-
Notifications
You must be signed in to change notification settings - Fork 1
/
util.py
369 lines (293 loc) · 12.3 KB
/
util.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
"""Utility classes and methods.
Author:
Chris Chute (CS224n teaching staff): https://github.com/chrischute/squad
"""
import logging
import os
import queue
import re
import shutil
import torch
import torch.nn.functional as F
import torch.utils.data as data
import tqdm
import numpy as np
import ujson as json
class AverageMeter:
"""Keep track of average values over time.
Adapted from:
> https://github.com/pytorch/examples/blob/master/imagenet/main.py
"""
def __init__(self):
self.avg = 0
self.sum = 0
self.count = 0
def reset(self):
"""Reset meter."""
self.__init__()
def update(self, val, num_samples=1):
"""Update meter with new value `val`, the average of `num` samples.
Args:
val (float): Average value to update the meter with.
num_samples (int): Number of samples that were averaged to
produce `val`.
"""
self.count += num_samples
self.sum += val * num_samples
self.avg = self.sum / self.count
class CheckpointSaver:
"""Class to save and load model checkpoints.
Save the best checkpoints as measured by a metric value passed into the
`save` method. Overwrite checkpoints with better checkpoints once
`max_checkpoints` have been saved.
Args:
save_dir (str): Directory to save checkpoints.
max_checkpoints (int): Maximum number of checkpoints to keep before
overwriting old ones.
metric_name (str): Name of metric used to determine best model.
maximize_metric (bool): If true, best checkpoint is that which maximizes
the metric value passed in via `save`. Otherwise, best checkpoint
minimizes the metric.
log (logging.Logger): Optional logger for printing information.
"""
def __init__(self, save_dir, max_checkpoints, metric_name,
maximize_metric=False, log=None):
super(CheckpointSaver, self).__init__()
self.save_dir = save_dir
self.max_checkpoints = max_checkpoints
self.metric_name = metric_name
self.maximize_metric = maximize_metric
self.best_val = None
self.ckpt_paths = queue.PriorityQueue() # store checkpoint in sorted metric order
self.log = log
self._print('Saver will {}imize {}...'.format('max' if maximize_metric else 'min', metric_name))
def is_best(self, metric_val):
"""Check whether `metric_val` is the best seen so far.
Args:
metric_val (float): Metric value to compare to prior checkpoints.
"""
if metric_val is None:
# No metric reported
return False
if self.best_val is None:
# No checkpoint saved yet
return True
return ((self.maximize_metric and self.best_val < metric_val)
or (not self.maximize_metric and self.best_val > metric_val))
def _print(self, message):
"""Print a message if logging is enabled."""
if self.log is not None:
self.log.info(message)
def save(self, step, model, metric_val, device):
"""Save model parameters to disk.
Args:
step (int): Total number of examples seen during training so far.
model (torch.nn.DataParallel): Model to save.
metric_val (float): Determines whether checkpoint is best so far.
device (torch.device): Device where model resides.
"""
ckpt_dict = {
'model_name': model.__class__.__name__,
'model_state': model.cpu().state_dict(),
'step': step
}
model.to(device)
checkpoint_path = os.path.join(self.save_dir,
'step_{}.pth.tar'.format(step))
torch.save(ckpt_dict, checkpoint_path)
self._print('Saved checkpoint: {}'.format(checkpoint_path))
if self.is_best(metric_val):
# Save the best model
self.best_val = metric_val
best_path = os.path.join(self.save_dir, 'best.pth.tar')
shutil.copy(checkpoint_path, best_path)
self._print('New best checkpoint at step {}...'.format(step))
# Add checkpoint path to priority queue (lowest priority removed first)
if self.maximize_metric:
priority_order = metric_val
else:
priority_order = -metric_val
self.ckpt_paths.put((priority_order, checkpoint_path))
# Remove a checkpoint if more than max_checkpoints have been saved
if self.ckpt_paths.qsize() > self.max_checkpoints:
_, worst_ckpt = self.ckpt_paths.get()
try:
os.remove(worst_ckpt)
self._print('Removed checkpoint: {}'.format(worst_ckpt))
except OSError:
# Avoid crashing if checkpoint has been removed or protected
pass
def load_model(model, checkpoint_path, gpu_ids, return_step=True):
"""Load model parameters from disk.
Args:
model (torch.nn.DataParallel): Load parameters into this model.
checkpoint_path (str): Path to checkpoint to load.
gpu_ids (list): GPU IDs for DataParallel.
return_step (bool): Also return the step at which checkpoint was saved.
Returns:
model (torch.nn.DataParallel): Model loaded from checkpoint.
step (int): Step at which checkpoint was saved. Only if `return_step`.
"""
device = 'cuda:{}'.format(gpu_ids[0]) if gpu_ids else 'cpu'
ckpt_dict = torch.load(checkpoint_path, map_location=device)
# Build model, load parameters
model.load_state_dict(ckpt_dict['model_state'])
if return_step:
step = ckpt_dict['step']
return model, step
return model
def get_available_devices():
"""Get IDs of all available GPUs.
Returns:
device (torch.device): Main device (GPU 0 or CPU).
gpu_ids (list): List of IDs of all GPUs that are available.
"""
gpu_ids = []
if torch.cuda.is_available():
gpu_ids += [gpu_id for gpu_id in range(torch.cuda.device_count())]
device = torch.device('cuda:{}'.format(gpu_ids[0]))
torch.cuda.set_device(device)
else:
device = torch.device('cpu')
return device, gpu_ids
def visualize(tbx, y_pred, step, split, num_visuals, data_loader):
"""Visualize image examples to TensorBoard.
