_base_ = ['../PixArt_xl2_internal.py']
data_root = 'data'
image_list_json = ['data_info.json',]

data = dict(type='InternalData', root='InternData', image_list_json=image_list_json, transform='default_train', load_vae_feat=True)
image_size = 1024

# model setting
model = 'PixArt_XL_2'
fp32_attention = True
load_from = None
vae_pretrained = "output/pretrained_models/sd-vae-ft-ema"
pe_interpolation = 2.0

# training setting
num_workers=10
train_batch_size = 2 # 32
num_epochs = 200 # 3
gradient_accumulation_steps = 1
grad_checkpointing = True
gradient_clip = 0.01
optimizer = dict(type='AdamW', lr=2e-5, weight_decay=3e-2, eps=1e-10)
lr_schedule_args = dict(num_warmup_steps=1000)

eval_sampling_steps = 200
log_interval = 20
save_model_epochs=1
save_model_steps=2000
work_dir = 'output/debug'