-
Notifications
You must be signed in to change notification settings - Fork 7
/
Solver.py
478 lines (417 loc) · 17 KB
/
Solver.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
import logging
import mxnet as mx
import numpy as np
def save_checkpoint(prefix, epoch, symbol, arg_params, aux_params):
"""Checkpoint the model data into file.
Parameters
----------
prefix : str
Prefix of model name.
epoch : int
The epoch number of the model.
symbol : Symbol
The input symbol
arg_params : dict of str to NDArray
Model parameter, dict of name to NDArray of net's weights.
aux_params : dict of str to NDArray
Model parameter, dict of name to NDArray of net's auxiliary states.
Notes
-----
- ``prefix-symbol.json`` will be saved for symbol.
- ``prefix-epoch.params`` will be saved for parameters.
"""
symbol.save('%s-symbol.json' % prefix)
save_dict = {('arg:%s' % k) : v.copyto(mx.cpu()) for k, v in arg_params.items()}
save_dict.update({('aux:%s' % k) : v.copyto(mx.cpu()) for k, v in aux_params.items()})
param_name = '%s-%04d.params' % (prefix, epoch)
mx.nd.save(param_name, save_dict)
logging.info('Saved checkpoint to (ctx=cpu) \"%s\"', param_name)
class CarReID_Solver(object):
def __init__(self, prefix='', symbol=None, ctx=None, data_shape=None, label_shape=None,
num_epoch=None, opt_method='sgd', **kwargs):
self.prefix = prefix
self.symbol = symbol
self.ctx = ctx
if self.ctx is None:
self.ctx = mx.cpu()
self.data_shape = data_shape
self.label_shape = label_shape
self.batchsize = data_shape[0]
self.num_epoch = num_epoch
self.update_params = None
self.arg_params = None
self.aux_params = None
self.grad_params = None
self.executor = None
self.opt_method = opt_method
self.optimizer = None
self.updater = None
self.kwargs = kwargs.copy()
self.initializer=mx.init.Xavier()
def init_args(self, args):
for key in args:
arr = args[key]
if key.endswith('_weight'):
self.initializer(key, arr)
if key.endswith('_bias'):
arr[:] = 0.0
if key.endswith('_gamma'):
arr[:] = 1.0
if key.endswith('_beta'):
arr[:] = 0.0
if key.endswith('_init_c'):
arr[:] = 0.0
if key.endswith('_init_h'):
arr[:] = 0.0
def get_params(self, grad_req):
arg_names = self.symbol.list_arguments()
arg_shapes, _, aux_shapes = \
self.symbol.infer_shape(part1_data=self.data_shape,
part2_data=self.data_shape,
label=self.label_shape)
self.arg_params = {}
self.update_params = {}
for name, shape in zip(arg_names, arg_shapes):
self.arg_params[name] = mx.nd.zeros(shape, self.ctx)
if name.endswith('weight') or name.endswith('bias') or name.endswith('gamma') or name.endswith('beta'):
# print name
self.update_params[name] = self.arg_params[name]
self.init_args(self.arg_params)
if grad_req != 'null':
self.grad_params = {}
for name, shape in zip(arg_names, arg_shapes):
self.grad_params[name] = mx.nd.zeros(shape, self.ctx)
aux_names = self.symbol.list_auxiliary_states()
self.aux_params = {k: mx.nd.zeros(s, self.ctx) for k, s in zip(aux_names, aux_shapes)}
def set_params(self, whichone):
logging.info('loading checkpoint from %s-->%d...', self.prefix, whichone)
loadfunc = mx.model.load_checkpoint
self.symbol, update_params, aux_params = loadfunc(self.prefix, whichone)
for name in self.update_params:
self.arg_params[name][:] = update_params[name]
for name in self.aux_params:
self.aux_params[name][:] = aux_params[name]
# name = 'PART2_COV_5_bn_moving_mean'
# print name, aux_params[name].asnumpy()
# exit()
return
def fit(self, train_data, grad_req='write', showperiod=100, whichone=None, logger=None):
if logger is not None:
logger.info('Start training with %s', str(self.ctx))
savefunc = save_checkpoint
self.get_params(grad_req)
if whichone is not None:
self.set_params(whichone)
