Open YerongLi opened 2 weeks ago
I found with original training workflow, the loss is not decling, I am not sure this is because I am using a subset of the training set.
# File modified by authors of InstructDiffusion from original (https://github.com/CompVis/stable-diffusion). # See more details in LICENSE. model: base_learning_rate: 1.0e-04 weight_decay: 0.01 target: ldm.models.diffusion.ddpm_edit.LatentDiffusion params: fp16: True deepspeed: 'deepspeed_1' ckpt_path: stable_diffusion/models/ldm/stable-diffusion-v1/v1-5-pruned-emaonly-adaption-task.ckpt linear_start: 0.00085 linear_end: 0.0120 num_timesteps_cond: 1 log_every_t: 200 timesteps: 1000 first_stage_key: edited cond_stage_key: edit image_size: 32 channels: 4 cond_stage_trainable: false # Note: different from the one we trained before conditioning_key: hybrid monitor: val/loss_simple_ema scale_factor: 0.18215 scheduler_config: # 10000 warmup steps target: ldm.lr_scheduler.LambdaLinearScheduler params: warm_up_steps: [ 0 ] cycle_lengths: [ 10000000000000 ] # incredibly large number to prevent corner cases f_start: [ 1.e-6 ] f_max: [ 1. ] f_min: [ 1. ] unet_config: target: ldm.modules.diffusionmodules.openaimodel.UNetModel params: image_size: 32 # unused in_channels: 8 out_channels: 4 model_channels: 320 attention_resolutions: [ 4, 2, 1 ] num_res_blocks: 2 channel_mult: [ 1, 2, 4, 4 ] num_heads: 8 use_spatial_transformer: True transformer_depth: 1 context_dim: 768 use_checkpoint: True legacy: False force_type_convert: True first_stage_config: target: ldm.models.autoencoder.AutoencoderKL params: embed_dim: 4 monitor: val/rec_loss ddconfig: double_z: true z_channels: 4 resolution: 256 in_channels: 3 out_ch: 3 ch: 128 ch_mult: - 1 - 2 - 4 - 4 num_res_blocks: 2 attn_resolutions: [] dropout: 0.0 lossconfig: target: torch.nn.Identity cond_stage_config: target: ldm.modules.encoders.modules.FrozenCLIPEmbedder data: target: main.DataModuleFromConfig params: batch_size: 2 num_workers: 4 train: - ds1: target: dataset.editing.edit_zip_dataset.GIERDataset params: path: data/GIER_editing_data/ split: train min_resize_res: 256 max_resize_res: 256 crop_res: 256 flip_prob: 0.0 zip_start_index: 0 zip_end_index: 100 sample_weight: 2.0 validation: target: dataset.pose.pose.COCODataset params: root: data/coco/ image_set: val2017 is_train: False max_prompt_num: 5 min_prompt_num: 1 radius: 10 trainer: initial_scale: 13 max_epochs: 200 save_freq: 20 accumulate_grad_batches: 1 clip_grad: 0.0 optimizer: adamw
I don't get it why you are not logging the loss in the https://github.com/cientgu/InstructDiffusion/blob/435f0c3e4bd4dce4223df6ad4de1b511498a6138/main.py#L276
I found with original training workflow, the loss is not decling, I am not sure this is because I am using a subset of the training set.