[convnext] add convnext distillation

dinov3/configs/ssl_default_config.yaml

@@ -169,6 +169,7 @@ crops:
   - 0.229
   - 0.224
   - 0.225
+  teacher_to_student_resolution_scale: 1.0
 evaluation:
   eval_period_iterations: 12500
   low_freq_every: 5

dinov3/configs/train/distillation_convnext/convnext_base_p16.yaml

@@ -0,0 +1,43 @@
+ibot:
+  loss_weight: 1.0
+  mask_sample_probability: 0.5
+  mask_ratio_min_max:
+  - 0.1
+  - 0.5
+  mask_random_circular_shift: false
+  force_masking_even_with_zero_weight: false
+  separate_head: true
+  head_norm_last_layer: false
+  head_nlayers: 3
+  head_hidden_dim: 2048
+student:
+  arch: convnext_base
+  patch_size: 16
+  drop_path_rate: 0.0
+  block_chunks: 4
+optim:
+  epochs: 500
+  clip_grad: 3.0
+  layerwise_decay: 1.0
+schedules:
+  lr:
+    start: 1e-6
+    peak: 1e-4
+    end: 1e-6
+    warmup_epochs: 80
+    freeze_last_layer_epochs: 1
+  weight_decay:
+    start: 0.02
+    end: 0.2
+    peak: 0.2
+    warmup_epochs: 500
+  teacher_temp:
+    start: 0.04
+    peak: 0.07
+    end: 0.07
+    warmup_epochs: 120
+  momentum:
+    start: 0.994
+    peak: 1.0
+    end: 1.0
+    warmup_epochs: 500

dinov3/configs/train/distillation_convnext/convnext_large_p16.yaml

@@ -0,0 +1,43 @@
+ibot:
+  loss_weight: 1.0
+  mask_sample_probability: 0.5
+  mask_ratio_min_max:
+  - 0.1
+  - 0.5
+  mask_random_circular_shift: false
+  force_masking_even_with_zero_weight: false
+  separate_head: true
+  head_norm_last_layer: false
+  head_nlayers: 3
+  head_hidden_dim: 2048
+student:
+  arch: convnext_large
+  patch_size: 16
+  drop_path_rate: 0.0
+  block_chunks: 4
+optim:
+  epochs: 500
+  clip_grad: 3.0
+  layerwise_decay: 1.0
+schedules:
+  lr:
+    start: 1e-6
+    peak: 1e-4
+    end: 1e-6
+    warmup_epochs: 80
+    freeze_last_layer_epochs: 1
+  weight_decay:
+    start: 0.04
+    end: 0.2
+    peak: 0.2
+    warmup_epochs: 500
+  teacher_temp:
+    start: 0.04
+    peak: 0.07
+    end: 0.07
+    warmup_epochs: 120
+  momentum:
+    start: 0.994
+    peak: 1.0
+    end: 1.0
+    warmup_epochs: 500

dinov3/configs/train/distillation_convnext/convnext_small_p16.yaml

@@ -0,0 +1,43 @@
+ibot:
+  loss_weight: 1.0
+  mask_sample_probability: 0.5
+  mask_ratio_min_max:
+  - 0.1
+  - 0.5
+  mask_random_circular_shift: false
+  force_masking_even_with_zero_weight: false
+  separate_head: true
+  head_norm_last_layer: false
+  head_nlayers: 3
+  head_hidden_dim: 2048
+student:
+  arch: convnext_small
+  patch_size: 16
+  drop_path_rate: 0.0
+  block_chunks: 4
+optim:
+  epochs: 500
+  clip_grad: 3.0
+  layerwise_decay: 1.0
+schedules:
+  lr:
+    start: 1e-6
+    peak: 2e-4
+    end: 1e-6
+    warmup_epochs: 80
+    freeze_last_layer_epochs: 1
+  weight_decay:
+    start: 0.04
+    end: 0.2
+    peak: 0.2
+    warmup_epochs: 500
+  teacher_temp:
+    start: 0.04
+    peak: 0.07
+    end: 0.07
+    warmup_epochs: 120
+  momentum:
+    start: 0.994
+    peak: 1.0
+    end: 1.0
+    warmup_epochs: 500

dinov3/configs/train/distillation_convnext/convnext_tiny_p16.yaml

@@ -0,0 +1,43 @@
+ibot:
+  loss_weight: 1.0
+  mask_sample_probability: 0.5
+  mask_ratio_min_max:
+  - 0.1
+  - 0.5
+  mask_random_circular_shift: false
+  force_masking_even_with_zero_weight: false
+  separate_head: true
+  head_norm_last_layer: false
+  head_nlayers: 3
+  head_hidden_dim: 2048
+student:
+  arch: convnext_tiny
+  patch_size: 16
+  drop_path_rate: 0.0
+  block_chunks: 4
+optim:
+  epochs: 500
+  clip_grad: 3.0
+  layerwise_decay: 1.0
+schedules:
+  lr:
+    start: 1e-6
+    peak: 2e-4
+    end: 1e-6
+    warmup_epochs: 80
+    freeze_last_layer_epochs: 1
+  weight_decay:
+    start: 0.04
+    end: 0.2
+    peak: 0.2
+    warmup_epochs: 500
+  teacher_temp:
+    start: 0.04
+    peak: 0.07
+    end: 0.07
+    warmup_epochs: 120
+  momentum:
+    start: 0.994
+    peak: 1.0
+    end: 1.0
+    warmup_epochs: 500

dinov3/configs/train/distillation_convnext/multi_distillation_convnext_test.yaml

@@ -0,0 +1,41 @@
+MODEL:
+  META_ARCHITECTURE: MultiDistillationMetaArch
+multidistillation:
+  enabled: true
+  global_batch_size: 32  # 4096 for 16 nodes
+  students:
+  - name: convnext_tiny
+    config_path: dinov3/configs/train/distillation_convnext/convnext_tiny_p16.yaml
+    ranks_range:
+    - 0
+    - 2
+  - name: convnext_small
+    config_path: dinov3/configs/train/distillation_convnext/convnext_small_p16.yaml
+    ranks_range:
+    - 2
+    - 4
+  - name: convnext_base
+    config_path: dinov3/configs/train/distillation_convnext/convnext_base_p16.yaml
+    ranks_range:
+    - 4
+    - 6
+  - name: convnext_large
+    config_path: dinov3/configs/train/distillation_convnext/convnext_large_p16.yaml
+    ranks_range:
+    - 6
+    - 8
+distillation:  # teacher
+  enabled: true
+  full_cfg_path: dinov3/configs/train/vitl_im1k_lin834.yaml
+  checkpoint_path: ignore
+crops:
+  global_crops_size: 512
+  local_crops_size: 224
+  teacher_to_student_resolution_scale: 2.0
+train:
+  dataset_path: ImageNet:split=TRAIN
+  cache_dataset: false
+  centering: "sinkhorn_knopp"
+  compile: true
+ibot:
+  separate_head: true