FHE-feasible diagnostic CNN: artefacts + measured metrics

README.md

@@ -23,14 +23,17 @@ a quantum-safe orchestration demo. This is the **plaintext** model for the
 
 | metric | value |
 |---|---|
-| `agreement` | 0.75 |
-| `disagreement_on_positives` | 0.078947 |
-| `metric_delta` | 0.134616 |
-| `metric_delta_pp` | 13.4616 |
+| `agreement` | 0.730769 |
+| `auc` | 0.609231 |
+| `disagreement_on_positives` | 0.157895 |
+| `majority_class_rate` | 0.730769 |
+| `metric_delta` | 0.166667 |
+| `metric_delta_pp` | 16.6667 |
+| `n_classes_predicted` | 2 |
 | `n_eval` | 156 |
-| `student_metric` | 0.730769 |
+| `student_metric` | 0.685897 |
 | `student_params` | 3226 |
-| `teacher_metric` | 0.865385 |
+| `teacher_metric` | 0.852564 |
 | `teacher_params` | 11171266 |
 
 ## Published baselines this is measured against

eval_sample.json

@@ -13,8 +13,8 @@
   "layout": "samples[i] is one 14x14 image, row-major, already normalised"
  },
  "classes": {
-  "0": "benign/normal",
-  "1": "malignant"
+  "0": "malignant",
+  "1": "normal/benign"
  },
  "samples": [
   [
@@ -31066,7 +31066,7 @@
  ],
  "expected": {
   "graph": "model.onnx",
-  "accuracy": 0.7307692307692307,
+  "accuracy": 0.6858974358974359,
   "majority_baseline": 0.7307692307692307,
   "note": "The whole official test split, and the same rows the registry's recorded student accuracy was measured on, so this reproduces that figure rather than resampling it."
  }

export_report.json

@@ -4,9 +4,9 @@
   "fp32_bytes": 14691,
   "int8_bytes": 7274,
   "opset": 17,
-  "fp32_max_abs_error": 1.4901161193847656e-08,
-  "int8_agreement": 1.0,
-  "int8_max_abs_error": 0.0008438974618911743,
+  "fp32_max_abs_error": 2.9802322387695312e-08,
+  "int8_agreement": 0.9935897435897436,
+  "int8_max_abs_error": 0.0021088868379592896,
   "n_fixtures": 156,
   "passed": true,
   "int8_skipped_reason": "",

manifest.json

@@ -1,11 +1,11 @@
 {
   "key": "medical_fl.student",
-  "metric_name": "agreement",
-  "best_metric": 0.7435897435897436,
-  "best_epoch": 0,
-  "epochs_run": 26,
-  "wall_clock_s": 0.5,
-  "git_sha": "8317cb2-dirty",
+  "metric_name": "auc",
+  "best_metric": 0.7319189971070394,
+  "best_epoch": 6,
+  "epochs_run": 32,
+  "wall_clock_s": 0.6,
+  "git_sha": "528d852-dirty",
   "seed": 42,
   "hyperparameters": {
     "epochs": 120,

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:052703bed0e734520026316e1804bd0b2f528d131756d7e12905b6d3b9d5815c
+oid sha256:935b67771bd2df324edda80aa81a2cf11d54ff31daf748cbf046f5a1a6e0712d
 size 13384

onnx/model.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d97a9bb42067da0a7ed71a56b132f71fff12183e241c51a76dac299b393a3e9a
+oid sha256:9ccd75172dc0e2be4b54a1849081f917032eb274133b90ea9d2e58f681ae418c
 size 14691

onnx/model_int8.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:992409de581e9e8a85440e8529830c0759333a7daf6b78cdf9949d0c117217e7
+oid sha256:8995e22180497040d324e37f77fa3f9c025be59752b77be5502514b1f00c95f8
 size 7274