-
Notifications
You must be signed in to change notification settings - Fork 1
/
utils.py
592 lines (505 loc) · 20.7 KB
/
utils.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
import os
import csv
import types
import ast
import cv2
import pandas as pd
from tqdm import tqdm
import torch
import torchvision
from torch.utils.data import Dataset
# DATASET_PATH = { # dataset : "/path/to/images/",
# "imagenet": "/path/to/imagenet_val/",
# "places365": "/path/to/places365_val",
# }
DATASET_PATH = {} # {dataset : "/path/to/images/",}
MEAN = (0.485, 0.456, 0.406)
STD = (0.229, 0.224, 0.225)
TRANSFORMS = {
"transform_imagenet": torchvision.transforms.Compose(
[
torchvision.transforms.ToPILImage(),
torchvision.transforms.Resize(224),
torchvision.transforms.CenterCrop((224, 224)),
torchvision.transforms.ToTensor(),
torchvision.transforms.Normalize(mean=MEAN, std=STD),
]
),
"transform_places365": torchvision.transforms.Compose(
[
torchvision.transforms.Resize((256, 256)),
torchvision.transforms.CenterCrop((224)),
torchvision.transforms.ToTensor(),
torchvision.transforms.Normalize(mean=MEAN, std=STD),
]
),
}
class ImageDataset(Dataset):
"""
Dataset class for loading images.
"""
def __init__(self, root, transform):
"""
Initialize the ImageDataset.
Args:
root (str): Root directory of the dataset.
transform (torchvision.transforms.Compose): Image transformation pipeline.
"""
self.root = root
self.transform = transform
self.image_names = [self.root + x for x in os.listdir(self.root)]
self.image_names.sort()
def __len__(self):
"""
Get the length of the dataset.
Returns:
int: Length of the dataset.
"""
return len(self.image_names)
def __getitem__(self, index):
"""
Get an item from the dataset.
Args:
index (int): Index of the item.
Returns:
torch.Tensor: Transformed image.
"""
image = cv2.imread(self.image_names[index])
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
image = self.transform(image)
return image
# Function to load pre-trained models
def get_target_model(target_name, device):
"""
returns target model in eval mode and its preprocess function
"""
if target_name == "resnet18":
weights = torchvision.models.ResNet18_Weights.IMAGENET1K_V1
preprocess = weights.transforms()
target_model = torchvision.models.resnet18(weights=weights).to(device).eval()
features_layer = target_model.avgpool
elif target_name == "resnet50":
weights = torchvision.models.ResNet50_Weights.IMAGENET1K_V1
preprocess = weights.transforms()
target_model = torchvision.models.resnet50(weights=weights).to(device).eval()
features_layer = target_model.avgpool
elif target_name == "resnet50_places":
arch = "resnet50"
# load the pre-trained weights
model_file = "%s_places365.pth.tar" % arch
if not os.access(model_file, os.W_OK):
weight_url = "http://places2.csail.mit.edu/models_places365/" + model_file
os.system("wget " + weight_url)
target_model = torchvision.models.__dict__[arch](num_classes=365)
checkpoint = torch.load(model_file, map_location=lambda storage, loc: storage)
state_dict = {
str.replace(k, "module.", ""): v
for k, v in checkpoint["state_dict"].items()
}
target_model.load_state_dict(state_dict)
target_model = target_model.eval().to(device)
preprocess = TRANSFORMS["transform_places365"]
features_layer = target_model.avgpool
elif target_name == "vit_b_16":
weights = torchvision.models.ViT_B_16_Weights.IMAGENET1K_V1
preprocess = weights.transforms()
target_model = torchvision.models.vit_b_16(weights=weights).to(device).eval()
features_layer = target_model.heads.head
elif "densenet161" in target_name:
weights = torchvision.models.DenseNet161_Weights.IMAGENET1K_V1
preprocess = weights.transforms()
target_model = torchvision.models.densenet161(weights=weights).to(device)
features_layer = target_model.classifier
elif "googlenet" in target_name:
weights = torchvision.models.GoogLeNet_Weights.IMAGENET1K_V1
preprocess = weights.transforms()
target_model = torchvision.models.googlenet(weights=weights).to(device)
features_layer = target_model.fc
elif "dino_vits8" in target_name:
weights = None
preprocess = TRANSFORMS["transform_imagenet"]
target_model = (
torch.hub.load("facebookresearch/dino:main", "dino_vits8").to(device).eval()
)
features_layer = target_model.blocks[11].mlp.fc1
