
|
download_size: 964606495 |
|
dataset_size: 2739746552 |
|
configs: |
|
- config_name: default |
|
data_files: |
|
- split: train |
|
path: data/train-* |
|
- split: validation |
|
path: data/validation-* |
|
--- |
|
|
|
# Experiments for training Auto Regressive models for text-to-image generation |
|
This dataset is derived from [conceptual captions](https://huggingface.co/datasets/pixparse/cc3m-wds) (CC3M) which contains roughly 3.3M image and caption pairs. For images we use [1d-tokenizer](https://github.com/bytedance/1d-tokenizer) by [bytedance](https://www.bytedance.com/en/) which tokenizes a 256 * 256 image into 32 tokens while still achieving SOTA fidelity ratio. For text we train a BPE based tokenizer on the image captions dataset with a vocab size set to 30K, where 4096 tokens where used to represent images, 9 to represent some special tokens and the remaining 25895 tokens for text |
|
|
|
# Visualization |
|
<table> |
|
<tr> |
|
<td><img src="vis_1.png" alt="example 1" width="200"/></td> |
|
<td><img src="vis_2.png" alt="example 2" width="200"/></td> |
|
<td><img src="vis_3.png" alt="example 3" width="200"/></td> |
|
<td><img src="vis_4.png" alt="example 4" width="200"/></td> |
|
</tr> |
|
</table> |
|
|
|
# Inference |
|
For generating images download and save the image_tokenizer and checkpoint-20000 in the root dir of this repo then run infer.py with your prompt |
|
|
|
|
|
## Training Procedure |
|
For training we prompt the model to generate an image based on a text such as: "a river has burst it 's banks and has spread out onto arable farmland alongside<|startofimage|><|image:2931|><|image:560|><|image:763|><|image:1539|><|image:3161|><|image:1997|><|image:3376|><|image:510|><|image:3036|><|image:1585|><|image:1853|><|image:1970|><|image:2687|><|image:1436|><|image:2213|><|image:3968|><|image:3999|><|image:877|><|image:725|><|image:3013|><|image:438|><|image:3159|><|image:2936|><|image:3003|><|image:2261|><|image:2137|><|image:3821|><|image:1513|><|image:3536|><|image:311|><|image:494|><|image:413|><|endofimage|>". We use use cross entropy loss with logits masked for the audio tokens as it showed performance improvements for speech-to-text tasks and employ the standard cross entorpy loss over the masked logits |
|
|
|
|
|
| Train Iter | hard rock artist performing music | football player during a match | concept vector illustration showing a flag | police officer and soldiers arrest military combatant | bird on a tree | |
|
| ---- | ---- | ---- | ---- | ---- | ---- | |
|
| 5000 |  |  |  |  |  | |
|
| 6000 |  |  |  |  |  | |
|
| 7000 |  |  |  |  |  | |
|
| 8000 |  |  |  |  |  | |
|
| 9000 |  |  |  |  |  | |
|
| 10000 |  |  |  |  |  | |
|
| 11000 |  |  |  |  |  | |