birder 0.3.2__tar.gz → 0.3.3__tar.gz

{birder-0.3.2 → birder-0.3.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: birder
-Version: 0.3.2
+Version: 0.3.3
 Summary: An open-source computer vision framework for wildlife image analysis, featuring state-of-the-art models for species classification and detection.
 Author: Ofer Hasson
 License-Expression: Apache-2.0

{birder-0.3.2 → birder-0.3.3}/birder/net/hiera.py

@@ -301,14 +301,14 @@ class HieraBlock(nn.Module):
         self.dim = dim
         self.dim_out = dim_out
 
-        self.norm1 = nn.LayerNorm(dim)
+        self.norm1 = nn.LayerNorm(dim, eps=1e-6)
         if dim != dim_out:
             self.proj = nn.Linear(dim, dim_out)
         else:
             self.proj = None
 
         self.attn = MaskUnitAttention(dim, dim_out, heads, q_stride, window_size, use_mask_unit_attn)
-        self.norm2 = nn.LayerNorm(dim_out)
+        self.norm2 = nn.LayerNorm(dim_out, eps=1e-6)
         self.mlp = MLP(dim_out, [int(dim_out * mlp_ratio), dim_out], activation_layer=nn.GELU)
         self.drop_path = StochasticDepth(drop_path, mode="row")
 
@@ -450,7 +450,7 @@ class Hiera(DetectorBackbone, PreTrainEncoder, MaskedTokenOmissionMixin):
         self.body = nn.Sequential(stages)
         self.features = nn.Sequential(
             attn_pool if attn_pool is not None else AvgTokens(),
-            nn.LayerNorm(embed_dim),
+            nn.LayerNorm(embed_dim, eps=1e-6),
             nn.Flatten(1),
         )
         self.return_channels = return_channels

{birder-0.3.2 → birder-0.3.3}/birder/net/hieradet.py

@@ -125,7 +125,7 @@ class MultiScaleBlock(nn.Module):
         self.dim = dim
         self.dim_out = dim_out
 
-        self.norm1 = nn.LayerNorm(dim)
+        self.norm1 = nn.LayerNorm(dim, eps=1e-6)
         if dim != dim_out:
             self.proj = nn.Linear(dim, dim_out)
         else:
@@ -144,7 +144,7 @@ class MultiScaleBlock(nn.Module):
             num_heads=num_heads,
             q_pool=copy.deepcopy(self.pool),
         )
-        self.norm2 = nn.LayerNorm(dim_out)
+        self.norm2 = nn.LayerNorm(dim_out, eps=1e-6)
         self.mlp = MLP(dim_out, [int(dim_out * mlp_ratio), dim_out], activation_layer=nn.GELU)
         self.drop_path = StochasticDepth(drop_path, mode="row")
 
@@ -173,11 +173,9 @@ class MultiScaleBlock(nn.Module):
         if self.q_stride is not None:
             # Shapes have changed due to Q pooling
             window_size = self.window_size // self.q_stride[0]
-            (H, W) = (shortcut.size(1), shortcut.size(2))
+            pad_hw = (pad_hw[0] // self.q_stride[0], pad_hw[1] // self.q_stride[1])
 
-            pad_h = (window_size - H % window_size) % window_size
-            pad_w = (window_size - W % window_size) % window_size
-            pad_hw = (H + pad_h, W + pad_w)
+            (H, W) = (shortcut.size(1), shortcut.size(2))
 
         # Reverse window partition
         if self.window_size > 0:
@@ -271,7 +269,7 @@ class HieraDet(DetectorBackbone, PreTrainEncoder, MaskedTokenRetentionMixin):
 
