File size: 9,412 Bytes
56ec3ba
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
# docs/app_notes.md
# RareNoduleFL β€” App Deployment Technical Notes

---

## What Was Added

The following files were added on top of the existing research pipeline
without modifying any research code:

```
app.py                     Gradio Blocks entrypoint β€” 6 tabs
requirements.txt           9 CPU-safe pip dependencies
README.md                  Updated: Gradio HF Spaces YAML + full docs
.gitignore                 Excludes checkpoints, NPY data, caches

app_core/
  __init__.py
  config.py                All paths, constants, known metrics, stat highlights
  model_registry.py        Lazy model loading, caching, graceful fallback
  preprocessing.py         Upload β†’ resize β†’ normalise β†’ tensor
  inference.py             Single + multi-model inference + summary formatting
  visualization.py         Mask, overlay, heatmap rendering
  results_loader.py        Load figures/CSVs/stats from results/
  sample_cases.py          Auto-discovers assets/samples + synthetic fallback generation

assets/
  css/app.css              Academic medical dashboard CSS
  samples/                 User-managed demo images for Inference/Compare dropdowns
                          (synthetic fallback set generated only if folder is empty)

docs/
  app_notes.md             This file
```

---

## How the App Layer Maps to the Research Layer

```
Research Layer                        App Layer
──────────────────────────────────────────────────────────────────
training/lidc_step2_step3_step4.py    ← architecture in model_registry.py
                                        (minimal duplication β€” exact
                                         channel/stride/norm to match checkpoints)

evaluation/complete_project_finish.py ← figures/CSVs consumed by results_loader.py
                                        (app reads from results/, does not re-run)

models/*.pth (from Google Drive)      ← loaded lazily by model_registry.py
                                        (app works without them β€” shows known metrics)

results/*.png, *.csv, *.txt           ← loaded at runtime by results_loader.py
                                        (with embedded fallbacks if files missing)
```

The app deliberately **does not import** from `preprocessing/`, `training/`,
`federated/`, or `evaluation/`. This keeps the deployment layer clean and
avoids pulling in training-time dependencies (Flower, Google Colab, etc.).

---

## Live Inference vs Precomputed Display

| Tab | Mode | Notes |
|-----|------|-------|
| Overview | Precomputed | Pipeline figure from `results/pipeline_overview.png` |
| Inference | Live (if checkpoint) | Falls back to known metrics if `.pth` missing |
| Compare | Live (if checkpoint) | Same fallback per model |
| Explainability | Precomputed | Loads from `results/xai_gradcam_panels.png` etc. |
| Results | Precomputed | Figures gallery + ablation CSV + stat tests text |
| About | Static | Hardcoded text |

**Live inference** = uploads image β†’ preprocessing β†’ model forward pass β†’ sigmoid β†’ mask/overlay.

**Precomputed display** = reads PNG/CSV/TXT from `results/` and renders directly.

The app is designed so that **all tabs provide meaningful content** even on a
Hugging Face Space that has no checkpoint files uploaded β€” because research
result figures and embedded known metrics cover every tab.

---

## App Assumptions

1. **`results/` folder exists and is populated** by running
   `evaluation/complete_project_finish.py` on Google Colab.
   If figures are missing, tabs display placeholder messages.

2. **`models/` folder** may be empty. The app checks each checkpoint path
   at startup via `check_availability()` and adjusts UI labels accordingly.

3. **CPU execution only.** No CUDA tensors. All `torch.load` calls use
   `map_location="cpu"`. Model `eval()` mode is enforced before inference.

4. **No Google Drive paths.** All paths are resolved relative to the repo root
   via `Path(__file__).parent.parent` in `config.py`.

5. **Sample source** is the `assets/samples/` folder. Any image placed there
   appears in Inference/Compare dropdowns. If the folder is empty, the app
   auto-generates a synthetic fallback set (CT-like grayscale nodules) so the
   inference pipeline remains demo-ready.

---

## Architecture Duplication Note

`app_core/model_registry.py` contains a minimal re-definition of the U-Net
architecture (`_build_unet()`) using MONAI. This is intentional β€” importing
from `training/lidc_step2_step3_step4.py` would pull in:
- `from google.colab import drive`
- `import flwr`
- MONAI training transforms not needed at inference time

The re-definition uses exactly the same hyperparameters as the training checkpoints:
```python
UNet(spatial_dims=2, in_channels=1, out_channels=1,
     channels=(32,64,128,256,512), strides=(2,2,2,2),
     num_res_units=2, norm=Norm.BATCH, dropout=0.1)
```
Any change to this in `training/` must be reflected here.

