Dataset Card for "pokemon-512-valid"

A cleaned + upsampled-to-512px-square version of https://www.kaggle.com/datasets/djilax/pkmn-image-dataset, suitable for training high-resolution unconditional image generators.

source from madebyollin/pokemon-512

80% train_dataset + 10% test_dataset + 10% valid_dataset

I use the following code to split it

from datasets import load_dataset, DatasetDict,Dataset
images_dataset = load_dataset('madebyollin/pokemon-512', split="train")
# 80% train_dataset + 20% train_testvalid
train_testvalid = images_dataset.train_test_split(test_size=0.2,shuffle=True,seed=2000)
# 10% test_dataset + 10% valid_dataset
test_valid = train_testvalid['test'].train_test_split(test_size=0.5,shuffle=True,seed=2000)

train_dev_test_dataset = DatasetDict({
    'train': train_testvalid['train'],
    'test': test_valid['train'],
    'validation': test_valid['test']})
print(train_dev_test_dataset)

train_dataset = train_dev_test_dataset["train"]
test_dataset = train_dev_test_dataset["test"]
valid_dataset = train_dev_test_dataset["validation"]

train_dataset.to_parquet("./data/train_dataset.parquet")
test_dataset.to_parquet("./data/test_dataset.parquet")
valid_dataset.to_parquet("./data/valid_dataset.parquet")

I customed a "train_unconditional.py" from diffusers,logging "validation_loss" while training, and added a module to caculate the FID score by using test_dataset.