checkpoint: update KhangTruong/Testing-model (periodic at epoch 4, step 7)

README.md

@@ -18,11 +18,11 @@ Official repository for **Testing-model** checkpoints, evaluated on AI-generated
 
 | Version | Best Metric Score | Best Epoch | Updated |
 | :--- | :--- | :--- | :--- |
-| `v1` | 0.6 | 1 | 2026-10-06 |
+| `v1` | 0.15213540285745106 | 10 | 2026-10-06 |
 
 - **Active Version:** `v1`
-- **Best Validation Score:** `0.6` (Epoch `1`)
-- **Total Training Epochs:** `4`
+- **Best Validation Score:** `0.15213540285745106` (Epoch `10`)
+- **Total Training Epochs:** `10`
 
 ---
 

checkpoints/checkpoint_best.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7b16091a0d393b82e510e0abd22b214f8ca99aacfc1bc6b0fcadee4c5ff5afdc
-size 218403
+oid sha256:7537716ccfb7cc5c6743eac53dcc70f8ccb9d968c1baaf5b9db58324f2030a71
+size 839778939

checkpoints/checkpoint_best_config.yaml

@@ -1,61 +1,92 @@
 project:
-  name: test_smoke
+  name: default
   seed: 42
-  device: cpu
-  output_dir: /tmp/tmp64tsilow/out/RUN/test_smoke
-  hub_repo: null
-  resume_repo: null
+  device: auto
+  output_dir: outputs/RUN/default
 data:
-  dataset_name: KhangTruong/IMD2020
+  dataset_name: KhangTruong/COCO-inpainted
   streaming: false
   image_size:
-  - 32
-  - 32
-  batch_size: 16
-  num_workers: 0
-  pin_memory: false
-  shuffle_buffer_size: 1000
+  - 256
+  - 256
+  batch_size: 8
+  num_workers: 2
+  pin_memory: true
+  shuffle_buffer_size: 32
   train_split: train
   val_split: validation
-  test_split: test
-  eval_split: test
-  train_samples_per_epoch: 2
+  test_split: validation
+  eval_split: validation
+  train_samples_per_epoch: -1
   val_samples_per_epoch: -1
-  val_samples: 2
-  test_samples: 2
+  val_samples: -1
+  test_samples: -1
   evaluate_on_test: true
   augmentations:
     horizontal_flip: 0.5
     vertical_flip: 0.0
-    random_rotate90: 0.0
+    random_rotate90: 0.25
     color_jitter: 0.0
-  base_batch_size: 16
-  num_gpus: 1
-  _batch_size_scaled: true
+  prefetch_batches: 48
 model:
-  name: unet
+  name: diffusion_diff_minimized
   in_channels: 3
   out_channels: 1
   features:
-  - 8
-  - 16
+  - 64
+  - 128
+  - 256
+  - 512
   bilinear: true
-  dropout: 0.0
-  aux_classifier: false
+  dropout: 0.1
+  aux_classifier: true
   num_classes: 3
   gradient_checkpointing: false
+  pretrained_model_name_or_path: segmind/tiny-sd
+  vae_subfolder: vae
+  unet_subfolder: unet
+  diffuser_fp16: true
+  device_map: cuda
+  freeze_encoder: true
+  use_encoder_skips: false
+  use_perpendicular_skips: true
+  timesteps:
+  - 250
+  timestep_embed_dim: 32
+  sigma_embed_dim: 32
+  include_noisy_latents: true
+  include_added_noise: true
+  include_predicted_noise: true
+  include_noise_diff: true
+  include_z0: true
+  scaling_factor: 0.18215
+  decoder:
+    channels:
+    - 256
+    - 128
+    - 64
+    - 32
+    upsample_mode: bilinear
+    norm_layer: batchnorm
+    z_norm: groupnorm
+    activation: silu
+    dropout: 0.1
+    num_res_blocks: 1
+  input_rescale: true
 loss:
   mask_loss_type: combined
   bce_weight: 0.5
   dice_weight: 0.5
   focal_weight: 0.5
   focal_gamma: 2.0
-  focal_alpha: 0.25
+  focal_alpha: 0.75
   aux_loss_type: cross_entropy
   aux_weight: 0.2
+  pos_weight: 1.0
+  aux_classifier: true
 training:
-  epochs: 1
-  learning_rate: 0.001
+  epochs: 10
+  learning_rate: 0.0003
   weight_decay: 0.0001
   optimizer: adamw
   use_8bit_optimizer: false
@@ -63,41 +94,27 @@ training:
   fallback_to_16bit: true
   fallback_on_unsupported_8bit: true
   scheduler: cosine
-  resume_lr_mode: auto
-  resume_lr: null
   warmup_epochs: 1
   min_lr: 1.0e-06
   grad_clip_norm: 1.0
-  amp: false
-  gradient_accumulation_steps: 1
-  data_parallel: false
-  use_hard_mining: false
+  amp: true
+  gradient_accumulation_steps: 4
   auto_batch_size: true
   empty_cache_per_epoch: true
   save_best: true
   save_latest: true
-  checkpoint_period: 3600
+  checkpoint_period: 1
   eval_interval: 1
-  early_stopping_patience: 2
+  early_stopping_patience: 4
   early_stopping_metric: val_iou
   early_stopping_mode: max
-  push_to_hub: true
-  hub_repo: KhangTruong/Testing-model
-  hub_version: v1
-  batch_size: 2
-  hf_repo: KhangTruong/Testing-model
-  num_gpus: 1
-hard_mining:
-  enabled: false
-  metric: median
-  reset_epochs: 5
 bootstrapping:
   enabled: false
   run_bootstrap: false
-  epochs: 5
+  epochs: 30
   num_samples: 512
   batch_size: null
-  learning_rate: null
+  learning_rate: 0.0005
   freeze_strategy: channel_stream
   stream_ratio: 0.5
   release_mode: restore
@@ -108,7 +125,7 @@ bootstrapping:
   early_stopping: false
   patience: 3
 logging:
-  log_interval: 1
+  log_interval: 20
   log_memory: true
   measure_network: true
 post_processing:

checkpoints/checkpoint_latest.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:365246b94cbc08ebd23281443e85ded06c1ab4486c1fb2a76f331088268b81a9
-size 221163
+oid sha256:91dc74c0b4bf4bdf528d475eabe0448157f00a7da3e9fc1a3822cd6cd364d465
+size 822846163

checkpoints/checkpoint_latest_config.yaml

@@ -1,61 +1,97 @@
 project:
-  name: test_smoke
+  name: default
   seed: 42
-  device: cpu
-  output_dir: /tmp/tmprsobsoue/out/RUN/test_smoke
+  device: auto
+  output_dir: outputs/RUN/default
   hub_repo: null
   resume_repo: null
 data:
-  dataset_name: KhangTruong/IMD2020
-  streaming: false
+  dataset_name: KhangTruong/COCO-inpainted
+  streaming: true
   image_size:
-  - 32
-  - 32
-  batch_size: 16
-  num_workers: 0
-  pin_memory: false
-  shuffle_buffer_size: 1000
+  - 256
+  - 256
+  batch_size: 128
+  num_workers: -1
+  pin_memory: true
+  shuffle_buffer_size: 32
   train_split: train
   val_split: validation
-  test_split: test
-  eval_split: test
-  train_samples_per_epoch: 2
+  test_split: validation
+  eval_split: validation
+  train_samples_per_epoch: -1
   val_samples_per_epoch: -1
-  val_samples: 2
-  test_samples: 2
+  val_samples: -1
+  test_samples: -1
   evaluate_on_test: true
   augmentations:
     horizontal_flip: 0.5
     vertical_flip: 0.0
-    random_rotate90: 0.0
+    random_rotate90: 0.25
     color_jitter: 0.0
-  base_batch_size: 16
-  num_gpus: 1
+  prefetch_batches: 48
+  base_batch_size: 128
+  num_gpus: 2
   _batch_size_scaled: true
 model:
-  name: unet
+  name: diffusion_diff_minimized
   in_channels: 3
   out_channels: 1
   features:
-  - 8
-  - 16
+  - 64
+  - 128
+  - 256
+  - 512
   bilinear: true
-  dropout: 0.0
-  aux_classifier: false
+  dropout: 0.1
+  aux_classifier: true
   num_classes: 3
   gradient_checkpointing: false
+  pretrained_model_name_or_path: segmind/tiny-sd
+  vae_subfolder: vae
+  unet_subfolder: unet
+  diffuser_fp16: true
+  device_map: cuda
+  freeze_encoder: true
+  use_encoder_skips: false
+  use_perpendicular_skips: true
+  timesteps:
+  - 250
+  timestep_embed_dim: 32
+  sigma_embed_dim: 32
+  include_noisy_latents: true
+  include_added_noise: true
+  include_predicted_noise: true
+  include_noise_diff: true
+  include_z0: true
+  scaling_factor: 0.18215
+  decoder:
+    channels:
+    - 256
+    - 128
+    - 64
+    - 32
+    upsample_mode: bilinear
+    norm_layer: batchnorm
+    z_norm: groupnorm
+    activation: silu
+    dropout: 0.1
+    num_res_blocks: 1
+  input_rescale: true
 loss:
   mask_loss_type: combined
   bce_weight: 0.5
   dice_weight: 0.5
   focal_weight: 0.5
   focal_gamma: 2.0
-  focal_alpha: 0.25
+  focal_alpha: 0.75
   aux_loss_type: cross_entropy
   aux_weight: 0.2
+  pos_weight: 1.0
+  aux_classifier: true
 training:
-  epochs: 1
-  learning_rate: 0.001
+  epochs: 40
+  learning_rate: 0.0003
   weight_decay: 0.0001
   optimizer: adamw
   use_8bit_optimizer: false
@@ -68,36 +104,35 @@ training:
   warmup_epochs: 1
   min_lr: 1.0e-06
   grad_clip_norm: 1.0
-  amp: false
-  gradient_accumulation_steps: 1
-  data_parallel: false
-  use_hard_mining: false
+  amp: true
+  gradient_accumulation_steps: 4
+  data_parallel: true
+  use_hard_mining: true
   auto_batch_size: true
   empty_cache_per_epoch: true
   save_best: true
   save_latest: true
-  checkpoint_period: 3600
+  checkpoint_period: 2m
   eval_interval: 1
-  early_stopping_patience: 2
+  early_stopping_patience: 4
   early_stopping_metric: val_iou
   early_stopping_mode: max
   push_to_hub: true
   hub_repo: KhangTruong/Testing-model
   hub_version: v1
-  batch_size: 2
   hf_repo: KhangTruong/Testing-model
-  num_gpus: 1
+  num_gpus: 2
 hard_mining:
-  enabled: false
+  enabled: true
   metric: median
   reset_epochs: 5
 bootstrapping:
   enabled: false
   run_bootstrap: false
-  epochs: 5
+  epochs: 30
   num_samples: 512
   batch_size: null
-  learning_rate: null
+  learning_rate: 0.0005
   freeze_strategy: channel_stream
   stream_ratio: 0.5
   release_mode: restore
@@ -108,7 +143,7 @@ bootstrapping:
   early_stopping: false
   patience: 3
 logging:
-  log_interval: 1
+  log_interval: 20
   log_memory: true
   measure_network: true
 post_processing:

checkpoints/checkpoint_periodic.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ad30390251f34136e39e38a260d2f7f3704b055caf400f940144ea86747bd3e2
-size 220211
+oid sha256:99c5b4384e04a2ed1d16b1246daf544da5140734f99ca2483e4554edcdb227ac
+size 822847589

