Upload GradeEye baseline-3ch checkpoints

README.md

@@ -1,3 +1,55 @@
 ---
 license: cc-by-nc-4.0
+tags:
+  - medical-imaging
+  - diabetic-retinopathy
+  - domain-generalization
+  - pytorch
+pipeline_tag: image-classification
+library_name: pytorch
 ---
+
+# GradeEye baseline-3ch
+
+This repository contains GradeEye CORN ordinal diabetic-retinopathy classifiers using three backbones (`convnext_tiny`, `deit3_small_patch16_384`, `maxvit_tiny_tf_384`) at 384x384 resolution, with 3-channel input and a 4-threshold ordinal head. The `_ema.safetensors` file is the **primary** artifact: the paper's reported evaluation metrics were generated using the EMA state dict. The unsuffixed `.safetensors` file is the corresponding raw `model_state_dict` secondary artifact.
+
+## Checkpoints
+
+| Primary EMA weights | Raw secondary weights | Architecture | Held-out fold | Best QWK | Epoch |
+|---|---|---|---|---:|---:|
+| `lodo_aptos_convnext_tiny_best_ema.safetensors` | `lodo_aptos_convnext_tiny_best.safetensors` | `convnext_tiny` | aptos | 0.7614 | 6 |
+| `lodo_aptos_deit3_small_patch16_384_best_ema.safetensors` | `lodo_aptos_deit3_small_patch16_384_best.safetensors` | `deit3_small_patch16_384` | aptos | 0.7064 | 6 |
+| `lodo_aptos_maxvit_tiny_tf_384_best_ema.safetensors` | `lodo_aptos_maxvit_tiny_tf_384_best.safetensors` | `maxvit_tiny_tf_384` | aptos | 0.6558 | 11 |
+| `lodo_ddr_convnext_tiny_best_ema.safetensors` | `lodo_ddr_convnext_tiny_best.safetensors` | `convnext_tiny` | ddr | 0.7506 | 3 |
+| `lodo_ddr_deit3_small_patch16_384_best_ema.safetensors` | `lodo_ddr_deit3_small_patch16_384_best.safetensors` | `deit3_small_patch16_384` | ddr | 0.7252 | 3 |
+| `lodo_ddr_maxvit_tiny_tf_384_best_ema.safetensors` | `lodo_ddr_maxvit_tiny_tf_384_best.safetensors` | `maxvit_tiny_tf_384` | ddr | 0.5483 | 14 |
+| `lodo_eyepacs_convnext_tiny_best_ema.safetensors` | `lodo_eyepacs_convnext_tiny_best.safetensors` | `convnext_tiny` | eyepacs | 0.7820 | 1 |
+| `lodo_eyepacs_deit3_small_patch16_384_best_ema.safetensors` | `lodo_eyepacs_deit3_small_patch16_384_best.safetensors` | `deit3_small_patch16_384` | eyepacs | 0.7684 | 8 |
+| `lodo_eyepacs_maxvit_tiny_tf_384_best_ema.safetensors` | `lodo_eyepacs_maxvit_tiny_tf_384_best.safetensors` | `maxvit_tiny_tf_384` | eyepacs | 0.6500 | 17 |
+| `lodo_messidor2_convnext_tiny_best_ema.safetensors` | `lodo_messidor2_convnext_tiny_best.safetensors` | `convnext_tiny` | messidor2 | 0.7615 | 10 |
+| `lodo_messidor2_deit3_small_patch16_384_best_ema.safetensors` | `lodo_messidor2_deit3_small_patch16_384_best.safetensors` | `deit3_small_patch16_384` | messidor2 | 0.7253 | 5 |
+| `lodo_messidor2_maxvit_tiny_tf_384_best_ema.safetensors` | `lodo_messidor2_maxvit_tiny_tf_384_best.safetensors` | `maxvit_tiny_tf_384` | messidor2 | 0.5182 | 8 |
+
+## Preprocessing
+
+1. Resize the RGB fundus image to 384x384 using the same offline preprocessing pipeline.
+2. Convert RGB to float in [0,1] and apply ImageNet normalization: mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225).
+
+## Loading
+
+```python
+import json
+from modeling import load_model
+
+config = json.load(open('config.json'))
+model = load_model('lodo_eyepacs_convnext_tiny_best_ema.safetensors', config)
+# model(x) returns CORN logits with shape (batch, 4)
+```
+
+Install `torch`, `timm`, and `safetensors`, and make the GradeEye source repository available on `PYTHONPATH`.
+
+## Intended use and limitations
+
+These weights are released for research and reproducibility only. They are not validated for clinical diagnosis or treatment decisions. Performance varies substantially by held-out dataset and should not be interpreted as clinical-grade generalization.
+
+Source code and paper materials: [https://github.com/ahmed-farhanur-rashid/gradeeye](https://github.com/ahmed-farhanur-rashid/gradeeye).

aptos/convnext_tiny/config.json

@@ -0,0 +1,45 @@
+{
+  "model_type": "DRGradingModel",
+  "task": "diabetic-retinopathy-grading",
+  "architecture": "convnext_tiny",
+  "backbone_timm_name": "convnext_tiny",
+  "in_channels": 3,
+  "image_size": 384,
+  "num_classes": 5,
+  "ordinal_head": "CORN",
+  "num_thresholds": 4,
+  "use_cbam": true,
+  "cbam_num_stages": 2,
+  "head_hidden_dim": 512,
+  "dropout": 0.4,
+  "channels_last": true,
+  "ema_decay": 0.999,
+  "loss_type": "corn",
+  "use_class_weighting": true,
+  "use_mixup": true,
+  "primary_weights": "lodo_aptos_convnext_tiny_best_ema.safetensors",
+  "secondary_weights": "lodo_aptos_convnext_tiny_best.safetensors",
+  "normalization": {
+    "rgb": {
+      "mean": [
+        0.485,
+        0.456,
+        0.406
+      ],
+      "std": [
+        0.229,
+        0.224,
+        0.225
+      ]
+    },
+    "extra_channel": "Already in [0,1]; appended after RGB normalization; not ImageNet-normalized."
+  },
+  "class_names": [
+    "No DR",
+    "Mild",
+    "Moderate",
+    "Severe",
+    "Proliferative DR"
+  ],
+  "held_out_fold": "aptos"
+}

