birder 0.3.2__tar.gz → 0.4.0__tar.gz

{birder-0.3.2 → birder-0.4.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: birder
-Version: 0.3.2
+Version: 0.4.0
 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-0.4.0}/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]:

{birder-0.3.2 → birder-0.4.0}/birder/introspection/attention_rollout.py

@@ -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:

{birder-0.3.2 → birder-0.4.0}/birder/introspection/transformer_attribution.py

@@ -132,7 +132,7 @@ class TransformerAttribution:
         net: nn.Module,
         device: torch.device,
         transform: Callable[..., torch.Tensor],
-        attention_layer_name: str = "self_attention",
+        attention_layer_name: str = "attn",
     ) -> None:
         self.net = net.eval()
         self.device = device

{birder-0.3.2 → birder-0.4.0}/birder/layers/layer_scale.py

@@ -23,7 +23,7 @@ class LayerScale2d(nn.Module):
     def __init__(self, dim: int, init_value: float, inplace: bool = False) -> None:
         super().__init__()
         self.inplace = inplace
-        self.gamma = nn.Parameter(init_value * torch.ones(dim, 1, 1), requires_grad=True)
+        self.gamma = nn.Parameter(init_value * torch.ones(dim, 1, 1))
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:
         if self.inplace is True:

{birder-0.3.2 → birder-0.4.0}/birder/net/__init__.py

@@ -55,8 +55,7 @@ from birder.net.metaformer import MetaFormer
 from birder.net.mnasnet import MNASNet
 from birder.net.mobilenet_v1 import MobileNet_v1
 from birder.net.mobilenet_v2 import MobileNet_v2
-from birder.net.mobilenet_v3_large import MobileNet_v3_Large
-from birder.net.mobilenet_v3_small import MobileNet_v3_Small
+from birder.net.mobilenet_v3 import MobileNet_v3
 from birder.net.mobilenet_v4 import MobileNet_v4
 from birder.net.mobilenet_v4_hybrid import MobileNet_v4_Hybrid
 from birder.net.mobileone import MobileOne
@@ -84,9 +83,6 @@ from birder.net.resnext import ResNeXt
 from birder.net.rope_deit3 import RoPE_DeiT3
 from birder.net.rope_flexivit import RoPE_FlexiViT
 from birder.net.rope_vit import RoPE_ViT
-from birder.net.se_resnet_v1 import SE_ResNet_v1
-from birder.net.se_resnet_v2 import SE_ResNet_v2
-from birder.net.se_resnext import SE_ResNeXt
 from birder.net.sequencer2d import Sequencer2d
 from birder.net.shufflenet_v1 import ShuffleNet_v1
 from birder.net.shufflenet_v2 import ShuffleNet_v2
@@ -171,8 +167,7 @@ __all__ = [
     "MNASNet",
     "MobileNet_v1",
     "MobileNet_v2",
-    "MobileNet_v3_Large",
-    "MobileNet_v3_Small",
+    "MobileNet_v3",
     "MobileNet_v4",
     "MobileNet_v4_Hybrid",
     "MobileOne",
@@ -200,9 +195,6 @@ __all__ = [
     "RoPE_DeiT3",
     "RoPE_FlexiViT",
     "RoPE_ViT",
-    "SE_ResNet_v1",
-    "SE_ResNet_v2",
-    "SE_ResNeXt",
     "Sequencer2d",
     "ShuffleNet_v1",
     "ShuffleNet_v2",

