birder 0.3.3__py3-none-any.whl → 0.4.1__py3-none-any.whl

birder/adversarial/base.py

@@ -56,7 +56,7 @@ def pixel_eps_to_normalized(
 
 
 def clamp_normalized(inputs: torch.Tensor, rgb_stats: RGBType) -> torch.Tensor:
-    (min_val, max_val) = normalized_bounds(rgb_stats, device=inputs.device, dtype=inputs.dtype)
+    min_val, max_val = normalized_bounds(rgb_stats, device=inputs.device, dtype=inputs.dtype)
     return torch.clamp(inputs, min=min_val, max=max_val)
 
 

birder/adversarial/simba.py

@@ -87,7 +87,7 @@ class SimBA:
         if self._is_successful(current_logits, label, target_label):
             return adv_inputs.detach(), num_queries
 
-        (_, channels, height, width) = adv_inputs.shape
+        _, channels, height, width = adv_inputs.shape
         num_dims = channels * height * width
         step = pixel_eps_to_normalized(self.step_size, self.rgb_stats, device=adv_inputs.device, dtype=adv_inputs.dtype)
         step_vals = step.view(-1)  # Per-channel steps
@@ -98,11 +98,11 @@ class SimBA:
 
         # Coordinate-wise search in random order
         for flat_idx in perm[:num_steps]:
-            (c, rem) = divmod(int(flat_idx.item()), stride)
-            (h, w) = divmod(rem, width)
+            c, rem = divmod(int(flat_idx.item()), stride)
+            h, w = divmod(rem, width)
             step_val = step_vals[c]
 
-            (candidate_inputs, candidate_logits, candidate_objective) = self._best_candidate(
+            candidate_inputs, candidate_logits, candidate_objective = self._best_candidate(
                 adv_inputs, c, h, w, step_val, label, target_label
             )
             num_queries += 2

birder/common/cli.py

@@ -49,7 +49,7 @@ class FlexibleDictAction(argparse.Action):
             new_dict = {}
             for pair in pairs:
                 # Split each pair into key and value
-                (key, value) = pair.split("=", 1)
+                key, value = pair.split("=", 1)
                 key = key.strip()
 
                 # Try to safely evaluate the value (handles ints and strings mostly)

birder/common/fs_ops.py

@@ -158,7 +158,7 @@ def model_path(
         file_name = f"{file_name}_quantized"
 
     if states is True:
-        file_name = f"{file_name}_states"
+        file_name = f"{file_name}_states.pt"
     elif lite is True:
         file_name = f"{file_name}.ptl"
     elif pt2 is True:
@@ -254,7 +254,7 @@ def clean_checkpoints(network_name: str, keep_last: int) -> None:
     models_glob = str(model_path(network_name, epoch=epoch))
     states_glob = str(model_path(network_name, epoch=epoch, states=True))
     model_pattern = re.compile(r".*_([1-9][0-9]*)\.pt$")
-    states_pattern = re.compile(r".*_([1-9][0-9]*)_states$")
+    states_pattern = re.compile(r".*_([1-9][0-9]*)_states\.pt$")
 
     model_paths = list(settings.BASE_DIR.glob(models_glob))
     for p in sorted(model_paths, key=lambda p: p.stat().st_mtime)[:-keep_last]:
@@ -384,7 +384,7 @@ def load_checkpoint(
     )
 
     # Initialize network and restore checkpoint state
-    net = registry.net_factory(network, input_channels, num_classes, config=config, size=size)
+    net = registry.net_factory(network, num_classes, input_channels, config=config, size=size)
 
     # When a checkpoint was trained with EMA:
     #   The primary weights in the checkpoint file are the EMA weights
@@ -437,7 +437,7 @@ def load_mim_checkpoint(
     size = lib.get_size_from_signature(signature)
 
