/
train.py
443 lines (351 loc) · 15.2 KB
/
train.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
# pylint: disable=W0221, C, R, W1202, E1101, E1102, W0401, W0614
import torch
import os
from shutil import copyfile
import argparse
import logging
from itertools import count
import copy
import numpy as np
import time_logging
from functions import *
import collections
def parse():
parser = argparse.ArgumentParser()
parser.add_argument("--device", type=str)
parser.add_argument("--log_dir", type=str, required=True)
parser.add_argument("--init_seed", type=int)
parser.add_argument("--data_seed", type=int)
parser.add_argument("--skip", type=to_bool, default="True")
parser.add_argument("--dataset", required=True)
parser.add_argument("--dim", type=int, required=True)
parser.add_argument("--p", type=parse_kmg, required=True)
parser.add_argument("--width", type=float, required=True)
parser.add_argument("--depth", type=int, required=True)
parser.add_argument("--rep", type=int, default=0)
parser.add_argument("--init", choices={"orth", "normal"}, default="orth", required=True)
parser.add_argument("--init_gain", type=float, default=1)
parser.add_argument("--optimizer", required=True)
parser.add_argument("--lr_width_exponent", type=float, default=0)
parser.add_argument("--n_steps_max", type=parse_kmg, required=True)
parser.add_argument("--compute_hessian", type=to_bool, default="False")
parser.add_argument("--compute_neff", type=to_bool, default="False")
parser.add_argument("--compute_activities", type=to_bool, default="False")
parser.add_argument("--compute_input_gradients", type=to_bool, default="False")
parser.add_argument("--compute_outputs", type=to_bool, default="False")
parser.add_argument("--subtract_init", type=to_bool, default="False")
parser.add_argument("--save_hessian", type=to_bool, default="False")
parser.add_argument("--checkpoints", type=int, nargs='+', default=[])
parser.add_argument("--nd_stop", type=int, default=0)
parser.add_argument("--losspp_stop", type=float, default=0)
parser.add_argument("--kappa", type=float, default=1)
parser.add_argument("--max_learning_rate", type=float)
parser.add_argument("--learning_rate", type=float)
parser.add_argument("--momentum", type=float, default=0.0)
parser.add_argument("--eps", type=float)
parser.add_argument("--train_last", type=to_bool, default="False")
parser.add_argument("--dropout", type=to_bool, default="False")
parser.add_argument("--batch_size", type=int)
parser.add_argument("--n_steps_bs_grow", type=int)
parser.add_argument("--bs_grow_factor", type=float)
parser.add_argument("--precision", choices={"f32", "f64"}, default="f64")
parser.add_argument("--chunk", type=int)
parser.add_argument("--activation", choices={"relu", "tanh"}, default="relu")
args = parser.parse_args()
if args.device is None:
args.device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
if args.optimizer == "sgd":
if args.batch_size is None:
args.batch_size = args.p
if args.optimizer == "adam":
if args.eps is None:
args.eps = 1e-8
if args.max_learning_rate is None:
args.max_learning_rate = 1e-4
if args.learning_rate is None:
args.learning_rate = 1e-4
if args.batch_size is None:
args.batch_size = args.p
if args.chunk is None:
args.chunk = args.p
return args
def init(args):
try:
os.mkdir(args.log_dir)
except FileExistsError:
pass
desc = {
"p": args.p,
"dim": args.dim,
"depth": args.depth,
"width": args.width,
"kappa": args.kappa,
"rep": args.rep,
}
if args.skip and desc in load_dir_desc2(args.log_dir):
print("{} skiped".format(repr(desc)))
return None
torch.backends.cudnn.benchmark = True
device = torch.device(args.device)
dtype = torch.float32 if args.precision == "f32" else torch.float64
torch.set_default_dtype(dtype)
logger = logging.getLogger("default")
logger.setLevel(logging.DEBUG)
logger.handlers = []
ch = logging.StreamHandler()
logger.addHandler(ch)
with FSLocker(os.path.join(args.log_dir, "output.pkl.lock")):
for i in count():
path_log = os.path.join(args.log_dir, "log_{:04d}".format(i))
if not os.path.isfile(path_log):
run_id = i
fh = logging.FileHandler(path_log)
break
logger.addHandler(fh)
copyfile(__file__, os.path.join(args.log_dir, "script_{:04d}.py".format(run_id)))