birder-0.4.0/birder/net/_rope_vit_configs.py

@@ -0,0 +1,430 @@
+"""
+RoPE ViT model configuration registrations
+
+This file contains *only* model variant definitions and their registration
+with the global model registry. The actual RoPE ViT implementation lives in rope_vit.py.
+"""
+
+from birder.model_registry import registry
+from birder.net._vit_configs import BASE
+from birder.net._vit_configs import GIANT
+from birder.net._vit_configs import GIGANTIC
+from birder.net._vit_configs import HUGE
+from birder.net._vit_configs import LARGE
+from birder.net._vit_configs import MEDIUM
+from birder.net._vit_configs import SMALL
+from birder.net._vit_configs import SO150
+from birder.net._vit_configs import SO400
+from birder.net._vit_configs import TINY
+from birder.net.base import BaseNet
+
+# Vision Transformer Model Naming Convention
+# ==========================================
+#
+# Model names follow a structured pattern to encode architectural choices:
+# [rope_]vit_[reg{N}_][size][patch_size][_components][_pooling][_c{N}]
+#
+# Core Components:
+# - rope_       : Rotary Position Embedding (RoPE) enabled
+# - rope_i_     : Rotary Position Embedding (RoPE) enabled with interleaved rotation - implies different temp, indexing
+# - vit_        : Vision Transformer base architecture
+# - reg{N}_     : Register tokens (N = number of register tokens, e.g., reg4, reg8)
+# - size        : Model size (s=small, b=base, l=large, or specific like so150m)
+# - patch_size  : Patch size (e.g., 14, 16, 32 for 14x14, 16x16, 32x32 patches)
+#
+# Optional Components:
+#     Position Embeddings:
+#     - nps         : No Position embedding on Special tokens
+#
+#     Normalization:
+#     - rms         : RMSNorm (instead of LayerNorm)
+#     - pn          : Pre-Norm (layer norm before the encoder) - implies norm eps of 1e-5
+#     - npn         : No Post Norm (disables post-normalization layer)
+#     - qkn         : QK Norm
+#
+#     Feed-Forward Network:
+#     - swiglu      : SwiGLU FFN layer type (instead of standard FFN)
+#
+#     Activation:
+#     - quick_gelu  : QuickGELU activation type
+#     - ...
+#
+#     Regularization:
+#     - ls          : Layer Scaling applied
+#
+#     Pooling/Reduction:
+#     - avg         : Average pooling for sequence reduction
+#     - ap          : Attention Pooling for sequence reduction
+#     - aps         : Attention Pooling inc. Special tokens for sequence reduction
+#
+#     Custom Variants:
+#     - c{N}        : Custom variant (N = version number) for models with fine-grained or non-standard
+#                     modifications not fully reflected in the name
+
+
+def register_rope_vit_configs(rope_vit: type[BaseNet]) -> None:
+    registry.register_model_config(
+        "rope_vit_t32",
+        rope_vit,
+        config={"patch_size": 32, **TINY},
+    )
+    registry.register_model_config(
+        "rope_vit_t16",
+        rope_vit,
+        config={"patch_size": 16, **TINY},
+    )
+    registry.register_model_config(
+        "rope_vit_t14",
+        rope_vit,
+        config={"patch_size": 14, **TINY},
+    )
+    registry.register_model_config(
+        "rope_vit_s32",
+        rope_vit,
+        config={"patch_size": 32, **SMALL},
+    )
+    registry.register_model_config(
+        "rope_vit_s16",
+        rope_vit,
+        config={"patch_size": 16, **SMALL},
+    )
+    registry.register_model_config(
+        "rope_i_vit_s16_pn_aps_c1",  # For PE Core - https://arxiv.org/abs/2504.13181
+        rope_vit,
+        config={
+            "patch_size": 16,
+            **SMALL,