aptos/convnext_tiny/lodo_aptos_convnext_tiny_best.meta.json

@@ -0,0 +1,38 @@
+{
+  "best_metric": 0.7613941888114648,
+  "checkpoint_source": "saved/checkpoints/baseline_3ch/aptos/convnext_tiny/lodo_aptos_convnext_tiny_best.pt",
+  "class_names": [
+    "No DR",
+    "Mild",
+    "Moderate",
+    "Severe",
+    "Proliferative DR"
+  ],
+  "ema_state_dict_present": true,
+  "epoch": 6,
+  "global_step": 8198,
+  "phase_name": "phase2_full_training",
+  "primary_weights": "lodo_aptos_convnext_tiny_best_ema.safetensors",
+  "primary_weights_kind": "ema",
+  "secondary_weights": "lodo_aptos_convnext_tiny_best.safetensors",
+  "secondary_weights_kind": "raw_model_state_dict",
+  "training_config": {
+    "arch": "convnext_tiny",
+    "cbam_num_stages": 2,
+    "channels_last": true,
+    "dropout": 0.4,
+    "ema_decay": 0.999,
+    "head_hidden_dim": 512,
+    "img_size": 384,
+    "in_chans": 3,
+    "loss_type": "corn",
+    "num_thresholds": 4,
+    "phase_batches": {
+      "phase1_frozen": 64,
+      "phase2_full_training": 24
+    },
+    "use_cbam": true,
+    "use_class_weighting": true,
+    "use_mixup": true
+  }
+}

aptos/convnext_tiny/lodo_aptos_convnext_tiny_best.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:69a26a60f469642a92c0b7c11da15fab85ffe1f7e30ee6ea8b5a05be03cb5d91
+size 113269368

aptos/convnext_tiny/lodo_aptos_convnext_tiny_best_ema.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:031562ecc8ce73e5a3ed18e2ffdbb921f48ae8c45db7791dbebd458a44adbc58
+size 113269368

aptos/deit3_small_patch16_384/config.json

@@ -0,0 +1,45 @@
+{
+  "model_type": "DRGradingModel",
+  "task": "diabetic-retinopathy-grading",
+  "architecture": "deit3_small_patch16_384",
+  "backbone_timm_name": "deit3_small_patch16_384",
+  "in_channels": 3,
+  "image_size": 384,
+  "num_classes": 5,
+  "ordinal_head": "CORN",
+  "num_thresholds": 4,
+  "use_cbam": true,
+  "cbam_num_stages": 2,
+  "head_hidden_dim": 384,
+  "dropout": 0.3,
+  "channels_last": false,
+  "ema_decay": 0.999,
+  "loss_type": "corn",
+  "use_class_weighting": true,
+  "use_mixup": true,
+  "primary_weights": "lodo_aptos_deit3_small_patch16_384_best_ema.safetensors",
+  "secondary_weights": "lodo_aptos_deit3_small_patch16_384_best.safetensors",
+  "normalization": {
+    "rgb": {
+      "mean": [
+        0.485,
+        0.456,
+        0.406
+      ],
+      "std": [
+        0.229,
+        0.224,
+        0.225
+      ]
+    },
+    "extra_channel": "Already in [0,1]; appended after RGB normalization; not ImageNet-normalized."
+  },
+  "class_names": [
+    "No DR",
+    "Mild",
+    "Moderate",
+    "Severe",
+    "Proliferative DR"
+  ],
+  "held_out_fold": "aptos"
+}

aptos/deit3_small_patch16_384/lodo_aptos_deit3_small_patch16_384_best.meta.json

@@ -0,0 +1,38 @@
+{
+  "best_metric": 0.7063935995279769,
+  "checkpoint_source": "saved/checkpoints/baseline_3ch/aptos/deit3_small_patch16_384/lodo_aptos_deit3_small_patch16_384_best.pt",
+  "class_names": [
+    "No DR",
+    "Mild",
+    "Moderate",
+    "Severe",
+    "Proliferative DR"
+  ],
+  "ema_state_dict_present": true,
+  "epoch": 6,
+  "global_step": 11432,
+  "phase_name": "phase2_full_training",
+  "primary_weights": "lodo_aptos_deit3_small_patch16_384_best_ema.safetensors",
+  "primary_weights_kind": "ema",
+  "secondary_weights": "lodo_aptos_deit3_small_patch16_384_best.safetensors",
+  "secondary_weights_kind": "raw_model_state_dict",
+  "training_config": {
+    "arch": "deit3_small_patch16_384",
+    "cbam_num_stages": 2,
+    "channels_last": false,
+    "dropout": 0.3,
+    "ema_decay": 0.999,
+    "head_hidden_dim": 384,
+    "img_size": 384,
+    "in_chans": 3,
+    "loss_type": "corn",
+    "num_thresholds": 4,
+    "phase_batches": {
+      "phase1_frozen": 64,
+      "phase2_full_training": 16
+    },
+    "use_cbam": true,
+    "use_class_weighting": true,
+    "use_mixup": true
+  }
+}

aptos/deit3_small_patch16_384/lodo_aptos_deit3_small_patch16_384_best.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7d245e0632601999c18b47e4a34dd875435e26eb10091f565fe9388a7ee308c9
+size 88057776

aptos/deit3_small_patch16_384/lodo_aptos_deit3_small_patch16_384_best_ema.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6391d54afb45e9150bfcf99507e1b594c3759588d31c99f8b4ffd1dd5ffc0a36
+size 88057776

