A pretrained chinese text-to-image generation model for Xingzhi competition.
| item | ver |
|---|---|
| OS | ubuntu 20.04 |
| GPU | NVIDIA RTX 3090 24GB x6 |
| python | 3.10.4 |
| cuda | 11.3 |
| torch | 1.12.1 |
Converting the text to tokens by cogview-data, then packing the ID and tokens to dataset/{train, val, test}.npy using make_dataset.py.
We use 6 Nvidia RTX 3090. Training process is about 20 days and inference is about 1 day.
Important!! You must put the train and val images into the dataset floder, and rename them into {train, val} before train.
Then
bash scripts/train.sh
bash scripts/predict.sh
dataset: https://pan.baidu.com/s/1FaE8O25wq-1YWmVVHIPq8A passwd: zHvD
FID:40.748
