-
Notifications
You must be signed in to change notification settings - Fork 11
/
trainers.py
347 lines (298 loc) · 12.4 KB
/
trainers.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
# copyright (c) 2017 NVIDIA CORPORATION. All rights reserved.
# See the LICENSE file for licensing terms (BSD-style).
"""A set of "trainers", classes that wrap around Torch models
and provide methods for training and evaluation."""
import time
import types
import platform
import numpy as np
import torch
from torch import autograd, nn, optim
from torch.autograd import Variable, Function
import torch.nn.functional as F
from scipy import ndimage
import helpers as dlh
def add_log(log, logname, **kw):
entry = dict(kw, __log__=logname, __at__=time.time(), __node__=platform.node())
log.append(entry)
def get_log(log, logname, **kw):
records = [x for x in log if x.get("__log__")==logname]
return records
def update_display():
from matplotlib import pyplot
from IPython import display
display.clear_output(wait=True)
display.display(pyplot.gcf())
class Weighted(Function):
def forward(self, x, weights):
self.saved_for_backward = [weights]
return x
def backward(self, grad_output):
weights, = self.saved_for_backward
grad_input = weights * grad_output
return grad_input
class BasicTrainer(object):
"""Trainers take care of bookkeeping for training models.
The basic method is `train_batch(inputs, targets)`. It catches errors
during forward propagation and reports the model and input shapes
(shape mismatches are the most common source of errors.
Trainers are just a temporary tool that's wrapped around a model
for training purposes, so you can create, use, and discard them
as convenient.
"""
def __init__(self, model, use_cuda=True,
fields = ("input", "output"),
input_axes = None,
output_axes = None):
self.use_cuda = use_cuda
self.model = self._cuda(model)
self.init_loss()
self.input_name, self.output_name = fields
self.no_display = False
self.current_lr = None
self.optimizer = None
self.weighted = Weighted()
self.ntrain = 0
self.log = []
def _cuda(self, x):
"""Convert object to CUDA if use_cuda==True."""
if self.use_cuda:
return x.cuda()
else:
return x.cpu()
def set_training(self, mode=True):
"""Set training or prediction mode."""
if mode:
if not self.model.training:
self.model.train()
self.cuinput = autograd.Variable(
torch.randn(1, 1, 100, 100).cuda())
self.cutarget = autograd.Variable(torch.randn(1, 11).cuda())
else:
if self.model.training:
self.model.eval()
self.cuinput = autograd.Variable(torch.randn(1, 1, 100, 100).cuda(),
volatile=True)
self.cutarget = autograd.Variable(torch.randn(1, 11).cuda(),
volatile=True)
def set_lr(self, lr, momentum=0.9, weight_decay=0.0):
"""Set the optimizer to SGD with the given parameters."""
self.current_lr = lr
self.optimizer = optim.SGD(self.model.parameters(),
lr=lr,
momentum=momentum,
weight_decay=weight_decay)
def get_outputs(self):
"""Performs any necessary transformations on the output tensor.
"""
return dlh.novar(self.cuoutput).cpu()
def set_inputs(self, batch):
"""Sets the cuinput variable from the input data.
"""
assert isinstance(batch, torch.Tensor)
dlh.assign(self.cuinput, batch)
def set_targets(self, targets, weights=None):
"""Sets the cutarget variable from the given tensor.