logger.info("%s", repr(args))
logger.info(desc)
init_seed = torch.randint(2 ** 62, (), dtype=torch.long).item() if args.init_seed is None else args.init_seed
data_seed = torch.randint(2 ** 62, (), dtype=torch.long).item() if args.data_seed is None else args.data_seed
trainset, testset = get_dataset(args.dataset, args.p, args.dim, data_seed)
trainset = (trainset[0].type(dtype).to(device), trainset[1].type(dtype).to(device))
if testset is not None:
testset = (testset[0].type(dtype).to(device), testset[1].type(dtype).to(device))
_x, y = trainset
n_classes = 1 if y.ndimension() == 1 else y.size(1)
torch.manual_seed(init_seed)
activation = F.relu if args.activation == "relu" else torch.tanh
trainset = (trainset[0].flatten(1), trainset[1])
if testset is not None:
testset = (testset[0].flatten(1), testset[1])
model = FC(args.dim, args.width, args.depth, activation, kappa=args.kappa, n_classes=n_classes, dropout=args.dropout)
for n, p in model.named_parameters():
if 'bias' in n:
nn.init.zeros_(p)
if 'weight' in n:
if args.init == "orth":
orthogonal_(p, gain=args.init_gain)
elif args.init == "normal":
nn.init.normal_(p, std=args.init_gain / p.size(1) ** 0.5)
else:
raise ValueError()
model.to(device)
model.type(dtype)
logger.info("N={}".format(model.N))
if args.train_last:
parameters = model.layers[-1].parameters()
else:
parameters = model.parameters()
if args.subtract_init:
f = model
f0 = copy.deepcopy(f)
for p in f0.parameters():
p.requires_grad = False
model = SumModules([f, f0], [1, -1])
model.kappa = f.kappa
model.preactivations = f.preactivations
model.act = f.act
model.N = f.N
scheduler = None
learning_rate = min(args.learning_rate * args.width ** args.lr_width_exponent, args.max_learning_rate)
logger.info("learning rate = {}".format(learning_rate))
if args.optimizer == "sgd":
optimizer = torch.optim.SGD(parameters, lr=learning_rate, momentum=args.momentum, weight_decay=0)
if args.optimizer == "adam":
optimizer = torch.optim.Adam(parameters, lr=learning_rate, eps=args.eps)
if args.optimizer == "fire":
from fire import FIRE
optimizer = FIRE(parameters, dt_max=learning_rate, a_start=1 - args.momentum)
return model, trainset, testset, logger, optimizer, scheduler, device, desc, init_seed, data_seed, run_id
def train(args, model, trainset, testset, logger, optimizer, scheduler, device, desc, init_seed, data_seed, run_id):
measure_points = set(intlogspace(1, args.n_steps_max, 150, with_zero=True, with_end=True))
dynamics = []
checkpoints = []
time_0 = time_1 = time_logging.start()
batch_size = args.batch_size
bins = np.logspace(-9, 4, 130)
bins = np.concatenate([[-1], bins])
init_state = copy.deepcopy(model.state_dict())
if args.compute_activities:
init_act = get_activities(model, trainset[0], 1024)
step = 0
while True:
if step > args.n_steps_max:
break
if step in measure_points or step % 1000 == 0:
data = {}
dynamics.append(data)
data['state'] = {
"norm": collections.OrderedDict([(n, p.norm().item()) for n, p in model.named_parameters()]),
"displacement": collections.OrderedDict([(n, (p - init_state[n]).norm().item()) for n, p in model.named_parameters()]),
}
data['outnorm'] = {
"train": get_outputs(model, trainset[0], 1024).pow(2).mean().item(),
"test": get_outputs(model, testset[0], 1024).pow(2).mean().item() if testset is not None else None,
}
data['step'] = step
data['train'] = error_loss_grad(model, *trainset)
logger.info("id={} P={} d={} L={} h={} step={} nd={:d} nd/P={:.1f}% Loss={:.2g} |Grad|={:.2g} |w-w0|={:.2g} |w|={:.2g}".format(
run_id,
desc['p'],
desc['dim'],
desc['depth'],
desc['width'],
step,
data['train'][0],
100 * data['train'][0] / args.p,
data['train'][1],
data['train'][2],
sum(p ** 2 for n, p in data['state']['displacement'].items()) ** 0.5,
sum(p ** 2 for n, p in data['state']['norm'].items()) ** 0.5,
)
)
if testset is not None:
data['test'] = error_loss_grad(model, *testset)
if args.compute_activities:
acti = get_activities(model, trainset[0], 1024)
data['activities'] = {
"continuous": [(a - a0).norm().div(a0.norm()).item() for a, a0 in zip(acti, init_act)],
"binary": [((a > 0) != (a0 > 0)).long().sum().item() for a, a0 in zip(acti, init_act)],
}
data['batch_size'] = batch_size
