Support optional fixed eval_size and update detection configs.

Add Resize(eval_size) for TAIC eval with correct COCO box scale_factor; sync README and local detection train/eval settings.

README.md

@@ -363,11 +363,7 @@ For C-TAIC, reported `bpp` is **extension-layer only** (base-layer rate is exclu
 
 ```bash
 python scripts/eval_taic.py -c configs/eval/taic_detection.yaml
-python scripts/eval_taic.py -c configs/eval/taic_detection.yaml --actual-bpp
-python scripts/eval_taic.py -c configs/eval/taic_detection.yaml --max-batches 10
-
 python scripts/eval_ctaic.py -c configs/eval/ctaic_s1.yaml
-python scripts/eval_ctaic.py -c configs/eval/ctaic_s1.yaml --no-condition
 ```
 
 ---
@@ -431,15 +427,15 @@ ann_file: "annotations/instances_val2017.json"
 ### Run codec + metrics
 
 ```bash
-python scripts/eval_taic.py -c configs/eval/taic_detection.yaml --with-metrics
-python scripts/eval_taic.py -c configs/eval/taic_instance.yaml --with-metrics
-python scripts/eval_taic.py -c configs/eval/taic_semantic.yaml --with-metrics
-python scripts/eval_taic.py -c configs/eval/taic_panoptic.yaml --with-metrics
-python scripts/eval_taic.py -c configs/eval/taic_pose.yaml --with-metrics
-
-python scripts/eval_ctaic.py -c configs/eval/ctaic_s1.yaml --with-metrics
-python scripts/eval_ctaic.py -c configs/eval/ctaic_s2.yaml --with-metrics
-python scripts/eval_ctaic.py -c configs/eval/ctaic_s3.yaml --with-metrics
+python scripts/eval_taic.py -c configs/eval/taic_detection.yaml 
+python scripts/eval_taic.py -c configs/eval/taic_instance.yaml
+python scripts/eval_taic.py -c configs/eval/taic_semantic.yaml
+python scripts/eval_taic.py -c configs/eval/taic_panoptic.yaml 
+python scripts/eval_taic.py -c configs/eval/taic_pose.yaml 
+
+python scripts/eval_ctaic.py -c configs/eval/ctaic_s1.yaml 
+python scripts/eval_ctaic.py -c configs/eval/ctaic_s2.yaml
+python scripts/eval_ctaic.py -c configs/eval/ctaic_s3.yaml 
 ```
 
 JSON results (codec + task metrics) are written under `logs/eval_taic/` or `logs/eval_ctaic/`.

configs/eval/taic_detection.yaml

@@ -5,8 +5,10 @@ split: "val2017"
 ann_file: "annotations/instances_val2017.json"
 quality_level: 1
 lmbda: 0.0035
-checkpoint: "./logs/taic_detection/1/checkpoint_0.1040.pth.tar"
+checkpoint: "./logs/taic_detection/1/checkpoint_0.1334.pth.tar"
 base_codec: "./checkpoints/base_codec/base_codec_1.pth.tar"
+# null = original COCO resolution; 256 = force 256x256 (aspect stretched)
+eval_size: null
 # official task network (required for --with-metrics)
 gpu_id: 0
 cuda: true

configs/taic/detection.yaml

@@ -4,18 +4,18 @@ exp_name: "taic_detection"
 task: "detection"
 dataset_path: "/data/Dataset/coco2017"
 base_codec: "./checkpoints/base_codec/base_codec_1.pth.tar"  # quality matches lmbda mapping
-checkpoint: null # "./logs/taic_detection/1/checkpoint_60.pth.tar"
+checkpoint:  "/data/Dataset/FlexICM/logs/taic_detection/1/checkpoint_best_loss.pth.tar"
 freeze_base_codec: false  # false = fine-tune the entire TIC + SFMA + Task Connector
 epochs: 999
 learning_rate: 1.0e-4
 gpu_id: 0
 quality_level: 1
-# paper lambda set: {0.0035, 0.0067, 0.0130, 0.0250}
+# paper lambda set: {0.0035, 0.0067, 0.0130, 0.0250, 0.2050}
 lmbda: 0.2050
 num_workers: 8
-batch_size: 32
+batch_size: 8
 test_batch_size: 1
-patch_size: 256
+patch_size: 512
 cuda: true
 save: true
 seed: 42

flexicm/data/datasets.py

@@ -28,8 +28,16 @@ def build_train_transform(patch_size: int = 256) -> Callable:
     )
 
