checkpoint: update KhangTruong/diffusion-diff-minimized-512 (periodic at epoch 3, step 3643)

checkpoints/checkpoint_latest.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e31721b8d4d4924f93783aa69ba2cf15e0509003b43ecc1f52ffda1840bb5ed8
+oid sha256:51e556e7f70ca958639cc30e28ebb9fbed47aa98b1d9fc0878b0132938df70c9
 size 207855059

checkpoints/checkpoint_periodic.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d1c0d17075c77b13de26aadb02f935799b336ea7bcd31bceee67a46eefe303fc
+size 207856373

checkpoints/checkpoint_periodic_config.yaml

@@ -0,0 +1,157 @@
+project:
+  name: diffusion_diff_minimized_512
+  seed: 42
+  device: auto
+  output_dir: outputs/RUN/diffusion_diff_minimized_512
+  hub_repo: null
+  resume_repo: null
+data:
+  dataset_name: KhangTruong/DEAL300K-30K
+  streaming: true
+  image_size:
+  - 512
+  - 512
+  batch_size: 16
+  num_workers: -1
+  pin_memory: true
+  shuffle_buffer_size: 32
+  train_split: train
+  val_split: validation
+  test_split: validation
+  eval_split: validation
+  train_samples_per_epoch: -1
+  val_samples_per_epoch: -1
+  val_samples: -1
+  test_samples: -1
+  evaluate_on_test: true
+  augmentations:
+    horizontal_flip: 0.5
+    vertical_flip: 0.0
+    random_rotate90: 0.25
+    color_jitter: 0.0
+  prefetch_batches: 48
+  save_to_disk: true
+  cached_hf_repo: KhangTruong/DEAL300K-30K-cache
+  base_batch_size: 64
+  num_gpus: 1
+  _batch_size_scaled: true
+model:
+  name: diffusion_diff_minimized
+  in_channels: 3
+  out_channels: 1
+  features:
+  - 64
+  - 128
+  - 256
+  - 512
+  bilinear: true
+  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: 0.75
+  aux_loss_type: cross_entropy
+  aux_weight: 0.2
+  pos_weight: 1.0
+  aux_classifier: true
+training:
+  epochs: 10
+  learning_rate: 0.0003
+  weight_decay: 0.0001
+  optimizer: adamw
+  use_8bit_optimizer: false
+  check_8bit_compatibility: true
+  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: true
+  gradient_accumulation_steps: 16
+  data_parallel: false
+  use_hard_mining: true
+  auto_batch_size: true
+  empty_cache_per_epoch: true
+  save_best: true
+  save_latest: true
+  checkpoint_period: 3600
+  eval_interval: 1
+  early_stopping_patience: 4
+  early_stopping_metric: val_iou
+  early_stopping_mode: max
+  push_to_hub: true
+  hub_repo: KhangTruong/diffusion-diff-minimized-512
+  hub_version: v1
+  hf_repo: KhangTruong/diffusion-diff-minimized-512
+  num_gpus: 1
+hard_mining:
+  enabled: true
+  metric: median
+  reset_epochs: 5
+bootstrapping:
+  enabled: false
+  run_bootstrap: false
+  epochs: 30
+  num_samples: 512
+  batch_size: null
+  learning_rate: 0.0005
+  freeze_strategy: channel_stream
+  stream_ratio: 0.5
+  release_mode: restore
+  freeze_modules: []
+  initialization: kaiming_normal
+  target_score: 0.5
+  min_loss_drop: 0.15
+  early_stopping: false
+  patience: 3
+logging:
+  log_interval: 20
+  log_memory: true
+  measure_network: true
+post_processing:
+  enabled: true
+  min_area: 64
+  fill_holes: true
+  max_hole_size: 256
+  morphology: open_close
+  morph_kernel_size: 3

manifest.json

@@ -9,10 +9,11 @@
       "best_epoch": "2",
       "total_epochs": "2",
       "step": "2813",
-      "updated_at": "2026-11-02T07:46:30.771857+00:00",
+      "updated_at": "2026-11-02T07:58:53.430514+00:00",
       "files": [
         "checkpoint_best.pt",
-        "checkpoint_latest.pt"
+        "checkpoint_latest.pt",
+        "checkpoint_periodic.pt"
       ]
     }
   },

