HaiwangYu commited on
Commit
1dd731b
·
1 Parent(s): 2db7897

test with a readme

Browse files
Files changed (1) hide show
  1. dnn-roi/pdhd/20260601/README.md +125 -0
dnn-roi/pdhd/20260601/README.md ADDED
@@ -0,0 +1,125 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # PDHD MobileNetV3-UNet baseline ROI models
2
+
3
+ ## Author
4
+ Hokyeong Nam, gkddk33@gmail.com
5
+
6
+ ## Description
7
+
8
+ This document records the CVMFS README metadata for three TorchScript
9
+ **MobileNetV3-UNet baseline** DNN-ROI models trained for ProtoDUNE-HD (PDHD)
10
+ signal-processing studies. It summarizes the training and preprocessing choices
11
+ needed by DUNE/Wire-Cell users to select and configure the correct model.
12
+
13
+ ## Model files
14
+
15
+ | File | Architecture | Training events | Time rebin | Truth threshold | Truth padding | Training planes | Input channels |
16
+ | --- | --- | ---: | ---: | ---: | --- | --- | ---: |
17
+ | `mobileunetv3-cosmic590-deposplat-pdhd-rebin6-pad1-uv.ts` | MobileNetV3-UNet baseline | 590 | 6 | 100 | enabled | U + V | 3 |
18
+ | `mobileunetv3-cosmic590-deposplat-pdhd-rebin4-pad1-uv.ts` | MobileNetV3-UNet baseline | 590 | 4 | 100 | enabled | U + V | 3 |
19
+ | `mobileunetv3-cosmic590-deposplat-pdhd-rebin10-pad0-uv.ts` | MobileNetV3-UNet baseline | 590 | 10 | 100 | disabled | U + V | 3 |
20
+
21
+ The filename tokens encode the model family, training/data variant, detector
22
+ sample, time-rebinning configuration, truth-label padding configuration, and
23
+ training plane set.
24
+
25
+ ## Training and preprocessing metadata
26
+
27
+ All three models were produced with the same baseline workflow except for the
28
+ `rebin` and padding choices listed above.
29
+
30
+ ### Detector sample, planes, and event count
31
+
32
+ - Training sample: **PDHD** training sample.
33
+ - Number of training events: **590**.
34
+ - Planes used for training: **U plane and V plane only** (`UV`).
35
+ - The models were trained as per-plane ROI predictors; the collection plane was
36
+ not included in this model set.
37
+
38
+ ### Input tensor
39
+
40
+ The network input has **three channels**:
41
+
42
+ 1. `loose_lf`
43
+ 2. `mp2_roi`
44
+ 3. `mp3_roi`
45
+
46
+ When connecting these models in a Wire-Cell graph, make sure the three input
47
+ frames/tensors are supplied to the model.
48
+
49
+ ### Truth-label construction
50
+
51
+ - Truth source: PDHD truth/deposition image used by the baseline preprocessing.
52
+ - Truth threshold: `100` for all three models.
53
+ - Thresholding was applied consistently for U-plane and V-plane training samples.
54
+
55
+ ### Padding configuration
56
+
57
+ The rebin-4 and rebin-6 models use padded truth labels:
58
+
59
+ ```text
60
+ padding = 1
61
+ padding_side = both
62
+ min_run = 1
63
+ avoid_merge = enabled
64
+ ```
65
+
66
+ The rebin-10 model does **not** add extra padding to the truth labels:
67
+
68
+ ```text
69
+ padding = 0
70
+ ```
71
+
72
+ ## Choosing a model
73
+
74
+ Use the model whose `rebin` value matches the preprocessing expected by the
75
+ Wire-Cell deployment or validation chain:
76
+
77
+ - Use `mobileunetv3-cosmic590-deposplat-pdhd-rebin4-pad1-uv.ts` for rebin-4 input.
78
+ - Use `mobileunetv3-cosmic590-deposplat-pdhd-rebin6-pad1-uv.ts` for rebin-6 input.
79
+ - Use `mobileunetv3-cosmic590-deposplat-pdhd-rebin10-pad0-uv.ts` for rebin-10 input.
80
+
81
+ Do not mix a model with a different time-rebinning configuration unless the
82
+ consumer explicitly converts the input tensors to the model's training binning.
83
+
84
+ ## Expected inference contract
85
+
86
+ - Format: PyTorch TorchScript (`.ts`).
87
+ - Model family: MobileNetV3 encoder with U-Net-style decoder.
88
+ - Inputs: one batch of three-channel ROI/SP feature images.
89
+ - Output: one ROI probability/mask image in the corresponding rebinned time
90
+ coordinate system.
91
+
92
+ ## Minimal provenance record
93
+
94
+ ```yaml
95
+ model_family: MobileNetV3-UNet baseline
96
+ format: torchscript
97
+ sample: PDHD
98
+ training_events: 590
99
+ training_variant: cosmic590-deposplat
100
+ planes: [U, V]
101
+ input_channels: [loose_lf, mp2_roi, mp3_roi]
102
+ truth_threshold: 100
103
+ models:
104
+ - file: mobileunetv3-cosmic590-deposplat-pdhd-rebin6-pad1-uv.ts
105
+ rebin: 6
106
+ padding:
107
+ enabled: true
108
+ padding: 1
109
+ padding_side: both
110
+ min_run: 1
111
+ avoid_merge: true
112
+ - file: mobileunetv3-cosmic590-deposplat-pdhd-rebin4-pad1-uv.ts
113
+ rebin: 4
114
+ padding:
115
+ enabled: true
116
+ padding: 1
117
+ padding_side: both
118
+ min_run: 1
119
+ avoid_merge: true
120
+ - file: mobileunetv3-cosmic590-deposplat-pdhd-rebin10-pad0-uv.ts
121
+ rebin: 10
122
+ padding:
123
+ enabled: false
124
+ padding: 0
125
+ ```