     # Initialize network and restore checkpoint state
-    net_encoder = registry.net_factory(encoder, input_channels, num_classes, config=encoder_config, size=size)
+    net_encoder = registry.net_factory(encoder, num_classes, input_channels, config=encoder_config, size=size)
     net = registry.mim_net_factory(
         network, net_encoder, config=config, size=size, mask_ratio=mask_ratio, min_mask_size=min_mask_size
     )
@@ -488,7 +488,7 @@ def load_detection_checkpoint(
     size = lib.get_size_from_signature(signature)
 
     # Initialize network and restore checkpoint state
-    net_backbone = registry.net_factory(backbone, input_channels, num_classes, config=backbone_config, size=size)
+    net_backbone = registry.net_factory(backbone, num_classes, input_channels, config=backbone_config, size=size)
     net = registry.detection_net_factory(network, num_classes, net_backbone, config=config, size=size)
 
     # When a checkpoint was trained with EMA:
@@ -584,7 +584,7 @@ def load_model(
             merged_config = None  # type: ignore[assignment]
 
         model_state: dict[str, Any] = safetensors.torch.load_file(path, device=device.type)
-        net = registry.net_factory(network, input_channels, num_classes, config=merged_config, size=size)
+        net = registry.net_factory(network, num_classes, input_channels, config=merged_config, size=size)
         if reparameterized is True:
             net.reparameterize_model()
 
@@ -611,7 +611,7 @@ def load_model(
         if len(merged_config) == 0:
             merged_config = None
 
-        net = registry.net_factory(network, input_channels, num_classes, config=merged_config, size=size)
+        net = registry.net_factory(network, num_classes, input_channels, config=merged_config, size=size)
         if reparameterized is True:
             net.reparameterize_model()
 
@@ -733,7 +733,7 @@ def load_detection_model(
 
         model_state: dict[str, Any] = safetensors.torch.load_file(path, device=device.type)
         net_backbone = registry.net_factory(
-            backbone, input_channels, num_classes, config=backbone_merged_config, size=size
+            backbone, num_classes, input_channels, config=backbone_merged_config, size=size
         )
         if backbone_reparameterized is True:
             net_backbone.reparameterize_model()
@@ -776,7 +776,7 @@ def load_detection_model(
             merged_config = None
 
         net_backbone = registry.net_factory(
-            backbone, input_channels, num_classes, config=backbone_merged_config, size=size
+            backbone, num_classes, input_channels, config=backbone_merged_config, size=size
         )
         if backbone_reparameterized is True:
             net_backbone.reparameterize_model()
@@ -959,7 +959,7 @@ def load_model_with_cfg(
             encoder_name = cfg["encoder"]
 
         encoder_config = cfg.get("encoder_config", None)
-        encoder = registry.net_factory(encoder_name, input_channels, num_classes=0, config=encoder_config, size=size)
+        encoder = registry.net_factory(encoder_name, 0, input_channels, config=encoder_config, size=size)
         net = registry.mim_net_factory(name, encoder, config=model_config, size=size)
 
     elif cfg["task"] == Task.OBJECT_DETECTION:
@@ -969,14 +969,14 @@ def load_model_with_cfg(
             backbone_name = cfg["backbone"]
 
         backbone_config = cfg.get("backbone_config", None)
-        backbone = registry.net_factory(backbone_name, input_channels, num_classes, config=backbone_config, size=size)
+        backbone = registry.net_factory(backbone_name, num_classes, input_channels, config=backbone_config, size=size)
         if cfg.get("backbone_reparameterized", False) is True:
             backbone.reparameterize_model()
 
         net = registry.detection_net_factory(name, num_classes, backbone, config=model_config, size=size)
 
     elif cfg["task"] == Task.IMAGE_CLASSIFICATION:
-        net = registry.net_factory(name, input_channels, num_classes, config=model_config, size=size)
+        net = registry.net_factory(name, num_classes, input_channels, config=model_config, size=size)
 
     else:
         raise ValueError(f"Configuration not supported: {cfg['task']}")
@@ -1019,7 +1019,7 @@ def download_model_by_weights(
             f"Requested format '{file_format}' not available for {weights}, available formats are: {available_formats}"
         )
 
