Skip to content
This repository has been archived by the owner on Jan 5, 2024. It is now read-only.

Additional pytorch models long promised by Joel #79

Open
wants to merge 2 commits into
base: master
Choose a base branch
from
Open

Additional pytorch models long promised by Joel #79

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
7228014
updated CM with a number of pytorch models and adversarially trained …
Mar 2, 2021
4ea2903
add few more layers to alexnet
Mar 2, 2021
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
93 changes: 83 additions & 10 deletions candidate_models/base_models/__init__.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -273,10 +273,46 @@ def vonecornet(model_name='cornets'):
wrapper.image_size = 224
return wrapper

def GoN_model(function, train, image_size):
from urllib import request
import torch
from model_tools.activations.pytorch import load_preprocess_images
module = import_module(f'torchvision.models')
model_ctr = getattr(module, function)
model = model_ctr()
preprocessing = functools.partial(load_preprocess_images, image_size=image_size)
# load weights
framework_home = os.path.expanduser(os.getenv('CM_HOME', '~/.candidate_models'))
weightsdir_path = os.getenv('CM_TSLIM_WEIGHTS_DIR',
os.path.join(framework_home, 'model-weights', 'resnet-50-robust'))
weights_id = 'resnet-50-' + train
weights_path = os.path.join(weightsdir_path, weights_id)
if not os.path.isfile(weights_path):
weight_urls = {
'resnet-50-GNTsig0.5': 'https://github.com/bethgelab/game-of-noise/releases/download/v1.0/Gauss_sigma_0.5_Model.pth',
'resnet-50-ANT3x3_SIN': 'https://github.com/bethgelab/game-of-noise/releases/download/v1.0/ANT3x3_SIN_Model.pth',
}
assert weights_id in weight_urls
url = weight_urls[weights_id]
_logger.debug(f"Downloading weights for resnet-50-GoN from {url} to {weights_path}")
os.makedirs(weightsdir_path, exist_ok=True)
request.urlretrieve(url, weights_path)

if torch.cuda.is_available():
checkpoint = torch.load(weights_path)
else:
checkpoint = torch.load(weights_path, map_location=torch.device('cpu'))

def robust_model(function, image_size):
# process weights -- remove the attacker and prepocessing weights
model.load_state_dict(checkpoint['model_state_dict'])
# wrap model with pytorch wrapper
wrapper = PytorchWrapper(identifier=weights_id, model=model, preprocessing=preprocessing)
wrapper.image_size = image_size
return wrapper

def robust_model(function, penalty, eps, image_size):
from urllib import request
import torch
import torch
from model_tools.activations.pytorch import load_preprocess_images
module = import_module(f'torchvision.models')
model_ctr = getattr(module, function)
Expand All @@ -286,20 +322,32 @@ def robust_model(function, image_size):
framework_home = os.path.expanduser(os.getenv('CM_HOME', '~/.candidate_models'))
weightsdir_path = os.getenv('CM_TSLIM_WEIGHTS_DIR',
os.path.join(framework_home, 'model-weights', 'resnet-50-robust'))
weights_path = os.path.join(weightsdir_path, 'resnet-50-robust')
weights_id = 'resnet-50-robust-' + penalty + '-' + eps
weights_path = os.path.join(weightsdir_path, weights_id)
if not os.path.isfile(weights_path):
url = 'http://andrewilyas.com/ImageNet.pt'
weight_urls = {
'resnet-50-robust-l2-3': 'https://www.dropbox.com/s/knf4uimlqsi1yz8/imagenet_l2_3_0.pt?dl=1',
'resnet-50-robust-linf-4': 'https://www.dropbox.com/s/axfuary2w1cnyrg/imagenet_linf_4.pt?dl=1',
'resnet-50-robust-linf-8': 'https://www.dropbox.com/s/yxn15a9zklz3s8q/imagenet_linf_8.pt?dl=1',
}
assert weights_id in weight_urls
url = weight_urls[weights_id]
_logger.debug(f"Downloading weights for resnet-50-robust from {url} to {weights_path}")
os.makedirs(weightsdir_path, exist_ok=True)
request.urlretrieve(url, weights_path)
checkpoint = torch.load(weights_path, map_location=torch.device('cpu'))

if torch.cuda.is_available():
checkpoint = torch.load(weights_path)
else:
checkpoint = torch.load(weights_path, map_location=torch.device('cpu'))