data['optimizer'] = {
'state': simplify(optimizer.state),
'param_groups': simplify(optimizer.param_groups),
}
with torch.no_grad():
deltas = get_deltas(model, *trainset, 1024)
h_pos = None
if data['train'][1] > 0:
x = deltas.clone()
x[x < 0] = 0
x = x / data['train'][1] ** 0.5
h_pos, _ = np.histogram(x.detach().cpu().numpy(), bins, density=True)
x = -deltas.clone()
x[x < 0] = 0
h_neg, _ = np.histogram(x.detach().cpu().numpy(), bins, density=True)
data['deltas'] = {
'bins': bins,
'positive': h_pos,
'negative': h_neg,
}
time_1 = time_logging.end("error and loss", time_1)
if data['train'][0] <= args.nd_stop: # with the hinge, no errors => finished
break
if data['train'][1] * args.p < args.losspp_stop * data['train'][0]:
break
if step in args.checkpoints:
logger.info("({}|{}) checkpoint".format(run_id, desc['p']))
hessian = None
if 8 * model.N**2 < 2e9 and args.compute_hessian:
logger.info("({}|{}) compute the hessian".format(run_id, desc['p']))
hess1, hess2, e, e1, e2 = compute_hessian_evalues(model, *trainset)
hessian = {
"hess_eval": e.cpu(),
"hess1_eval": e1.cpu(),
"hess2_eval": e2.cpu(),
}
if args.save_hessian:
hessian["hess1"] = hess1.cpu()
hessian["hess2"] = hess2.cpu()
time_1 = time_logging.end("hessian", time_1)
with torch.no_grad():
deltas = get_deltas(model, *trainset, 1024)
error_loss = error_loss_grad(model, *trainset)
checkpoints.append({
"step": step,
"train": error_loss,
"state": copy.deepcopy(model.cpu().state_dict()),
"deltas": deltas.cpu(),
"hessian": hessian,
})
model.to(device)
if args.n_steps_bs_grow and step > 0 and step % args.n_steps_bs_grow == 0:
batch_size = min(args.p, int(batch_size * args.bs_grow_factor))
logger.info("({}|{}) batch size set to {}".format(run_id, desc['p'], batch_size))
time_1 = time_logging.end("load data", time_1)
make_a_step(model, optimizer, *trainset, batch_size)
time_1 = time_logging.end("make a step", time_1)
step += 1
del optimizer
run = {
"id": run_id,
"desc": desc,
"args": args,
"init_seed": init_seed,
"data_seed": data_seed,
"N": model.N,
"dynamics": dynamics,
"checkpoints": checkpoints,
}
error_loss = error_loss_grad(model, *trainset)
with torch.no_grad():
deltas = get_deltas(model, *trainset, 1024)
run["init"] = {
"state": collections.OrderedDict([(n, p.cpu()) for n, p in init_state.items()]),
}
grads = None
if args.compute_input_gradients:
logger.info("({}|{}) compute input gradients".format(run_id, desc['p']))
grads = {
"train": get_gradients(model, trainset[0]).cpu(),
"test": get_gradients(model, testset[0]).cpu() if testset is not None else None,
}
time_1 = time_logging.end("input gradients", time_1)
outputs = None
if args.compute_outputs:
outputs = {
"train": get_outputs(model, trainset[0], 1024).cpu(),
"test": get_outputs(model, testset[0], 1024).cpu() if testset is not None else None,
}
run["last"] = {
"train": error_loss,
"state": None,
"deltas": deltas.cpu(),
"hessian": None,
"Neff": None,
"grads": grads,
"outputs": outputs,
}
if 8 * model.N**2 < 1e9 and args.compute_neff:
try:
run['last']['Neff'] = n_effective(model, trainset[0], n_derive=1)
except RuntimeError:
pass
if testset is not None:
run['last']['test'] = error_loss_grad(model, *testset)
with torch.no_grad():
deltas_test = get_deltas(model, *testset, 1024)
run['last']["deltas_test"] = deltas_test.cpu(),
run['p_test'] = len(testset[0])
if 8 * model.N**2 < 2e9 and args.compute_hessian:
try:
logger.info("({}|{}) compute the hessian".format(run_id, desc['p']))
hess1, hess2, e, e1, e2 = compute_hessian_evalues(model, *trainset)
hessian = {
"hess_eval": e.cpu(),
"hess1_eval": e1.cpu(), # H0
"hess2_eval": e2.cpu(), # Hp
}
if args.save_hessian:
hessian["hess1"] = hess1.cpu() # H0
hessian["hess2"] = hess2.cpu() # Hp
run["last"]["hessian"] = hessian
del hess1, hess2, e, e1, e2
time_1 = time_logging.end("hessian", time_1)
except RuntimeError:
pass
run["last"]["state"] = model.cpu().state_dict()
dump_run2(args.log_dir, run)
time_logging.end("total", time_0)
logger.info(time_logging.text_statistics())
def main():
args = parse()
objs = init(args)
if objs is None:
return
time_0 = time_logging.start()
train(args, *objs)
time_logging.end("run", time_0)
if __name__ == '__main__':
main()