 
-def build_test_transform() -> Callable:
-    return transforms.ToTensor()
+def build_test_transform(eval_size: Optional[int] = None) -> Callable:
+    """Eval preprocess. If ``eval_size`` is set (e.g. 256), force HxW = size×size."""
+    if eval_size is None:
+        return transforms.ToTensor()
+    return transforms.Compose(
+        [
+            transforms.Resize((int(eval_size), int(eval_size))),
+            transforms.ToTensor(),
+        ]
+    )
 
 
 class ImageFolderDataset(Dataset):

flexicm/tasks/metric_eval.py

@@ -55,6 +55,12 @@ def run_task_metric_eval(
         for image, meta in zip(images_list, metas):
             image = image.unsqueeze(0).to(device)
             _, _, H, W = image.shape
+            ori_h = int(meta["height"])
+            ori_w = int(meta["width"])
+            # If the image was resized (e.g. eval_size=256), boxes must be
+            # mapped back with the correct scale_factor for COCO mAP.
+            scale_w = float(W) / float(ori_w) if ori_w > 0 else 1.0
+            scale_h = float(H) / float(ori_h) if ori_h > 0 else 1.0
             x, _ = pad_for_codec(image, divisor=align_divisor, device=device)
 
             y_b = None
@@ -77,6 +83,7 @@ def run_task_metric_eval(
             meta = dict(meta)
             meta["pad_height"] = int(x.shape[-2])
             meta["pad_width"] = int(x.shape[-1])
+            meta["scale_factor"] = (scale_w, scale_h)
             pred = runner.predict_from_h(h, meta)
             predictions.append(pred)
 

flexicm/tasks/metric_runners.py

@@ -204,6 +204,11 @@ class DetectionMetricRunner(TaskMetricRunner):
         # FPN strides line up; boxes are rescaled back to ori_shape.
         pad_h = int(img_meta.get("pad_height", h.shape[-2] * 4))
         pad_w = int(img_meta.get("pad_width", h.shape[-1] * 4))
+        sf = img_meta.get("scale_factor", (1.0, 1.0))
+        if isinstance(sf, (int, float)):
+            scale_factor = (float(sf), float(sf))
+        else:
+            scale_factor = (float(sf[0]), float(sf[1]))
         try:
             # MMDet 3.x style
             from mmdet.structures import DetDataSample
@@ -214,7 +219,7 @@ class DetectionMetricRunner(TaskMetricRunner):
                     img_shape=(pad_h, pad_w),
                     ori_shape=(ori_h, ori_w),
                     pad_shape=(pad_h, pad_w),
-                    scale_factor=(1.0, 1.0),
+                    scale_factor=scale_factor,
                     img_id=img_meta.get("image_id"),
                 )
             )
@@ -241,7 +246,7 @@ class DetectionMetricRunner(TaskMetricRunner):
                     img_shape=(pad_h, pad_w, 3),
                     ori_shape=(ori_h, ori_w, 3),
                     pad_shape=(pad_h, pad_w, 3),
-                    scale_factor=1.0,
+                    scale_factor=scale_factor,
                     flip=False,
                 )
             ]

scripts/eval_taic.py

@@ -65,6 +65,12 @@ def parse_args(argv):
     parser.add_argument("--with-metrics", action="store_true", help="Run full task-network metrics")
     parser.add_argument("--max-batches", type=int, default=None)
     parser.add_argument("--split", type=str, default=None)
+    parser.add_argument(
+        "--eval-size",
+        type=int,
+        default=cfg.get("eval_size", None),
+        help="If set (e.g. 256), resize eval images to eval_size×eval_size",
+    )
     args = parser.parse_args(remaining)
     args.config = cfg_path
     if "--actual-bpp" in argv:
@@ -76,7 +82,8 @@ def parse_args(argv):
 
 def build_codec_loader(args, device):
     split = args.split or getattr(args, "split", None) or "val2017"
-    tf = build_test_transform()
+    eval_size = getattr(args, "eval_size", None)
+    tf = build_test_transform(eval_size=eval_size)
     root = args.dataset_path
     split_dir = os.path.join(root, split)
     if os.path.isdir(split_dir):
@@ -101,11 +108,12 @@ def build_metric_loader(args, device):
     if ann_rel is None:
         raise ValueError(f"No ann_file for task={args.task}")
     ann_file = ann_rel if os.path.isabs(ann_rel) else os.path.join(args.dataset_path, ann_rel)
+    eval_size = getattr(args, "eval_size", None)
     dataset = COCOEvalDataset(
         args.dataset_path,
         ann_file=ann_file,
         image_prefix=split,
-        transform=build_test_transform(),
+        transform=build_test_transform(eval_size=eval_size),
     )
     loader = DataLoader(
         dataset,
@@ -152,7 +160,9 @@ def main(argv):
     teacher = teacher.to(device).eval()
     criterion = TAICCriterion(lmbda=lmbda, align_mode=align_mode)
     codec_loader = build_codec_loader(args, device)
-    print(f"[codec] test set size: {len(codec_loader.dataset)}  device={device}  task={task}")
+    eval_size = getattr(args, "eval_size", None)
+    size_msg = f"{eval_size}x{eval_size}" if eval_size else "original"
+    print(f"[codec] test set size: {len(codec_loader.dataset)}  device={device}  task={task}  input={size_msg}")
 
     codec_result = test_taic_loader(
         net,

scripts/train_taic.py

@@ -210,7 +210,7 @@ def main(argv):
     for epoch in range(args.epochs):
         logging.info(f"===== Epoch {epoch}/{args.epochs} =====")
         train_stats = train_one_epoch(net, teacher, train_loader, optimizer, criterion, device)
-        if epoch % 10 == 0:
+        if epoch % 2 == 0:
             val_stats = validate(net, teacher, val_loader, criterion, device)
             logging.info(f"train={train_stats} val={val_stats}")
             is_best = val_stats["loss"] < best