# process weights -- remove the attacker and prepocessing weights
weights = checkpoint['model']
weights = {k[len('module.model.'):]: v for k, v in weights.items() if 'attacker' not in k}
weights = {k: weights[k] for k in list(weights.keys())[2:]}
model.load_state_dict(weights)
# wrap model with pytorch wrapper
wrapper = PytorchWrapper(identifier=function+'-robust', model=model, preprocessing=preprocessing)
wrapper = PytorchWrapper(identifier=weights_id, model=model, preprocessing=preprocessing)
wrapper.image_size = image_size
return wrapper

Expand Down Expand Up @@ -382,12 +430,37 @@ def __init__(self):

_key_functions = {
'alexnet': lambda: torchvision_model('alexnet', image_size=224),
'vgg-11-pt': lambda: torchvision_model('vgg11', image_size=224),
'vgg-11-bn-pt': lambda: torchvision_model('vgg11_bn', image_size=224),
'vgg-13-pt': lambda: torchvision_model('vgg13', image_size=224),
'vgg-13-bn-pt': lambda: torchvision_model('vgg13_bn', image_size=224),
'vgg-16-pt': lambda: torchvision_model('vgg16', image_size=224),
'vgg-16-bn-pt': lambda: torchvision_model('vgg16_bn', image_size=224),
'vgg-19-pt': lambda: torchvision_model('vgg19', image_size=224),
'vgg-19-bn-pt': lambda: torchvision_model('vgg19_bn', image_size=224),
'squeezenet1_0': lambda: torchvision_model('squeezenet1_0', image_size=224),
'squeezenet1_1': lambda: torchvision_model('squeezenet1_1', image_size=224),
'resnet-18': lambda: torchvision_model('resnet18', image_size=224),
'resnet-34': lambda: torchvision_model('resnet34', image_size=224),
'resnet-50-pytorch': lambda: torchvision_model('resnet50', image_size=224),
'resnet-50-robust': lambda: robust_model('resnet50', image_size=224),
'resnet-18-pt': lambda: torchvision_model('resnet18', image_size=224),
'resnet-34-pt': lambda: torchvision_model('resnet34', image_size=224),
'resnet-50-pt': lambda: torchvision_model('resnet50', image_size=224),
'resnet-101-pt': lambda: torchvision_model('resnet101', image_size=224),
'resnet-152-pt': lambda: torchvision_model('resnet152', image_size=224),
'densenet-121-pt': lambda: torchvision_model('densenet121', image_size=224),
'densenet-169-pt': lambda: torchvision_model('densenet169', image_size=224),
'densenet-201-pt': lambda: torchvision_model('densenet201', image_size=224),
'densenet-161-pt': lambda: torchvision_model('densenet161', image_size=224),
'resnext-50-32x4d-pt': lambda: torchvision_model('resnext50_32x4d', image_size=224),
'resnext-101-32x8d-pt': lambda: torchvision_model('resnext101_32x8d', image_size=224),
'wide-resnet-50-pt': lambda: torchvision_model('wide_resnet50_2', image_size=224),
'wide-resnet-101-pt': lambda: torchvision_model('wide_resnet101_2', image_size=224),

'resnet-50-robust-l2-3': lambda: robust_model('resnet50', penalty='l2', eps='3', image_size=224),
'resnet-50-robust-linf-4': lambda: robust_model('resnet50', penalty='linf', eps='4', image_size=224),
'resnet-50-robust-linf-8': lambda: robust_model('resnet50', penalty='linf', eps='8', image_size=224),

'resnet-50-GNTsig0.5': lambda: GoN_model('resnet50', train='GNTsig0.5', image_size=224),
'resnet-50-ANT3x3_SIN': lambda: GoN_model('resnet50', train='ANT3x3_SIN', image_size=224),

'voneresnet-50': lambda: voneresnet(model_name='resnet50'),
'voneresnet-50-robust': lambda: voneresnet(model_name='resnet50_at'),

