update liconn

Author: Donglai Wei

tutorials/neuron_nisb_40nm_common.yaml

@@ -1,195 +1,49 @@
-_base_: bases/mednext.yaml
+_base_: neuron_nisb_common.yaml
 
 experiment_name: neuron_nisb_40nm_common_mednext_b_sdt
 description: NISB neuron instance segmentation (BANIS-style) with MedNeXt-B, affinity + SDT (40nm)
 
-system:
-  training:
-    num_gpus: -1
-    num_workers: -1
-    batch_size: 4
-  inference:
-    num_gpus: 1
-    num_workers: 4
-    batch_size: 4
-  seed: 42
 
 model:
   input_size: [128, 128, 128]
   output_size: [128, 128, 128]
-  in_channels: 1
-  out_channels: 7
-  mednext_size: B
-  mednext_kernel_size: 3
-  mednext_dim: 3d
-  mednext_checkpoint_style: outside_block
-  deep_supervision: false
 
-  loss_functions:
-    - WeightedBCEWithLogitsLoss
-    - WeightedMSELoss
-  loss_weights: [1.0, 1.0]
-  loss_kwargs:
-    - {}
-    - {tanh: true}
-
-  multi_task_config:
-    - [0, 6, affinity, [0]]
-    - [6, 7, sdt, [1]]
 
 data:
-  # BANIS command mapping:
-  # --base_data_path /projects/weilab/dataset/nisb
-  # --data_setting base
-
-  # BANIS data.zarr arrays are under each seed directory.
   train_image: seed*/data.zarr/img_40-36-36nm.h5
   train_label: seed*/data.zarr/seg_40-36-36nm.h5
   val_image: seed*/data.zarr/img_40-36-36nm.h5
   val_label: seed*/data.zarr/seg_40-36-36nm.h5
-  iter_num_per_epoch: 200
 
   train_resolution: [36, 36, 40]
   val_resolution: [36, 36, 40]
 
   patch_size: [128, 128, 128]
-  use_preloaded_cache_train: true
-  use_preloaded_cache_val: true
-  use_cache: false
-  persistent_workers: true
 
-  image_transform:
-    normalize: "0-1"
-    clip_percentile_low: 0.0
-    clip_percentile_high: 1.0
 
   label_transform:
     targets:
       - name: affinity
         kwargs:
-          long_range: 10
+          long_range: 3
       - name: skeleton_aware_edt
         kwargs:
           resolution: [36, 36, 40]
           alpha: 0.8
           bg_value: -1.0
           relabel: true
 
-  augmentation:
-    preset: some
-
-    flip:
-      enabled: true
-      prob: 0.5
-
-    rotate:
-      enabled: true
-      prob: 0.5
-
-    affine:
-      enabled: true
-      prob: 0.5
-      rotate_range: [3.1416, 3.1416, 3.1416]
-      scale_range: [0.2, 0.2, 0.2]
-      shear_range: [0.5, 0.5, 0.5]
-
-    intensity:
-      enabled: true
-      gaussian_noise_prob: 0.5
-      gaussian_noise_std: 0.5
-      shift_intensity_prob: 0.5
-      shift_intensity_offset: 0.1
-      contrast_prob: 0.5
-      contrast_range: [0.9, 1.1]
-
-
-optimization:
-  max_epochs: 500
-  gradient_clip_val: 1.0
-  accumulate_grad_batches: 1
-  precision: "16-mixed"
-  log_every_n_steps: 100
-  val_check_interval: 1
-  num_sanity_val_steps: 0
-
-  optimizer:
-    name: AdamW
-    lr: 1.0e-3
-    weight_decay: 1.0e-2
-    betas: [0.9, 0.999]
-    eps: 1.0e-8
-
-  scheduler:
-    name: CosineAnnealingLR
-    t_max: 50000
-    interval: step
-    frequency: 1
-
+  
 monitor:
-  logging:
-    scalar:
-      loss: [train_loss_total_epoch, val_loss_total, train_loss_affinity_total, train_loss_sdt_total]
-      loss_every_n_steps: 100
-      val_check_interval: 1.0
-    images:
-      enabled: true
-      max_images: 8
-      num_slices: 8
-      log_every_n_epochs: 1
-      channel_mode: all
-
   checkpoint:
-    monitor: val_loss_total
-    mode: min
-    save_top_k: 100
-    save_last: true
-    save_every_n_epochs: 5
     dirpath: outputs/neuron_nisb_base_40nm_mednext_b_sdt/checkpoints/
-    use_timestamp: true
-
-  early_stopping:
-    enabled: false
-
+  
 test:
   data:
     test_image: seed101/data.zarr/img_40-36-36nm.h5
     test_label: seed101/data.zarr/seg_40-36-36nm.h5
     test_resolution: [36, 36, 40]
-    output_path: outputs/neuron_nisb_base_40nm_mednext_b_sdt/results_seed101/
-
-  decoding:
-    - name: decode_affinity_cc
-      kwargs:
-        threshold: 0.95        
-
-  evaluation:
-    enabled: false
 