checkpoints/checkpoint_periodic_config.yaml

@@ -1,61 +1,97 @@
 project:
-  name: test_smoke
+  name: default
   seed: 42
-  device: cpu
-  output_dir: /tmp/tmprsobsoue/out/RUN/test_smoke
+  device: auto
+  output_dir: outputs/RUN/default
   hub_repo: null
   resume_repo: null
 data:
-  dataset_name: KhangTruong/IMD2020
-  streaming: false
+  dataset_name: KhangTruong/COCO-inpainted
+  streaming: true
   image_size:
-  - 32
-  - 32
-  batch_size: 16
-  num_workers: 0
-  pin_memory: false
-  shuffle_buffer_size: 1000
+  - 256
+  - 256
+  batch_size: 128
+  num_workers: -1
+  pin_memory: true
+  shuffle_buffer_size: 32
   train_split: train
   val_split: validation
-  test_split: test
-  eval_split: test
-  train_samples_per_epoch: 2
+  test_split: validation
+  eval_split: validation
+  train_samples_per_epoch: -1
   val_samples_per_epoch: -1
-  val_samples: 2
-  test_samples: 2
+  val_samples: -1
+  test_samples: -1
   evaluate_on_test: true
   augmentations:
     horizontal_flip: 0.5
     vertical_flip: 0.0
-    random_rotate90: 0.0
+    random_rotate90: 0.25
     color_jitter: 0.0
-  base_batch_size: 16
-  num_gpus: 1
+  prefetch_batches: 48
+  base_batch_size: 128
+  num_gpus: 2
   _batch_size_scaled: true
 model:
-  name: unet
+  name: diffusion_diff_minimized
   in_channels: 3
   out_channels: 1
   features:
-  - 8
-  - 16
+  - 64
+  - 128
+  - 256
+  - 512
   bilinear: true
-  dropout: 0.0
-  aux_classifier: false
+  dropout: 0.1
+  aux_classifier: true
   num_classes: 3
   gradient_checkpointing: false
+  pretrained_model_name_or_path: segmind/tiny-sd
+  vae_subfolder: vae
+  unet_subfolder: unet
+  diffuser_fp16: true
+  device_map: cuda
+  freeze_encoder: true
+  use_encoder_skips: false
+  use_perpendicular_skips: true
+  timesteps:
+  - 250
+  timestep_embed_dim: 32
+  sigma_embed_dim: 32
+  include_noisy_latents: true
+  include_added_noise: true
+  include_predicted_noise: true
+  include_noise_diff: true
+  include_z0: true
+  scaling_factor: 0.18215
+  decoder:
+    channels:
+    - 256
+    - 128
+    - 64
+    - 32
+    upsample_mode: bilinear
+    norm_layer: batchnorm
+    z_norm: groupnorm
+    activation: silu
+    dropout: 0.1
+    num_res_blocks: 1
+  input_rescale: true
 loss:
   mask_loss_type: combined
   bce_weight: 0.5
   dice_weight: 0.5
   focal_weight: 0.5
   focal_gamma: 2.0
-  focal_alpha: 0.25
+  focal_alpha: 0.75
   aux_loss_type: cross_entropy
   aux_weight: 0.2
+  pos_weight: 1.0
+  aux_classifier: true
 training:
-  epochs: 1
-  learning_rate: 0.001
+  epochs: 40
+  learning_rate: 0.0003
   weight_decay: 0.0001
   optimizer: adamw
   use_8bit_optimizer: false
@@ -68,36 +104,35 @@ training:
   warmup_epochs: 1
   min_lr: 1.0e-06
   grad_clip_norm: 1.0
-  amp: false
-  gradient_accumulation_steps: 1
-  data_parallel: false
-  use_hard_mining: false
+  amp: true
+  gradient_accumulation_steps: 4
+  data_parallel: true
+  use_hard_mining: true
   auto_batch_size: true
   empty_cache_per_epoch: true
   save_best: true
   save_latest: true
-  checkpoint_period: 3600
+  checkpoint_period: 2m
   eval_interval: 1
-  early_stopping_patience: 2
+  early_stopping_patience: 4
   early_stopping_metric: val_iou
   early_stopping_mode: max
   push_to_hub: true
   hub_repo: KhangTruong/Testing-model
   hub_version: v1
-  batch_size: 2
   hf_repo: KhangTruong/Testing-model
-  num_gpus: 1
+  num_gpus: 2
 hard_mining:
-  enabled: false
+  enabled: true
   metric: median
   reset_epochs: 5
 bootstrapping:
   enabled: false
   run_bootstrap: false
-  epochs: 5
+  epochs: 30
   num_samples: 512
   batch_size: null
-  learning_rate: null
+  learning_rate: 0.0005
   freeze_strategy: channel_stream
   stream_ratio: 0.5
   release_mode: restore
@@ -108,7 +143,7 @@ bootstrapping:
   early_stopping: false
   patience: 3
 logging:
-  log_interval: 1
+  log_interval: 20
   log_memory: true
   measure_network: true
 post_processing:

checkpoints/effective_config.yaml

@@ -1,58 +1,94 @@
 project:
-  name: test_smoke
+  name: default
   seed: 42
-  device: cpu
-  output_dir: /tmp/tmprsobsoue/out/RUN/test_smoke
+  device: auto
+  output_dir: outputs/RUN/default
   hub_repo: null
   resume_repo: null
 data:
-  dataset_name: KhangTruong/IMD2020
-  streaming: false
+  dataset_name: KhangTruong/COCO-inpainted
+  streaming: true
   image_size:
-  - 32
-  - 32
-  batch_size: 16
-  num_workers: 0
-  pin_memory: false
-  shuffle_buffer_size: 1000
+  - 256
+  - 256
+  batch_size: 128
+  num_workers: -1
+  pin_memory: true
+  shuffle_buffer_size: 32
   train_split: train
   val_split: validation
-  test_split: test
-  eval_split: test
-  train_samples_per_epoch: 2
+  test_split: validation
+  eval_split: validation
+  train_samples_per_epoch: -1
   val_samples_per_epoch: -1
-  val_samples: 2
-  test_samples: 2
+  val_samples: -1
+  test_samples: -1
   evaluate_on_test: true
   augmentations:
     horizontal_flip: 0.5
     vertical_flip: 0.0
-    random_rotate90: 0.0
+    random_rotate90: 0.25
     color_jitter: 0.0
+  prefetch_batches: 48
 model:
-  name: unet
+  name: diffusion_diff_minimized
   in_channels: 3
   out_channels: 1
   features:
-  - 8
-  - 16
+  - 64
+  - 128
+  - 256
+  - 512
   bilinear: true
-  dropout: 0.0
-  aux_classifier: false
+  dropout: 0.1
+  aux_classifier: true
   num_classes: 3
   gradient_checkpointing: false
+  pretrained_model_name_or_path: segmind/tiny-sd
+  vae_subfolder: vae
+  unet_subfolder: unet
+  diffuser_fp16: true
+  device_map: cuda
+  freeze_encoder: true
+  use_encoder_skips: false
+  use_perpendicular_skips: true
+  timesteps:
+  - 250
+  timestep_embed_dim: 32
+  sigma_embed_dim: 32
+  include_noisy_latents: true
+  include_added_noise: true
+  include_predicted_noise: true
+  include_noise_diff: true
+  include_z0: true
+  scaling_factor: 0.18215
+  decoder:
+    channels:
+    - 256
+    - 128
+    - 64
+    - 32
+    upsample_mode: bilinear
+    norm_layer: batchnorm
+    z_norm: groupnorm
+    activation: silu
+    dropout: 0.1
+    num_res_blocks: 1
+  input_rescale: true
 loss:
   mask_loss_type: combined
   bce_weight: 0.5
   dice_weight: 0.5
   focal_weight: 0.5
   focal_gamma: 2.0
-  focal_alpha: 0.25
+  focal_alpha: 0.75
   aux_loss_type: cross_entropy
   aux_weight: 0.2
+  pos_weight: 1.0
+  aux_classifier: true
 training:
-  epochs: 1
-  learning_rate: 0.001
+  epochs: 40
+  learning_rate: 0.0003
   weight_decay: 0.0001
   optimizer: adamw
   use_8bit_optimizer: false
@@ -65,35 +101,34 @@ training:
   warmup_epochs: 1
   min_lr: 1.0e-06
   grad_clip_norm: 1.0
-  amp: false
-  gradient_accumulation_steps: 1
+  amp: true
+  gradient_accumulation_steps: 4
   data_parallel: true
-  use_hard_mining: false
+  use_hard_mining: true
   auto_batch_size: true
   empty_cache_per_epoch: true
   save_best: true
   save_latest: true
-  checkpoint_period: 3600
+  checkpoint_period: 2m
   eval_interval: 1
-  early_stopping_patience: 2
+  early_stopping_patience: 4
   early_stopping_metric: val_iou
   early_stopping_mode: max
   push_to_hub: true
   hub_repo: KhangTruong/Testing-model
   hub_version: v1
-  batch_size: 2
   hf_repo: KhangTruong/Testing-model
 hard_mining:
-  enabled: false
+  enabled: true
   metric: median
   reset_epochs: 5
 bootstrapping:
   enabled: false
   run_bootstrap: false
-  epochs: 5
+  epochs: 30
   num_samples: 512
   batch_size: null
-  learning_rate: null
+  learning_rate: 0.0005
   freeze_strategy: channel_stream
   stream_ratio: 0.5
   release_mode: restore
@@ -104,7 +139,7 @@ bootstrapping:
   early_stopping: false
   patience: 3
 logging:
-  log_interval: 1
+  log_interval: 20
   log_memory: true
   measure_network: true
 post_processing:

config.yaml

@@ -1,61 +1,97 @@
 project:
-  name: test_smoke
+  name: default
   seed: 42
-  device: cpu
-  output_dir: /tmp/tmprsobsoue/out/RUN/test_smoke
+  device: auto
+  output_dir: outputs/RUN/default
   hub_repo: null
   resume_repo: null
 data:
-  dataset_name: KhangTruong/IMD2020
-  streaming: false
+  dataset_name: KhangTruong/COCO-inpainted
+  streaming: true
   image_size:
-  - 32
-  - 32
-  batch_size: 16
-  num_workers: 0
-  pin_memory: false
-  shuffle_buffer_size: 1000
+  - 256
+  - 256
+  batch_size: 128
+  num_workers: -1
+  pin_memory: true
+  shuffle_buffer_size: 32
   train_split: train
   val_split: validation
-  test_split: test
-  eval_split: test
-  train_samples_per_epoch: 2
+  test_split: validation
+  eval_split: validation
+  train_samples_per_epoch: -1
   val_samples_per_epoch: -1
-  val_samples: 2
-  test_samples: 2
+  val_samples: -1
+  test_samples: -1
   evaluate_on_test: true
   augmentations:
     horizontal_flip: 0.5
     vertical_flip: 0.0
-    random_rotate90: 0.0
+    random_rotate90: 0.25
     color_jitter: 0.0
-  base_batch_size: 16
-  num_gpus: 1
+  prefetch_batches: 48
+  base_batch_size: 128
+  num_gpus: 2
   _batch_size_scaled: true
 model:
-  name: unet
+  name: diffusion_diff_minimized
   in_channels: 3
   out_channels: 1
   features:
-  - 8
-  - 16
+  - 64
+  - 128
+  - 256
+  - 512
   bilinear: true
-  dropout: 0.0
-  aux_classifier: false
+  dropout: 0.1
+  aux_classifier: true
   num_classes: 3
   gradient_checkpointing: false
+  pretrained_model_name_or_path: segmind/tiny-sd
+  vae_subfolder: vae
+  unet_subfolder: unet
+  diffuser_fp16: true
+  device_map: cuda
+  freeze_encoder: true
+  use_encoder_skips: false
+  use_perpendicular_skips: true
+  timesteps:
+  - 250
+  timestep_embed_dim: 32
+  sigma_embed_dim: 32
+  include_noisy_latents: true
+  include_added_noise: true
+  include_predicted_noise: true
+  include_noise_diff: true
+  include_z0: true
+  scaling_factor: 0.18215
+  decoder:
+    channels:
+    - 256
+    - 128
+    - 64
+    - 32
+    upsample_mode: bilinear
+    norm_layer: batchnorm
+    z_norm: groupnorm
+    activation: silu
+    dropout: 0.1
+    num_res_blocks: 1
+  input_rescale: true
 loss:
   mask_loss_type: combined
   bce_weight: 0.5
   dice_weight: 0.5
   focal_weight: 0.5
   focal_gamma: 2.0
-  focal_alpha: 0.25
+  focal_alpha: 0.75
   aux_loss_type: cross_entropy
   aux_weight: 0.2
+  pos_weight: 1.0
+  aux_classifier: true
 training:
-  epochs: 1
-  learning_rate: 0.001
+  epochs: 40
+  learning_rate: 0.0003
   weight_decay: 0.0001
   optimizer: adamw
   use_8bit_optimizer: false
@@ -68,36 +104,35 @@ training:
   warmup_epochs: 1
   min_lr: 1.0e-06
   grad_clip_norm: 1.0
-  amp: false
-  gradient_accumulation_steps: 1
-  data_parallel: false
-  use_hard_mining: false
+  amp: true
+  gradient_accumulation_steps: 4
+  data_parallel: true
+  use_hard_mining: true
   auto_batch_size: true
   empty_cache_per_epoch: true
   save_best: true
   save_latest: true
-  checkpoint_period: 3600
+  checkpoint_period: 2m
   eval_interval: 1
-  early_stopping_patience: 2
+  early_stopping_patience: 4
   early_stopping_metric: val_iou
   early_stopping_mode: max
   push_to_hub: true
   hub_repo: KhangTruong/Testing-model
   hub_version: v1
-  batch_size: 2
   hf_repo: KhangTruong/Testing-model
-  num_gpus: 1
+  num_gpus: 2
 hard_mining:
-  enabled: false
+  enabled: true
   metric: median
   reset_epochs: 5
 bootstrapping:
   enabled: false
   run_bootstrap: false
-  epochs: 5
+  epochs: 30
   num_samples: 512
   batch_size: null
-  learning_rate: null
+  learning_rate: 0.0005
   freeze_strategy: channel_stream
   stream_ratio: 0.5
   release_mode: restore
@@ -108,7 +143,7 @@ bootstrapping:
   early_stopping: false
   patience: 3
 logging:
-  log_interval: 1
+  log_interval: 20
   log_memory: true
   measure_network: true
 post_processing:

manifest.json

@@ -5,12 +5,13 @@
   "versions_meta": {
     "v1": {
       "version": "v1",
-      "best_score": "0.6",
-      "best_epoch": "1",
-      "total_epochs": "4",
-      "step": "7",
-      "updated_at": "2026-10-06T13:14:49.747047+00:00",
+      "best_score": "0.15213540285745106",
+      "best_epoch": "10",
+      "total_epochs": "10",
+      "step": "152730",
+      "updated_at": "2026-10-06T13:24:10.086971+00:00",
       "files": [
+        "checkpoint_best.pt",
         "checkpoint_latest.pt",
         "checkpoint_periodic.pt"
       ]

versions/v1/checkpoint_best_config.yaml

@@ -1,61 +1,92 @@
 project:
-  name: test_smoke
+  name: default
   seed: 42
-  device: cpu
-  output_dir: /tmp/tmp64tsilow/out/RUN/test_smoke
-  hub_repo: null
-  resume_repo: null
+  device: auto
+  output_dir: outputs/RUN/default
 data:
-  dataset_name: KhangTruong/IMD2020
+  dataset_name: KhangTruong/COCO-inpainted
   streaming: false
   image_size:
-  - 32
-  - 32
-  batch_size: 16
-  num_workers: 0
-  pin_memory: false
-  shuffle_buffer_size: 1000
+  - 256
+  - 256
+  batch_size: 8
+  num_workers: 2
+  pin_memory: true
+  shuffle_buffer_size: 32
   train_split: train
   val_split: validation
-  test_split: test
-  eval_split: test
-  train_samples_per_epoch: 2
+  test_split: validation
+  eval_split: validation
+  train_samples_per_epoch: -1
   val_samples_per_epoch: -1
-  val_samples: 2
-  test_samples: 2
+  val_samples: -1
+  test_samples: -1
   evaluate_on_test: true
   augmentations:
     horizontal_flip: 0.5
     vertical_flip: 0.0
-    random_rotate90: 0.0
+    random_rotate90: 0.25
     color_jitter: 0.0
-  base_batch_size: 16
-  num_gpus: 1
-  _batch_size_scaled: true
+  prefetch_batches: 48
 model:
-  name: unet
+  name: diffusion_diff_minimized
   in_channels: 3
   out_channels: 1
   features:
-  - 8
-  - 16
+  - 64
+  - 128
+  - 256
+  - 512
   bilinear: true
-  dropout: 0.0
-  aux_classifier: false
+  dropout: 0.1
+  aux_classifier: true
   num_classes: 3
   gradient_checkpointing: false
+  pretrained_model_name_or_path: segmind/tiny-sd
+  vae_subfolder: vae
+  unet_subfolder: unet
+  diffuser_fp16: true
+  device_map: cuda
+  freeze_encoder: true
+  use_encoder_skips: false
+  use_perpendicular_skips: true
+  timesteps:
+  - 250
+  timestep_embed_dim: 32
+  sigma_embed_dim: 32
+  include_noisy_latents: true
+  include_added_noise: true
+  include_predicted_noise: true
+  include_noise_diff: true
+  include_z0: true
+  scaling_factor: 0.18215
+  decoder:
+    channels:
+    - 256
+    - 128
+    - 64
+    - 32
+    upsample_mode: bilinear
+    norm_layer: batchnorm
+    z_norm: groupnorm
+    activation: silu
+    dropout: 0.1
+    num_res_blocks: 1
+  input_rescale: true
 loss:
   mask_loss_type: combined
   bce_weight: 0.5
   dice_weight: 0.5
   focal_weight: 0.5
   focal_gamma: 2.0
-  focal_alpha: 0.25
+  focal_alpha: 0.75
   aux_loss_type: cross_entropy
   aux_weight: 0.2
+  pos_weight: 1.0
+  aux_classifier: true
 training:
-  epochs: 1
-  learning_rate: 0.001
+  epochs: 10
+  learning_rate: 0.0003
   weight_decay: 0.0001
   optimizer: adamw
   use_8bit_optimizer: false
@@ -63,41 +94,27 @@ training:
   fallback_to_16bit: true
   fallback_on_unsupported_8bit: true
   scheduler: cosine
-  resume_lr_mode: auto
-  resume_lr: null
   warmup_epochs: 1
   min_lr: 1.0e-06
   grad_clip_norm: 1.0
-  amp: false
-  gradient_accumulation_steps: 1
-  data_parallel: false
-  use_hard_mining: false
+  amp: true
+  gradient_accumulation_steps: 4
   auto_batch_size: true
   empty_cache_per_epoch: true
   save_best: true
   save_latest: true
-  checkpoint_period: 3600
+  checkpoint_period: 1
   eval_interval: 1
-  early_stopping_patience: 2
+  early_stopping_patience: 4
   early_stopping_metric: val_iou
   early_stopping_mode: max
-  push_to_hub: true
-  hub_repo: KhangTruong/Testing-model
-  hub_version: v1
-  batch_size: 2
-  hf_repo: KhangTruong/Testing-model
-  num_gpus: 1
-hard_mining:
-  enabled: false
-  metric: median
-  reset_epochs: 5
 bootstrapping:
   enabled: false
   run_bootstrap: false
-  epochs: 5
+  epochs: 30
   num_samples: 512
   batch_size: null
-  learning_rate: null
+  learning_rate: 0.0005
   freeze_strategy: channel_stream
   stream_ratio: 0.5
   release_mode: restore
@@ -108,7 +125,7 @@ bootstrapping:
   early_stopping: false
   patience: 3
 logging:
-  log_interval: 1
+  log_interval: 20
   log_memory: true
   measure_network: true
 post_processing:

versions/v1/checkpoint_latest_config.yaml

@@ -1,61 +1,97 @@
 project:
-  name: test_smoke
+  name: default
   seed: 42
-  device: cpu
-  output_dir: /tmp/tmprsobsoue/out/RUN/test_smoke
+  device: auto
+  output_dir: outputs/RUN/default
   hub_repo: null
   resume_repo: null
 data:
-  dataset_name: KhangTruong/IMD2020
-  streaming: false
+  dataset_name: KhangTruong/COCO-inpainted
+  streaming: true
   image_size:
-  - 32
-  - 32
-  batch_size: 16
-  num_workers: 0
-  pin_memory: false
-  shuffle_buffer_size: 1000
+  - 256
+  - 256
+  batch_size: 128
+  num_workers: -1
+  pin_memory: true
+  shuffle_buffer_size: 32
   train_split: train
   val_split: validation
-  test_split: test
-  eval_split: test
-  train_samples_per_epoch: 2
+  test_split: validation
+  eval_split: validation
+  train_samples_per_epoch: -1
   val_samples_per_epoch: -1
-  val_samples: 2
-  test_samples: 2
+  val_samples: -1
+  test_samples: -1
   evaluate_on_test: true
   augmentations:
     horizontal_flip: 0.5
     vertical_flip: 0.0
-    random_rotate90: 0.0
+    random_rotate90: 0.25
     color_jitter: 0.0
-  base_batch_size: 16
-  num_gpus: 1
+  prefetch_batches: 48
+  base_batch_size: 128
+  num_gpus: 2
   _batch_size_scaled: true
 model:
-  name: unet
+  name: diffusion_diff_minimized
   in_channels: 3
   out_channels: 1
   features:
-  - 8
-  - 16
+  - 64
+  - 128
+  - 256
+  - 512
   bilinear: true
-  dropout: 0.0
-  aux_classifier: false
+  dropout: 0.1
+  aux_classifier: true
   num_classes: 3
   gradient_checkpointing: false
+  pretrained_model_name_or_path: segmind/tiny-sd
+  vae_subfolder: vae
+  unet_subfolder: unet
+  diffuser_fp16: true
+  device_map: cuda
+  freeze_encoder: true
+  use_encoder_skips: false
+  use_perpendicular_skips: true
+  timesteps:
+  - 250
+  timestep_embed_dim: 32
+  sigma_embed_dim: 32
+  include_noisy_latents: true
+  include_added_noise: true
+  include_predicted_noise: true
+  include_noise_diff: true
+  include_z0: true
+  scaling_factor: 0.18215
+  decoder:
+    channels:
+    - 256
+    - 128
+    - 64
+    - 32
+    upsample_mode: bilinear
+    norm_layer: batchnorm
+    z_norm: groupnorm
+    activation: silu
+    dropout: 0.1
+    num_res_blocks: 1
+  input_rescale: true
 loss:
   mask_loss_type: combined
   bce_weight: 0.5
   dice_weight: 0.5
   focal_weight: 0.5
   focal_gamma: 2.0
-  focal_alpha: 0.25
+  focal_alpha: 0.75
   aux_loss_type: cross_entropy
   aux_weight: 0.2
+  pos_weight: 1.0
+  aux_classifier: true
 training:
-  epochs: 1
-  learning_rate: 0.001
+  epochs: 40
+  learning_rate: 0.0003
   weight_decay: 0.0001
   optimizer: adamw
   use_8bit_optimizer: false
@@ -68,36 +104,35 @@ training:
   warmup_epochs: 1
   min_lr: 1.0e-06
   grad_clip_norm: 1.0
-  amp: false
-  gradient_accumulation_steps: 1
-  data_parallel: false
-  use_hard_mining: false
+  amp: true
+  gradient_accumulation_steps: 4
+  data_parallel: true
+  use_hard_mining: true
   auto_batch_size: true
   empty_cache_per_epoch: true
   save_best: true
   save_latest: true
-  checkpoint_period: 3600
+  checkpoint_period: 2m
   eval_interval: 1
-  early_stopping_patience: 2
+  early_stopping_patience: 4
   early_stopping_metric: val_iou
   early_stopping_mode: max
   push_to_hub: true
   hub_repo: KhangTruong/Testing-model
   hub_version: v1
-  batch_size: 2
   hf_repo: KhangTruong/Testing-model
-  num_gpus: 1
+  num_gpus: 2
 hard_mining:
-  enabled: false
+  enabled: true
   metric: median
   reset_epochs: 5
 bootstrapping:
   enabled: false
   run_bootstrap: false
-  epochs: 5
+  epochs: 30
   num_samples: 512
   batch_size: null
-  learning_rate: null
+  learning_rate: 0.0005
   freeze_strategy: channel_stream
   stream_ratio: 0.5
   release_mode: restore
@@ -108,7 +143,7 @@ bootstrapping:
   early_stopping: false
   patience: 3
 logging:
-  log_interval: 1
+  log_interval: 20
   log_memory: true
   measure_network: true
 post_processing:

versions/v1/checkpoint_periodic_config.yaml

@@ -1,61 +1,97 @@
 project:
-  name: test_smoke
+  name: default
   seed: 42
-  device: cpu
-  output_dir: /tmp/tmprsobsoue/out/RUN/test_smoke
+  device: auto
+  output_dir: outputs/RUN/default
   hub_repo: null
   resume_repo: null
 data:
-  dataset_name: KhangTruong/IMD2020
-  streaming: false
+  dataset_name: KhangTruong/COCO-inpainted
+  streaming: true
   image_size:
-  - 32
-  - 32
-  batch_size: 16
-  num_workers: 0
-  pin_memory: false
-  shuffle_buffer_size: 1000
+  - 256
+  - 256
+  batch_size: 128
+  num_workers: -1
+  pin_memory: true
+  shuffle_buffer_size: 32
   train_split: train
   val_split: validation
-  test_split: test
-  eval_split: test
-  train_samples_per_epoch: 2
+  test_split: validation
+  eval_split: validation
+  train_samples_per_epoch: -1
   val_samples_per_epoch: -1
-  val_samples: 2
-  test_samples: 2
+  val_samples: -1
+  test_samples: -1
   evaluate_on_test: true
   augmentations:
     horizontal_flip: 0.5
     vertical_flip: 0.0
-    random_rotate90: 0.0
+    random_rotate90: 0.25
     color_jitter: 0.0
-  base_batch_size: 16
-  num_gpus: 1
+  prefetch_batches: 48
+  base_batch_size: 128
+  num_gpus: 2
   _batch_size_scaled: true
 model:
-  name: unet
+  name: diffusion_diff_minimized
   in_channels: 3
   out_channels: 1
   features:
-  - 8
-  - 16
+  - 64
+  - 128
+  - 256
+  - 512
   bilinear: true
-  dropout: 0.0
-  aux_classifier: false
+  dropout: 0.1
+  aux_classifier: true
   num_classes: 3
   gradient_checkpointing: false
+  pretrained_model_name_or_path: segmind/tiny-sd
+  vae_subfolder: vae
+  unet_subfolder: unet
+  diffuser_fp16: true
+  device_map: cuda
+  freeze_encoder: true
+  use_encoder_skips: false
+  use_perpendicular_skips: true
+  timesteps:
+  - 250
+  timestep_embed_dim: 32
+  sigma_embed_dim: 32
+  include_noisy_latents: true
+  include_added_noise: true
+  include_predicted_noise: true
+  include_noise_diff: true
+  include_z0: true
+  scaling_factor: 0.18215
+  decoder:
+    channels:
+    - 256
+    - 128
+    - 64
+    - 32
+    upsample_mode: bilinear
+    norm_layer: batchnorm
+    z_norm: groupnorm
+    activation: silu
+    dropout: 0.1
+    num_res_blocks: 1
+  input_rescale: true
 loss:
   mask_loss_type: combined
   bce_weight: 0.5
   dice_weight: 0.5
   focal_weight: 0.5
   focal_gamma: 2.0
-  focal_alpha: 0.25
+  focal_alpha: 0.75
   aux_loss_type: cross_entropy
   aux_weight: 0.2
+  pos_weight: 1.0
+  aux_classifier: true
 training:
-  epochs: 1
-  learning_rate: 0.001
+  epochs: 40
+  learning_rate: 0.0003
   weight_decay: 0.0001
   optimizer: adamw
   use_8bit_optimizer: false
@@ -68,36 +104,35 @@ training:
   warmup_epochs: 1
   min_lr: 1.0e-06
   grad_clip_norm: 1.0
-  amp: false
-  gradient_accumulation_steps: 1
-  data_parallel: false
-  use_hard_mining: false
+  amp: true
+  gradient_accumulation_steps: 4
+  data_parallel: true
+  use_hard_mining: true
   auto_batch_size: true
   empty_cache_per_epoch: true
   save_best: true
   save_latest: true
-  checkpoint_period: 3600
+  checkpoint_period: 2m
   eval_interval: 1
-  early_stopping_patience: 2
+  early_stopping_patience: 4
   early_stopping_metric: val_iou
   early_stopping_mode: max
   push_to_hub: true
   hub_repo: KhangTruong/Testing-model
   hub_version: v1
-  batch_size: 2
   hf_repo: KhangTruong/Testing-model
-  num_gpus: 1
+  num_gpus: 2
 hard_mining:
-  enabled: false
+  enabled: true
   metric: median
   reset_epochs: 5
 bootstrapping:
   enabled: false
   run_bootstrap: false
-  epochs: 5
+  epochs: 30
   num_samples: 512
   batch_size: null
-  learning_rate: null
+  learning_rate: 0.0005
   freeze_strategy: channel_stream
   stream_ratio: 0.5
   release_mode: restore
@@ -108,7 +143,7 @@ bootstrapping:
   early_stopping: false
   patience: 3
 logging:
-  log_interval: 1
+  log_interval: 20
   log_memory: true
   measure_network: true
 post_processing:

versions/v1/checkpoints/checkpoint_best.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7b16091a0d393b82e510e0abd22b214f8ca99aacfc1bc6b0fcadee4c5ff5afdc
-size 218403
+oid sha256:7537716ccfb7cc5c6743eac53dcc70f8ccb9d968c1baaf5b9db58324f2030a71
+size 839778939