aptos/maxvit_tiny_tf_384/config.json

@@ -0,0 +1,45 @@
+{
+  "model_type": "DRGradingModel",
+  "task": "diabetic-retinopathy-grading",
+  "architecture": "maxvit_tiny_tf_384",
+  "backbone_timm_name": "maxvit_tiny_tf_384",
+  "in_channels": 3,
+  "image_size": 384,
+  "num_classes": 5,
+  "ordinal_head": "CORN",
+  "num_thresholds": 4,
+  "use_cbam": true,
+  "cbam_num_stages": 2,
+  "head_hidden_dim": 512,
+  "dropout": 0.4,
+  "channels_last": false,
+  "ema_decay": 0.999,
+  "loss_type": "corn",
+  "use_class_weighting": true,
+  "use_mixup": true,
+  "primary_weights": "lodo_aptos_maxvit_tiny_tf_384_best_ema.safetensors",
+  "secondary_weights": "lodo_aptos_maxvit_tiny_tf_384_best.safetensors",
+  "normalization": {
+    "rgb": {
+      "mean": [
+        0.485,
+        0.456,
+        0.406
+      ],
+      "std": [
+        0.229,
+        0.224,
+        0.225
+      ]
+    },
+    "extra_channel": "Already in [0,1]; appended after RGB normalization; not ImageNet-normalized."
+  },
+  "class_names": [
+    "No DR",
+    "Mild",
+    "Moderate",
+    "Severe",
+    "Proliferative DR"
+  ],
+  "held_out_fold": "aptos"
+}

aptos/maxvit_tiny_tf_384/lodo_aptos_maxvit_tiny_tf_384_best.meta.json

@@ -0,0 +1,38 @@
+{
+  "best_metric": 0.6557566734353691,
+  "checkpoint_source": "saved/checkpoints/baseline_3ch/aptos/maxvit_tiny_tf_384/lodo_aptos_maxvit_tiny_tf_384_best.pt",
+  "class_names": [
+    "No DR",
+    "Mild",
+    "Moderate",
+    "Severe",
+    "Proliferative DR"
+  ],
+  "ema_state_dict_present": true,
+  "epoch": 11,
+  "global_step": 36729,
+  "phase_name": "phase2_full_training",
+  "primary_weights": "lodo_aptos_maxvit_tiny_tf_384_best_ema.safetensors",
+  "primary_weights_kind": "ema",
+  "secondary_weights": "lodo_aptos_maxvit_tiny_tf_384_best.safetensors",
+  "secondary_weights_kind": "raw_model_state_dict",
+  "training_config": {
+    "arch": "maxvit_tiny_tf_384",
+    "cbam_num_stages": 2,
+    "channels_last": false,
+    "dropout": 0.4,
+    "ema_decay": 0.999,
+    "head_hidden_dim": 512,
+    "img_size": 384,
+    "in_chans": 3,
+    "loss_type": "corn",
+    "num_thresholds": 4,
+    "phase_batches": {
+      "phase1_frozen": 32,
+      "phase2_full_training": 8
+    },
+    "use_cbam": true,
+    "use_class_weighting": true,
+    "use_mixup": true
+  }
+}

aptos/maxvit_tiny_tf_384/lodo_aptos_maxvit_tiny_tf_384_best.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e5200f8349c6a52ed12fc3908a831a4efa27b78298322544a5711ab3c0265d15
+size 122302616

aptos/maxvit_tiny_tf_384/lodo_aptos_maxvit_tiny_tf_384_best_ema.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ca69365547272c6ffb9ad2419ef903838d713e7dad4d788c89949d21b5a8dae6
+size 122302616

ddr/convnext_tiny/config.json

@@ -0,0 +1,45 @@
+{
+  "model_type": "DRGradingModel",
+  "task": "diabetic-retinopathy-grading",
+  "architecture": "convnext_tiny",
+  "backbone_timm_name": "convnext_tiny",
+  "in_channels": 3,
+  "image_size": 384,
+  "num_classes": 5,
+  "ordinal_head": "CORN",
+  "num_thresholds": 4,
+  "use_cbam": true,
+  "cbam_num_stages": 2,
+  "head_hidden_dim": 512,
+  "dropout": 0.4,
+  "channels_last": true,
+  "ema_decay": 0.999,
+  "loss_type": "corn",
+  "use_class_weighting": true,
+  "use_mixup": true,
+  "primary_weights": "lodo_ddr_convnext_tiny_best_ema.safetensors",
+  "secondary_weights": "lodo_ddr_convnext_tiny_best.safetensors",
+  "normalization": {
+    "rgb": {
+      "mean": [
+        0.485,
+        0.456,
+        0.406
+      ],
+      "std": [
+        0.229,
+        0.224,
+        0.225
+      ]
+    },
+    "extra_channel": "Already in [0,1]; appended after RGB normalization; not ImageNet-normalized."
+  },
+  "class_names": [
+    "No DR",
+    "Mild",
+    "Moderate",
+    "Severe",
+    "Proliferative DR"
+  ],
+  "held_out_fold": "ddr"
+}

ddr/convnext_tiny/lodo_ddr_convnext_tiny_best.meta.json

@@ -0,0 +1,38 @@
+{
+  "best_metric": 0.7506140888561054,
+  "checkpoint_source": "saved/checkpoints/baseline_3ch/ddr/convnext_tiny/lodo_ddr_convnext_tiny_best.pt",
+  "class_names": [
+    "No DR",
+    "Mild",
+    "Moderate",
+    "Severe",
+    "Proliferative DR"
+  ],
+  "ema_state_dict_present": true,
+  "epoch": 3,
+  "global_step": 5208,
+  "phase_name": "phase2_full_training",
+  "primary_weights": "lodo_ddr_convnext_tiny_best_ema.safetensors",
+  "primary_weights_kind": "ema",
+  "secondary_weights": "lodo_ddr_convnext_tiny_best.safetensors",
+  "secondary_weights_kind": "raw_model_state_dict",
+  "training_config": {
+    "arch": "convnext_tiny",
+    "cbam_num_stages": 2,
+    "channels_last": true,
+    "dropout": 0.4,
+    "ema_decay": 0.999,
+    "head_hidden_dim": 512,
+    "img_size": 384,
+    "in_chans": 3,
+    "loss_type": "corn",
+    "num_thresholds": 4,
+    "phase_batches": {
+      "phase1_frozen": 64,
+      "phase2_full_training": 24
+    },
+    "use_cbam": true,
+    "use_class_weighting": true,
+    "use_mixup": true
+  }
+}