Expand Down
157 changes: 97 additions & 60 deletions candidate_models/model_commitments/model_layer_def.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,6 +2,13 @@

from brainscore.submission.ml_pool import ModelLayers

def resnet_pt_layers(units):
return (['relu', 'maxpool'] +
[f"layer1.{i}" for i in range(units[0])] +
[f"layer2.{i}" for i in range(units[1])] +
[f"layer3.{i}" for i in range(units[2])] +
[f"layer4.{i}" for i in range(units[3])] +
["avgpool"])

def resnet50_layers(bottleneck_version):
return resnet_layers(bottleneck_version=bottleneck_version, units=[3, 4, 6, 3])
Expand Down Expand Up @@ -66,12 +73,35 @@ def prednet():


layers = {
'alexnet':
[ # conv-relu-[pool]{1,2,3,4,5}
'features.2', 'features.5', 'features.7', 'features.9', 'features.12',
'classifier.2', 'classifier.5'], # fc-[relu]{6,7,8}
'alexnet':
[f'features.{i}' for i in [1,2,4,5,7,9,11,12]] +
[f'classifier.{i}' for i in [2,5]], # fc-[relu]{6,7,8}
'vgg-16': [f'block{i + 1}_pool' for i in range(5)] + ['fc1', 'fc2'],
'vgg-19': [f'block{i + 1}_pool' for i in range(5)] + ['fc1', 'fc2'],
'vgg-11-pt':
[f'features.{i}' for i in [1,2,4,5,7,9,10,12,14,15,17,19,20]] +
['classifier.1', 'classifier.4'],
'vgg-11-bn-pt':
[f'features.{i}' for i in [2,3,6,7,10,13,14,17,20,21,24,27,28]] +
['classifier.1', 'classifier.4'],
'vgg-13-pt':
[f'features.{i}' for i in [1,3,4,6,8,9,11,13,14,16,18,19,21,23,24]] +
['classifier.1', 'classifier.4'],
'vgg-13-bn-pt':
[f'features.{i}' for i in [2,5,6,9,12,13,16,19,20,23,26,27,30,33,34]] +
['classifier.1', 'classifier.4'],
'vgg-16-pt':
[f'features.{i}' for i in [1,3,4,6,8,9,11,13,15,16,18,20,21,23,25,27,29,30]] +
['classifier.1', 'classifier.4'],
'vgg-16-bn-pt':
[f'features.{i}' for i in [2,5,6,9,12,13,16,19,22,23,26,29,32,33,36,39,42,43]] +
['classifier.1', 'classifier.4'],
'vgg-19-pt':
[f'features.{i}' for i in [1,3,4,6,8,9,11,13,15,17,18,20,22,24,26,27,29,31,33,35,36]] +
['classifier.1', 'classifier.4'],
'vgg-19-bn-pt':
[f'features.{i}' for i in [2,5,6,9,12,13,16,19,22,25,26,29,32,35,38,39,42,45,48,51,52]] +
['classifier.1', 'classifier.4'],
'squeezenet1_0':
['features.' + layer for layer in
# max pool + fire outputs (ignoring pools)
Expand Down Expand Up @@ -116,61 +146,75 @@ def prednet():
[f'block13_sepconv{i + 1}_act' for i in range(2)] +
[f'block14_sepconv{i + 1}_act' for i in range(2)] +
['avg_pool'],
'resnet-18':
['conv1'] +
['layer1.0.relu', 'layer1.1.relu'] +
['layer2.0.relu', 'layer2.0.downsample.0', 'layer2.1.relu'] +
['layer3.0.relu', 'layer3.0.downsample.0', 'layer3.1.relu'] +
['layer4.0.relu', 'layer4.0.downsample.0', 'layer4.1.relu'] +
['avgpool'],
'resnet-34':
['conv1'] +
['layer1.0.conv2', 'layer1.1.conv2', 'layer1.2.conv2'] +
['layer2.0.downsample.0', 'layer2.1.conv2', 'layer2.2.conv2', 'layer2.3.conv2'] +
['layer3.0.downsample.0', 'layer3.1.conv2', 'layer3.2.conv2', 'layer3.3.conv2',
'layer3.4.conv2', 'layer3.5.conv2'] +