versions/v1/checkpoints/checkpoint_latest.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:365246b94cbc08ebd23281443e85ded06c1ab4486c1fb2a76f331088268b81a9
-size 221163
+oid sha256:91dc74c0b4bf4bdf528d475eabe0448157f00a7da3e9fc1a3822cd6cd364d465
+size 115763584

versions/v1/checkpoints/checkpoint_periodic.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ad30390251f34136e39e38a260d2f7f3704b055caf400f940144ea86747bd3e2
-size 220211
+oid sha256:99c5b4384e04a2ed1d16b1246daf544da5140734f99ca2483e4554edcdb227ac
+size 822847589

versions/v1/effective_config.yaml

@@ -1,58 +1,94 @@
 project:
-  name: test_smoke
+  name: default
   seed: 42
-  device: cpu
-  output_dir: /tmp/tmprsobsoue/out/RUN/test_smoke
+  device: auto
+  output_dir: outputs/RUN/default
   hub_repo: null
   resume_repo: null
 data:
-  dataset_name: KhangTruong/IMD2020
-  streaming: false
+  dataset_name: KhangTruong/COCO-inpainted
+  streaming: true
   image_size:
-  - 32
-  - 32
-  batch_size: 16
-  num_workers: 0
-  pin_memory: false
-  shuffle_buffer_size: 1000
+  - 256
+  - 256
+  batch_size: 128
+  num_workers: -1
+  pin_memory: true
+  shuffle_buffer_size: 32
   train_split: train
   val_split: validation
-  test_split: test
-  eval_split: test
-  train_samples_per_epoch: 2
+  test_split: validation
+  eval_split: validation
+  train_samples_per_epoch: -1
   val_samples_per_epoch: -1
-  val_samples: 2
-  test_samples: 2
+  val_samples: -1
+  test_samples: -1
   evaluate_on_test: true
   augmentations:
     horizontal_flip: 0.5
     vertical_flip: 0.0
-    random_rotate90: 0.0
+    random_rotate90: 0.25
     color_jitter: 0.0
+  prefetch_batches: 48
 model:
-  name: unet
+  name: diffusion_diff_minimized
   in_channels: 3
   out_channels: 1
   features:
-  - 8
-  - 16
+  - 64
+  - 128
+  - 256
+  - 512
   bilinear: true
-  dropout: 0.0
-  aux_classifier: false
+  dropout: 0.1
+  aux_classifier: true
   num_classes: 3
   gradient_checkpointing: false
+  pretrained_model_name_or_path: segmind/tiny-sd
+  vae_subfolder: vae
+  unet_subfolder: unet
+  diffuser_fp16: true
+  device_map: cuda
+  freeze_encoder: true
+  use_encoder_skips: false
+  use_perpendicular_skips: true
+  timesteps:
+  - 250
+  timestep_embed_dim: 32
+  sigma_embed_dim: 32
+  include_noisy_latents: true
+  include_added_noise: true
+  include_predicted_noise: true
+  include_noise_diff: true
+  include_z0: true
+  scaling_factor: 0.18215
+  decoder:
+    channels:
+    - 256
+    - 128
+    - 64
+    - 32
+    upsample_mode: bilinear
+    norm_layer: batchnorm
+    z_norm: groupnorm
+    activation: silu
+    dropout: 0.1
+    num_res_blocks: 1
+  input_rescale: true
 loss:
   mask_loss_type: combined
   bce_weight: 0.5
   dice_weight: 0.5
   focal_weight: 0.5
   focal_gamma: 2.0
-  focal_alpha: 0.25
+  focal_alpha: 0.75
   aux_loss_type: cross_entropy
   aux_weight: 0.2
+  pos_weight: 1.0
+  aux_classifier: true
 training:
-  epochs: 1
-  learning_rate: 0.001
+  epochs: 40
+  learning_rate: 0.0003
   weight_decay: 0.0001
   optimizer: adamw
   use_8bit_optimizer: false
@@ -65,35 +101,34 @@ training:
   warmup_epochs: 1
   min_lr: 1.0e-06
   grad_clip_norm: 1.0
-  amp: false
-  gradient_accumulation_steps: 1
+  amp: true
+  gradient_accumulation_steps: 4
   data_parallel: true
-  use_hard_mining: false
+  use_hard_mining: true
   auto_batch_size: true
   empty_cache_per_epoch: true
   save_best: true
   save_latest: true
-  checkpoint_period: 3600
+  checkpoint_period: 2m
   eval_interval: 1
-  early_stopping_patience: 2
+  early_stopping_patience: 4
   early_stopping_metric: val_iou
   early_stopping_mode: max
   push_to_hub: true
   hub_repo: KhangTruong/Testing-model
   hub_version: v1
-  batch_size: 2
   hf_repo: KhangTruong/Testing-model
 hard_mining:
-  enabled: false
+  enabled: true
   metric: median
   reset_epochs: 5
 bootstrapping:
   enabled: false
   run_bootstrap: false
-  epochs: 5
+  epochs: 30
   num_samples: 512
   batch_size: null
-  learning_rate: null
+  learning_rate: 0.0005
   freeze_strategy: channel_stream
   stream_ratio: 0.5
   release_mode: restore
@@ -104,7 +139,7 @@ bootstrapping:
   early_stopping: false
   patience: 3
 logging:
-  log_interval: 1
+  log_interval: 20
   log_memory: true
   measure_network: true
 post_processing:

versions/v1/manifest.json

@@ -5,12 +5,13 @@
   "versions_meta": {
     "v1": {
       "version": "v1",
-      "best_score": "0.6",
-      "best_epoch": "1",
-      "total_epochs": "4",
-      "step": "7",
-      "updated_at": "2026-10-06T13:14:49.747047+00:00",
+      "best_score": "0.15213540285745106",
+      "best_epoch": "10",
+      "total_epochs": "10",
+      "step": "152730",
+      "updated_at": "2026-10-06T13:24:10.086971+00:00",
       "files": [
+        "checkpoint_best.pt",
         "checkpoint_latest.pt",
         "checkpoint_periodic.pt"
       ]