ddr/convnext_tiny/lodo_ddr_convnext_tiny_best.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:068bc43c1d122536cfb58a2a222886b5f97dd36f7b58ca40b330f22129ac7f20
+size 113269368

ddr/convnext_tiny/lodo_ddr_convnext_tiny_best_ema.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0cde14e317e2630a6f525757b37669d26d2b0c31680c7f926f505030cbc85a47
+size 113269368

ddr/deit3_small_patch16_384/config.json

@@ -0,0 +1,45 @@
+{
+  "model_type": "DRGradingModel",
+  "task": "diabetic-retinopathy-grading",
+  "architecture": "deit3_small_patch16_384",
+  "backbone_timm_name": "deit3_small_patch16_384",
+  "in_channels": 3,
+  "image_size": 384,
+  "num_classes": 5,
+  "ordinal_head": "CORN",
+  "num_thresholds": 4,
+  "use_cbam": true,
+  "cbam_num_stages": 2,
+  "head_hidden_dim": 384,
+  "dropout": 0.3,
+  "channels_last": false,
+  "ema_decay": 0.999,
+  "loss_type": "corn",
+  "use_class_weighting": true,
+  "use_mixup": true,
+  "primary_weights": "lodo_ddr_deit3_small_patch16_384_best_ema.safetensors",
+  "secondary_weights": "lodo_ddr_deit3_small_patch16_384_best.safetensors",
+  "normalization": {
+    "rgb": {
+      "mean": [
+        0.485,
+        0.456,
+        0.406
+      ],
+      "std": [
+        0.229,
+        0.224,
+        0.225
+      ]
+    },
+    "extra_channel": "Already in [0,1]; appended after RGB normalization; not ImageNet-normalized."
+  },
+  "class_names": [
+    "No DR",
+    "Mild",
+    "Moderate",
+    "Severe",
+    "Proliferative DR"
+  ],
+  "held_out_fold": "ddr"
+}

ddr/deit3_small_patch16_384/lodo_ddr_deit3_small_patch16_384_best.meta.json

@@ -0,0 +1,38 @@
+{
+  "best_metric": 0.7252182450988855,
+  "checkpoint_source": "saved/checkpoints/baseline_3ch/ddr/deit3_small_patch16_384/lodo_ddr_deit3_small_patch16_384_best.pt",
+  "class_names": [
+    "No DR",
+    "Mild",
+    "Moderate",
+    "Severe",
+    "Proliferative DR"
+  ],
+  "ema_state_dict_present": true,
+  "epoch": 3,
+  "global_step": 6984,
+  "phase_name": "phase2_full_training",
+  "primary_weights": "lodo_ddr_deit3_small_patch16_384_best_ema.safetensors",
+  "primary_weights_kind": "ema",
+  "secondary_weights": "lodo_ddr_deit3_small_patch16_384_best.safetensors",
+  "secondary_weights_kind": "raw_model_state_dict",
+  "training_config": {
+    "arch": "deit3_small_patch16_384",
+    "cbam_num_stages": 2,
+    "channels_last": false,
+    "dropout": 0.3,
+    "ema_decay": 0.999,
+    "head_hidden_dim": 384,
+    "img_size": 384,
+    "in_chans": 3,
+    "loss_type": "corn",
+    "num_thresholds": 4,
+    "phase_batches": {
+      "phase1_frozen": 64,
+      "phase2_full_training": 16
+    },
+    "use_cbam": true,
+    "use_class_weighting": true,
+    "use_mixup": true
+  }
+}

ddr/deit3_small_patch16_384/lodo_ddr_deit3_small_patch16_384_best.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cffbaa74cb5ab4a400930992e696dd631a10253b97fbe54cbf4c7a1f064a4a9f
+size 88057776

ddr/deit3_small_patch16_384/lodo_ddr_deit3_small_patch16_384_best_ema.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cec6f5796b0076caaabdc1a08b4d2beec8bd5ca2ad41ce06ddb080064b4f5706
+size 88057776

ddr/maxvit_tiny_tf_384/config.json

@@ -0,0 +1,45 @@
+{
+  "model_type": "DRGradingModel",
+  "task": "diabetic-retinopathy-grading",
+  "architecture": "maxvit_tiny_tf_384",
+  "backbone_timm_name": "maxvit_tiny_tf_384",
+  "in_channels": 3,
+  "image_size": 384,
+  "num_classes": 5,
+  "ordinal_head": "CORN",
+  "num_thresholds": 4,
+  "use_cbam": true,
+  "cbam_num_stages": 2,
+  "head_hidden_dim": 512,
+  "dropout": 0.4,
+  "channels_last": false,
+  "ema_decay": 0.999,
+  "loss_type": "corn",
+  "use_class_weighting": true,
+  "use_mixup": true,
+  "primary_weights": "lodo_ddr_maxvit_tiny_tf_384_best_ema.safetensors",
+  "secondary_weights": "lodo_ddr_maxvit_tiny_tf_384_best.safetensors",
+  "normalization": {
+    "rgb": {
+      "mean": [
+        0.485,
+        0.456,
+        0.406
+      ],
+      "std": [
+        0.229,
+        0.224,
+        0.225
+      ]
+    },
+    "extra_channel": "Already in [0,1]; appended after RGB normalization; not ImageNet-normalized."
+  },
+  "class_names": [
+    "No DR",
+    "Mild",
+    "Moderate",
+    "Severe",
+    "Proliferative DR"
+  ],
+  "held_out_fold": "ddr"
+}