         self.body = nn.Sequential(stages)
         self.features = nn.Sequential(
-            nn.LayerNorm(embed_dim),
+            nn.LayerNorm(embed_dim, eps=1e-6),
             Permute([0, 3, 1, 2]),  # B H W C -> B C H W
             nn.AdaptiveAvgPool2d(output_size=(1, 1)),
             nn.Flatten(1),
@@ -415,7 +413,7 @@ registry.register_model_config(
         "num_heads": 1,
         "global_pos_size": (7, 7),
         "global_att_blocks": [5, 7, 9],
-        "window_spec": [8, 4, -16, -32],
+        "window_spec": [8, 4, 14, 7],
         "drop_path_rate": 0.1,
     },
 )
@@ -428,7 +426,7 @@ registry.register_model_config(
         "num_heads": 1,
         "global_pos_size": (7, 7),
         "global_att_blocks": [7, 10, 13],
-        "window_spec": [8, 4, -16, -32],
+        "window_spec": [8, 4, 14, 7],
         "drop_path_rate": 0.1,
     },
 )
@@ -441,7 +439,7 @@ registry.register_model_config(
         "num_heads": 1,
         "global_pos_size": (14, 14),
         "global_att_blocks": [12, 16, 20],
-        "window_spec": [8, 4, -16, -32],
+        "window_spec": [8, 4, 14, 7],
         "drop_path_rate": 0.1,
     },
 )
@@ -454,7 +452,7 @@ registry.register_model_config(
         "num_heads": 2,
         "global_pos_size": (14, 14),
         "global_att_blocks": [12, 16, 20],
-        "window_spec": [8, 4, -16, -32],
+        "window_spec": [8, 4, 14, 7],
         "drop_path_rate": 0.1,
     },
 )
@@ -467,17 +465,84 @@ registry.register_model_config(
         "num_heads": 2,
         "global_pos_size": (7, 7),
         "global_att_blocks": [23, 33, 43],
-        "window_spec": [8, 4, -16, -32],
+        "window_spec": [8, 4, 14, 7],
+        "drop_path_rate": 0.2,
+    },
+)
+
+# Dynamic window size
+registry.register_model_config(
+    "hieradet_d_tiny",
+    HieraDet,
+    config={
+        "depths": [1, 2, 7, 2],
+        "embed_dim": 96,
+        "num_heads": 1,
+        "global_pos_size": (7, 7),
+        "global_att_blocks": [5, 7, 9],
+        "window_spec": [8, 4, 0, 0],
+        "drop_path_rate": 0.1,
+    },
+)
+registry.register_model_config(
+    "hieradet_d_small",
+    HieraDet,
+    config={
+        "depths": [1, 2, 11, 2],
+        "embed_dim": 96,
+        "num_heads": 1,
+        "global_pos_size": (7, 7),
+        "global_att_blocks": [7, 10, 13],
+        "window_spec": [8, 4, 0, 0],
+        "drop_path_rate": 0.1,
+    },
+)
+registry.register_model_config(
+    "hieradet_d_base",
+    HieraDet,
+    config={
+        "depths": [2, 3, 16, 3],
+        "embed_dim": 96,
+        "num_heads": 1,
+        "global_pos_size": (14, 14),
+        "global_att_blocks": [12, 16, 20],
+        "window_spec": [8, 4, 0, 0],
+        "drop_path_rate": 0.1,
+    },
+)
+registry.register_model_config(
+    "hieradet_d_base_plus",
+    HieraDet,
+    config={
+        "depths": [2, 3, 16, 3],
+        "embed_dim": 112,
+        "num_heads": 2,
+        "global_pos_size": (14, 14),
+        "global_att_blocks": [12, 16, 20],
+        "window_spec": [8, 4, 0, 0],
+        "drop_path_rate": 0.1,
+    },
+)
+registry.register_model_config(
+    "hieradet_d_large",
+    HieraDet,
+    config={
+        "depths": [2, 6, 36, 4],
+        "embed_dim": 144,
+        "num_heads": 2,
+        "global_pos_size": (7, 7),
+        "global_att_blocks": [23, 33, 43],
+        "window_spec": [8, 4, 0, 0],
         "drop_path_rate": 0.2,
     },
 )
 
