search

whaohan / desigen

Official code for paper: Desigen: A Pipeline for Controllable Design Template Generation [CVPR'24]
44 stars 3 forks source link

Training code error #6

Closed aaghawaheed closed 1 day ago

aaghawaheed commented 2 weeks ago

accelerate launch train_background.py \ --pretrained_model_name_or_path=$MODEL_NAME \ --train_text_encoder \ --instance_data_dir=$INSTANCE_DIR \ --output_dir=$OUTPUT_DIR \ --resolution=512 \ --train_batch_size=3 \ --gradient_checkpointing \ --learning_rate=1e-5 \ --lr_scheduler="constant" \ --lr_warmup_steps=500 \ --num_train_epochs=50 \ --with_spatial_loss \ --checkpointing_steps=10000

"I'm encountering an error while running my training script (train_background.py). The error message states:

Traceback (most recent call last): File "train_background.py", line 968, in main(args) File "train_background.py", line 908, in main attn = torch.stack(unet.module.attn_store) # [5, 8*bs, 77, 16, 16] RuntimeError: stack expects a non-empty TensorList Steps: 0%|

If I turn off the with_spatial_loss parameter its working fine.

whaohan commented 2 weeks ago

Hi, it looks like the saliency maps of background images are not correctly processed and resized. Did you preprocess the saliency map correctly based on the provided scripts here?