-    (model_file, url) = get_pretrained_model_url(weights, file_format)
+    model_file, url = get_pretrained_model_url(weights, file_format)
     if dst is None:
         dst = settings.MODELS_DIR.joinpath(model_file)
 

birder/common/lib.py

@@ -157,6 +157,6 @@ def get_pretrained_model_url(weights: str, file_format: str) -> tuple[str, str]:
 
 def format_duration(seconds: float) -> str:
     s = int(seconds)
-    (mm, ss) = divmod(s, 60)
-    (hh, mm) = divmod(mm, 60)
+    mm, ss = divmod(s, 60)
+    hh, mm = divmod(mm, 60)
     return f"{hh:d}:{mm:02d}:{ss:02d}"

birder/common/masking.py

@@ -16,7 +16,7 @@ def _mask_token_omission(
     Parameters
     ----------
     x
-        Tensor of shape (N, L, D), where N is the batch size, L is the sequence length, and D is the feature dimension.
+        Tensor of shape (N, L, D), where N is the batch size, L is the sequence length and D is the feature dimension.
     mask_ratio
         The ratio of the sequence length to be masked. This value should be between 0 and 1.
     kept_mask_ratio
@@ -48,7 +48,7 @@ def _mask_token_omission(
     # Masking: length -> length * mask_ratio
     # Perform per-sample random masking by per-sample shuffling.
     # Per-sample shuffling is done by argsort random noise.
-    (N, L, D) = x.size()  # batch, length, dim
+    N, L, D = x.size()  # batch, length, dim
     len_keep = int(L * (1 - mask_ratio))
     len_masked = int(L * (mask_ratio - kept_mask_ratio))
 
@@ -82,7 +82,7 @@ def mask_tensor(
     if channels_last is False:
         x = x.permute(0, 2, 3, 1)
 
-    (B, H, W, _) = x.size()
+    B, H, W, _ = x.size()
 
     shaped_mask = mask.reshape(B, H // patch_factor, W // patch_factor)
     shaped_mask = shaped_mask.repeat_interleave(patch_factor, dim=1).repeat_interleave(patch_factor, dim=2)

birder/common/training_cli.py

@@ -13,6 +13,7 @@ from birder.conf import settings
 from birder.data.datasets.coco import MosaicType
 from birder.data.transforms.classification import AugType
 from birder.data.transforms.classification import RGBMode
+from birder.data.transforms.detection import MULTISCALE_STEP
 from birder.data.transforms.detection import AugType as DetAugType
 
 logger = logging.getLogger(__name__)
@@ -199,10 +200,16 @@ def add_detection_input_args(parser: argparse.ArgumentParser) -> None:
         action="store_true",
         help="enable random square resize once per batch (capped by max(--size))",
     )
+    group.add_argument(
+        "--multiscale-step",
+        type=int,
+        default=MULTISCALE_STEP,
+        help="step size for multiscale size lists and collator padding divisibility (size_divisible)",
+    )
     group.add_argument(
         "--multiscale-min-size",
         type=int,
-        help="minimum short-edge size for multiscale lists (rounded up to nearest multiple of 32)",
+        help="minimum short-edge size for multiscale lists (rounded up to nearest multiple of --multiscale-step)",
     )
 