self.optimizer = mx.optimizer.create(self.opt_method, rescale_grad=(1.0 / self.batchsize), **self.kwargs)
self.updater = mx.optimizer.get_updater(self.optimizer)
self.executor = self.symbol.bind(self.ctx, self.arg_params, args_grad=self.grad_params,
grad_req=grad_req, aux_states=self.aux_params)
update_dict = self.update_params
# epoch_end_callback = mx.callback.do_checkpoint(self.prefix)
# begin training
for epoch in range(0, self.num_epoch):
nbatch = 0
train_data.reset()
num_batches = train_data.num_batches
cost = []
for databatch in train_data:
nbatch += 1
for k, v in databatch.data.items():
# print k, v.shape
self.arg_params[k][:] = mx.nd.array(v, self.ctx)
for k, v in databatch.label.items():
# print k, v.shape
self.arg_params[k][:] = mx.nd.array(v, self.ctx)
output_dict = {name: nd for name, nd in zip(self.symbol.list_outputs(), self.executor.outputs)}
self.executor.forward(is_train=True)
self.executor.backward()
for key in update_dict:
# print key, np.sum(arr.asnumpy())
arr = self.grad_params[key]
self.updater(key, arr, self.arg_params[key])
outval = output_dict['reid_loss_output'].asnumpy()
outval = np.mean(outval)
cost.append(outval)
lrsch = self.optimizer.lr_scheduler
step = lrsch.step
nowlr = lrsch.base_lr
num_update = self.optimizer.num_update
if num_update % showperiod == 0:
print num_update, 'cost:', np.mean(cost), 'lr:', nowlr, num_batches
cost = []
# epoch_end_callback(epoch, self.symbol, self.update_params, self.aux_params)
savefunc(self.prefix, epoch%10, self.symbol, self.update_params, self.aux_params)
# print databatch.label['label'].T
class CarReID_Proxy_Solver(object):
def __init__(self, prefix='', symbol=None, ctx=None, data_shape=None, label_shape=None,
num_epoch=None, falsebigbatch=1, opt_method='sgd', **kwargs):
self.prefix = prefix
self.symbol = symbol
self.ctx = ctx
if self.ctx is None:
self.ctx = mx.cpu()
self.data_shape = data_shape
self.label_shape = label_shape
self.batchsize = data_shape[0]
self.num_epoch = num_epoch
self.update_params = None
self.falsebigbatch = falsebigbatch
print 'false big batch size:%d*%d=%d'%(falsebigbatch, self.batchsize, falsebigbatch * self.batchsize)
self.arg_params = None
self.aux_params = None
self.grad_params = None
self.executor = None
self.opt_method = opt_method
self.optimizer = None
self.updater = None
self.kwargs = kwargs.copy()
self.initializer=mx.init.Xavier()
def init_args(self, args):
for key in args:
arr = args[key]
if key.endswith('_weight'):
# self.initializer(mx.init.InitDesc(key), arr)
self.initializer(key, arr)
if key.endswith('_bias'):
arr[:] = 0.0
if key.endswith('_gamma'):
arr[:] = 1.0
if key.endswith('_beta'):
arr[:] = 0.0
if key.endswith('_init_c'):
arr[:] = 0.0
if key.endswith('_init_h'):
arr[:] = 0.0
def get_params(self, grad_req):
arg_names = self.symbol.list_arguments()
arg_shapes, _, aux_shapes = \
self.symbol.infer_shape(data=self.data_shape,
**self.label_shape)
self.arg_params = {}
self.update_params = {}
for name, shape in zip(arg_names, arg_shapes):
# print name, shape
self.arg_params[name] = mx.nd.zeros(shape, self.ctx)
if name.endswith('weight') or name.endswith('bias') or \
name.endswith('gamma') or name.endswith('beta') or \
name=='proxy_Z':
# print name
self.update_params[name] = self.arg_params[name]
self.init_args(self.arg_params)
if grad_req != 'null':
self.grad_params = {}
self.bigbatch_grads = {}
for name, shape in zip(arg_names, arg_shapes):
self.grad_params[name] = mx.nd.zeros(shape, self.ctx)
self.bigbatch_grads[name] = mx.nd.zeros(shape, self.ctx)
aux_names = self.symbol.list_auxiliary_states()
self.aux_params = {k: mx.nd.zeros(s, self.ctx) for k, s in zip(aux_names, aux_shapes)}
def set_params(self, whichone):
logging.info('loading checkpoint from %s-->%d...', self.prefix, whichone)