ddr/maxvit_tiny_tf_384/lodo_ddr_maxvit_tiny_tf_384_best.meta.json

@@ -0,0 +1,38 @@
+{
+  "best_metric": 0.5482642270604744,
+  "checkpoint_source": "saved/checkpoints/baseline_3ch/ddr/maxvit_tiny_tf_384/lodo_ddr_maxvit_tiny_tf_384_best.pt",
+  "class_names": [
+    "No DR",
+    "Mild",
+    "Moderate",
+    "Severe",
+    "Proliferative DR"
+  ],
+  "ema_state_dict_present": true,
+  "epoch": 14,
+  "global_step": 43255,
+  "phase_name": "phase2_full_training",
+  "primary_weights": "lodo_ddr_maxvit_tiny_tf_384_best_ema.safetensors",
+  "primary_weights_kind": "ema",
+  "secondary_weights": "lodo_ddr_maxvit_tiny_tf_384_best.safetensors",
+  "secondary_weights_kind": "raw_model_state_dict",
+  "training_config": {
+    "arch": "maxvit_tiny_tf_384",
+    "cbam_num_stages": 2,
+    "channels_last": false,
+    "dropout": 0.4,
+    "ema_decay": 0.999,
+    "head_hidden_dim": 512,
+    "img_size": 384,
+    "in_chans": 3,
+    "loss_type": "corn",
+    "num_thresholds": 4,
+    "phase_batches": {
+      "phase1_frozen": 32,
+      "phase2_full_training": 8
+    },
+    "use_cbam": true,
+    "use_class_weighting": true,
+    "use_mixup": true
+  }
+}

ddr/maxvit_tiny_tf_384/lodo_ddr_maxvit_tiny_tf_384_best.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2332c0861549725a05e6d3bd404dc502a2596cf64fad99022970f9527c55f719
+size 122302616

ddr/maxvit_tiny_tf_384/lodo_ddr_maxvit_tiny_tf_384_best_ema.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c64c1e1ac3ee1253aa3c86e5ce2354d322d38d64bc2362f4faba04f25fdb3180
+size 122302616

eyepacs/convnext_tiny/config.json

@@ -0,0 +1,45 @@
+{
+  "model_type": "DRGradingModel",
+  "task": "diabetic-retinopathy-grading",
+  "architecture": "convnext_tiny",
+  "backbone_timm_name": "convnext_tiny",
+  "in_channels": 3,
+  "image_size": 384,
+  "num_classes": 5,
+  "ordinal_head": "CORN",
+  "num_thresholds": 4,
+  "use_cbam": true,
+  "cbam_num_stages": 2,
+  "head_hidden_dim": 512,
+  "dropout": 0.4,
+  "channels_last": true,
+  "ema_decay": 0.999,
+  "loss_type": "corn",
+  "use_class_weighting": true,
+  "use_mixup": true,
+  "primary_weights": "lodo_eyepacs_convnext_tiny_best_ema.safetensors",
+  "secondary_weights": "lodo_eyepacs_convnext_tiny_best.safetensors",
+  "normalization": {
+    "rgb": {
+      "mean": [
+        0.485,
+        0.456,
+        0.406
+      ],
+      "std": [
+        0.229,
+        0.224,
+        0.225
+      ]
+    },
+    "extra_channel": "Already in [0,1]; appended after RGB normalization; not ImageNet-normalized."
+  },
+  "class_names": [
+    "No DR",
+    "Mild",
+    "Moderate",
+    "Severe",
+    "Proliferative DR"
+  ],
+  "held_out_fold": "eyepacs"
+}

eyepacs/convnext_tiny/lodo_eyepacs_convnext_tiny_best.meta.json

@@ -0,0 +1,38 @@
+{
+  "best_metric": 0.7820317857888303,
+  "checkpoint_source": "saved/checkpoints/baseline_3ch/eyepacs/convnext_tiny/lodo_eyepacs_convnext_tiny_best.pt",
+  "class_names": [
+    "No DR",
+    "Mild",
+    "Moderate",
+    "Severe",
+    "Proliferative DR"
+  ],
+  "ema_state_dict_present": true,
+  "epoch": 1,
+  "global_step": 786,
+  "phase_name": "phase2_full_training",
+  "primary_weights": "lodo_eyepacs_convnext_tiny_best_ema.safetensors",
+  "primary_weights_kind": "ema",
+  "secondary_weights": "lodo_eyepacs_convnext_tiny_best.safetensors",
+  "secondary_weights_kind": "raw_model_state_dict",
+  "training_config": {
+    "arch": "convnext_tiny",
+    "cbam_num_stages": 2,
+    "channels_last": true,
+    "dropout": 0.4,
+    "ema_decay": 0.999,
+    "head_hidden_dim": 512,
+    "img_size": 384,
+    "in_chans": 3,
+    "loss_type": "corn",
+    "num_thresholds": 4,
+    "phase_batches": {
+      "phase1_frozen": 64,
+      "phase2_full_training": 24
+    },
+    "use_cbam": true,
+    "use_class_weighting": true,
+    "use_mixup": true
+  }
+}

eyepacs/convnext_tiny/lodo_eyepacs_convnext_tiny_best.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ab3775b5d113b362464de844d11ce0b7ab543b1247dc46ee341659c69709b3b0
+size 113269368

eyepacs/convnext_tiny/lodo_eyepacs_convnext_tiny_best_ema.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e95a4a0735ba9dfd09f34911035fd844a536230bda67741e8ae6bc70b668eb60
+size 113269368

eyepacs/deit3_small_patch16_384/config.json

@@ -0,0 +1,45 @@
+{
+  "model_type": "DRGradingModel",
+  "task": "diabetic-retinopathy-grading",
+  "architecture": "deit3_small_patch16_384",
+  "backbone_timm_name": "deit3_small_patch16_384",
+  "in_channels": 3,
+  "image_size": 384,
+  "num_classes": 5,
+  "ordinal_head": "CORN",
+  "num_thresholds": 4,
+  "use_cbam": true,
+  "cbam_num_stages": 2,
+  "head_hidden_dim": 384,
+  "dropout": 0.3,
+  "channels_last": false,
+  "ema_decay": 0.999,
+  "loss_type": "corn",
+  "use_class_weighting": true,
+  "use_mixup": true,
+  "primary_weights": "lodo_eyepacs_deit3_small_patch16_384_best_ema.safetensors",
+  "secondary_weights": "lodo_eyepacs_deit3_small_patch16_384_best.safetensors",
+  "normalization": {
+    "rgb": {
+      "mean": [
+        0.485,
+        0.456,
+        0.406
+      ],
+      "std": [
+        0.229,
+        0.224,
+        0.225
+      ]
+    },
+    "extra_channel": "Already in [0,1]; appended after RGB normalization; not ImageNet-normalized."
+  },
+  "class_names": [
+    "No DR",
+    "Mild",
+    "Moderate",
+    "Severe",
+    "Proliferative DR"
+  ],
+  "held_out_fold": "eyepacs"
+}

eyepacs/deit3_small_patch16_384/lodo_eyepacs_deit3_small_patch16_384_best.meta.json

