-
Notifications
You must be signed in to change notification settings - Fork 102
/
get_signals.py
206 lines (186 loc) · 7.59 KB
/
get_signals.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
import os.path
from typing import Optional
import numpy as np
import torch
from torch.nn import functional as F
from tqdm import tqdm
from transformers import PreTrainedModel, AutoTokenizer
from dataset.utils import load_dataset_subsets
def get_softmax(
model: PreTrainedModel,
samples: torch.Tensor,
labels: torch.Tensor,
batch_size: int,
device: str,
temp: float = 1.0,
pad_token_id: Optional[int] = None,
) -> np.ndarray:
"""
Get the model's softmax probabilities for the given inputs and expected outputs.
Args:
model (PreTrainedModel or torch.nn.Module): Model instance.
samples (torch.Tensor): Model input.
labels (torch.Tensor): Model expected output.
batch_size (int): Batch size for getting signals.
device (str): Device used for computing signals.
temp (float): Temperature used in softmax computation.
pad_token_id (Optional[int]): Padding token ID to ignore in aggregation.
Returns:
all_softmax_list (np.array): softmax value of all samples
"""
model.to(device)
model.eval()
with torch.no_grad():
softmax_list = []
batched_samples = torch.split(samples, batch_size)
batched_labels = torch.split(labels, batch_size)
for x, y in tqdm(
zip(batched_samples, batched_labels),
total=len(batched_samples),
desc="Computing softmax",
):
x = x.to(device)
y = y.to(device)
pred = model(x)
if isinstance(model, PreTrainedModel):
logits = pred.logits
logit_signals = torch.div(logits, temp)
softmax_probs = torch.log_softmax(logit_signals, dim=-1)
true_class_probs = softmax_probs.gather(2, y.unsqueeze(-1)).squeeze(-1)
# Mask out padding tokens
mask = (
y != pad_token_id
if pad_token_id is not None
else torch.ones_like(y, dtype=torch.bool)
)
true_class_probs = true_class_probs * mask
sequence_probs = torch.exp(
(true_class_probs * mask).sum(1) / mask.sum(1)
)
softmax_list.append(sequence_probs.to("cpu").view(-1, 1))
else:
logit_signals = torch.div(pred, temp)
max_logit_signals, _ = torch.max(logit_signals, dim=1)
logit_signals = torch.sub(
logit_signals, max_logit_signals.reshape(-1, 1)
)
exp_logit_signals = torch.exp(logit_signals)
exp_logit_sum = exp_logit_signals.sum(dim=1).reshape(-1, 1)
true_exp_logit = exp_logit_signals.gather(1, y.reshape(-1, 1))
softmax_list.append(torch.div(true_exp_logit, exp_logit_sum).to("cpu"))
all_softmax_list = np.concatenate(softmax_list)
model.to("cpu")
return all_softmax_list
def get_loss(
model: PreTrainedModel,
samples: torch.Tensor,
labels: torch.Tensor,
batch_size: int,
device: str,
pad_token_id: Optional[int] = None,
) -> np.ndarray:
"""
Get the model's loss for the given inputs and expected outputs.
Args:
model (PreTrainedModel or torch.nn.Module): Model instance.
samples (torch.Tensor): Model input.
labels (torch.Tensor): Model expected output.
batch_size (int): Batch size for getting signals.
device (str): Device used for computing signals.
pad_token_id (Optional[int]): Padding token ID to ignore in aggregation.
Returns:
all_loss_list (np.array): Loss value of all samples
"""
model.to(device)
model.eval()
with torch.no_grad():
loss_list = []
batched_samples = torch.split(samples, batch_size)
batched_labels = torch.split(labels, batch_size)
for x, y in zip(batched_samples, batched_labels):
x = x.to(device)
y = y.to(device)
if isinstance(model, PreTrainedModel):
logit_signals = model(x).logits
loss = torch.nn.CrossEntropyLoss(
reduction="none", ignore_index=pad_token_id
)(logit_signals.transpose(1, 2), y)
mask = loss != 0
loss = (loss * mask).sum(1) / mask.sum(1)
loss_list.append(loss.cpu().detach().numpy().reshape(batch_size, -1))
else:
logit_signals = model(x)
loss_list.append(
F.cross_entropy(logit_signals, y.ravel(), reduction="none").to(
"cpu"
)
)
all_loss_list = np.concatenate(loss_list).reshape((-1, 1))
model.to("cpu")
return all_loss_list
def get_model_signals(models_list, dataset, configs, logger):
"""Function to get models' signals (softmax, loss, logits) on a given dataset.
Args:
models_list (list): List of models for computing (softmax, loss, logits) signals from them.
dataset (torchvision.datasets): The whole dataset.
configs (dict): Configurations of the tool.
logger (logging.Logger): Logger object for the current run.
Returns:
signals (np.array): Signal value for all samples in all models
"""
# Check if signals are available on disk
if os.path.exists(
f"{configs['run']['log_dir']}/signals/{configs['audit']['algorithm'].lower()}_signals.npy",
):
signals = np.load(
f"{configs['run']['log_dir']}/signals/{configs['audit']['algorithm'].lower()}_signals.npy",
)
if signals.shape[0] == len(dataset):
logger.info("Signals loaded from disk.")
return signals
else:
logger.warning(
"Signals shape does not match the audit data size. This is probably due to a different audit data size than the training data size."
)
logger.info("Ignoring the signals on disk and recomputing.")
batch_size = configs["audit"]["batch_size"] # Batch size used for inferring signals
model_name = configs["train"]["model_name"] # Algorithm used for training models
device = configs["audit"]["device"] # GPU device used for inferring signals
if "tokenizer" in configs["data"].keys():
tokenizer = AutoTokenizer.from_pretrained(
configs["data"]["tokenizer"], clean_up_tokenization_spaces=True
)
if tokenizer.pad_token is None:
tokenizer.pad_token = tokenizer.eos_token
pad_token_id = tokenizer.pad_token_id
else:
pad_token_id = None
dataset_samples = np.arange(len(dataset))
data, targets = load_dataset_subsets(
dataset, dataset_samples, model_name, batch_size, device
)
signals = []
for model in models_list:
if configs["audit"]["algorithm"] == "RMIA":
signals.append(
get_softmax(
model, data, targets, batch_size, device, pad_token_id=pad_token_id
)
)
elif configs["audit"]["algorithm"] == "LOSS":
signals.append(
get_loss(
model, data, targets, batch_size, device, pad_token_id=pad_token_id
)
)
else:
raise NotImplementedError(
f"{configs['audit']['algorithm']} is not implemented"
)
signals = np.concatenate(signals, axis=1)
np.save(
f"{configs['run']['log_dir']}/signals/{configs['audit']['algorithm'].lower()}_signals.npy",
signals,
)
logger.info("Signals saved to disk.")
return signals