+            "pre_norm": True,
+            "attn_pool_head": True,
+            "attn_pool_num_heads": 8,
+            "attn_pool_special_tokens": True,
+            "norm_layer_eps": 1e-5,
+            "rope_rot_type": "interleaved",
+            "rope_grid_indexing": "xy",
+            "rope_grid_offset": 1,
+            "rope_temperature": 10000.0,
+        },
+    )
+    registry.register_model_config(
+        "rope_vit_s14",
+        rope_vit,
+        config={"patch_size": 14, **SMALL},
+    )
+    registry.register_model_config(
+        "rope_vit_m32",
+        rope_vit,
+        config={"patch_size": 32, **MEDIUM},
+    )
+    registry.register_model_config(
+        "rope_vit_m16",
+        rope_vit,
+        config={"patch_size": 16, **MEDIUM},
+    )
+    registry.register_model_config(
+        "rope_vit_m14",
+        rope_vit,
+        config={"patch_size": 14, **MEDIUM},
+    )
+    registry.register_model_config(
+        "rope_vit_b32",
+        rope_vit,
+        config={"patch_size": 32, **BASE, "drop_path_rate": 0.0},  # Override the BASE definition
+    )
+    registry.register_model_config(
+        "rope_vit_b16",
+        rope_vit,
+        config={"patch_size": 16, **BASE},
+    )
+    registry.register_model_config(
+        "rope_vit_b16_qkn_ls",
+        rope_vit,
+        config={"patch_size": 16, **BASE, "layer_scale_init_value": 1e-5, "qk_norm": True},
+    )
+    registry.register_model_config(
+        "rope_i_vit_b16_pn_aps_c1",  # For PE Core - https://arxiv.org/abs/2504.13181
+        rope_vit,
+        config={
+            "patch_size": 16,
+            **BASE,
+            "pre_norm": True,
+            "attn_pool_head": True,
+            "attn_pool_num_heads": 8,
+            "attn_pool_special_tokens": True,
+            "norm_layer_eps": 1e-5,
+            "rope_rot_type": "interleaved",
+            "rope_grid_indexing": "xy",
+            "rope_grid_offset": 1,
+            "rope_temperature": 10000.0,
+        },
+    )
+    registry.register_model_config(
+        "rope_vit_b14",
+        rope_vit,
+        config={"patch_size": 14, **BASE},
+    )
+    registry.register_model_config(
+        "rope_vit_so150m_p14_ap",
+        rope_vit,
+        config={"patch_size": 14, **SO150, "class_token": False, "attn_pool_head": True},
+    )
+    registry.register_model_config(
+        "rope_vit_l32",
+        rope_vit,
+        config={"patch_size": 32, **LARGE},
+    )
+    registry.register_model_config(
+        "rope_vit_l16",
+        rope_vit,
+        config={"patch_size": 16, **LARGE},
+    )
+    registry.register_model_config(
+        "rope_vit_l14",
+        rope_vit,
+        config={"patch_size": 14, **LARGE},
+    )
+    registry.register_model_config(
+        "rope_i_vit_l14_pn_aps_c1",  # For PE Core - https://arxiv.org/abs/2504.13181
+        rope_vit,
+        config={
+            "patch_size": 14,
+            **LARGE,
+            "pre_norm": True,
+            "attn_pool_head": True,
+            "attn_pool_num_heads": 8,
+            "attn_pool_special_tokens": True,
+            "norm_layer_eps": 1e-5,
+            "rope_rot_type": "interleaved",
+            "rope_grid_indexing": "xy",
+            "rope_grid_offset": 1,
+            "rope_temperature": 10000.0,
+        },
+    )
+    registry.register_model_config(
+        "rope_vit_so400m_p14_ap",
+        rope_vit,
+        config={"patch_size": 14, **SO400, "class_token": False, "attn_pool_head": True},
+    )
+    registry.register_model_config(
+        "rope_vit_h16",
+        rope_vit,