@@ -0,0 +1,38 @@
+{
+  "best_metric": 0.7684374648510071,
+  "checkpoint_source": "saved/checkpoints/baseline_3ch/eyepacs/deit3_small_patch16_384/lodo_eyepacs_deit3_small_patch16_384_best.pt",
+  "class_names": [
+    "No DR",
+    "Mild",
+    "Moderate",
+    "Severe",
+    "Proliferative DR"
+  ],
+  "ema_state_dict_present": true,
+  "epoch": 8,
+  "global_step": 3125,
+  "phase_name": "phase2_full_training",
+  "primary_weights": "lodo_eyepacs_deit3_small_patch16_384_best_ema.safetensors",
+  "primary_weights_kind": "ema",
+  "secondary_weights": "lodo_eyepacs_deit3_small_patch16_384_best.safetensors",
+  "secondary_weights_kind": "raw_model_state_dict",
+  "training_config": {
+    "arch": "deit3_small_patch16_384",
+    "cbam_num_stages": 2,
+    "channels_last": false,
+    "dropout": 0.3,
+    "ema_decay": 0.999,
+    "head_hidden_dim": 384,
+    "img_size": 384,
+    "in_chans": 3,
+    "loss_type": "corn",
+    "num_thresholds": 4,
+    "phase_batches": {
+      "phase1_frozen": 64,
+      "phase2_full_training": 16
+    },
+    "use_cbam": true,
+    "use_class_weighting": true,
+    "use_mixup": true
+  }
+}

eyepacs/deit3_small_patch16_384/lodo_eyepacs_deit3_small_patch16_384_best.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:07ae1653d9e64e27f2e07c0d8bbdac09ee2f649829e87092857fe46da0baa23a
+size 88057776

eyepacs/deit3_small_patch16_384/lodo_eyepacs_deit3_small_patch16_384_best_ema.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:888875e04a4be230fcdba117bc902081cc8b490e31d8742765722297f048af88
+size 88057776

eyepacs/maxvit_tiny_tf_384/config.json

@@ -0,0 +1,45 @@
+{
+  "model_type": "DRGradingModel",
+  "task": "diabetic-retinopathy-grading",
+  "architecture": "maxvit_tiny_tf_384",
+  "backbone_timm_name": "maxvit_tiny_tf_384",
+  "in_channels": 3,
+  "image_size": 384,
+  "num_classes": 5,
+  "ordinal_head": "CORN",
+  "num_thresholds": 4,
+  "use_cbam": true,
+  "cbam_num_stages": 2,
+  "head_hidden_dim": 512,
+  "dropout": 0.4,
+  "channels_last": false,
+  "ema_decay": 0.999,
+  "loss_type": "corn",
+  "use_class_weighting": true,
+  "use_mixup": true,
+  "primary_weights": "lodo_eyepacs_maxvit_tiny_tf_384_best_ema.safetensors",
+  "secondary_weights": "lodo_eyepacs_maxvit_tiny_tf_384_best.safetensors",
+  "normalization": {
+    "rgb": {
+      "mean": [
+        0.485,
+        0.456,
+        0.406
+      ],
+      "std": [
+        0.229,
+        0.224,
+        0.225
+      ]
+    },
+    "extra_channel": "Already in [0,1]; appended after RGB normalization; not ImageNet-normalized."
+  },
+  "class_names": [
+    "No DR",
+    "Mild",
+    "Moderate",
+    "Severe",
+    "Proliferative DR"
+  ],
+  "held_out_fold": "eyepacs"
+}

eyepacs/maxvit_tiny_tf_384/lodo_eyepacs_maxvit_tiny_tf_384_best.meta.json

@@ -0,0 +1,38 @@
+{
+  "best_metric": 0.6500080983317436,
+  "checkpoint_source": "saved/checkpoints/baseline_3ch/eyepacs/maxvit_tiny_tf_384/lodo_eyepacs_maxvit_tiny_tf_384_best.pt",
+  "class_names": [
+    "No DR",
+    "Mild",
+    "Moderate",
+    "Severe",
+    "Proliferative DR"
+  ],
+  "ema_state_dict_present": true,
+  "epoch": 17,
+  "global_step": 11740,
+  "phase_name": "phase2_full_training",
+  "primary_weights": "lodo_eyepacs_maxvit_tiny_tf_384_best_ema.safetensors",
+  "primary_weights_kind": "ema",
+  "secondary_weights": "lodo_eyepacs_maxvit_tiny_tf_384_best.safetensors",
+  "secondary_weights_kind": "raw_model_state_dict",
+  "training_config": {
+    "arch": "maxvit_tiny_tf_384",
+    "cbam_num_stages": 2,
+    "channels_last": false,
+    "dropout": 0.4,
+    "ema_decay": 0.999,
+    "head_hidden_dim": 512,
+    "img_size": 384,
+    "in_chans": 3,
+    "loss_type": "corn",
+    "num_thresholds": 4,
+    "phase_batches": {
+      "phase1_frozen": 32,
+      "phase2_full_training": 8
+    },
+    "use_cbam": true,
+    "use_class_weighting": true,
+    "use_mixup": true
+  }
+}

eyepacs/maxvit_tiny_tf_384/lodo_eyepacs_maxvit_tiny_tf_384_best.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eb76b503d8cb46d75ffca7a630bea0a40b6d8398e1993d43ee69814f6f2b1e4f
+size 122302616

eyepacs/maxvit_tiny_tf_384/lodo_eyepacs_maxvit_tiny_tf_384_best_ema.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0edf62751f8fc5b52c5affc7027b7ebfb1c71db4556b9d28c2bcf197199cc157
+size 122302616