# elif target_name == "densenet161":
# weights = torchvision.models.DenseNet161_Weights.IMAGENET1K_V1
# target_model = torchvision.models.densenet161(weights=weights).to(device).eval()
# preprocess = weights.transforms()
# target_model.features.add_module("relu", torch.nn.ReLU(inplace=False))
# target_model.features.add_module(
# "adaptive_avgpool", torch.nn.AdaptiveAvgPool2d((1, 1))
# )
# target_model.features.add_module("flatten", torch.nn.Flatten(1))
# def new_forward(self, x: torch.Tensor):
# features = self.features(x)
# out = self.classifier(features)
# return out
# target_model.forward = types.MethodType(new_forward, target_model)
# features_layer = target_model.features.to(device).eval()
elif "densenet161_places" in target_name:
arch = "densenet161"
# load the pre-trained weights
model_file = "%s_places365.pth.tar" % arch
if not os.access(model_file, os.W_OK):
weight_url = "http://places2.csail.mit.edu/models_places365/" + model_file
os.system("wget " + weight_url)
target_model = torchvision.models.__dict__[arch](num_classes=365)
checkpoint = torch.load(model_file, map_location=lambda storage, loc: storage)
state_dict = {
str.replace(k, "module.", ""): v
for k, v in checkpoint["state_dict"].items()
}
if arch == "densenet161":
state_dict = {
str.replace(k, "norm.", "norm"): v for k, v in state_dict.items()
}
state_dict = {
str.replace(k, "conv.", "conv"): v for k, v in state_dict.items()
}
state_dict = {
str.replace(k, "normweight", "norm.weight"): v
for k, v in state_dict.items()
}
state_dict = {
str.replace(k, "normrunning", "norm.running"): v
for k, v in state_dict.items()
}
state_dict = {
str.replace(k, "normbias", "norm.bias"): v
for k, v in state_dict.items()
}
state_dict = {
str.replace(k, "convweight", "conv.weight"): v
for k, v in state_dict.items()
}
target_model.load_state_dict(state_dict)
preprocess = TRANSFORMS["transform_places365"]
# second to last layer:
target_model.features.add_module("relu", torch.nn.ReLU(inplace=False))
target_model.features.add_module(
"adaptive_avgpool", torch.nn.AdaptiveAvgPool2d((1, 1))
)
target_model.features.add_module("flatten", torch.nn.Flatten(1))
def new_forward(self, x: torch.Tensor):
features = self.features(x)
out = self.classifier(features)
return out
target_model.forward = types.MethodType(new_forward, target_model)
target_model = target_model.to(device).eval()
features_layer = target_model.features.to(device).eval()
elif target_name == "vit_b_16":
weights = torchvision.models.ViT_B_16_Weights.IMAGENET1K_V1
target_model = torchvision.models.vit_b_16(weights=weights).eval()
preprocess = weights.transforms()
# see here https://pytorch.org/vision/stable/_modules/torchvision/models/vision_transformer.html#vit_b_16
index = torch.zeros([1]).long()
setattr(
target_model,
"subset",
torch.nn.Sequential(Subset(index), torch.nn.Flatten()),
)
def new_forward(self, x: torch.Tensor):
# Reshape and permute the input tensor
x = self._process_input(x)
n = x.shape[0]
# Expand the class token to the full batch
batch_class_token = self.class_token.expand(n, -1, -1)
x = torch.cat([batch_class_token, x], dim=1)
x = self.encoder(x)
x = self.subset(x)
x = self.heads(x)
return x
target_model.forward = types.MethodType(new_forward, target_model)
target_model = target_model.to(device)
features_layer = target_model.subset
features_layer = features_layer.eval()
target_model.eval()
return target_model, features_layer, preprocess
def get_data_path(dataset_name):
"""
Get the data path for a given dataset name.
Args:
dataset_name (str): Name of the dataset.
Returns:
str: Data path for the dataset.
Raises:
ValueError: If the dataset name is not supported.
"""
if dataset_name in DATASET_PATH.keys():
path = DATASET_PATH[dataset_name]
else:
raise ValueError("Unsupported dataset_name")
return path
def get_transform(transform_type):
"""
Get the transform for a given transform type.
Args:
transform_type (str): Type of the transform.
Returns:
torchvision.transforms.Compose: Transform for the given type.
Raises:
ValueError: If the transform type is not supported.
"""
if transform_type in TRANSFORMS.keys():
path = TRANSFORMS[transform_type]
else:
raise ValueError("Unsupported transform_type")
return path
def get_dataset(dataset_name):
"""
Get the dataset for a given dataset name.
Args:
dataset_name (str): Name of the dataset.
Returns:
ImageDataset: Dataset for the given name.
