output_lr1e4 / readme.txt
jaruce's picture
End of training
df2120e verified
前面2500个step是用lr 1e-4 训的,accumulation step = 4,loss持续下降(wandb 23)
Namespace(adam_beta1=0.9, adam_beta2=0.999, adam_epsilon=1e-08, adam_weight_decay=0.01, allow_tf32=False, cache_dir=None, center_crop=False, checkpointing_steps=500, checkpoints_total_limit=None, conditioning_dropout_prob=None, data_file_records='/home/ach17654gk/code/train_sdxl_gg/metadata.jsonl', dataloader_num_workers=0, dataset_config_name=None, dataset_name=None, edit_prompt_column='text', edited_image_column='conditioning_image', enable_xformers_memory_efficient_attention=False, gradient_accumulation_steps=4, gradient_checkpointing=False, hub_model_id=None, hub_token=None, learning_rate=0.0001, local_rank=2, logging_dir='logs', lr_scheduler='constant', lr_warmup_steps=500, max_grad_norm=1.0, max_train_samples=None, max_train_steps=2500, mixed_precision=None, non_ema_revision=None, num_train_epochs=100, num_validation_images=4, original_image_column='file_name', output_dir='output_lr1e4', pretrained_model_name_or_path='/home/ach17654gk/.cache/huggingface/hub/models--runwayml--stable-diffusion-v1-5/snapshots/451f4fe16113bff5a5d2269ed5ad43b0592e9a14', push_to_hub=True, random_flip=False, report_to='wandb', resolution=512, resume_from_checkpoint='latest', revision=None, scale_lr=False, seed=0, train_batch_size=16, train_data_dir='/home/ach17654gk/download/dataset/DFC/Track_1', use_8bit_adam=False, use_ema=False, val_image_url='/home/ach17654gk/download/dataset/DFC/Track_1/png/rgb_images/TrainArea_1551.png', validation_epochs=1, validation_prompt='A high-resolution synthetic aperture radar (SAR) satellite image, heavy speckle noise, monochrome, moderate contrast, detailed geometric patterns, realistic radar texture, remote sensing style, urban area, distinct linear and rectangular structures', variant=None)
再用这套参数训2500