A `FallbackUNet` pure-PyTorch implementation is also provided in case MONAI
is unavailable, though the `requirements.txt` includes `monai>=1.3.0`.

---

## How to Extend Later

### Add a new model
1. Add the checkpoint path to `CHECKPOINT_PATHS` in `app_core/config.py`
2. Add the state dict key to `CHECKPOINT_SD_KEY`
3. Add known metrics to `KNOWN_METRICS`
4. Add the model key to `MODEL_KEYS`
5. If the architecture differs, add a branch in `model_registry._build_unet()`

### Add a new result figure
1. Add the figure path to `FIGURES` in `app_core/config.py`
2. Add a `(key, caption)` entry in `results_loader.load_all_figures()`
3. It will appear automatically in the Results tab gallery

### Enable Grad-CAM live in the app
Current: Explainability tab shows precomputed figures.
To enable live: add `torchcam` to `requirements.txt`, import `GradCAMpp`
in `inference.py`, and add a `run_gradcam()` function following the pattern
in `evaluation/complete_project_finish.py` (lines 570-610).
Note: GradCAMpp requires `inp.requires_grad_(True)` and a forward pass
outside `torch.no_grad()`.

### Deploy with GPU on HF Spaces
Change the Space hardware to T4 in the HF Spaces settings.
No code changes required β€” the model_registry already runs on CPU,
which is correct for T4 (`map_location="cpu"` β†’ then `.to(device)` if needed).
Add `device = torch.device("cuda" if torch.cuda.is_available() else "cpu")`
in `app_core/config.py` and update `model_registry.load_model()` to use it.

---

## Local Run Checklist

```bash
# 1. Clone repo
git clone https://huggingface.co/spaces/YOUR_HF_USERNAME/RareNoduleFL
cd RareNoduleFL

# 2. Install dependencies
pip install -r requirements.txt

# 3. (Optional) Add model checkpoints
mkdir -p models
# Copy step4.1_unet_best.pth, fl_iid_global_best.pth,
#       fl_noniid_rus_global_best.pth from Google Drive β†’ models/

# 4. (Optional) Add result figures
# Copy results/ folder from Google Drive β†’ results/

# 5. Run
python app.py
# Open http://localhost:7860
```

---

## Hugging Face Spaces Deployment Checklist

```
βœ… README.md has valid YAML front matter:
     sdk: gradio
     app_file: app.py

βœ… app.py ends with demo.launch() (no share=True needed on Spaces)

βœ… requirements.txt lists all dependencies
   (Spaces installs from this file automatically)

βœ… All paths in app_core/config.py are repo-relative
   (REPO_ROOT = Path(__file__).parent.parent)

βœ… No Colab / Google Drive imports in app layer

βœ… No training-time imports (flwr, Colab, etc.) in app layer

βœ… App works without model checkpoints
   (precomputed results cover all tabs)

βœ… App works without results/ files
   (embedded fallback values in results_loader.py)

βœ… Sample images auto-generated at startup
   (ensure_samples_exist() called in app.py)
```

---

## Missing Assets Checklist

Items that must be added to the repo manually (not auto-generated):

| Item | Source | Required for |
|------|--------|-------------|
| `models/step4.1_unet_best.pth` | Google Drive (training output) | Live inference |
| `models/fl_iid_global_best.pth` | Google Drive (training output) | Live inference |
| `models/fl_noniid_rus_global_best.pth` | Google Drive (training output) | Live inference |
| `results/pipeline_overview.png` | Google Drive (evaluation output) | Overview tab figure |
| `results/fl_learning_curves.png` | Google Drive (evaluation output) | Results tab |
| `results/fl_client_drift.png` | Google Drive (evaluation output) | Results tab |
| `results/fl_ece_comparison.png` | Google Drive (evaluation output) | Results tab |
| `results/xai_entropy_heatmaps.png` | Google Drive (evaluation output) | Explainability tab |
| `results/xai_gradcam_panels.png` | Google Drive (evaluation output) | Explainability tab |
| `results/fl_heterogeneity_bars.png` | Google Drive (evaluation output) | Results tab |
| `results/froc_curve_fixed.png` | Google Drive (evaluation output) | Results tab |
| `results/calibration_analysis.png` | Google Drive (evaluation output) | Results tab |
| `results/fl_ablation_table.csv` | Google Drive (evaluation output) | Results tab table |
| `results/fl_statistical_tests.txt` | Google Drive (evaluation output) | Results tab stats |

All evaluation outputs are generated by:
```bash
# On Google Colab T4 (~1 compute unit)
python evaluation/complete_project_finish.py
```