+        config={"patch_size": 16, **HUGE},
+    )
+    registry.register_model_config(
+        "rope_vit_h14",
+        rope_vit,
+        config={"patch_size": 14, **HUGE},
+    )
+    registry.register_model_config(  # From "Scaling Vision Transformers"
+        "rope_vit_g16",
+        rope_vit,
+        config={"patch_size": 16, **GIANT},
+    )
+    registry.register_model_config(  # From "Scaling Vision Transformers"
+        "rope_vit_g14",
+        rope_vit,
+        config={"patch_size": 14, **GIANT},
+    )
+    registry.register_model_config(  # From "Scaling Vision Transformers"
+        "rope_vit_gigantic14",
+        rope_vit,
+        config={"patch_size": 14, **GIGANTIC},
+    )
+
+    # With registers
+    ####################
+
+    registry.register_model_config(
+        "rope_vit_reg1_t32",
+        rope_vit,
+        config={"patch_size": 32, **TINY, "num_reg_tokens": 1},
+    )
+    registry.register_model_config(
+        "rope_vit_reg1_t16",
+        rope_vit,
+        config={"patch_size": 16, **TINY, "num_reg_tokens": 1},
+    )
+    registry.register_model_config(
+        "rope_vit_reg1_t14",
+        rope_vit,
+        config={"patch_size": 14, **TINY, "num_reg_tokens": 1},
+    )
+    registry.register_model_config(
+        "rope_vit_reg1_s32",
+        rope_vit,
+        config={"patch_size": 32, **SMALL, "num_reg_tokens": 1},
+    )
+    registry.register_model_config(
+        "rope_vit_reg1_s16",
+        rope_vit,
+        config={"patch_size": 16, **SMALL, "num_reg_tokens": 1},
+    )
+    registry.register_model_config(
+        "rope_i_vit_reg1_s16_pn_npn_avg_c1",  # For PE Spatial - https://arxiv.org/abs/2504.13181
+        rope_vit,
+        config={
+            "patch_size": 16,
+            **SMALL,
+            "num_reg_tokens": 1,
+            "class_token": False,
+            "pre_norm": True,
+            "post_norm": False,
+            "norm_layer_eps": 1e-5,
+            "rope_rot_type": "interleaved",
+            "rope_grid_indexing": "xy",
+            "rope_grid_offset": 1,
+            "rope_temperature": 10000.0,
+        },
+    )
+    registry.register_model_config(
+        "rope_vit_reg1_s14",
+        rope_vit,
+        config={"patch_size": 14, **SMALL, "num_reg_tokens": 1},
+    )
+    registry.register_model_config(
+        "rope_vit_reg4_m32",
+        rope_vit,
+        config={"patch_size": 32, **MEDIUM, "num_reg_tokens": 4},
+    )
+    registry.register_model_config(
+        "rope_vit_reg4_m16",
+        rope_vit,
+        config={"patch_size": 16, **MEDIUM, "num_reg_tokens": 4},
+    )
+    registry.register_model_config(
+        "rope_vit_reg4_m16_rms_avg",
+        rope_vit,
+        config={"patch_size": 16, **MEDIUM, "num_reg_tokens": 4, "class_token": False, "norm_layer_type": "RMSNorm"},
+    )
+    registry.register_model_config(
+        "rope_vit_reg4_m14",
+        rope_vit,
+        config={"patch_size": 14, **MEDIUM, "num_reg_tokens": 4},
+    )
+    registry.register_model_config(
+        "rope_vit_reg4_m14_avg",
+        rope_vit,
+        config={"patch_size": 14, **MEDIUM, "num_reg_tokens": 4, "class_token": False},
+    )
+    registry.register_model_config(
+        "rope_vit_reg4_b32",
+        rope_vit,
+        config={"patch_size": 32, **BASE, "num_reg_tokens": 4, "drop_path_rate": 0.0},  # Override the BASE definition
+    )
+    registry.register_model_config(
+        "rope_vit_reg4_b16",
+        rope_vit,
+        config={"patch_size": 16, **BASE, "num_reg_tokens": 4},
+    )
+    registry.register_model_config(
+        "rope_vit_reg4_b14",