messidor2/convnext_tiny/config.json

@@ -0,0 +1,45 @@
+{
+  "model_type": "DRGradingModel",
+  "task": "diabetic-retinopathy-grading",
+  "architecture": "convnext_tiny",
+  "backbone_timm_name": "convnext_tiny",
+  "in_channels": 3,
+  "image_size": 384,
+  "num_classes": 5,
+  "ordinal_head": "CORN",
+  "num_thresholds": 4,
+  "use_cbam": true,
+  "cbam_num_stages": 2,
+  "head_hidden_dim": 512,
+  "dropout": 0.4,
+  "channels_last": true,
+  "ema_decay": 0.999,
+  "loss_type": "corn",
+  "use_class_weighting": true,
+  "use_mixup": true,
+  "primary_weights": "lodo_messidor2_convnext_tiny_best_ema.safetensors",
+  "secondary_weights": "lodo_messidor2_convnext_tiny_best.safetensors",
+  "normalization": {
+    "rgb": {
+      "mean": [
+        0.485,
+        0.456,
+        0.406
+      ],
+      "std": [
+        0.229,
+        0.224,
+        0.225
+      ]
+    },
+    "extra_channel": "Already in [0,1]; appended after RGB normalization; not ImageNet-normalized."
+  },
+  "class_names": [
+    "No DR",
+    "Mild",
+    "Moderate",
+    "Severe",
+    "Proliferative DR"
+  ],
+  "held_out_fold": "messidor2"
+}

messidor2/convnext_tiny/lodo_messidor2_convnext_tiny_best.meta.json

@@ -0,0 +1,38 @@
+{
+  "best_metric": 0.7614860286460639,
+  "checkpoint_source": "saved/checkpoints/baseline_3ch/messidor2/convnext_tiny/lodo_messidor2_convnext_tiny_best.pt",
+  "class_names": [
+    "No DR",
+    "Mild",
+    "Moderate",
+    "Severe",
+    "Proliferative DR"
+  ],
+  "ema_state_dict_present": true,
+  "epoch": 10,
+  "global_step": 12653,
+  "phase_name": "phase2_full_training",
+  "primary_weights": "lodo_messidor2_convnext_tiny_best_ema.safetensors",
+  "primary_weights_kind": "ema",
+  "secondary_weights": "lodo_messidor2_convnext_tiny_best.safetensors",
+  "secondary_weights_kind": "raw_model_state_dict",
+  "training_config": {
+    "arch": "convnext_tiny",
+    "cbam_num_stages": 2,
+    "channels_last": true,
+    "dropout": 0.4,
+    "ema_decay": 0.999,
+    "head_hidden_dim": 512,
+    "img_size": 384,
+    "in_chans": 3,
+    "loss_type": "corn",
+    "num_thresholds": 4,
+    "phase_batches": {
+      "phase1_frozen": 64,
+      "phase2_full_training": 24
+    },
+    "use_cbam": true,
+    "use_class_weighting": true,
+    "use_mixup": true
+  }
+}

messidor2/convnext_tiny/lodo_messidor2_convnext_tiny_best.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eb2f04801b0d49d163a4ec40a5f7192fce57f324ae0b678e2653467b17ea8648
+size 113269368

messidor2/convnext_tiny/lodo_messidor2_convnext_tiny_best_ema.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:33b2f5b98efded1ba159228b0bc71db7acfa7fa2b9fbeb7400d7403fb909d0cc
+size 113269368

messidor2/deit3_small_patch16_384/config.json

@@ -0,0 +1,45 @@
+{
+  "model_type": "DRGradingModel",
+  "task": "diabetic-retinopathy-grading",
+  "architecture": "deit3_small_patch16_384",
+  "backbone_timm_name": "deit3_small_patch16_384",
+  "in_channels": 3,
+  "image_size": 384,
+  "num_classes": 5,
+  "ordinal_head": "CORN",
+  "num_thresholds": 4,
+  "use_cbam": true,
+  "cbam_num_stages": 2,
+  "head_hidden_dim": 384,
+  "dropout": 0.3,
+  "channels_last": false,
+  "ema_decay": 0.999,
+  "loss_type": "corn",
+  "use_class_weighting": true,
+  "use_mixup": true,
+  "primary_weights": "lodo_messidor2_deit3_small_patch16_384_best_ema.safetensors",
+  "secondary_weights": "lodo_messidor2_deit3_small_patch16_384_best.safetensors",
+  "normalization": {
+    "rgb": {
+      "mean": [
+        0.485,
+        0.456,
+        0.406
+      ],
+      "std": [
+        0.229,
+        0.224,
+        0.225
+      ]
+    },
+    "extra_channel": "Already in [0,1]; appended after RGB normalization; not ImageNet-normalized."
+  },
+  "class_names": [
+    "No DR",
+    "Mild",
+    "Moderate",
+    "Severe",
+    "Proliferative DR"
+  ],
+  "held_out_fold": "messidor2"
+}

messidor2/deit3_small_patch16_384/lodo_messidor2_deit3_small_patch16_384_best.meta.json

@@ -0,0 +1,38 @@
+{
+  "best_metric": 0.7253012991702625,
+  "checkpoint_source": "saved/checkpoints/baseline_3ch/messidor2/deit3_small_patch16_384/lodo_messidor2_deit3_small_patch16_384_best.pt",
+  "class_names": [
+    "No DR",
+    "Mild",
+    "Moderate",
+    "Severe",
+    "Proliferative DR"
+  ],
+  "ema_state_dict_present": true,
+  "epoch": 5,
+  "global_step": 10690,
+  "phase_name": "phase2_full_training",
+  "primary_weights": "lodo_messidor2_deit3_small_patch16_384_best_ema.safetensors",
+  "primary_weights_kind": "ema",
+  "secondary_weights": "lodo_messidor2_deit3_small_patch16_384_best.safetensors",
+  "secondary_weights_kind": "raw_model_state_dict",
+  "training_config": {
+    "arch": "deit3_small_patch16_384",
+    "cbam_num_stages": 2,
+    "channels_last": false,
+    "dropout": 0.3,
+    "ema_decay": 0.999,
+    "head_hidden_dim": 384,
+    "img_size": 384,
+    "in_chans": 3,
+    "loss_type": "corn",
+    "num_thresholds": 4,
+    "phase_batches": {
+      "phase1_frozen": 64,
+      "phase2_full_training": 16
+    },
+    "use_cbam": true,
+    "use_class_weighting": true,
+    "use_mixup": true
+  }
+}