['layer4.0.downsample.0', 'layer4.1.conv2', 'layer4.2.conv2'] +
['avgpool'],
'resnet-50':
['conv1'] +
['layer1.0.conv3', 'layer1.1.conv3', 'layer1.2.conv3'] +
['layer2.0.downsample.0', 'layer2.1.conv3', 'layer2.2.conv3', 'layer2.3.conv3'] +
['layer3.0.downsample.0', 'layer3.1.conv3', 'layer3.2.conv3', 'layer3.3.conv3',
'layer3.4.conv3', 'layer3.5.conv3'] +
['layer4.0.downsample.0', 'layer4.1.conv3', 'layer4.2.conv3'] +
['avgpool'],
'resnet-18-pt': resnet_pt_layers([2,2,2,2]),
'resnet-34-pt': resnet_pt_layers([3,4,6,3]),
'resnet-50-pt': resnet_pt_layers([3,4,6,3]),
'wide-resnet-50-pt': resnet_pt_layers([3,4,6,3]),
'resnet-101-pt': resnet_pt_layers([3,4,23,3]),
'wide-resnet-101-pt': resnet_pt_layers([3,4,23,3]),
'resnet-152-pt': resnet_pt_layers([3,8,36,3]),
'resnext-50-32x4d-pt': resnet_pt_layers([3,4,6,3]),
'resnext-101-32x8d-pt': resnet_pt_layers([3,4,23,3]),
'resnet-50-robust-l2-3': resnet_pt_layers([3,4,6,3]),
'resnet-50-robust-linf-4': resnet_pt_layers([3,4,6,3]),
'resnet-50-robust-linf-8': resnet_pt_layers([3,4,6,3]),
'resnet-50-GNTsig0.5': resnet_pt_layers([3,4,6,3]),
'resnet-50-ANT3x3_SIN': resnet_pt_layers([3,4,6,3]),
'resnet-50-SIN': resnet_pt_layers([3,4,6,3]),
'resnet-50-SIN_IN': resnet_pt_layers([3,4,6,3]),
'resnet-50-SIN_IN_IN': resnet_pt_layers([3,4,6,3]),
'voneresnet-50':
['vone_block'] +
['model.layer1.0.conv3', 'model.layer1.1.conv3', 'model.layer1.2.conv3'] +
['model.layer2.0.downsample.0', 'model.layer2.1.conv3', 'model.layer2.2.conv3', 'model.layer2.3.conv3'] +
['model.layer3.0.downsample.0', 'model.layer3.1.conv3', 'model.layer3.2.conv3', 'model.layer3.3.conv3',
'model.layer3.4.conv3', 'model.layer3.5.conv3'] +
['model.layer4.0.downsample.0', 'model.layer4.1.conv3', 'model.layer4.2.conv3'] +
['model.layer1.0', 'model.layer1.1', 'model.layer1.2'] +
['model.layer2.0', 'model.layer2.1', 'model.layer2.2', 'model.layer2.3'] +
['model.layer3.0', 'model.layer3.1', 'model.layer3.2', 'model.layer3.3',
'model.layer3.4', 'model.layer3.5'] +
['model.layer4.0', 'model.layer4.1', 'model.layer4.2'] +
['model.avgpool'],
'voneresnet-50-robust':
['vone_block'] +
['model.layer1.0.conv3', 'model.layer1.1.conv3', 'model.layer1.2.conv3'] +
['model.layer2.0.downsample.0', 'model.layer2.1.conv3', 'model.layer2.2.conv3', 'model.layer2.3.conv3'] +
['model.layer3.0.downsample.0', 'model.layer3.1.conv3', 'model.layer3.2.conv3', 'model.layer3.3.conv3',
'model.layer3.4.conv3', 'model.layer3.5.conv3'] +
['model.layer4.0.downsample.0', 'model.layer4.1.conv3', 'model.layer4.2.conv3'] +
['model.layer1.0', 'model.layer1.1', 'model.layer1.2'] +
['model.layer2.0', 'model.layer2.1', 'model.layer2.2', 'model.layer2.3'] +
['model.layer3.0', 'model.layer3.1', 'model.layer3.2', 'model.layer3.3',
'model.layer3.4', 'model.layer3.5'] +
['model.layer4.0', 'model.layer4.1', 'model.layer4.2'] +
['model.avgpool'],
'resnet-50-robust':
['conv1'] +
['layer1.0.conv3', 'layer1.1.conv3', 'layer1.2.conv3'] +
