/
kfac_adam.py
executable file
·307 lines (245 loc) · 9.01 KB
/
kfac_adam.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
#!/usr/bin/env python
import argparse
import os
import sys
import time
use_kfac = True
parser = argparse.ArgumentParser()
parser.add_argument('-m', '--mode', type=str, default='run', help='record to record test data, test to perform test, run to run training for longer')
args = parser.parse_args()
LR=0.02
LAMBDA=1e-1
use_tikhonov=False
if args.mode == 'record' or args.mode == 'test':
num_steps = 10
else:
num_steps = 100
use_fixed_labels = True
hack_global_init_dict = {}
prefix="kfac_adam"
# Test implementation of KFAC on MNIST
import load_MNIST
import util as u
import util
from util import t # transpose
import kfac as kfac_lib
from kfac import Model
from kfac import Kfac
from kfac import IndexedGrad
import kfac
import sys
import tensorflow as tf
import numpy as np
# TODO: get rid of this
purely_linear = False # convert sigmoids into linear nonlinearities
purely_relu = False # convert sigmoids into ReLUs
regularized_svd = True # kfac_lib.regularized_svd # TODO: delete this
# TODO: get rid
def W_uniform(s1, s2): # uniform weight init from Ng UFLDL
r = np.sqrt(6) / np.sqrt(s1 + s2 + 1)
result = np.random.random(2*s2*s1)*2*r-r
return result
def ng_init(rows, cols):
# creates uniform initializer using Ng's formula
# TODO: turn into TF
r = np.sqrt(6) / np.sqrt(rows + cols + 1)
result = np.random.random(rows*cols)*2*r-r
return result.reshape((rows, cols))
def model_creator(batch_size, name='defaultmodel', dtype=np.float32):
"""Create MNIST autoencoder model. Dataset is part of model."""
global hack_global_init_dict
model = Model(name)
# TODO: actually use batch_size
init_dict = {} # todo: rename to feed_dict?
global_vars = []
local_vars = []
# TODO: rename to make_var
def init_var(val, name, is_global=False):
"""Helper to create variables with numpy or TF initial values."""
if isinstance(val, tf.Tensor):
var = u.get_variable(name=name, initializer=val, reuse=is_global)
else:
val = np.array(val)
assert u.is_numeric(val), "Non-numeric type."
var_struct = u.get_var(name=name, initializer=val, reuse=is_global)
holder = var_struct.val_
init_dict[holder] = val
var = var_struct.var
if is_global:
global_vars.append(var)
else:
local_vars.append(var)
return var
# TODO: get rid of purely_relu
def nonlin(x):
if purely_relu:
return tf.nn.relu(x)
elif purely_linear:
return tf.identity(x)
else:
return tf.sigmoid(x)
# TODO: rename into "nonlin_d"
def d_nonlin(y):
if purely_relu:
return u.relu_mask(y)
elif purely_linear:
return 1
else:
return y*(1-y)
train_images = load_MNIST.load_MNIST_images('data/train-images-idx3-ubyte').astype(dtype)
patches = train_images[:,:batch_size];
fs = [batch_size, 28*28, 196, 28*28]
def f(i): return fs[i+1] # W[i] has shape f[i] x f[i-1]
n = len(fs) - 2
X = init_var(patches, "X", is_global=False)
W = [None]*n
W.insert(0, X)
A = [None]*(n+2)
A[1] = W[0]
W0f_old = W_uniform(fs[2],fs[3]).astype(dtype) # to match previous generation
W0s_old = u.unflatten(W0f_old, fs[1:]) # perftodo: this creates transposes
for i in range(1, n+1):
# temp = init_var(ng_init(f(i), f(i-1)), "W_%d"%(i,), is_global=True)
# init_val1 = W0s_old[i-1]
init_val = ng_init(f(i), f(i-1)).astype(dtype)
W[i] = init_var(init_val, "W_%d"%(i,),
is_global=True)
A[i+1] = nonlin(kfac_lib.matmul(W[i], A[i]))
err = A[n+1] - A[1]
# manually compute backprop to use for sanity checking
B = [None]*(n+1)
B2 = [None]*(n+1)
B[n] = err*d_nonlin(A[n+1])
_sampled_labels_live = tf.random_normal((f(n), f(-1)), dtype=dtype, seed=0)
if use_fixed_labels:
_sampled_labels_live = tf.ones(shape=(f(n), f(-1)), dtype=dtype)
_sampled_labels = init_var(_sampled_labels_live, "to_be_deleted",
is_global=False)
B2[n] = _sampled_labels*d_nonlin(A[n+1])
for i in range(n-1, -1, -1):
backprop = t(W[i+1]) @ B[i+1]
B[i] = backprop*d_nonlin(A[i+1])
backprop2 = t(W[i+1]) @ B2[i+1]
B2[i] = backprop2*d_nonlin(A[i+1])
cov_A = [None]*(n+1) # covariance of activations[i]
cov_B2 = [None]*(n+1) # covariance of synthetic backprops[i]
vars_svd_A = [None]*(n+1)
vars_svd_B2 = [None]*(n+1)
dW = [None]*(n+1)
dW2 = [None]*(n+1)
pre_dW = [None]*(n+1) # preconditioned dW
