vid_dropout_variance

[wr0124]

For video generation with Canny conditions, the option --alg_diffusion_vid_canny_dropout provides a range of dropout probabilities for the Canny dropout effect applied to each frame.

Example of execution. It works with:

python3 -W ignore::UserWarning  train.py \
--dataroot /path/to/online_mario2sonic_full_mario  \
--checkpoints_dir  /path/to/checkpoints \
--name mario_vid  \
--gpu_ids 3    \
--model_type palette \
--output_print_freq 1   \
--output_display_freq 1   \
--data_dataset_mode  self_supervised_temporal_labeled_mask_online  \
--train_batch_size 2  \
--train_iter_size 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_crop_size_A 32  \
--data_online_creation_crop_size_B 32 \
--data_crop_size 32 \
--data_load_size 32  \
--data_online_creation_rand_mask_A \
--train_G_lr 0.0001 \
--dataaug_no_rotate \
--G_diff_n_timestep_train  6  \
--G_diff_n_timestep_test  4  \
--data_temporal_number_frames 8  \
--data_temporal_frame_step 1 \
--data_online_creation_mask_delta_A_ratio 0.12 0.12 \
--alg_diffusion_cond_image_creation    computed_sketch  \
--alg_diffusion_cond_computed_sketch_list canny \
--alg_diffusion_vid_canny_dropout 0.1 0.8