jolibrain / joliGEN

Generative AI Image Toolset with GANs and Diffusion for Real-World Applications
https://www.joligen.com
Other
243 stars 32 forks source link

fix vid dataloader frame selection #705

Closed wr0124 closed 3 weeks ago

wr0124 commented 3 weeks ago

This dataloader generates a pool of available starting frame numbers for each video series and selects one from that pool. It works with: python3 -W ignore::UserWarning train.py \ --dataroot /path/to/data \ --checkpoints_dir /path/to/checkpoints \ --name test_debug \ --gpu_ids 0 \ --data_relative_paths \ --model_type palette \ --data_dataset_mode self_supervised_vid_mask_online \ --train_batch_size 1 \ --train_iter_size 1 \ --data_num_threads 1 \ --model_input_nc 3 \ --model_output_nc 3 \ --data_relative_paths \ --train_G_ema \ --train_optim adamw \ --G_netG unet_vid \ --data_online_creation_rand_mask_A \ --train_G_lr 0.0002 \ --dataaug_no_rotate \ --G_diff_n_timestep_train 8 \ --G_diff_n_timestep_test 3 \ --data_temporal_frame_step 1 \ --alg_diffusion_cond_image_creation computed_sketch \ --alg_diffusion_cond_computed_sketch_list canny \ --diffusion_cond_sketch_canny_range 0 150 \ --alg_diffusion_vid_canny_dropout 0.0 0.0 \ --output_print_freq 1 \ --output_display_freq 1 \ --data_temporal_number_frames 8 \ --data_online_creation_crop_size_A 32 \ --data_online_creation_crop_size_B 32 \ --data_crop_size 32 \ --data_load_size 32 \ --train_compute_metrics_test \ --train_metrics_every 1 \ --train_metrics_list PSNR LPIPS SSIM \