search

johannakarras / DreamPose

Official implementation of "DreamPose: Fashion Image-to-Video Synthesis via Stable Diffusion"
MIT License
962 stars 73 forks source link

CUDA out of memory #42

Open abdelrahmanabdelghany opened 1 year ago

abdelrahmanabdelghany commented 1 year ago

runing finetune-unet.py on colab but i get cuda out of memory issue. Is it possible to run this on colab ?

chenbolin-master commented 1 year ago

when running demo.py ,I also out of memory。

yuchen1984 commented 1 year ago

Same problem, what's the minimal GPU memory requirement for running the inference? I'm on 11GB GPU RAM. I also tried hacking the imSize to e.g 384x480 or 480x600 to work around the memory problem, but it doesn't work as the tensor dimensions in the demo model seems to be binded to 512x640

abdelrahmanabdelghany commented 1 year ago

running on colab pro A100, I think the minimum is 18 GB

kobybibas commented 11 months ago

When adding --gradient_checkpointing --use_8bit_adam the model training consumes 15GB although I'm not sure how it affects the results

accelerate launch finetune-unet.py --pretrained_model_name_or_path="CompVis/stable-diffusion-v1-4" --instance_data_dir=demo/sample/train --output_dir=demo/custom-chkpts --resolution=512 --train_batch_size=1 --gradient_accumulation_steps=1 --learning_rate=1e-5 --num_train_epochs=500 --dropout_rate=0.0 --custom_chkpt=checkpoints/unet_epoch_20.pth --revision "ebb811dd71cdc38a204ecbdd6ac5d580f529fd8c" --gradient_checkpointing --use_8bit_adam