"""
dlh.assign(self.cutarget, targets, False)
assert self.cuoutput.size() == self.cutargets.size()
if weights is not None:
dlh.assign(self.cuweights, weights, False)
assert self.cuoutput.size() == self.cuweights.size()
else:
self.cuweights = None
def init_loss(self, loss=nn.MSELoss()):
self.criterion = self._cuda(loss)
def compute_loss(self, targets, weights=None):
self.set_targets(targets, weights=weights)
return self.criterion(self.cuoutput, self.cutarget)
def forward(self):
try:
self.cuoutput = self.model(self.cuinput)
except RuntimeError, err:
print "runtime error in forward step:"
print "input", self.cuinput.size()
raise err
def train_batch(self, inputs, targets, weights=None, update=True, logname="train"):
if update:
self.set_training(True)
self.optimizer.zero_grad()
else:
self.set_training(False)
self.set_inputs(inputs)
self.forward()
if weights is not None:
self.cuweights = autograd.Variable(torch.randn(1, 1).cuda())
dlh.assign(self.cuweights, weights, False)
self.cuoutput = self.weighted(self.cuoutput, self.cuweights)
culoss = self.compute_loss(targets, weights=weights)
if update:
culoss.backward()
self.optimizer.step()
ploss = dlh.novar(culoss)[0]
self.ntrain += dlh.size(inputs, 0)
add_log(self.log, logname, loss=ploss, ntrain=self.ntrain, lr=self.current_lr)
return self.get_outputs(), ploss
def eval_batch(self, inputs, targets):
return self.train_batch(inputs, targets, update=False, logname="eval")
def predict_batch(self, inputs):
self.set_training(False)
self.set_inputs(inputs)
self.forward()
return self.get_outputs()
def loss_curve(self, logname):
records = get_log(self.log, logname)
records = [(x["ntrain"], x["loss"]) for x in records]
records = sorted(records)
if len(records)==0:
return [], []
else:
return zip(*records)
def plot_loss(self, every=100, smooth=1e-2, yscale=None):
if self.no_display: return
# we import these locally to avoid dependence on display
# functions for training
import matplotlib as mpl
from matplotlib import pyplot
from scipy.ndimage import filters
x, y = self.loss_curve("train")
pyplot.plot(x, y)
x, y = self.loss_curve("test")
pyplot.plot(x, y)
def display_loss(self, *args, **kw):
pyplot.clf()
self.plot_loss(*args, **kw)
update_display()
def set_sample_fields(self, input_name, output_name):
self.input_name = input_name
self.output_name = output_name
def train_for(self, training, training_size=1e99):
if isinstance(training, types.FunctionType):
training = training()
count = 0
losses = []
for batch in training:
if count >= training_size: break
input_tensor = batch[self.input_name]
output_tensor = batch[self.output_name]
_, loss = self.train_batch(input_tensor, output_tensor)
count += len(input_tensor)
losses.append(loss)
loss = np.mean(losses)
return loss, count
def eval_for(self, testset, testset_size=1e99):
if isinstance(testset, types.FunctionType):
testset = testset()
count = 0
losses = []
for batch in testset:
if count >= testset_size: break
input_tensor = batch[self.input_name]
output_tensor = batch[self.output_name]
_, loss = self.eval_batch(input_tensor, output_tensor)
count += len(input_tensor)
losses.append(loss)
loss = np.mean(losses)
return loss, count
class ImageClassifierTrainer(BasicTrainer):
def __init__(self, *args, **kw):
BasicTrainer.__init__(self, *args, **kw)
def set_inputs(self, images, depth1=False):
dlh.assign(self.cuinput, images, transpose_on_convert=(0, 3, 1, 2))
def set_targets(self, targets, weights=None):
assert weights is None, "weights not implemented"
if isinstance(targets, list):
targets = np.array(targets)
if dlh.rank(targets) == 1:
targets = dlh.as_torch(targets)
targets = targets.unsqueeze(1)
b, c = dlh.shp(self.cuoutput)
onehot = torch.zeros(b, c)
onehot.scatter_(1, targets, 1)
dlh.assign(self.cutarget, onehot)
else:
assert dlh.shp(targets) == dlh.shp(self.cuoutput)
dlh.assign(self.cutarget, targets)
def zoom_like(batch, target_shape, order=0):
assert isinstance(batch, np.ndarray)
scales = [r * 1.0 / b for r, b in zip(target_shape, batch.shape)]
result = np.zeros(target_shape)
ndimage.zoom(batch, scales, order=order, output=result)
return result
def pixels_to_batch(x):
b, d, h, w = x.size()
return x.permute(0, 2, 3, 1).contiguous().view(b*h*w, d)
class Image2ImageTrainer(BasicTrainer):
"""Train image to image models."""