 
@@ -515,7 +522,10 @@ def add_distributed_args(parser: argparse.ArgumentParser) -> None:
 
 
 def add_logging_and_debug_args(
-    parser: argparse.ArgumentParser, default_log_interval: int = 50, fake_data: bool = True
+    parser: argparse.ArgumentParser,
+    default_log_interval: int = 50,
+    fake_data: bool = True,
+    classification: bool = False,
 ) -> None:
     group = parser.add_argument_group("Logging and debugging parameters")
     group.add_argument(
@@ -525,6 +535,11 @@ def add_logging_and_debug_args(
         metavar="NAME",
         help="experiment name for logging (creates dedicated directory for the run)",
     )
+    if classification is True:
+        group.add_argument(
+            "--top-k", type=int, metavar="K", help="additional top-k accuracy value to track (top-1 is always tracked)"
+        )
+
     group.add_argument(
         "--log-interval",
         type=int,
@@ -746,3 +761,10 @@ def common_args_validation(args: argparse.Namespace) -> None:
     # Precision_args, shared by all scripts
     if args.amp is True and args.model_dtype != "float32":
         raise ValidationError("--amp can only be used with --model-dtype float32")
+
+    if hasattr(args, "top_k") is True and args.top_k is not None:
+        if args.top_k == 1:
+            raise ValidationError("Top-1 accuracy is tracked by default, please remove 1 from --top-k argument")
+
+        if args.top_k <= 0:
+            raise ValidationError("--top-k value must be a positive integer")

birder/common/training_utils.py

@@ -11,6 +11,7 @@ from collections import deque
 from collections.abc import Callable
 from collections.abc import Generator
 from collections.abc import Iterator
+from collections.abc import Sequence
 from datetime import datetime
 from pathlib import Path
 from typing import Any
@@ -361,7 +362,7 @@ def optimizer_parameter_groups(
     Return parameter groups for optimizers with per-parameter group weight decay.
 
     This function creates parameter groups with customizable weight decay, layer-wise
-    learning rate scaling, and special handling for different parameter types. It supports
+    learning rate scaling and special handling for different parameter types. It supports
     advanced optimization techniques like layer decay and custom weight decay rules.
 
     Referenced from https://github.com/pytorch/vision/blob/main/references/classification/utils.py and from
@@ -450,7 +451,7 @@ def optimizer_parameter_groups(
     visited_modules = []
     while len(module_stack_with_prefix) > 0:  # pylint: disable=too-many-nested-blocks
         skip_module = False
-        (module, prefix) = module_stack_with_prefix.pop()
+        module, prefix = module_stack_with_prefix.pop()
         if id(module) in visited_modules:
             skip_module = True
 
@@ -884,6 +885,11 @@ class SmoothedValue:
         self.total: torch.Tensor | float = 0.0
         self.count: int = 0
 
+    def clear(self) -> None:
+        self.deque.clear()
+        self.total = 0.0
+        self.count = 0
+
     def update(self, value: torch.Tensor | float, n: int = 1) -> None:
         self.deque.append(value)
         self.count += n
@@ -927,14 +933,32 @@ class SmoothedValue:
         return to_tensor(v, torch.device("cpu")).item()  # type: ignore[no-any-return]
 
 
-def accuracy(y_true: torch.Tensor, y_pred: torch.Tensor) -> float:
+@torch.no_grad()  # type: ignore[untyped-decorator]
+def accuracy(y_true: torch.Tensor, y_pred: torch.Tensor) -> torch.Tensor:
     if y_pred.dim() > 1 and y_pred.size(1) > 1:
         y_pred = y_pred.argmax(dim=1)
 
     y_true = y_true.flatten()
     y_pred = y_pred.flatten()
 
-    return (y_true == y_pred).float().mean().item()  # type: ignore[no-any-return]
+    return (y_true == y_pred).sum() / y_true.numel()
+
+
+@torch.no_grad()  # type: ignore[untyped-decorator]
+def topk_accuracy(y_true: torch.Tensor, y_pred: torch.Tensor, topk: Sequence[int]) -> list[torch.Tensor]:
+    maxk = min(max(topk), y_pred.size(1))
+    batch_size = y_true.size(0)
+
+    _, pred = y_pred.topk(maxk, dim=1, largest=True, sorted=True)
+    correct = pred.eq(y_true.unsqueeze(1))
+
+    res: list[torch.Tensor] = []
+    for k in topk:
+        k = min(k, maxk)
+        correct_k = correct[:, :k].any(dim=1).sum(dtype=torch.float32)
+        res.append((correct_k / batch_size))
+
+    return res
 