 registry.register_weights(
-    "hieradet_small_dino-v2",
+    "hieradet_d_small_dino-v2",
     {
-        "url": "https://huggingface.co/birder-project/hieradet_small_dino-v2/resolve/main",
+        "url": "https://huggingface.co/birder-project/hieradet_d_small_dino-v2/resolve/main",
         "description": (
-            "HieraDet small image encoder pre-trained using DINOv2. "
+            "HieraDet (d) small image encoder pre-trained using DINOv2. "
             "This model has not been fine-tuned for a specific classification task"
         ),
         "resolution": (224, 224),
@@ -487,14 +552,16 @@ registry.register_weights(
                 "sha256": "eb41b8a35445e7f350797094d5e365306b29351e64edd4a316420c23d1e17073",
             }
         },
-        "net": {"network": "hieradet_small", "tag": "dino-v2"},
+        "net": {"network": "hieradet_d_small", "tag": "dino-v2"},
     },
 )
 registry.register_weights(
-    "hieradet_small_dino-v2-inat21-256px",
+    "hieradet_d_small_dino-v2-inat21-256px",
     {
-        "url": "https://huggingface.co/birder-project/hieradet_small_dino-v2-inat21/resolve/main",
-        "description": "HieraDet small model pre-trained using DINOv2, then fine-tuned on the iNaturalist 2021 dataset",
+        "url": "https://huggingface.co/birder-project/hieradet_d_small_dino-v2-inat21/resolve/main",
+        "description": (
+            "HieraDet (d) small model pre-trained using DINOv2, then fine-tuned on the iNaturalist 2021 dataset"
+        ),
         "resolution": (256, 256),
         "formats": {
             "pt": {
@@ -502,14 +569,16 @@ registry.register_weights(
                 "sha256": "e1bdeba97eae816ec3ab9b3238d97decf2c34d29b70f9291116ce962b9a4f9df",
             }
         },
-        "net": {"network": "hieradet_small", "tag": "dino-v2-inat21-256px"},
+        "net": {"network": "hieradet_d_small", "tag": "dino-v2-inat21-256px"},
     },
 )
 registry.register_weights(
-    "hieradet_small_dino-v2-inat21",
+    "hieradet_d_small_dino-v2-inat21",
     {
-        "url": "https://huggingface.co/birder-project/hieradet_small_dino-v2-inat21/resolve/main",
-        "description": "HieraDet small model pre-trained using DINOv2, then fine-tuned on the iNaturalist 2021 dataset",
+        "url": "https://huggingface.co/birder-project/hieradet_d_small_dino-v2-inat21/resolve/main",
+        "description": (
+            "HieraDet (d) small model pre-trained using DINOv2, then fine-tuned on the iNaturalist 2021 dataset"
+        ),
         "resolution": (384, 384),
         "formats": {
             "pt": {
@@ -517,14 +586,14 @@ registry.register_weights(
                 "sha256": "271fa9ed6a9aa1f4d1fc8bbb4c4cac9d15b264f2ac544efb5cd971412691880d",
             }
         },
-        "net": {"network": "hieradet_small", "tag": "dino-v2-inat21"},
+        "net": {"network": "hieradet_d_small", "tag": "dino-v2-inat21"},
     },
 )
 registry.register_weights(
-    "hieradet_small_dino-v2-imagenet12k",
+    "hieradet_d_small_dino-v2-imagenet12k",
     {
-        "url": "https://huggingface.co/birder-project/hieradet_small_dino-v2-imagenet12k/resolve/main",
-        "description": "HieraDet small model pre-trained using DINOv2, then fine-tuned on the ImageNet-12K dataset",
+        "url": "https://huggingface.co/birder-project/hieradet_d_small_dino-v2-imagenet12k/resolve/main",
+        "description": "HieraDet (d) small model pre-trained using DINOv2, then fine-tuned on the ImageNet-12K dataset",
         "resolution": (256, 256),
         "formats": {
             "pt": {
@@ -532,6 +601,25 @@ registry.register_weights(
                 "sha256": "b89dd6c13d061fe8a09d051bb3d76e632e650067ca71578e37b02033107c9963",
             }
         },
-        "net": {"network": "hieradet_small", "tag": "dino-v2-imagenet12k"},
+        "net": {"network": "hieradet_d_small", "tag": "dino-v2-imagenet12k"},
+    },
+)
+
+registry.register_weights(  # SAM v2: https://arxiv.org/abs/2408.00714
+    "hieradet_small_sam2_1",
+    {
+        "url": "https://huggingface.co/birder-project/hieradet_small_sam2_1/resolve/main",
+        "description": (
+            "HieraDet small image encoder pre-trained by Meta AI using SAM v2. "
+            "This model has not been fine-tuned for a specific classification task"
+        ),
+        "resolution": (224, 224),
+        "formats": {
+            "pt": {
+                "file_size": 129.6,
+                "sha256": "79b6ffdfd4ea9f3b1489ce5a229fe9756b215fc3b52640d01d64136560c1d341",
+            }
+        },
+        "net": {"network": "hieradet_small", "tag": "sam2_1"},
     },
 )

