pytorch_connectomics/tutorials/waterz_decoding_large_abiss.yaml at master · PytorchConnectomics/pytorch_connectomics · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
experiment_name: waterz_decode_large_abiss

description: >
  Large-volume chunked decoding using the vendored ABISS pipeline.

  Usage:
    - Prepare local precomputed affinity + WS/SEG outputs only:
      python scripts/run_abiss_large.py --config tutorials/waterz_decoding_large_abiss.yaml --prepare-only
    - Run the full ABISS flow:
      python scripts/run_abiss_large.py --config tutorials/waterz_decoding_large_abiss.yaml
    - Run selected stages after prepare:
      python scripts/run_abiss_large.py --config tutorials/waterz_decoding_large_abiss.yaml --skip-prepare --stages watershed remap_watershed

  This workflow does not use scripts/decode_large.py. ABISS already manages its own
  chunk hierarchy and chunked watershed / agglomeration.

abiss_large:
  abiss_home: /projects/weilab/weidf/lib/pytorch_connectomics/lib/abiss
  workdir: /projects/weilab/weidf/lib/pytorch_connectomics/outputs/waterz_decode_large_abiss/run
  secrets_dir: /projects/weilab/weidf/lib/pytorch_connectomics/outputs/waterz_decode_large_abiss/secrets
  param_path: /projects/weilab/weidf/lib/pytorch_connectomics/outputs/waterz_decode_large_abiss/secrets/param

  source_affinity_h5: /projects/weilab/dataset/liconn/mansour/DL288B_251222S_cond5_40x_12tiles_round1_fused_488_crop512x1024x1024_ffn_sharp_tta_x8_prediction_uint8.h5
  source_dataset: main

  # H5 -> precomputed conversion settings. These are storage chunks, not ABISS logical chunks.
  copy_block_shape_xyz: [512, 512, 64]
  aff_chunk_size_xyz: [512, 512, 64]
  seg_chunk_size_xyz: [512, 512, 80]
  resolution_xyz: [1, 1, 1]

  # ABISS runtime settings.
  root_tag: "4_0_0_0"
  top_mip: 4
  overlap_mode: 0
  meta_dirs: []

  param:
    NAME: DL288B_251222S_cond5_abiss_large

    # All ABISS paths are CloudFiles / CloudVolume paths. Use file:// for local runs.
    AFF_PATH: file:///projects/weilab/weidf/lib/pytorch_connectomics/outputs/waterz_decode_large_abiss/precomputed/affinity
    WS_PATH: file:///projects/weilab/weidf/lib/pytorch_connectomics/outputs/waterz_decode_large_abiss/precomputed/ws
    SEG_PATH: file:///projects/weilab/weidf/lib/pytorch_connectomics/outputs/waterz_decode_large_abiss/precomputed/seg
    SCRATCH_PATH: file:///projects/weilab/weidf/lib/pytorch_connectomics/outputs/waterz_decode_large_abiss/scratch
    CHUNKMAP_OUTPUT: file:///projects/weilab/weidf/lib/pytorch_connectomics/outputs/waterz_decode_large_abiss/chunkmap
    AFF_RESOLUTION: 0

    # Current full volume in XYZ coordinates.
    BBOX: [0, 0, 0, 2066, 2066, 1040]

    # ABISS logical chunk size in XYZ. This is intentionally large in XY to reduce
    # cross-face stitching while keeping Z chunking for memory.
    CHUNK_SIZE: [2066, 2066, 80]

    # Conservative watershed first; chunked agglomeration handles later reconnection.
    WS_HIGH_THRESHOLD: 0.9
    WS_LOW_THRESHOLD: 0.1
    WS_SIZE_THRESHOLD: 400
    WS_DUST_THRESHOLD: 200

    # Mean-edge agglomeration threshold.
    AGG_THRESHOLD: 0.20

    PARANOID: false
    CHUNKED_AGG_OUTPUT: false

notes:
  - "If you only want chunked watershed fragments, stop after remap_watershed."
  - "AFF_PATH, WS_PATH, and SEG_PATH are local precomputed volumes prepared by scripts/run_abiss_large.py."
  - "CHUNK_SIZE is the ABISS chunking grid; aff_chunk_size_xyz and seg_chunk_size_xyz are the on-disk precomputed storage chunks."
  - "seg_chunk_size_xyz must align with every internal CHUNK_SIZE boundary on axes that have multiple chunks; for this config the Z chunk size must be 80, not 64."
  - "With the upstream chunk_iterator package, root_tag should point at the top-level chunk. For this config that is 4_0_0_0."
  - "With BBOX=[0,0,0,2066,2066,1040] and CHUNK_SIZE=[2066,2066,80], the chunk grid is 1 x 1 x 13, so top_mip=4."