loadfunc = mx.model.load_checkpoint
_, update_params, aux_params = loadfunc(self.prefix, whichone)
for name in self.update_params:
self.arg_params[name][:] = update_params[name]
for name in self.aux_params:
self.aux_params[name][:] = aux_params[name]
# name = 'PART2_COV_5_bn_moving_mean'
# print name, aux_params[name].asnumpy()
# exit()
return
def fit(self, train_data, grad_req='write', showperiod=100, whichone=None, logger=None):
if logger is not None:
logger.info('Start training with %s', str(self.ctx))
savefunc = save_checkpoint
self.get_params(grad_req)
if whichone is not None:
self.set_params(whichone)
else:
import DataGenerator as dg
proxyfn = 'proxy.bin'
proxy_set = dg.get_proxyset(proxyfn, self.arg_params['proxy_Z'].shape)
self.arg_params['proxy_Z'][:] = mx.nd.array(proxy_set, self.ctx)
self.optimizer = mx.optimizer.create(self.opt_method, rescale_grad=(1.0 / self.batchsize), **self.kwargs)
self.updater = mx.optimizer.get_updater(self.optimizer)
self.executor = self.symbol.bind(self.ctx, self.arg_params, args_grad=self.grad_params,
grad_req=grad_req, aux_states=self.aux_params)
update_dict = self.update_params
# epoch_end_callback = mx.callback.do_checkpoint(self.prefix)
# begin training
cost = []
nbatch = 0
for epoch in range(0, self.num_epoch):
train_data.reset()
num_batches = train_data.num_batches
for databatch in train_data:
nbatch += 1
for ks, v in zip(train_data.provide_data, databatch.data):
k = ks[0]
self.arg_params[k][:] = v
for ks, v in zip(train_data.provide_label, databatch.label):
k = ks[0]
self.arg_params[k][:] = v
output_dict = {name: nd for name, nd in zip(self.symbol.list_outputs(), self.executor.outputs)}
self.executor.forward(is_train=True)
self.executor.backward()
for key in update_dict:
arr = self.grad_params[key]
self.bigbatch_grads[key][:] = self.bigbatch_grads[key] + arr
if nbatch%self.falsebigbatch==0:
# print nbatch, self.bigbatchnum
for key in update_dict:
arr = self.grad_params[key]
arr[:] = self.bigbatch_grads[key] / self.falsebigbatch
self.updater(key, arr, self.arg_params[key])
self.bigbatch_grads[key][:] = 0
outval = output_dict['proxy_nca_loss_output'].asnumpy()
# print np.mean(outval)
cost.append(np.mean(outval))
lrsch = self.optimizer.lr_scheduler
step = lrsch.step
nowlr = lrsch.base_lr
num_update = nbatch
if num_update % showperiod == 0:
print num_update, 'cost:', np.mean(cost), 'lr:', nowlr, num_batches
cost = []
savefunc(self.prefix, epoch%10, self.symbol, self.update_params, self.aux_params)
## print databatch.label['label'].T
class CarReID_Softmax_Solver(object):
def __init__(self, prefix='', symbol=None, ctx=None, data_shape=None, label_shape=None,
num_epoch=None, falsebigbatch=1, opt_method='sgd', **kwargs):
self.prefix = prefix
self.symbol = symbol
self.ctx = ctx
if self.ctx is None:
self.ctx = mx.cpu()
self.data_shape = data_shape
self.label_shape = label_shape
self.batchsize = data_shape[0]
self.num_epoch = num_epoch
self.update_params = None
self.arg_params = None
self.aux_params = None
self.grad_params = None
self.executor = None
self.opt_method = opt_method
self.optimizer = None
self.falsebigbatch = falsebigbatch
print 'false big batch size:%d*%d=%d'%(falsebigbatch, self.batchsize, falsebigbatch * self.batchsize)
self.bigbatch_grads = None
self.updater = None
self.kwargs = kwargs.copy()
self.initializer=mx.init.Xavier()
def init_args(self, args):
for key in args:
arr = args[key]
if key.endswith('_weight'):
self.initializer(key, arr)
if key.endswith('_bias'):
arr[:] = 0.0
if key.endswith('_gamma'):
arr[:] = 1.0
if key.endswith('_beta'):
arr[:] = 0.0
if key.endswith('_init_c'):
arr[:] = 0.0
if key.endswith('_init_h'):
arr[:] = 0.0