messidor2/deit3_small_patch16_384/lodo_messidor2_deit3_small_patch16_384_best.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1a2ecf42b571d01e82f81c38c42b138195f024e47ed7765a174489b5d03f783e
+size 88057776

messidor2/deit3_small_patch16_384/lodo_messidor2_deit3_small_patch16_384_best_ema.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:64ff4841c17d38c1911778a4f2fd9cf5c85d1ab9319a1da9bf333724ce410cb1
+size 88057776

messidor2/maxvit_tiny_tf_384/config.json

@@ -0,0 +1,45 @@
+{
+  "model_type": "DRGradingModel",
+  "task": "diabetic-retinopathy-grading",
+  "architecture": "maxvit_tiny_tf_384",
+  "backbone_timm_name": "maxvit_tiny_tf_384",
+  "in_channels": 3,
+  "image_size": 384,
+  "num_classes": 5,
+  "ordinal_head": "CORN",
+  "num_thresholds": 4,
+  "use_cbam": true,
+  "cbam_num_stages": 2,
+  "head_hidden_dim": 512,
+  "dropout": 0.4,
+  "channels_last": false,
+  "ema_decay": 0.999,
+  "loss_type": "corn",
+  "use_class_weighting": true,
+  "use_mixup": true,
+  "primary_weights": "lodo_messidor2_maxvit_tiny_tf_384_best_ema.safetensors",
+  "secondary_weights": "lodo_messidor2_maxvit_tiny_tf_384_best.safetensors",
+  "normalization": {
+    "rgb": {
+      "mean": [
+        0.485,
+        0.456,
+        0.406
+      ],
+      "std": [
+        0.229,
+        0.224,
+        0.225
+      ]
+    },
+    "extra_channel": "Already in [0,1]; appended after RGB normalization; not ImageNet-normalized."
+  },
+  "class_names": [
+    "No DR",
+    "Mild",
+    "Moderate",
+    "Severe",
+    "Proliferative DR"
+  ],
+  "held_out_fold": "messidor2"
+}

messidor2/maxvit_tiny_tf_384/lodo_messidor2_maxvit_tiny_tf_384_best.meta.json

@@ -0,0 +1,38 @@
+{
+  "best_metric": 0.5182327564528264,
+  "checkpoint_source": "saved/checkpoints/baseline_3ch/messidor2/maxvit_tiny_tf_384/lodo_messidor2_maxvit_tiny_tf_384_best.pt",
+  "class_names": [
+    "No DR",
+    "Mild",
+    "Moderate",
+    "Severe",
+    "Proliferative DR"
+  ],
+  "ema_state_dict_present": true,
+  "epoch": 8,
+  "global_step": 30244,
+  "phase_name": "phase2_full_training",
+  "primary_weights": "lodo_messidor2_maxvit_tiny_tf_384_best_ema.safetensors",
+  "primary_weights_kind": "ema",
+  "secondary_weights": "lodo_messidor2_maxvit_tiny_tf_384_best.safetensors",
+  "secondary_weights_kind": "raw_model_state_dict",
+  "training_config": {
+    "arch": "maxvit_tiny_tf_384",
+    "cbam_num_stages": 2,
+    "channels_last": false,
+    "dropout": 0.4,
+    "ema_decay": 0.999,
+    "head_hidden_dim": 512,
+    "img_size": 384,
+    "in_chans": 3,
+    "loss_type": "corn",
+    "num_thresholds": 4,
+    "phase_batches": {
+      "phase1_frozen": 32,
+      "phase2_full_training": 8
+    },
+    "use_cbam": true,
+    "use_class_weighting": true,
+    "use_mixup": true
+  }
+}

messidor2/maxvit_tiny_tf_384/lodo_messidor2_maxvit_tiny_tf_384_best.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3ca317c23e61438d8a53ba1613872241ad3334204005052c5def4f020f1c0c2b
+size 122302616

messidor2/maxvit_tiny_tf_384/lodo_messidor2_maxvit_tiny_tf_384_best_ema.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3f4962a5800ab99a859682d606f90f2e43c3edc53563fca2359ff2b0d3567ee6
+size 122302616

modeling.py

@@ -0,0 +1,44 @@
+"""Minimal GradeEye classifier loader for Hugging Face Hub.
+
+Requires the GradeEye source package on PYTHONPATH plus torch, timm, and
+safetensors. The architecture is the same DRGradingModel used during training.
+EMA files are the primary weights and reproduce the paper evaluation protocol.
+"""
+from __future__ import annotations
+
+from pathlib import Path
+import sys
+import torch
+
+# For local source checkout usage. Users may instead install the GradeEye package.
+try:
+    from src.models.dr_model import DRGradingModel
+except ImportError as exc:
+    raise ImportError(
+        "Install/clone GradeEye and make its repository root available on PYTHONPATH."
+    ) from exc
+from safetensors.torch import load_file
+
+
+def load_model(weights_path: str | Path, config: dict, device: str = "cpu") -> DRGradingModel:
+    """Instantiate DRGradingModel and strictly load a .safetensors state dict."""
+    model = DRGradingModel(
+        pretrained=False,
+        use_cbam=config["use_cbam"],
+        cbam_num_stages=config["cbam_num_stages"],
+        num_thresholds=config["num_thresholds"],
+        head_hidden_dim=config["head_hidden_dim"],
+        dropout=config["dropout"],
+        output_mode="corn",
+        arch=config["architecture"],
+        in_chans=config["in_channels"],
+        img_size=config["image_size"],
+    )
+    state_dict = load_file(str(weights_path), device="cpu")
+    result = model.load_state_dict(state_dict, strict=True)
+    if result.missing_keys or result.unexpected_keys:
+        raise RuntimeError(
+            f"State-dict mismatch: missing={result.missing_keys}, "
+            f"unexpected={result.unexpected_keys}"
+        )
+    return model.to(device).eval()