{birder-0.3.2 → birder-0.3.3}/birder/scripts/train.py

@@ -762,47 +762,49 @@ def get_args_parser() -> argparse.ArgumentParser:
         epilog=(
             "Usage examples\n"
             "==============\n"
-            "Training on image directory in the default location can be as simple as:\n"
-            "python train.py --network densenet_161 --lr-scheduler cosine --batch-size 64 --smoothing-alpha 0.1\n"
+            "Training on an image directory using default locations:\n"
+            "python train.py --network densenet_161 --batch-size 64 --lr-scheduler cosine --smoothing-alpha 0.1\n"
             "\n"
-            "A more complicated ImageNet example on 2 GPUs, using a remote webdataset:\n"
+            "A more advanced ImageNet example on 2 GPUs using a remote WebDataset:\n"
             "torchrun --nproc_per_node=2 train.py \\\n"
             "    --network resnet_v2_50 \\\n"
             "    --tag imagenet1k \\\n"
+            "    --bce-loss --bce-threshold 0.2 \\\n"
+            "    --batch-size 128 \\\n"
             "    --opt lamb \\\n"
+            "    --grad-accum-steps 4 \\\n"
             "    --lr 0.005 \\\n"
+            "    --wd 0.02 \\\n"
             "    --lr-scheduler cosine \\\n"
             "    --lr-cosine-min 1e-7 \\\n"
-            "    --warmup-epochs 5 \\\n"
             "    --epochs 300 \\\n"
-            "    --wd 0.02 \\\n"
-            "    --grad-accum-steps 4 \\\n"
-            "    --mixup-alpha 0.1 \\\n"
-            "    --cutmix \\\n"
+            "    --warmup-epochs 5 \\\n"
             "    --aug-type ra \\\n"
             "    --re-prob 0.25 \\\n"
+            "    --mixup-alpha 0.1 \\\n"
+            "    --cutmix \\\n"
             "    --rgb-mode imagenet \\\n"
-            "    --bce-loss --bce-threshold 0.2 \\\n"
-            "    --amp \\\n"
+            "    --amp --amp-dtype bfloat16 \\\n"
             "    --compile \\\n"
             "    --wds \\\n"
-            "    --wds-class-file public_datasets_metadata/imagenet-1k-classes.txt \\\n"
             "    --wds-info https://huggingface.co/datasets/timm/imagenet-1k-wds/resolve/main/_info.json \\\n"
+            "    --wds-class-file public_datasets_metadata/imagenet-1k-classes.txt \\\n"
             "    --wds-training-split train\n"
             "\n"
-            "The script is also available as module:\n"
+            "The script can also be invoked as a module:\n"
             "torchrun --nproc_per_node=2 -m birder.scripts.train \\\n"
             "    --network regnet_y_8g \\\n"
+            "    --batch-size 64 \\\n"
             "    --lr 0.4 \\\n"
+            "    --wd 0.00005 \\\n"
             "    --lr-scheduler cosine \\\n"
             "    --warmup-epochs 5 \\\n"
-            "    --wd 0.00005 \\\n"
+            "    --aug-level 8 \\\n"
             "    --smoothing-alpha 0.1 \\\n"
             "    --mixup-alpha 0.2 \\\n"
             "    --cutmix \\\n"
-            "    --aug-level 8 \\\n"
             "    --ra-sampler --ra-reps 2 \\\n"
-            "    --amp --compile\n"
+            "    --compile\n"
         ),
         formatter_class=cli.ArgumentHelpFormatter,
     )