 
 ###############################################################################

birder/data/collators/detection.py

@@ -70,13 +70,21 @@ class BatchRandomResizeCollator(DetectionCollator):
         size: tuple[int, int],
         size_divisible: int = 32,
         multiscale_min_size: Optional[int] = None,
+        multiscale_step: Optional[int] = None,
     ) -> None:
         super().__init__(input_offset, size_divisible=size_divisible)
         if size is None:
             raise ValueError("size must be provided for batch multiscale")
 
         max_side = max(size)
-        sizes = [side for side in build_multiscale_sizes(multiscale_min_size) if side <= max_side]
+        if multiscale_step is None:
+            multiscale_step = size_divisible
+
+        sizes = []
+        for side in build_multiscale_sizes(multiscale_min_size, multiscale_step=multiscale_step):
+            if side <= max_side:
+                sizes.append(side)
+
         if len(sizes) == 0:
             sizes = [max_side]
 

birder/data/transforms/detection.py

@@ -17,17 +17,20 @@ DEFAULT_MULTISCALE_MAX_SIZE = 800
 
 
 def build_multiscale_sizes(
-    min_size: Optional[int] = None, max_size: int = DEFAULT_MULTISCALE_MAX_SIZE
+    min_size: Optional[int] = None, max_size: int = DEFAULT_MULTISCALE_MAX_SIZE, multiscale_step: int = MULTISCALE_STEP
 ) -> tuple[int, ...]:
+    if multiscale_step <= 0:
+        raise ValueError("multiscale_step must be positive")
+
     if min_size is None:
         min_size = DEFAULT_MULTISCALE_MIN_SIZE
 
-    start = int(math.ceil(min_size / MULTISCALE_STEP) * MULTISCALE_STEP)
-    end = int(math.floor(max_size / MULTISCALE_STEP) * MULTISCALE_STEP)
+    start = int(math.ceil(min_size / multiscale_step) * multiscale_step)
+    end = int(math.floor(max_size / multiscale_step) * multiscale_step)
     if end < start:
         return (start,)
 
-    return tuple(range(start, end + 1, MULTISCALE_STEP))
+    return tuple(range(start, end + 1, multiscale_step))
 
 
 class ResizeWithRandomInterpolation(nn.Module):
@@ -59,6 +62,7 @@ def get_birder_augment(
     multiscale: bool,
     max_size: Optional[int],
     multiscale_min_size: Optional[int],
+    multiscale_step: int = MULTISCALE_STEP,
     post_mosaic: bool = False,
 ) -> Callable[..., torch.Tensor]:
     if dynamic_size is True:
@@ -98,7 +102,10 @@ def get_birder_augment(
     # Resize
     if multiscale is True:
         transformations.append(
-            v2.RandomShortestSize(min_size=build_multiscale_sizes(multiscale_min_size), max_size=max_size or 1333),
+            v2.RandomShortestSize(
+                min_size=build_multiscale_sizes(multiscale_min_size, multiscale_step=multiscale_step),
+                max_size=max_size or 1333,
+            ),
         )
     else:
         transformations.append(
@@ -160,6 +167,7 @@ def training_preset(
     multiscale: bool = False,
     max_size: Optional[int] = None,
     multiscale_min_size: Optional[int] = None,
+    multiscale_step: int = MULTISCALE_STEP,
     post_mosaic: bool = False,
 ) -> Callable[..., torch.Tensor]:
     mean = rgv_values["mean"]
@@ -180,7 +188,15 @@ def training_preset(
             [
                 v2.ToImage(),
                 get_birder_augment(
-                    size, level, fill_value, dynamic_size, multiscale, max_size, multiscale_min_size, post_mosaic
+                    size,
+                    level,
+                    fill_value,
+                    dynamic_size,
+                    multiscale,
+                    max_size,
+                    multiscale_min_size,
+                    multiscale_step,
+                    post_mosaic,
                 ),
                 v2.ToDtype(torch.float32, scale=True),
                 v2.Normalize(mean=mean, std=std),
@@ -212,7 +228,10 @@ def training_preset(
         return v2.Compose(  # type: ignore
             [
                 v2.ToImage(),
-                v2.RandomShortestSize(min_size=build_multiscale_sizes(multiscale_min_size), max_size=max_size or 1333),
+                v2.RandomShortestSize(
+                    min_size=build_multiscale_sizes(multiscale_min_size, multiscale_step=multiscale_step),
+                    max_size=max_size or 1333,
+                ),
                 v2.RandomHorizontalFlip(0.5),
                 v2.SanitizeBoundingBoxes(),
                 v2.ToDtype(torch.float32, scale=True),
@@ -284,7 +303,7 @@ def training_preset(
         )
 
