Training on a custom dataset

[ys830]

I appreciate your great work in zero123.

I want to retrain zero123 on medical data. My dataset contains about 700 samples, using the same data processing method as in the paper. Each sample has 12 different perspective images. I trained using a single A100 GPU. Below are my configuration file settings and training logs.

During inference, I used the saved 'last.ckpt' from training, but the generated image perspective control isn't very accurate. There's a deviation in perspective compared to the ground truth images. What could be the reason for this?

Could there be an issue with my configuration settings leading to a higher loss during training convergence?

model:
  base_learning_rate: 0.0001
  target: ldm.models.diffusion.ddpm.LatentDiffusion
  params:
    linear_start: 0.00085
    linear_end: 0.012
    num_timesteps_cond: 1
    log_every_t: 200
    timesteps: 1000
    first_stage_key: image_target
    cond_stage_key: image_cond
    image_size: 32
    channels: 4
    cond_stage_trainable: false
    conditioning_key: hybrid
    monitor: val/loss_simple_ema
    scale_factor: 0.18215
    scheduler_config:
      target: ldm.lr_scheduler.LambdaLinearScheduler
      params:
        warm_up_steps:
        - 100
        cycle_lengths:
        - 10000000000000
        f_start:
        - 1.0e-06
        f_max:
        - 1.0
        f_min:
        - 1.0
    unet_config:
      target: ldm.modules.diffusionmodules.openaimodel.UNetModel
      params:
        image_size: 32
        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
    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.FrozenCLIPImageEmbedder
data:
  target: ldm.data.simple.ObjaverseDataModuleFromConfig
  params:
    root_dir: /data/yisi/mywork/zero123/diffusioncardiac
    batch_size: 16
    num_workers: 8
    total_view: 4
    train:
      validation: false
      image_transforms:
        size: 256
    validation:
      validation: true
      image_transforms:
        size: 256
lightning:
  find_unused_parameters: false
  metrics_over_trainsteps_checkpoint: true
  modelcheckpoint:
    params:
      every_n_train_steps: 5000
  callbacks:
    image_logger:
      target: main.ImageLogger
      params:
        batch_frequency: 500
        max_images: 32
        increase_log_steps: false
        log_first_step: true
        log_images_kwargs:
          use_ema_scope: false
          inpaint: false
          plot_progressive_rows: false
          plot_diffusion_rows: false
          'N': 32
          unconditional_guidance_scale: 3.0
          unconditional_guidance_label:
          - ''
  trainer:
    benchmark: true
    val_check_interval: 5000000
    num_sanity_val_steps: 0
    accumulate_grad_batches: 1
    accelerator: ddp
    check_val_every_n_epoch: 10
    gpus: 0,

[trltzy]

I have the same problem, how to fix it?

[EduardoABarrera]

I am commenting for visibility since I am also having a related issue.

[Genshin-Impact-king]

can you help me how to prepare my own data to train?what data structure should I process?very thanks!!!

[Genshin-Impact-king]

I am commenting for visibility since I am also having a related issue.

can you help me how to prepare my own data to train?what data structure should I process?very thanks!!!

[kada0720]

@gitsawww Do you know how to use custom data to train now?

[BaochaoZhu]

@ys830 @trltzy @gitsawww @kada0720 Could you please assist me with preparing my data for training? What data structure should I be processing? Thank you.

[sungmin9939]

total view should be 12 since you prepared 12 different images for each object

[kuailexiaohunzi]

@ys830 @trltzy @gitsawww @kada0720 Could you please assist me with preparing my data for training? What data structure should I be processing? Thank you.

Have you solved this problem? I also have the same question