{birder-0.3.2 → birder-0.3.3}/birder/scripts/train_kd.py

@@ -356,7 +356,7 @@ def train(args: argparse.Namespace) -> None:
 
     # Distillation
     if distillation_type == "soft":
-        distillation_criterion = torch.nn.KLDivLoss(reduction="batchmean", log_target=False)
+        distillation_criterion = torch.nn.KLDivLoss(reduction="batchmean", log_target=True)
     elif distillation_type == "hard":
         distillation_criterion = torch.nn.CrossEntropyLoss()
     elif distillation_type == "deit":
@@ -625,7 +625,7 @@ def train(args: argparse.Namespace) -> None:
                     with torch.no_grad():
                         teacher_outputs = teacher(inputs)
                         if distillation_type == "soft":
-                            teacher_targets = F.softmax(teacher_outputs / args.temperature, dim=-1)
+                            teacher_targets = F.log_softmax(teacher_outputs / args.temperature, dim=-1)
                         else:
                             teacher_targets = teacher_outputs.argmax(dim=-1)
 
@@ -900,43 +900,44 @@ def get_args_parser() -> argparse.ArgumentParser:
             "A typical 'soft' distillation:\n"
             "torchrun --nproc_per_node=2 train_kd.py \\\n"
             "    --type soft \\\n"
-            "    --temperature 1 \\\n"
             "    --teacher vit_l16 \\\n"
             "    --student tiny_vit_5m \\\n"
+            "    --temperature 3.5 \\\n"
+            "    --batch-size 32 \\\n"
             "    --opt adamw \\\n"
+            "    --clip-grad-norm 5 \\\n"
             "    --lr 0.002 \\\n"
+            "    --wd 0.01 \\\n"
+            "    --norm-wd 0 \\\n"
             "    --lr-scheduler cosine \\\n"
             "    --lr-cosine-min 1e-7 \\\n"
-            "    --batch-size 64 \\\n"
             "    --warmup-epochs 5 \\\n"
-            "    --wd 0.01 \\\n"
-            "    --norm-wd 0 \\\n"
             "    --smoothing-alpha 0.1 \\\n"
-            "    --clip-grad-norm 5 \\\n"
-            "    --amp \\\n"
+            "    --amp --amp-dtype bfloat16 \\\n"
             "    --compile \\\n"
             "    --wds \\\n"
-            "    --wds-class-file data/intermediate_packed/classes.txt \\\n"
-            "    --wds-info data/intermediate_packed/_info.json\n"
+            "    --wds-info data/intermediate_packed/_info.json \\\n"
+            "    --wds-class-file data/intermediate_packed/classes.txt\n"
             "\n"
-            "DeiT style distillation:\n"
+            "DeiT-style distillation:\n"
             "torchrun --nproc_per_node=2 train_kd.py \\\n"
             "    --type deit \\\n"
             "    --teacher regnet_y_8g \\\n"
             "    --student deit_s16 \\\n"
+            "    --batch-size 64 \\\n"
             "    --opt adamw \\\n"
+            "    --clip-grad-norm 1 \\\n"
             "    --lr 0.0005 \\\n"
-            "    --lr-scheduler cosine \\\n"
-            "    --warmup-epochs 5 \\\n"
-            "    --epochs 300 \\\n"
             "    --wd 0.05 \\\n"
             "    --norm-wd 0 \\\n"
+            "    --lr-scheduler cosine \\\n"
+            "    --epochs 300 \\\n"
+            "    --warmup-epochs 5 \\\n"
+            "    --aug-level 8 \\\n"
             "    --smoothing-alpha 0.1 \\\n"
             "    --mixup-alpha 0.8 \\\n"
-            "    --aug-level 8 \\\n"
             "    --model-ema \\\n"
             "    --ra-sampler --ra-reps 2 \\\n"
-            "    --clip-grad-norm 1 \\\n"
             "    --amp \\\n"
             "    --compile\n"
         ),