     if aug_type == "detr":
-        multiscale_sizes = build_multiscale_sizes(multiscale_min_size)
+        multiscale_sizes = build_multiscale_sizes(multiscale_min_size, multiscale_step=multiscale_step)
         return v2.Compose(  # type: ignore
             [
                 v2.ToImage(),

birder/data/transforms/mosaic.py

@@ -19,7 +19,7 @@ def mosaic_random_center(
     Create a mosaic augmentation by combining 4 images into a single image.
 
     This augmentation places 4 images on a canvas, meeting at a randomly selected
-    center point. Each image is scaled to fit, cropped as needed, and their bounding
+    center point. Each image is scaled to fit, cropped as needed and their bounding
     boxes are transformed accordingly.
 
     Parameters

birder/datahub/classification.py

@@ -63,7 +63,7 @@ class TestDataset(ImageFolder):
         super().__init__(self._root.joinpath(split), transform, target_transform, loader, is_valid_file)
 
     def __getitem__(self, index: int) -> tuple[str, torch.Tensor, Any]:
-        (path, target) = self.samples[index]
+        path, target = self.samples[index]
         sample = self.loader(path)
         if self.transform is not None:
             sample = self.transform(sample)
@@ -122,7 +122,7 @@ class Flowers102(ImageFolder):
         super().__init__(self._root.joinpath(split), transform, target_transform, loader, is_valid_file)
 
     def __getitem__(self, index: int) -> tuple[str, torch.Tensor, Any]:
-        (path, target) = self.samples[index]
+        path, target = self.samples[index]
         sample = self.loader(path)
         if self.transform is not None:
             sample = self.transform(sample)
@@ -182,7 +182,7 @@ class CUB_200_2011(ImageFolder):
         super().__init__(self._root.joinpath(split), transform, target_transform, loader, is_valid_file)
 
     def __getitem__(self, index: int) -> tuple[str, torch.Tensor, Any]:
-        (path, target) = self.samples[index]
+        path, target = self.samples[index]
         sample = self.loader(path)
         if self.transform is not None:
             sample = self.transform(sample)

birder/inference/classification.py

@@ -75,7 +75,7 @@ def infer_batch(
         embedding = embedding_tensor.cpu().float().numpy()
 
     elif tta is True:
-        (_, _, H, W) = inputs.size()
+        _, _, H, W = inputs.size()
         crop_h = int(H * 0.8)
         crop_w = int(W * 0.8)
         tta_inputs = five_crop(inputs, size=[crop_h, crop_w])
@@ -137,7 +137,7 @@ def infer_dataloader_iter(
             inputs = inputs.to(device, dtype=model_dtype)
 
             with torch.amp.autocast(device.type, enabled=amp, dtype=amp_dtype):
-                (out, embedding) = infer_batch(
+                out, embedding = infer_batch(
                     net, inputs, return_embedding=return_embedding, tta=tta, return_logits=return_logits, **kwargs
                 )
 