for i in range(1,n+1):
if regularized_svd:
cov_A[i] = init_var(A[i]@t(A[i])/batch_size+LAMBDA*u.Identity(f(i-1)), "cov_A%d"%(i,))
cov_B2[i] = init_var(B2[i]@t(B2[i])/batch_size+LAMBDA*u.Identity(f(i)), "cov_B2%d"%(i,))
else:
cov_A[i] = init_var(A[i]@t(A[i])/batch_size, "cov_A%d"%(i,))
cov_B2[i] = init_var(B2[i]@t(B2[i])/batch_size, "cov_B2%d"%(i,))
vars_svd_A[i] = u.SvdWrapper(cov_A[i],"svd_A_%d"%(i,))
vars_svd_B2[i] = u.SvdWrapper(cov_B2[i],"svd_B2_%d"%(i,))
if use_tikhonov:
whitened_A = u.regularized_inverse3(vars_svd_A[i],L=LAMBDA) @ A[i]
whitened_B2 = u.regularized_inverse3(vars_svd_B2[i],L=LAMBDA) @ B[i]
else:
whitened_A = u.pseudo_inverse2(vars_svd_A[i]) @ A[i]
whitened_B2 = u.pseudo_inverse2(vars_svd_B2[i]) @ B[i]
dW[i] = (B[i] @ t(A[i]))/batch_size
dW2[i] = B[i] @ t(A[i])
pre_dW[i] = (whitened_B2 @ t(whitened_A))/batch_size
# model.extra['A'] = A
# model.extra['B'] = B
# model.extra['B2'] = B2
# model.extra['cov_A'] = cov_A
# model.extra['cov_B2'] = cov_B2
# model.extra['vars_svd_A'] = vars_svd_A
# model.extra['vars_svd_B2'] = vars_svd_B2
# model.extra['W'] = W
# model.extra['dW'] = dW
# model.extra['dW2'] = dW2
# model.extra['pre_dW'] = pre_dW
model.loss = u.L2(err) / (2 * batch_size)
sampled_labels_live = A[n+1] + tf.random_normal((f(n), f(-1)),
dtype=dtype, seed=0)
if use_fixed_labels:
sampled_labels_live = A[n+1]+tf.ones(shape=(f(n), f(-1)), dtype=dtype)
sampled_labels = init_var(sampled_labels_live, "sampled_labels", is_global=False)
err2 = A[n+1] - sampled_labels
model.loss2 = u.L2(err2) / (2 * batch_size)
model.global_vars = global_vars
model.local_vars = local_vars
model.trainable_vars = W[1:]
def advance_batch():
sess = tf.get_default_session()
# TODO: get rid of _sampled_labels
sess.run([sampled_labels.initializer, _sampled_labels.initializer])
model.advance_batch = advance_batch
global_init_op = tf.group(*[v.initializer for v in global_vars])
def initialize_global_vars():
sess = tf.get_default_session()
sess.run(global_init_op, feed_dict=init_dict)
model.initialize_global_vars = initialize_global_vars
local_init_op = tf.group(*[v.initializer for v in local_vars])
def initialize_local_vars():
sess = tf.get_default_session()
sess.run(X.initializer, feed_dict=init_dict) # A's depend on X
sess.run(_sampled_labels.initializer, feed_dict=init_dict)
sess.run(local_init_op, feed_dict=init_dict)
model.initialize_local_vars = initialize_local_vars
hack_global_init_dict = init_dict
return model
if __name__ == '__main__':
np.random.seed(0)
tf.set_random_seed(0)
dsize = 1000
sess = tf.InteractiveSession()
model = model_creator(dsize) # TODO: share dataset between models?
model.initialize_global_vars()
model.initialize_local_vars()
kfac = Kfac(model_creator, dsize) # creates another copy of model, initializes
kfac.model.initialize_local_vars()
kfac.reset() # resets optimization variables (not model variables)
kfac.lr.set(LR)
kfac.Lambda.set(LAMBDA)
with u.capture_vars() as opt_vars:
if use_kfac:
opt = tf.train.AdamOptimizer(0.1)
else:
opt = tf.train.AdamOptimizer()
grads_and_vars = opt.compute_gradients(model.loss,
var_list=model.trainable_vars)
grad = IndexedGrad.from_grads_and_vars(grads_and_vars)
grad_new = kfac.correct(grad)
# grad_new = kfac.correct_normalized(grad)
train_op = opt.apply_gradients(grad_new.to_grads_and_vars())
[v.initializer.run() for v in opt_vars]
losses = []
u.record_time()
start_time = time.time()
for step in range(num_steps):
loss0 = model.loss.eval()
losses.append(loss0)
elapsed = time.time()-start_time
print("%d sec, step %d, loss %.2f" %(elapsed, step, loss0))
if use_kfac:
kfac.model.advance_batch()
kfac.update_stats()
model.advance_batch()
grad.update()
grad_new.update()
train_op.run()
u.record_time()
u.summarize_time()
losses_fn = '%s_losses_test.csv' %(prefix,)
if args.mode == 'record':
if os.path.exists('data/'+losses_fn):
answer = input("%s exists, overwrite? (Y/n) "%(losses_fn,))
if not answer:
answer = "y"
if answer.lower() != "y":
print("Exiting")
sys.exit()
u.dump(losses, losses_fn)
elif args.mode == 'test':
targets = np.loadtxt("data/"+losses_fn, delimiter=",")
u.check_equal(losses, targets, rtol=1e-2)
u.summarize_difference(losses, targets)