Args:
tbx (tensorboardX.SummaryWriter): Summary writer.
y_pred (list of int): List of predicted classes.
step (int): Number of examples seen so far during training.
split (str): Name of data split being visualized.
num_visuals (int): Number of visuals to select at random from preds.
data_loader (DataLoader): DataLoader object for the given split.
"""
if num_visuals <= 0:
return
if num_visuals > len(y_pred):
num_visuals = len(y_pred)
# sample 'num_visuals' random examples from 'split' for visualization
visual_ids = np.random.choice(list(range(len(y_pred))), size=num_visuals, replace=False)
for i, id_ in enumerate(visual_ids):
# get example pair
x, y = data_loader.dataset.__getitem__(id_, visualize=True)
# convert to numpy array
x = x.numpy()
# send image to Tensorboard
tbx.add_image('{}set/{}_of_{} : correct label {} , predicted label {}'.format(split, i + 1, num_visuals, y, y_pred[id_]),
x,
global_step=step)
def evaluate_preds(y_true, y_pred):
"""Evaluate the predicted classes from the model with the true classes.
Return a dictionary listing the metrics' values.
Args:
'y_true' (list of int): list of true class labels
'y_pred' (list of int): list of predicted class labels
Return:
'results' (dict): dictionary listing the metrics' values
Remark:
Implemented metrics: Accuracy, macro F1-score, Mean Absolute Error (MAE)
"""
def accuracy(y_true, y_pred):
n_examples = y_true.size
n_correct = np.sum(y_true == y_pred)
accuracy = n_correct / n_examples *100
return accuracy
def MAE(y_true, y_pred):
n_examples = y_true.size
diff = np.abs(y_true - y_pred)
MAE = np.sum(diff)/n_examples
return MAE
def F1(y_true, y_pred, weights={}):
F1s = {}
# compute F1-score per class
for c in set(y_true):
# class mask
mask = y_true == c
mask_pred = y_pred == c
# compute class F1-score
recall = np.sum(y_pred[mask] == c)/np.sum(mask)
precision = np.sum(y_pred[mask] == c)/np.sum(mask_pred)
if (recall == 0) or (precision == 0):
F1s[c] = 0
else:
F1s[c] = 2*(recall*precision)/(recall + precision)
F1 = np.sum([weights.get(c, 1/len(F1s))*F1 for c, F1 in F1s.items()]) * 100
return F1
y_true = np.array(y_true)
y_pred = np.array(y_pred)
results = {'Acc': accuracy(y_true, y_pred), 'MAE': MAE(y_true, y_pred), 'F1': F1(y_true, y_pred)}
return results
def get_save_dir(base_dir, name, training, id_max=100):
"""Get a unique save directory by appending the smallest positive integer
`id < id_max` that is not already taken (i.e., no dir exists with that id).
Args:
base_dir (str): Base directory in which to make save directories.
name (str): Name to identify this training run. Need not be unique.
training (bool): Save dir. is for training (determines subdirectory).
id_max (int): Maximum ID number before raising an exception.
Returns:
save_dir (str): Path to a new directory with a unique name.
"""
for uid in range(1, id_max):
subdir = 'train' if training else 'test'
save_dir = os.path.join(base_dir, subdir, '{}-{:02d}'.format(name, uid))
if not os.path.exists(save_dir):
os.makedirs(save_dir)
return save_dir
raise RuntimeError('Too many save directories created with the same name. \
Delete old save directories or use another name.')
def get_logger(log_dir, name):
"""Get a `logging.Logger` instance that prints to the console
and an auxiliary file.
Args:
log_dir (str): Directory in which to create the log file.
name (str): Name to identify the logs.
Returns:
logger (logging.Logger): Logger instance for logging events.
"""
class StreamHandlerWithTQDM(logging.Handler):
"""Let `logging` print without breaking `tqdm` progress bars.
See Also:
> https://stackoverflow.com/questions/38543506
"""
def emit(self, record):
try:
msg = self.format(record)
tqdm.tqdm.write(msg)
self.flush()
except (KeyboardInterrupt, SystemExit):
raise
except:
self.handleError(record)
# Create logger
logger = logging.getLogger(name)
logger.setLevel(logging.DEBUG)
# Log everything (i.e., DEBUG level and above) to a file
log_path = os.path.join(log_dir, 'log.txt')
file_handler = logging.FileHandler(log_path)
file_handler.setLevel(logging.DEBUG)
# Log everything except DEBUG level (i.e., INFO level and above) to console
console_handler = StreamHandlerWithTQDM()
console_handler.setLevel(logging.INFO)
# Create format for the logs
file_formatter = logging.Formatter('[%(asctime)s] %(message)s',
datefmt='%m.%d.%y %H:%M:%S')
file_handler.setFormatter(file_formatter)
console_formatter = logging.Formatter('[%(asctime)s] %(message)s',
datefmt='%m.%d.%y %H:%M:%S')
console_handler.setFormatter(console_formatter)
# add the handlers to the logger
logger.addHandler(file_handler)
logger.addHandler(console_handler)
return logger
def torch_from_json(path, dtype=torch.float32):
"""Load a PyTorch Tensor from a JSON file.
Args:
path (str): Path to the JSON file to load.
dtype (torch.dtype): Data type of loaded array.
Returns:
tensor (torch.Tensor): Tensor loaded from JSON file.
"""
with open(path, 'r') as fh:
array = np.array(json.load(fh))
tensor = torch.from_numpy(array).type(dtype)
return tensor