 inference:
   sliding_window:
     window_size: [128, 128, 128]
-    sw_batch_size: 4
-    overlap: 0.5
-    blending: gaussian
-    sigma_scale: 0.25
-    padding_mode: replicate
-    keep_input_on_cpu: true
-    sw_device: cuda
-    output_device: cpu
-
-  test_time_augmentation:
-    enabled: false
-    flip_axes: null
-    rotation90_axes: null
-    channel_activations:
-      - [0, 6, sigmoid]
-      - [6, 7, tanh]
-    select_channel: [0,1,2,6]
-    ensemble_mode: mean
-    apply_mask: false
-
-  save_prediction:
-    enabled: true
-    intensity_scale: -1.0
-    intensity_dtype: float32
-    output_formats: [h5]

tutorials/neuron_nisb_40nm_liconn.yaml

@@ -8,7 +8,6 @@ system:
     batch_size: 4
 
 
-  
 data:
   train_path: /projects/weilab/dataset/nisb/liconn/train
   val_path: /projects/weilab/dataset/nisb/liconn/val
@@ -20,7 +19,7 @@ data:
           long_range: 3
       - name: skeleton_aware_edt
         kwargs:
-          resolution: [36, 36, 40]
+          resolution: [36, 36, 24]
           alpha: 0.8
           bg_value: -1.0
           relabel: true

tutorials/neuron_nisb_base.yaml tutorials/neuron_nisb_9nm_base.yamltutorials/neuron_nisb_base.yaml renamed to tutorials/neuron_nisb_9nm_base.yaml

@@ -0,0 +1,53 @@
+_base_: neuron_nisb_common.yaml
+
+experiment_name: neuron_nisb_9nm_common_mednext_b_sdt
+description: NISB neuron instance segmentation (BANIS-style) with MedNeXt-B, affinity + SDT (9nm)
+
+
+model:
+  input_size: [32, 256, 256]
+  output_size: [32, 256, 256]
+
+
+data:
+  train_image: seed*/data.zarr/img
+  train_label: seed*/data.zarr/seg
+  val_image: seed*/data.zarr/img
+  val_label: seed*/data.zarr/seg
+  patch_size: [32, 256, 256]
+  train_resolution: [20, 9, 9]
+  val_resolution: [20, 9, 9]
+  # NISB zarr arrays are stored as XYZ; model/data pipeline expects ZYX.
+  train_transpose: [2, 1, 0]
+  val_transpose: [2, 1, 0]
+  
+  
+  use_lazy_zarr: true
+  use_preloaded_cache_train: false
+  use_preloaded_cache_val: false
+  
+  label_transform:
+    targets:
+      - name: affinity
+        kwargs:
+          long_range: 3
+      - name: skeleton_aware_edt
+        kwargs:
+          resolution: [20, 9, 9]
+          alpha: 0.8
+          bg_value: -1.0
+          relabel: true
+
+monitor:
+  checkpoint:
+    dirpath: outputs/neuron_nisb_9nm_mednext_b_sdt/checkpoints/
+  
+test:
+  data:
+    test_image: seed101/data.zarr/img
+    test_label: seed101/data.zarr/seg
+    test_resolution: [20, 9, 9]
+
+inference:
+  sliding_window:
+    window_size: [32, 256, 256]

tutorials/neuron_nisb_9nm_common.yaml

@@ -0,0 +1,53 @@
+_base_: neuron_nisb_9nm_common.yaml
+
+experiment_name: neuron_nisb_9nm_liconn_mednext_b_sdt
+description: NISB Liconn neuron instance segmentation (BANIS-style) with MedNeXt-B, affinity + SDT (9nm)
+
+system:
+  training:
+    batch_size: 1
+
+
+
+model:
+  input_size: [192, 192, 192]
+  output_size: [192, 192, 192]
+
+data:
+  train_path: /projects/weilab/dataset/nisb/liconn/train
+  val_path: /projects/weilab/dataset/nisb/liconn/val
+
+  train_resolution: [12, 9, 9]
+  val_resolution: [12, 9, 9]  
+  patch_size: [192, 192, 192]
+  
+  
+
+  image_transform:
+    normalize: "0-1"
+    clip_percentile_low: 0.0
+    clip_percentile_high: 1.0
+
+  label_transform:
+    targets:
+      - name: affinity
+        kwargs:
+          long_range: 3
+      - name: skeleton_aware_edt
+        kwargs:
+          resolution: [12, 9, 9]
+          alpha: 0.8
+          bg_value: -1.0
+          relabel: true
+optimization:
+  accumulate_grad_batches: 2
+
+monitor:
+  checkpoint:
+    dirpath: outputs/neuron_nisb_9nm_liconn_mednext_b_sdt/checkpoints/
+
+test:
+  data:
+    test_path: /projects/weilab/dataset/nisb/liconn/test/
+    test_resolution: [12, 9, 9]
+    output_path: outputs/neuron_nisb_9nm_liconn_mednext_b_sdt/results_seed101/