Ep/add fixres resnext101 32x48d wsl

brainscore_vision/models/fixres_resnext101_32x48d_wsl/__init__.py

@@ -0,0 +1,7 @@
+from brainscore_vision import model_registry
+from brainscore_vision.model_helpers.brain_transformation import ModelCommitment
+from .model import get_model, get_layers
+
+model_registry['fixres_resnext101_32x48d_wsl'] = lambda: ModelCommitment(identifier='fixres_resnext101_32x48d_wsl',
+                                                               activations_model=get_model(),
+                                                               layers=get_layers('fixres_resnext101_32x48d_wsl'))

brainscore_vision/models/fixres_resnext101_32x48d_wsl/model.py

@@ -0,0 +1,57 @@
+from brainscore_vision.model_helpers.activations.pytorch import PytorchWrapper
+from fixres.hubconf import load_state_dict_from_url
+from fixres.transforms_v2 import get_transforms
+from model_helpers.activations.pytorch import load_images
+import numpy as np
+from importlib import import_module
+import ssl
+
+
+ssl._create_default_https_context = ssl._create_unverified_context
+
+
+def get_model():
+    module = import_module('fixres.imnet_evaluate.resnext_wsl')
+    model_ctr = getattr(module, 'resnext101_32x48d_wsl')
+    model = model_ctr(pretrained=False)  # the pretrained flag here corresponds to standard resnext weights
+    pretrained_dict = load_state_dict_from_url('https://dl.fbaipublicfiles.com/FixRes_data/FixRes_Pretrained_Models/ResNeXt_101_32x48d.pth',
+                                               map_location=lambda storage, loc: storage)['model']
+    model_dict = model.state_dict()
+    for k in model_dict.keys():
+        assert ('module.' + k) in pretrained_dict.keys()
+        model_dict[k] = pretrained_dict.get(('module.' + k))
+    model.load_state_dict(model_dict)
+
+    # preprocessing
+    # 320 for ResNeXt:
+    # https://github.com/mschrimpf/FixRes/tree/4ddcf11b29c118dfb8a48686f75f572450f67e5d#example-evaluation-procedure
+    input_size = 320
+    # https://github.com/mschrimpf/FixRes/blob/0dc15ab509b9cb9d7002ca47826dab4d66033668/fixres/imnet_evaluate/train.py#L159-L160
+    transformation = get_transforms(input_size=input_size, test_size=input_size,
+                                    kind='full', need=('val',),
+                                    # this is different from standard ImageNet evaluation to show the whole image
+                                    crop=False,
+                                    # no backbone parameter for ResNeXt following
+                                    # https://github.com/mschrimpf/FixRes/blob/0dc15ab509b9cb9d7002ca47826dab4d66033668/fixres/imnet_evaluate/train.py#L154-L156
+                                    backbone=None)
+    transform = transformation['val']
+
+    def load_preprocess_images(image_filepaths):
+        images = load_images(image_filepaths)
+        images = [transform(image) for image in images]
+        images = [image.unsqueeze(0) for image in images]
+        images = np.concatenate(images)
+        return images
+
+    wrapper = PytorchWrapper(identifier='resnext101_32x48d_wsl', model=model, preprocessing=load_preprocess_images,
+                             batch_size=4)  # doesn't fit into 12 GB GPU memory otherwise
+    wrapper.image_size = input_size
+    return wrapper
+
+
+def get_layers(name):
+    return (['conv1'] +
+            # note that while relu is used multiple times, by default the last one will overwrite all previous ones
+            [f"layer{block + 1}.{unit}.relu"
+             for block, block_units in enumerate([3, 4, 23, 3]) for unit in range(block_units)] +
+            ['avgpool'])

brainscore_vision/models/fixres_resnext101_32x48d_wsl/requirements.txt

@@ -0,0 +1,5 @@
+torchvision
+torch
+numpy
+importlib
+Fixing-the-train-test-resolution-discrepancy-scripts@ git+https://github.com/mschrimpf/FixRes.git

brainscore_vision/models/fixres_resnext101_32x48d_wsl/test.py

@@ -0,0 +1,7 @@
+import pytest
+import brainscore_vision
+
+@pytest.mark.travis_slow
+def test_has_identifier():
+    model = brainscore_vision.load_model('fixres_resnext101_32x48d_wsl')
+    assert model.identifier == 'fixres_resnext101_32x48d_wsl'