-
Notifications
You must be signed in to change notification settings - Fork 26
/
adding.py
184 lines (136 loc) · 6.36 KB
/
adding.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
from __future__ import print_function
import sys
import argparse
import os
import time
import lasagne
import theano
import numpy as np
import theano.tensor as T
from lasagne.layers import InputLayer,ReshapeLayer,DimshuffleLayer,Gate
from lasagne.layers import Conv2DLayer
from lasagne.layers import Pool2DLayer
from lasagne.layers import DenseLayer
from lasagne.layers import ConcatLayer,NonlinearityLayer,DropoutLayer
from lasagne.layers.normalization import batch_norm
from lasagne.nonlinearities import softmax, rectify,tanh
from lasagne.layers import RecurrentLayer,LSTMLayer
from IndRNN import IndRNNLayer as indrnn
parser = argparse.ArgumentParser(description='IndRNN solving the adding problem')
parser.add_argument('--model', type=str, default='indrnn', help='models')
parser.add_argument('--hidden_units', type=int, default=128, help='humber of hidden units per layer')
parser.add_argument('--batch_size', type=int, default=32)
parser.add_argument('--seq_len', type=int, default=500)
parser.add_argument('--MAG', type=int, default=2)
parser.add_argument('--act', type=str, default='relu')
parser.add_argument('--lr', type=np.float32, default=2e-4, help='lr')
parser.add_argument('--in2hidW', type=np.float32, default=0.001)
parser.add_argument('--gradclipvalue', type=np.float32, default=10)
args = parser.parse_args()
print(args)
batch_size = args.batch_size
seq_len=args.seq_len
hidden_units=args.hidden_units
feature_size=2
outputclass=1
U_bound=pow(args.MAG, 1.0 / seq_len)
U_lowbound=pow(1.0/args.MAG, 1.0 / seq_len)
if args.act=='tanh':
U_bound=pow(args.MAG/(pow(0.9,seq_len/10.0)), 1.0 / seq_len)
act=rectify
if args.act=='tanh':
act=tanh
lr=args.lr
opti=lasagne.updates.adam#rmsprop
def build_rnn_network(rnnmodel):
net = {}
net['input'] = InputLayer((batch_size, seq_len, feature_size))
if rnnmodel==LSTMLayer:
net['rnn']=rnnmodel(net['input'],hidden_units,forgetgate=lasagne.layers.Gate(b=lasagne.init.Constant(1.)),only_return_final=True,grad_clipping=args.gradclipvalue )
elif act==rectify:
net['rnn']=rnnmodel(net['input'],hidden_units,W_in_to_hid=lasagne.init.Normal(args.in2hidW), W_hid_to_hid=lambda shape: np.identity(hidden_units,dtype=np.float32),nonlinearity=act,only_return_final=True,grad_clipping=args.gradclipvalue )
elif act==tanh:
net['rnn']=rnnmodel(net['input'],hidden_units,W_in_to_hid=lasagne.init.Normal(args.in2hidW), nonlinearity=act,only_return_final=True,grad_clipping=args.gradclipvalue )
net['out']=DenseLayer(net['rnn'],outputclass,nonlinearity=None)
print (lasagne.layers.get_output_shape(net['out']))
return net
def build_indrnn_network(res_rnnmodel):
net = {}
net['input'] = InputLayer((batch_size, seq_len, feature_size))
if act==rectify:
net['rnn0']=res_rnnmodel(net['input'],hidden_units,W_in_to_hid=lasagne.init.Normal(args.in2hidW), nonlinearity=act,W_hid_to_hid=lasagne.init.Uniform(range=(0,U_bound)),grad_clipping=args.gradclipvalue)
net['rnn']=res_rnnmodel(net['rnn0'],hidden_units,W_in_to_hid=lasagne.init.Normal(args.in2hidW), nonlinearity=act,W_hid_to_hid=lasagne.init.Uniform(range=(U_lowbound,U_bound)),only_return_final=True,grad_clipping=args.gradclipvalue)
elif act==tanh:
net['rnn0']=res_rnnmodel(net['input'],hidden_units,W_in_to_hid=lasagne.init.Normal(args.in2hidW), nonlinearity=act,W_hid_to_hid=lasagne.init.Uniform(range=(U_bound)),grad_clipping=args.gradclipvalue)
net['rnn']=res_rnnmodel(net['rnn0'],hidden_units,W_in_to_hid=lasagne.init.Normal(args.in2hidW), nonlinearity=act,W_hid_to_hid=lasagne.init.Uniform(range=(U_bound)),only_return_final=True,grad_clipping=args.gradclipvalue)
net['out']=DenseLayer(net['rnn'],outputclass,nonlinearity=None)
return net
def generate_data(time_steps, n_data):
x = np.asarray(np.zeros((time_steps, int(n_data), 2)),
dtype=theano.config.floatX)
x[:,:,0] = np.asarray(np.random.uniform(low=0.,
high=1.,
size=(time_steps, n_data)),
dtype=theano.config.floatX)
inds = np.asarray(np.random.randint(time_steps//2, size=(n_data, 2)))
inds[:, 1] += time_steps//2
for i in range(int(n_data)):
x[inds[i, 0], i, 1] = 1.0
x[inds[i, 1], i, 1] = 1.0
y = (x[:,:,0] * x[:,:,1]).sum(axis=0)
y = np.reshape(y, (n_data, 1))
x=np.transpose(x, (1, 0, 2))
return x, y
if args.model=='rnn':
learn_net=build_rnn_network(RecurrentLayer)
elif args.model=='lstm':
learn_net=build_rnn_network(LSTMLayer)
elif args.model=='indrnn':
learn_net=build_indrnn_network(indrnn)
X_sym = T.tensor3('inputs',dtype=theano.config.floatX)
y_sym = T.matrix('label',dtype=theano.config.floatX)
prediction = lasagne.layers.get_output(learn_net['out'], X_sym,deterministic=False)
loss = lasagne.objectives.squared_error(prediction, y_sym).mean()
params = lasagne.layers.get_all_params(learn_net['out'], trainable=True)
learning_ratetrain = T.scalar(name='learning_ratetrain',dtype=theano.config.floatX)
grads = theano.grad(loss, params)
#grads = [T.clip(g, -1, 1) for g in grads]
updates = opti( grads, params, learning_rate=learning_ratetrain)#nesterov_momentum( loss, params, learning_rate=learning_ratetrain)#
print('Compiling')
train_fn = theano.function([X_sym, y_sym,learning_ratetrain], [loss,], updates=updates)
test_fn = theano.function([X_sym, y_sym], [loss,])
learning_rate=np.float32(lr)
tmse=0
lastmse=100
count=0
for batchi in range(1,10000000):
x,y=generate_data(seq_len, batch_size)
if args.model=='indrnn':
i=0
for para in params:
if para.name=='hidden_to_hidden.W':
para.set_value(np.clip(para.get_value(),-1*U_bound,U_bound))
i+=1
mse,=train_fn(x, y,learning_rate)
if np.isnan(mse):
print ('NaN detected in cost')
assert(2==3)
if np.isinf(mse):
print ('INF detected in cost')
assert(2==3)
tmse+=mse
if batchi%100==0:
print ('training', tmse/100.0)
count+=1
x,y=generate_data(seq_len, 1000)
mse,=test_fn(x, y)
print ('accuracy:', mse)
if (count>200):
learning_rate=np.float32(learning_rate*0.1)
print ('learning rate',learning_rate)
count=0
if learning_rate<1e-6:
break
tmse=0
save_name=args.model+str(seq_len)
np.savez(save_name, *lasagne.layers.get_all_param_values(learn_net['out']))