Raises:
ValueError: If the dataset name is not supported.
"""
if dataset_name == "imagenet" or dataset_name == "imagenet_train":
data_path = DATASET_PATH[dataset_name]
data_transform = TRANSFORMS["transform_imagenet"]
dataset = ImageDataset(root=data_path, transform=data_transform)
elif dataset_name == "places365":
data_path = DATASET_PATH[dataset_name]
data_transform = TRANSFORMS["transform_places365"]
dataset = ImageDataset(root=data_path, transform=data_transform)
else:
raise ValueError("Unsupported dataset_name")
return dataset
def get_n_neurons(model_layer):
"""
Get the number of neurons for a given model layer.
Args:
model_layer (str): Name of the model layer.
Returns:
int: Number of neurons in the layer.
Raises:
ValueError: If the model layer is not supported.
"""
if (
model_layer == "resnet18-fc"
or model_layer == "densenet161-fc"
or model_layer == "googlenet-fc"
or model_layer == "vit_b_16-head"
):
neurons = 1000
elif model_layer == "resnet18-layer4" or model_layer == "resnet18-avgpool":
neurons = 512
elif model_layer == "resnet18-layer3":
neurons = 256
elif model_layer == "resnet18-layer2":
neurons = 128
elif model_layer == "resnet18-layer1":
neurons = 64
elif model_layer == "resnet50_places-avgpool" or model_layer == "resnet50-avgpool":
neurons = 2048
elif (
model_layer == "densenet161-features"
or model_layer == "densenet161_places-features"
):
neurons = 2208
elif model_layer == "vit_b_16-features":
neurons = 768
else:
raise ValueError("Unsupported model_layer")
return neurons
def get_activations(
model,
model_name,
tensor_path,
dataset,
dataloader,
n_neurons,
device,
preprocess=None,
):
"""
Get the activations of a model for a given dataset.
Args:
model (torch.nn.Module): Model to get activations from.
model_name (str): Name of the model.
tensor_path (str): Path to save the activations tensor.
dataset (torch.utils.data.Dataset): Dataset to get activations for.
dataloader (torch.utils.data.DataLoader): DataLoader for the dataset.
n_neurons (int): Number of neurons in the model layer.
device (torch.device): Device to load the model on.
preprocess (torchvision.transforms.Compose, optional): Preprocess function for the model.
Returns:
torch.Tensor: Activations tensor of shape [len(dataset), n_neurons].
"""
activation = {}
def get_activation(name):
def hook(model, input, output):
activation[name] = output
return hook
if model_name == "resnet18-fc" or model_name == "googlenet-fc":
model.fc.register_forward_hook(get_activation("fc"))
if (
model_name == "resnet50_places-avgpool"
or model_name == "resnet18-avgpool"
or model_name == "resnet50-avgpool"
):
model.avgpool.register_forward_hook(get_activation("avgpool"))
elif model_name == "resnet18-layer4":
model.layer4.register_forward_hook(get_activation("layer4"))
elif model_name == "resnet18-layer3":
model.layer3.register_forward_hook(get_activation("layer3"))
elif model_name == "resnet18-layer2":
model.layer2.register_forward_hook(get_activation("layer2"))
elif model_name == "resnet18-layer1":
model.layer1.register_forward_hook(get_activation("layer1"))
elif model_name == "densenet161-denseblock4":
model.features.denseblock4.register_forward_hook(get_activation("denseblock4"))
elif (
model_name == "densenet161-features"
or model_name == "densenet161_places-features"
):
model.register_forward_hook(get_activation("features"))
elif model_name == "densenet161-fc":
model.classifier.register_forward_hook(get_activation("classifier"))
elif model_name == "googlenet-inception5b":
model.inception5b.register_forward_hook(get_activation("inception5b"))
elif model_name == "vit_b_16-features":
model.subset.register_forward_hook(get_activation("heads"))
elif model_name == "vit_b_16-head":
model.heads.head.register_forward_hook(get_activation("head"))
elif model_name == "vit_b_16-layer11":
model.encoder.layers[11].register_forward_hook(get_activation("layer11"))
elif model_name == "vit_b_16-ln":
model.encoder.ln.register_forward_hook(get_activation("ln"))
model_features = torch.zeros([len(dataset), n_neurons]).to(device)
with torch.no_grad():
for i, x in tqdm(enumerate(dataloader), total=len(dataloader)):
torch.cuda.empty_cache()
x = x.float().to(device)
_ = model(x)
# if i == 0:
# # Print the shape of the activation for debugging purposes
# print(
# f"Activation shape for {model_name}: {activation[model_name.split('-')[-1]].shape}"
# )
if model_name in [