def __init__(self, *args, **kw):
BasicTrainer.__init__(self, *args, **kw)
def compute_loss(self, targets, weights=None):
self.set_targets(targets, weights=weights)
return self.criterion(pixels_to_batch(self.cuoutput),
pixels_to_batch(self.cutarget))
def set_inputs(self, images):
dlh.assign(self.cuinput, images, (0, 3, 1, 2))
def get_outputs(self):
return dlh.as_nda(self.cuoutput, (0, 2, 3, 1))
def set_targets(self, targets, weights=None):
b, d, h, w = tuple(self.cuoutput.size())
targets = dlh.as_nda(targets, (0, 2, 3, 1))
targets = zoom_like(targets, (b, h, w, d))
dlh.assign(self.cutarget, targets, (0, 3, 1, 2))
assert self.cutarget.size() == self.cuoutput.size()
if weights is not None:
weights = dlh.as_nda(weights, (0, 2, 3, 1))
weights = zoom_like(weights, (b, h, w, d))
dlh.assign(self.cuweights, weights, (0, 3, 1, 2))
def ctc_align(prob, target):
"""Perform CTC alignment on torch sequence batches (using ocrolstm).
Inputs are in BDL format.
"""
import cctc
assert dlh.sequence_is_normalized(prob), prob
assert dlh.sequence_is_normalized(target), target
# inputs are BDL
prob_ = dlh.novar(prob).permute(0, 2, 1).cpu().contiguous()
target_ = dlh.novar(target).permute(0, 2, 1).cpu().contiguous()
# prob_ and target_ are both BLD now
assert prob_.size(0) == target_.size(0), (prob_.size(), target_.size())
assert prob_.size(2) == target_.size(2), (prob_.size(), target_.size())
assert prob_.size(1) >= target_.size(1), (prob_.size(), target_.size())
result = torch.rand(1)
cctc.ctc_align_targets_batch(result, prob_, target_)
return dlh.typeas(result.permute(0, 2, 1).contiguous(), prob)
def sequence_softmax(seq):
"""Given a BDL sequence, computes the softmax for each time step."""
b, d, l = seq.size()
batch = seq.permute(0, 2, 1).contiguous().view(b*l, d)
smbatch = F.softmax(batch)
result = smbatch.view(b, l, d).permute(0, 2, 1).contiguous()
return result
class Image2SeqTrainer(BasicTrainer):
"""Train image to sequence models using CTC.
This takes images in BHWD order, plus output sequences
consisting of lists of integers.
"""
def __init__(self, *args, **kw):
BasicTrainer.__init__(self, *args, **kw)
def init_loss(self, loss=None):
assert loss is None, "Image2SeqTrainer must be trained with BCELoss (default)"
self.criterion = nn.BCELoss(size_average=False)
def compute_loss(self, targets, weights=None):
self.cutargets = None # not used
assert weights is None
logits = self.cuoutput
b, d, l = logits.size()
probs = sequence_softmax(logits)
assert dlh.sequence_is_normalized(probs), probs
ttargets = torch.FloatTensor(targets)
target_b, target_d, target_l = ttargets.size()
assert b == target_b, (b, target_b)
assert dlh.sequence_is_normalized(ttargets), ttargets
aligned = ctc_align(probs.cpu(), ttargets.cpu())
assert dlh.sequence_is_normalized(aligned)
return self.criterion(probs, Variable(self._cuda(aligned)))
def set_inputs(self, images):
dlh.assign(self.cuinput, images, (0, 3, 1, 2))
def set_targets(self, targets, outputs, weights=None):
raise Exception("overridden by compute_loss")