brain-score.org submission (user:822) | (public:False)

brainscore_vision/models/resnet50_marrenj/__init__.py

@@ -0,0 +1,11 @@
+from brainscore_vision import model_registry
+from brainscore_vision.model_helpers.brain_transformation import ModelCommitment
+from .model import get_model, LAYERS, REGION_LAYER_MAP, BEHAVIORAL_READOUT_LAYER
+
+model_registry['resnet50_marrenj'] = lambda: ModelCommitment(
+    identifier='resnet50_marrenj',
+    activations_model=get_model(),
+    layers=LAYERS,
+    region_layer_map=REGION_LAYER_MAP,
+    behavioral_readout_layer=BEHAVIORAL_READOUT_LAYER,
+)

brainscore_vision/models/resnet50_marrenj/model.py

@@ -0,0 +1,70 @@
+"""ResNet50 Vanderbilt — final checkpoint of our supervised ImageNet-1k baseline
+(epoch 89, 90-epoch budget). Direct counterpart to Schrimpf/Gokce 2025's
+`resnet50_imagenet_full`, evaluated through the same Brain-Score pipeline.
+
+Brain-Score sandbox runs this in a fresh Python env that only has the deps in
+requirements.txt; weights are downloaded from a public HuggingFace model repo.
+"""
+import functools
+import torch
+import torchvision.models                  # NOTE: torchvision, not timm — this
+                                            # matches Schrimpf's exact module-naming
+                                            # convention (avgpool, layerN.M.relu) so
+                                            # his fitted commitments apply verbatim.
+                                            # State-dict keys are identical between
+                                            # timm and torchvision resnet50, so the
+                                            # checkpoint loads either way.
+from huggingface_hub import hf_hub_download
+from brainscore_vision.model_helpers.activations.pytorch import (
+    PytorchWrapper, load_preprocess_images,
+)
+
+# === EDIT BEFORE SUBMITTING: point at YOUR HuggingFace repo holding the .ckpt ===
+HF_REPO_ID = "marrenj/temporal-dynamics-baselines"     # <- change to your HF username
+HF_FILENAME = "resnet50_imagenet_baseline_ep089.ckpt"
+
+# Schrimpf/Gokce 2025's fitted region-layer commitments for resnet50_imagenet_full
+# (from scaling_primate_vvs/brainscore/artifacts/commitments.json). These are
+# TORCHVISION module names (not timm's — timm uses `act3`/`global_pool` where
+# torchvision uses `relu`/`avgpool`). That's why we build with torchvision above.
+REGION_LAYER_MAP = {
+    "V1": "layer1.0.conv1",
+    "V2": "layer3.5.bn3",
+    "V4": "layer3.0.conv1",
+    "IT": "layer4.0.relu",
+}
+BEHAVIORAL_READOUT_LAYER = "avgpool"
+LAYERS = list(set(REGION_LAYER_MAP.values())) + [BEHAVIORAL_READOUT_LAYER]
+
+
+BIBTEX = """@misc{marrenj_temporal_dynamics_2026,
+  title={Temporal Dynamics of Human Behavioral Alignment in ImageNet-trained Models},
+  author={Wallace Lab},
+  year={2026},
+  note={Supervised ResNet50, 90-epoch ImageNet-1k recipe},
+}"""
+
+
+def get_model():
+    weights_path = hf_hub_download(repo_id=HF_REPO_ID, filename=HF_FILENAME)
+    raw = torch.load(weights_path, map_location="cpu", weights_only=True)
+    state_dict = raw.get("state_dict", raw)
+    # SupervisedLightningModule wraps the timm classifier as `self.model`; strip prefix.
+    state_dict = {
+        k[len("model."):]: v
+        for k, v in state_dict.items()
+        if k.startswith("model.")
+    }
+    model = torchvision.models.resnet50(weights=None)
+    model.load_state_dict(state_dict)
+    model.eval()
+    preprocessing = functools.partial(load_preprocess_images, image_size=224)
+    wrapper = PytorchWrapper(
+        identifier="resnet50_marrenj", model=model, preprocessing=preprocessing,
+    )
+    wrapper.image_size = 224
+    return wrapper
+
+
+def get_bibtex(model_identifier):
+    return BIBTEX

brainscore_vision/models/resnet50_marrenj/requirements.txt

@@ -0,0 +1,3 @@
+torch==2.2.1
+torchvision==0.17.1
+huggingface_hub>=0.25

brainscore_vision/models/resnet50_marrenj/test.py

@@ -0,0 +1,9 @@
+"""Minimum sanity test: plugin registers + loads. Brain-Score CI runs this."""
+import pytest
+import brainscore_vision
+
+
+@pytest.mark.private_access
+def test_has_identifier():
+    model = brainscore_vision.load_model('resnet50_marrenj')
+    assert model.identifier == 'resnet50_marrenj'