@@ -394,7 +394,7 @@ def evaluate(
     num_samples: Optional[int] = None,
     sparse: bool = False,
 ) -> Results | SparseResults:
-    (sample_paths, outs, labels, _) = infer_dataloader(
+    sample_paths, outs, labels, _ = infer_dataloader(
         device, net, dataloader, tta=tta, model_dtype=model_dtype, amp=amp, amp_dtype=amp_dtype, num_samples=num_samples
     )
     if sparse is True:

birder/inference/data_parallel.py

@@ -253,7 +253,7 @@ class InferenceDataParallel(nn.Module):
 
         This allows custom methods (e.g., model.embedding()) to be called
         on the InferenceDataParallel instance, which then scatters inputs,
-        calls the method on each replica, and gathers the results.
+        calls the method on each replica and gathers the results.
 
         Parameters
         ----------

birder/inference/detection.py

@@ -20,7 +20,7 @@ def _normalize_image_sizes(inputs: torch.Tensor, image_sizes: Optional[list[list
     if image_sizes is not None:
         return image_sizes
 
-    (_, _, height, width) = inputs.shape
+    _, _, height, width = inputs.shape
     return [[height, width] for _ in range(inputs.size(0))]
 
 
@@ -149,20 +149,20 @@ def infer_batch(
     **kwargs: Any,
 ) -> list[dict[str, torch.Tensor]]:
     if tta is False:
-        (detections, _) = net(inputs, masks=masks, image_sizes=image_sizes, **kwargs)
+        detections, _ = net(inputs, masks=masks, image_sizes=image_sizes, **kwargs)
         return detections  # type: ignore[no-any-return]
 
     normalized_sizes = _normalize_image_sizes(inputs, image_sizes)
     detections_list: list[list[dict[str, torch.Tensor]]] = []
 
     for scale in (0.8, 1.0, 1.2):
-        (scaled_inputs, scaled_masks, scaled_sizes) = _resize_batch(inputs, normalized_sizes, scale, size_divisible=32)
-        (detections, _) = net(scaled_inputs, masks=scaled_masks, image_sizes=scaled_sizes, **kwargs)
+        scaled_inputs, scaled_masks, scaled_sizes = _resize_batch(inputs, normalized_sizes, scale, size_divisible=32)
+        detections, _ = net(scaled_inputs, masks=scaled_masks, image_sizes=scaled_sizes, **kwargs)
         detections = _rescale_detections(detections, scaled_sizes, normalized_sizes)
         detections_list.append(detections)
 
         flipped_inputs = _hflip_inputs(scaled_inputs, scaled_sizes)
-        (flipped_detections, _) = net(flipped_inputs, masks=scaled_masks, image_sizes=scaled_sizes, **kwargs)
+        flipped_detections, _ = net(flipped_inputs, masks=scaled_masks, image_sizes=scaled_sizes, **kwargs)
         flipped_detections = _invert_detections(flipped_detections, scaled_sizes)
         flipped_detections = _rescale_detections(flipped_detections, scaled_sizes, normalized_sizes)
         detections_list.append(flipped_detections)

birder/inference/wbf.py

@@ -182,7 +182,7 @@ def fuse_detections_wbf_single(
     scores_list = [detection["scores"] for detection in detections]
     labels_list = [detection["labels"] for detection in detections]
 
-    (boxes, scores, labels) = weighted_boxes_fusion(
+    boxes, scores, labels = weighted_boxes_fusion(
         boxes_list,
         scores_list,
         labels_list,

birder/introspection/attention_rollout.py

@@ -70,7 +70,7 @@ def compute_rollout(
                     num_to_discard = int(num_allowed * discard_ratio)
                     if num_to_discard > 0:
                         # Drop the smallest allowed values
-                        (_, low_idx) = torch.topk(allowed_values, num_to_discard, largest=False)
+                        _, low_idx = torch.topk(allowed_values, num_to_discard, largest=False)
                         allowed_values[low_idx] = 0
                         attn[allow] = allowed_values
                         attention_heads_fused[0] = attn
@@ -97,7 +97,7 @@ def compute_rollout(
 
