Closed chavinlo closed 2 years ago
When executing train.sh, it outputs the following:
root@centro:/glid-3-xl-stable# ./train.sh usage: image_train_stable.py [-h] [--data_dir DATA_DIR] [--schedule_sampler SCHEDULE_SAMPLER] [--lr LR] [--weight_decay WEIGHT_DECAY] [--lr_anneal_steps LR_ANNEAL_STEPS] [--batch_size BATCH_SIZE] [--microbatch MICROBATCH] [--ema_rate EMA_RATE] [--log_interval LOG_INTERVAL] [--save_interval SAVE_INTERVAL] [--resume_checkpoint RESUME_CHECKPOINT] [--use_fp16 USE_FP16] [--fp16_scale_growth FP16_SCALE_GROWTH] [--kl_model KL_MODEL] [--actual_image_size ACTUAL_IMAGE_SIZE] [--image_size IMAGE_SIZE] [--num_channels NUM_CHANNELS] [--num_res_blocks NUM_RES_BLOCKS] [--num_heads NUM_HEADS] [--num_heads_upsample NUM_HEADS_UPSAMPLE] [--num_head_channels NUM_HEAD_CHANNELS] [--attention_resolutions ATTENTION_RESOLUTIONS] [--channel_mult CHANNEL_MULT] [--dropout DROPOUT] [--class_cond CLASS_COND] [--use_checkpoint USE_CHECKPOINT] [--use_scale_shift_norm USE_SCALE_SHIFT_NORM] [--resblock_updown RESBLOCK_UPDOWN] [--use_spatial_transformer USE_SPATIAL_TRANSFORMER] [--context_dim CONTEXT_DIM] [--clip_embed_dim CLIP_EMBED_DIM] [--image_condition IMAGE_CONDITION] [--super_res_condition SUPER_RES_CONDITION] [--learn_sigma LEARN_SIGMA] [--diffusion_steps DIFFUSION_STEPS] [--noise_schedule NOISE_SCHEDULE] [--timestep_respacing TIMESTEP_RESPACING] [--use_kl USE_KL] [--predict_xstart PREDICT_XSTART] [--rescale_timesteps RESCALE_TIMESTEPS] [--rescale_learned_sigmas RESCALE_LEARNED_SIGMAS] image_train_stable.py: error: unrecognized arguments: --lr_warmup_steps 10000 root@centro:/glid-3-xl-stable#
Is this expected? Should I just remove that argument?
just pushed a change that should fix it. In my experience warmup is needed to stabilize training with fp16
Works now.
When executing train.sh, it outputs the following:
Is this expected? Should I just remove that argument?