+        rope_vit,
+        config={"patch_size": 14, **BASE, "num_reg_tokens": 4},
+    )
+    registry.register_model_config(
+        "rope_vit_reg8_nps_b14_ap",
+        rope_vit,
+        config={
+            "pos_embed_special_tokens": False,
+            "patch_size": 14,
+            **BASE,
+            "num_reg_tokens": 8,
+            "class_token": False,
+            "attn_pool_head": True,
+        },
+    )
+    registry.register_model_config(
+        "rope_vit_reg4_so150m_p14_ap",
+        rope_vit,
+        config={"patch_size": 14, **SO150, "num_reg_tokens": 4, "class_token": False, "attn_pool_head": True},
+    )
+    registry.register_model_config(
+        "rope_vit_reg8_so150m_p14_ap",
+        rope_vit,
+        config={"patch_size": 14, **SO150, "num_reg_tokens": 8, "class_token": False, "attn_pool_head": True},
+    )
+    registry.register_model_config(
+        "rope_vit_reg8_so150m_p14_swiglu_rms_avg",
+        rope_vit,
+        config={
+            "patch_size": 14,
+            **SO150,
+            "num_reg_tokens": 8,
+            "class_token": False,
+            "norm_layer_type": "RMSNorm",
+            "mlp_layer_type": "SwiGLU_FFN",
+        },
+    )
+    registry.register_model_config(
+        "rope_vit_reg8_so150m_p14_swiglu_rms_ap",
+        rope_vit,
+        config={
+            "patch_size": 14,
+            **SO150,
+            "num_reg_tokens": 8,
+            "class_token": False,
+            "attn_pool_head": True,
+            "norm_layer_type": "RMSNorm",
+            "mlp_layer_type": "SwiGLU_FFN",
+        },
+    )
+    registry.register_model_config(
+        "rope_vit_reg8_so150m_p14_swiglu_rms_aps",
+        rope_vit,
+        config={
+            "patch_size": 14,
+            **SO150,
+            "num_reg_tokens": 8,
+            "class_token": False,
+            "attn_pool_head": True,
+            "attn_pool_special_tokens": True,
+            "norm_layer_type": "RMSNorm",
+            "mlp_layer_type": "SwiGLU_FFN",
+        },
+    )
+    registry.register_model_config(
+        "rope_vit_reg4_l32",
+        rope_vit,
+        config={"patch_size": 32, **LARGE, "num_reg_tokens": 4},
+    )
+    registry.register_model_config(
+        "rope_vit_reg4_l16",
+        rope_vit,
+        config={"patch_size": 16, **LARGE, "num_reg_tokens": 4},
+    )
+    registry.register_model_config(
+        "rope_vit_reg4_l14",
+        rope_vit,
+        config={"patch_size": 14, **LARGE, "num_reg_tokens": 4},
+    )
+    registry.register_model_config(
+        "rope_vit_reg8_l14_rms_ap",
+        rope_vit,
+        config={
+            "patch_size": 14,
+            **LARGE,
+            "num_reg_tokens": 8,
+            "class_token": False,
+            "attn_pool_head": True,
+            "norm_layer_type": "RMSNorm",
+        },
+    )
+    registry.register_model_config(
+        "rope_vit_reg8_so400m_p14_ap",
+        rope_vit,
+        config={"patch_size": 14, **SO400, "num_reg_tokens": 8, "class_token": False, "attn_pool_head": True},
+    )
+    registry.register_model_config(
+        "rope_vit_reg4_h16",
+        rope_vit,
+        config={"patch_size": 16, **HUGE, "num_reg_tokens": 4},
+    )
+    registry.register_model_config(
+        "rope_vit_reg4_h14",
+        rope_vit,
+        config={"patch_size": 14, **HUGE, "num_reg_tokens": 4},
+    )
+    registry.register_model_config(  # From "Scaling Vision Transformers"
+        "rope_vit_reg4_g14",
+        rope_vit,
+        config={"patch_size": 14, **GIANT, "num_reg_tokens": 4},
+    )