def get_params(self, grad_req):
arg_names = self.symbol.list_arguments()
arg_shapes, _, aux_shapes = \
self.symbol.infer_shape(data=self.data_shape,
label=self.label_shape)
self.arg_params = {}
self.update_params = {}
for name, shape in zip(arg_names, arg_shapes):
self.arg_params[name] = mx.nd.zeros(shape, self.ctx)
if name.endswith('weight') or name.endswith('bias') or name.endswith('gamma') or name.endswith('beta'):
# print name
self.update_params[name] = self.arg_params[name]
self.init_args(self.arg_params)
if grad_req != 'null':
self.grad_params = {}
self.bigbatch_grads = {}
for name, shape in zip(arg_names, arg_shapes):
self.grad_params[name] = mx.nd.zeros(shape, self.ctx)
self.bigbatch_grads[name] = mx.nd.zeros(shape, self.ctx)
aux_names = self.symbol.list_auxiliary_states()
self.aux_params = {k: mx.nd.zeros(s, self.ctx) for k, s in zip(aux_names, aux_shapes)}
def set_params(self, whichone):
logging.info('loading checkpoint from %s-->%d...', self.prefix, whichone)
loadfunc = mx.model.load_checkpoint
self.symbol, update_params, aux_params = loadfunc(self.prefix, whichone)
for name in self.update_params:
self.arg_params[name][:] = update_params[name]
for name in self.aux_params:
self.aux_params[name][:] = aux_params[name]
# name = 'PART2_COV_5_bn_moving_mean'
# print name, aux_params[name].asnumpy()
# exit()
return
def fit(self, train_data, grad_req='write', showperiod=100, whichone=None, logger=None):
if logger is not None:
logger.info('Start training with %s', str(self.ctx))
savefunc = save_checkpoint
self.get_params(grad_req)
if whichone is not None:
self.set_params(whichone)
self.optimizer = mx.optimizer.create(self.opt_method, rescale_grad=(1.0 / self.batchsize), **self.kwargs)
self.updater = mx.optimizer.get_updater(self.optimizer)
self.executor = self.symbol.bind(self.ctx, self.arg_params, args_grad=self.grad_params,
grad_req=grad_req, aux_states=self.aux_params)
update_dict = self.update_params
# epoch_end_callback = mx.callback.do_checkpoint(self.prefix)
# begin training
cost = []
accus = []
nbatch = 0
for epoch in range(0, self.num_epoch):
train_data.reset()
num_batches = train_data.num_batches
for databatch in train_data:
nbatch += 1
for k, v in databatch.data.items():
# print k, v.shape
self.arg_params[k][:] = mx.nd.array(v, self.ctx)
for k, v in databatch.label.items():
# print k, v
self.arg_params[k][:] = mx.nd.array(v, self.ctx)
output_dict = {name: nd for name, nd in zip(self.symbol.list_outputs(), self.executor.outputs)}
self.executor.forward(is_train=True)
self.executor.backward()
# print '--------------------'
for key in update_dict:
arr = self.grad_params[key]
self.bigbatch_grads[key][:] = self.bigbatch_grads[key] + arr
if nbatch%self.falsebigbatch==0:
# print nbatch, self.bigbatchnum
for key in update_dict:
arr = self.grad_params[key]
arr[:] = self.bigbatch_grads[key] / self.falsebigbatch
self.updater(key, arr, self.arg_params[key])
self.bigbatch_grads[key][:] = 0
outval = output_dict['cls_output'].asnumpy()
label = databatch.label['label']
cls_predict = np.argmax(outval, axis=1)
accone = np.mean(cls_predict!=label)
accus.append(accone)
# print label
# print outval
costone = []
for oi, ov in enumerate(outval):
# print np.min(ov), np.max(ov)
costtmp = 1.0 - ov[label[oi]]
costone.append(costtmp)
# print costone
outval = np.mean(costone)
cost.append(outval)
lrsch = self.optimizer.lr_scheduler
step = lrsch.step
nowlr = lrsch.base_lr
num_update = nbatch
if num_update % showperiod == 0:
print num_update, 'cost:', np.mean(cost), 'lr:', nowlr, num_batches, 'errate_accu:', np.mean(accus)
cost = []
accus = []
savefunc(self.prefix, epoch%10, self.symbol, self.update_params, self.aux_params)
## print databatch.label['label'].T