{birder-0.3.2 → birder-0.3.3}/birder/tools/show_iterator.py

@@ -140,10 +140,16 @@ def show_iterator(args: argparse.Namespace) -> None:
         mask_size = (args.size[0] // args.patch_size, args.size[1] // args.patch_size)
         mask_generator: Optional[masking.Masking]
         if args.masking == "uniform":
-            mask_generator = masking.UniformMasking(mask_size, args.mask_ratio)
+            mask_generator = masking.UniformMasking(mask_size, args.mask_ratio, min_mask_size=args.min_mask_size)
         elif args.masking == "block":
             max_patches = int(args.mask_ratio * mask_size[0] * mask_size[1])
             mask_generator = masking.BlockMasking(mask_size, 4, max_patches, 0.33, 3.33)
+        elif args.masking == "roll-block":
+            num_masking_patches = int(args.mask_ratio * mask_size[0] * mask_size[1])
+            mask_generator = masking.RollBlockMasking(mask_size, num_masking_patches=num_masking_patches)
+        elif args.masking == "inverse-roll":
+            num_masking_patches = int(args.mask_ratio * mask_size[0] * mask_size[1])
+            mask_generator = masking.InverseRollBlockMasking(mask_size, num_masking_patches=num_masking_patches)
         else:
             mask_generator = None
 
@@ -206,8 +212,16 @@ def set_parser(subparsers: Any) -> None:
     )
     subparser.add_argument("--mixup-alpha", type=float, help="mixup alpha")
     subparser.add_argument("--cutmix", default=False, action="store_true", help="enable cutmix")
-    subparser.add_argument("--masking", type=str, choices=["uniform", "block"], help="enable masking")
+    subparser.add_argument(
+        "--masking",
+        type=str,
+        choices=["uniform", "block", "roll-block", "inverse-roll"],
+        help="masking strategy to apply",
+    )
     subparser.add_argument("--mask-ratio", type=float, default=0.5, help="mask ratio")
+    subparser.add_argument(
+        "--min-mask-size", type=int, default=1, help="minimum mask unit size in patches (uniform only)"
+    )
     subparser.add_argument("--patch-size", type=int, default=16, help="mask base patch size")
     subparser.add_argument(
         "--data-path", type=str, default=str(settings.TRAINING_DATA_PATH), help="image directory path"

birder-0.3.3/birder/version.py

@@ -0,0 +1 @@
+__version__ = "v0.3.3"

{birder-0.3.2 → birder-0.3.3}/birder.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: birder
-Version: 0.3.2
+Version: 0.3.3
 Summary: An open-source computer vision framework for wildlife image analysis, featuring state-of-the-art models for species classification and detection.
 Author: Ofer Hasson
 License-Expression: Apache-2.0

{birder-0.3.2 → birder-0.3.3}/tests/test_net.py

@@ -71,6 +71,7 @@ NET_TEST_CASES = [
     ("hiera_abswin_tiny"),  # No bfloat16 support
     ("hiera_abswin_base_plus_ap"),  # No bfloat16 support
     ("hieradet_tiny"),
+    ("hieradet_d_tiny"),
     ("hornet_tiny_7x7"),
     ("hornet_tiny_gf"),  # PT2 fails, no bfloat16 support
     ("iformer_s"),
@@ -400,6 +401,7 @@ class TestNet(unittest.TestCase):
             ("hiera_abswin_tiny"),
             ("hiera_abswin_base_plus_ap"),
             ("hieradet_tiny"),
+            ("hieradet_d_tiny"),
             ("hornet_tiny_7x7"),
             ("hornet_tiny_gf"),
             ("iformer_s"),
@@ -575,6 +577,7 @@ class TestNet(unittest.TestCase):
             ("focalnet_t_srf"),
             ("gc_vit_xxt"),
             ("hieradet_tiny"),
+            ("hieradet_d_tiny"),
             ("iformer_s"),
             ("maxvit_t"),
             ("poolformer_v1_s12"),
@@ -793,6 +796,7 @@ class TestNonSquareNet(unittest.TestCase):
             ("hiera_abswin_tiny"),
             ("hiera_abswin_base_plus_ap"),
             ("hieradet_tiny"),
+            ("hieradet_d_tiny"),
             ("hornet_tiny_7x7"),
             ("hornet_tiny_gf"),
             ("iformer_s"),

birder-0.3.2/birder/version.py

@@ -1 +0,0 @@
-__version__ = "v0.3.2"