['layer2.0.downsample.0', 'layer2.1.conv3', 'layer2.2.conv3', 'layer2.3.conv3'] +
['layer3.0.downsample.0', 'layer3.1.conv3', 'layer3.2.conv3', 'layer3.3.conv3',
'layer3.4.conv3', 'layer3.5.conv3'] +
['layer4.0.downsample.0', 'layer4.1.conv3', 'layer4.2.conv3'] +
['avgpool'],
'resnet-50-pytorch':
['conv1'] +
['layer1.0.conv3', 'layer1.1.conv3', 'layer1.2.conv3'] +
['layer2.0.downsample.0', 'layer2.1.conv3', 'layer2.2.conv3', 'layer2.3.conv3'] +
['layer3.0.downsample.0', 'layer3.1.conv3', 'layer3.2.conv3', 'layer3.3.conv3',
'layer3.4.conv3', 'layer3.5.conv3'] +
['layer4.0.downsample.0', 'layer4.1.conv3', 'layer4.2.conv3'] +
['avgpool'],
'densenet-121-pt':
['features.relu0', 'features.pool0'] +
[f'features.denseblock1.denselayer{i + 1}.relu1' for i in range(6)] +
['features.transition1.relu'] +
[f'features.denseblock2.denselayer{i + 1}.relu1' for i in range(12)] +
['features.transition2.relu'] +
[f'features.denseblock3.denselayer{i + 1}.relu1' for i in range(24)] +
['features.transition3.relu'] +
[f'features.denseblock4.denselayer{i + 1}.relu1' for i in range(16)],
'densenet-169-pt':
['features.relu0', 'features.pool0'] +
[f'features.denseblock1.denselayer{i + 1}.relu1' for i in range(6)] +
['features.transition1.relu'] +
[f'features.denseblock2.denselayer{i + 1}.relu1' for i in range(12)] +
['features.transition2.relu'] +
[f'features.denseblock3.denselayer{i + 1}.relu1' for i in range(32)] +
['features.transition3.relu'] +
[f'features.denseblock4.denselayer{i + 1}.relu1' for i in range(32)],
'densenet-201-pt':
['features.relu0', 'features.pool0'] +
[f'features.denseblock1.denselayer{i + 1}.relu1' for i in range(6)] +
['features.transition1.relu'] +
[f'features.denseblock2.denselayer{i + 1}.relu1' for i in range(12)] +
['features.transition2.relu'] +
[f'features.denseblock3.denselayer{i + 1}.relu1' for i in range(48)] +
['features.transition3.relu'] +
[f'features.denseblock4.denselayer{i + 1}.relu1' for i in range(32)],
'densenet-161-pt':
['features.relu0', 'features.pool0'] +
[f'features.denseblock1.denselayer{i + 1}.relu2' for i in range(6)] +
['features.transition1.relu'] +
[f'features.denseblock2.denselayer{i + 1}.relu2' for i in range(12)] +
['features.transition2.relu'] +
[f'features.denseblock3.denselayer{i + 1}.relu2' for i in range(36)] +
['features.transition3.relu'] +
[f'features.denseblock4.denselayer{i + 1}.relu2' for i in range(24)],
# Slim
'inception_v1':
['MaxPool_2a_3x3'] +
Expand Down Expand Up @@ -251,13 +295,6 @@ def prednet():

model_layers = ModelLayers(layers)
model_layers['vggface'] = model_layers['vgg-16']
for sin_model in ['resnet50-SIN', 'resnet50-SIN_IN', 'resnet50-SIN_IN_IN']:
model_layers[sin_model] = \
['conv1'] + \
[f'layer{seq}.{bottleneck}.relu'
for seq, bottlenecks in enumerate([3, 4, 6, 3], start=1)
for bottleneck in range(bottlenecks)] + \
['avgpool']

for version, multiplier, image_size in [
# v1
Expand Down