     # Normalize and reshape to 2D map using actual patch grid dimensions
     mask = mask / (mask.max() + 1e-8)
-    (grid_h, grid_w) = patch_grid_shape
+    grid_h, grid_w = patch_grid_shape
     mask = mask.reshape(grid_h, grid_w)
 
     return mask
@@ -141,7 +141,7 @@ class AttentionRollout:
         net: nn.Module,
         device: torch.device,
         transform: Callable[..., torch.Tensor],
-        attention_layer_name: str = "self_attention",
+        attention_layer_name: str = "attn",
         discard_ratio: float = 0.9,
         head_fusion: Literal["mean", "max", "min"] = "max",
     ) -> None:
@@ -156,11 +156,11 @@ class AttentionRollout:
         self.attention_gatherer = AttentionGatherer(net, attention_layer_name)
 
     def __call__(self, image: str | Path | Image.Image, target_class: Optional[int] = None) -> InterpretabilityResult:
-        (input_tensor, rgb_img) = preprocess_image(image, self.transform, self.device)
+        input_tensor, rgb_img = preprocess_image(image, self.transform, self.device)
 
-        (attentions, logits) = self.attention_gatherer(input_tensor)
+        attentions, logits = self.attention_gatherer(input_tensor)
 
-        (_, _, H, W) = input_tensor.shape
+        _, _, H, W = input_tensor.shape
         patch_grid_shape = (H // self.net.stem_stride, W // self.net.stem_stride)
 
         attention_map = compute_rollout(

birder/introspection/feature_pca.py

@@ -17,7 +17,7 @@ class FeaturePCA:
     Visualizes feature maps using Principal Component Analysis
 
     This method extracts feature maps from a specified stage of a DetectorBackbone model,
-    applies PCA to reduce the channel dimension to 3 components, and visualizes them as an RGB image where:
+    applies PCA to reduce the channel dimension to 3 components and visualizes them as an RGB image where:
     - R channel = 1st principal component (most important)
     - G channel = 2nd principal component
     - B channel = 3rd principal component
@@ -40,7 +40,7 @@ class FeaturePCA:
         self.stage = stage
 
     def __call__(self, image: str | Path | Image.Image) -> InterpretabilityResult:
-        (input_tensor, rgb_img) = preprocess_image(image, self.transform, self.device)
+        input_tensor, rgb_img = preprocess_image(image, self.transform, self.device)
 
         with torch.inference_mode():
             features_dict = self.net.detection_features(input_tensor)
@@ -54,11 +54,11 @@ class FeaturePCA:
 
         # Handle channels_last format (B, H, W, C) vs channels_first (B, C, H, W)
         if self.channels_last is True:
-            (B, H, W, C) = features_np.shape
+            B, H, W, C = features_np.shape
             # Already in (B, H, W, C), just reshape to (B*H*W, C)
             features_reshaped = features_np.reshape(-1, C)
         else:
-            (B, C, H, W) = features_np.shape
+            B, C, H, W = features_np.shape
             # Reshape to (spatial_points, channels) for PCA
             features_reshaped = features_np.reshape(B, C, -1)
             features_reshaped = features_reshaped.transpose(0, 2, 1)  # (B, H*W, C)

birder/introspection/gradcam.py

@@ -98,7 +98,7 @@ class GradCAM:
         self.activation_capture = ActivationCapture(net, target_layer, reshape_transform)
 
     def __call__(self, image: str | Path | Image.Image, target_class: Optional[int] = None) -> InterpretabilityResult:
-        (input_tensor, rgb_img) = preprocess_image(image, self.transform, self.device)
+        input_tensor, rgb_img = preprocess_image(image, self.transform, self.device)
         input_tensor.requires_grad_(True)
 
         # Forward pass