versions/v1/checkpoint_periodic_config.yaml

@@ -0,0 +1,157 @@
+project:
+  name: diffusion_diff_minimized_512
+  seed: 42
+  device: auto
+  output_dir: outputs/RUN/diffusion_diff_minimized_512
+  hub_repo: null
+  resume_repo: null
+data:
+  dataset_name: KhangTruong/DEAL300K-30K
+  streaming: true
+  image_size:
+  - 512
+  - 512
+  batch_size: 16
+  num_workers: -1
+  pin_memory: true
+  shuffle_buffer_size: 32
+  train_split: train
+  val_split: validation
+  test_split: validation
+  eval_split: validation
+  train_samples_per_epoch: -1
+  val_samples_per_epoch: -1
+  val_samples: -1
+  test_samples: -1
+  evaluate_on_test: true
+  augmentations:
+    horizontal_flip: 0.5
+    vertical_flip: 0.0
+    random_rotate90: 0.25
+    color_jitter: 0.0
+  prefetch_batches: 48
+  save_to_disk: true
+  cached_hf_repo: KhangTruong/DEAL300K-30K-cache
+  base_batch_size: 64
+  num_gpus: 1
+  _batch_size_scaled: true
+model:
+  name: diffusion_diff_minimized
+  in_channels: 3
+  out_channels: 1
+  features:
+  - 64
+  - 128
+  - 256
+  - 512
+  bilinear: true
+  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: 0.75
+  aux_loss_type: cross_entropy
+  aux_weight: 0.2
+  pos_weight: 1.0
+  aux_classifier: true
+training:
+  epochs: 10
+  learning_rate: 0.0003
+  weight_decay: 0.0001
+  optimizer: adamw
+  use_8bit_optimizer: false
+  check_8bit_compatibility: true
+  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: true
+  gradient_accumulation_steps: 16
+  data_parallel: false
+  use_hard_mining: true
+  auto_batch_size: true
+  empty_cache_per_epoch: true
+  save_best: true
+  save_latest: true
+  checkpoint_period: 3600
+  eval_interval: 1
+  early_stopping_patience: 4
+  early_stopping_metric: val_iou
+  early_stopping_mode: max
+  push_to_hub: true
+  hub_repo: KhangTruong/diffusion-diff-minimized-512
+  hub_version: v1
+  hf_repo: KhangTruong/diffusion-diff-minimized-512
+  num_gpus: 1
+hard_mining:
+  enabled: true
+  metric: median
+  reset_epochs: 5
+bootstrapping:
+  enabled: false
+  run_bootstrap: false
+  epochs: 30
+  num_samples: 512
+  batch_size: null
+  learning_rate: 0.0005
+  freeze_strategy: channel_stream
+  stream_ratio: 0.5
+  release_mode: restore
+  freeze_modules: []
+  initialization: kaiming_normal
+  target_score: 0.5
+  min_loss_drop: 0.15
+  early_stopping: false
+  patience: 3
+logging:
+  log_interval: 20
+  log_memory: true
+  measure_network: true
+post_processing:
+  enabled: true
+  min_area: 64
+  fill_holes: true
+  max_hole_size: 256
+  morphology: open_close
+  morph_kernel_size: 3

versions/v1/checkpoints/checkpoint_latest.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e31721b8d4d4924f93783aa69ba2cf15e0509003b43ecc1f52ffda1840bb5ed8
+oid sha256:51e556e7f70ca958639cc30e28ebb9fbed47aa98b1d9fc0878b0132938df70c9
 size 207855059

versions/v1/checkpoints/checkpoint_periodic.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d1c0d17075c77b13de26aadb02f935799b336ea7bcd31bceee67a46eefe303fc
+size 207856373

versions/v1/manifest.json

@@ -9,10 +9,11 @@
       "best_epoch": "2",
       "total_epochs": "2",
       "step": "2813",
-      "updated_at": "2026-11-02T07:46:30.771857+00:00",
+      "updated_at": "2026-11-02T07:58:53.430514+00:00",
       "files": [
         "checkpoint_best.pt",
-        "checkpoint_latest.pt"
+        "checkpoint_latest.pt",
+        "checkpoint_periodic.pt"
       ]
     }
   },