"resnet18-fc",
"googlenet-fc",
"vit_b_16-head",
]:
model_features[i * x.size(0) : (i + 1) * x.size(0), :] = activation[
model_name.split("-")[-1]
].data
elif model_name == "densenet161-fc":
model_features[i * x.size(0) : (i + 1) * x.size(0), :] = activation[
"classifier"
].data
elif model_name == "vit_b_16-features":
model_features[i * x.size(0) : (i + 1) * x.size(0), :] = activation[
"heads"
].data
elif model_name in [
"resnet18-avgpool",
"resnet50_places-avgpool",
"resnet50-avgpool",
]:
model_features[i * x.size(0) : (i + 1) * x.size(0), :] = activation[
"avgpool"
][:, :, 0, 0].data
elif model_name == "resnet18-layer4":
model_features[i * x.size(0) : (i + 1) * x.size(0), :] = (
activation["layer4"].mean(dim=[2, 3]).data
)
elif model_name == "resnet18-layer3":
model_features[i * x.size(0) : (i + 1) * x.size(0), :] = (
activation["layer3"].mean(axis=[2, 3]).data
)
elif model_name == "resnet18-layer2":
model_features[i * x.size(0) : (i + 1) * x.size(0), :] = (
activation["layer2"].mean(axis=[2, 3]).data
)
elif model_name == "resnet18-layer1":
model_features[i * x.size(0) : (i + 1) * x.size(0), :] = (
activation["layer1"].mean(axis=[2, 3]).data
)
elif model_name in ["densenet161-features", "densenet161_places-features"]:
model_features[i * x.size(0) : (i + 1) * x.size(0), :] = activation[
"features"
].data
torch.save(model_features, tensor_path)
return model_features
def load_explanations(path, name, image_path, neuron_ids):
"""
Load explanations based on the given parameters.
Args:
path (str): The path to the CSV file containing the explanations.
name (str): The name of the explanation method.
image_path (str): The path to the directory containing the explanation images.
neuron_ids (list): A list of neuron IDs for which explanations are needed.
Returns:
tuple: A tuple containing two lists - explanations and explanations_filtered.
- explanations: A list of explanations for the given neuron IDs.
- explanations_filtered: A list of explanations that are missing corresponding images.
Raises:
FileNotFoundError: If the CSV file or the image directory does not exist.
"""
try:
df = pd.read_csv(path)
except FileNotFoundError:
raise FileNotFoundError("The CSV file does not exist.")
if name == "INVERT":
explanations = []
for neuron_id in neuron_ids:
explanation = df.loc[df["neuron"] == neuron_id, "concept"].values[0]
explanations.append(explanation)
elif name == "CLIP-Dissect":
explanations = []
for neuron_id in neuron_ids:
explanation = df.loc[df["unit"] == neuron_id, "description"].values[0]
explanations.append(explanation)
elif name == "MILAN":
explanations = []
for neuron_id in neuron_ids:
explanation = df.loc[df["unit"] == neuron_id, "description"].values[0]
explanations.append(explanation.lower())
elif name == "FALCON":
falcon_concept_list = []
for i in range(len(df)):
falcon_concepts_all = ast.literal_eval(df["concept_set_noun_phrases"][i])
falcon_concept = falcon_concepts_all[0][0]
falcon_concept_list.append(falcon_concept)
falcon_neuron_ids = df["group"].to_list()
falcon_concept_ids = dict(zip(falcon_neuron_ids, falcon_concept_list))
filtered_dict = {
key: value for key, value in falcon_concept_ids.items() if key in neuron_ids
}
explanations = list(filtered_dict.values())
# Check which explanation images are already existing and output missing ones
explanations_set = set(explanations)
explanations_set = list(explanations_set)
image_directories = [i.replace("_", " ") for i in os.listdir(image_path)]
missing_items = list(set(explanations_set) - set(image_directories))
explanations_filtered = missing_items
return explanations, explanations_filtered
def create_csv(filename, headers):
"""
Create a new CSV file with the given filename and write the headers to it.
Args:
filename (str): The name of the CSV file to create.
headers (list): A list of strings representing the column headers.
Returns:
None
"""
with open(filename, mode="w", newline="") as file:
writer = csv.writer(file)
writer.writerow(headers)
def add_rows_to_csv(filename, rows):
"""
Appends rows to a CSV file.
Args:
filename (str): The path to the CSV file.
rows (list): A list of rows to be added to the CSV file.
Returns:
None
"""
with open(filename, mode="a", newline="") as file:
writer = csv.writer(file)
for row in rows:
writer.writerow(row)