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
|
#!/usr/bin/env python
import logging
import numpy
import sys
import importlib
from contextlib import closing
import theano
from theano import tensor
from theano.tensor.shared_randomstreams import RandomStreams
from blocks.serialization import load_parameter_values, secure_dump, BRICK_DELIMITER
from blocks.extensions import Printing, SimpleExtension
from blocks.extensions.monitoring import DataStreamMonitoring, TrainingDataMonitoring
from blocks.extras.extensions.plot import Plot
from blocks.extensions.saveload import Checkpoint, Load
from blocks.graph import ComputationGraph
from blocks.main_loop import MainLoop
from blocks.model import Model
from blocks.algorithms import GradientDescent
import datastream
# from apply_model import Apply
logging.basicConfig(level='INFO')
logger = logging.getLogger(__name__)
sys.setrecursionlimit(1500)
if __name__ == "__main__":
if len(sys.argv) != 2:
print >> sys.stderr, 'Usage: %s config' % sys.argv[0]
sys.exit(1)
model_name = sys.argv[1]
config = importlib.import_module('%s' % model_name)
class GenText(SimpleExtension):
def __init__(self, model, init_text, max_bytes, **kwargs):
super(GenText, self).__init__(**kwargs)
self.init_text = init_text
self.max_bytes = max_bytes
out = model.out[:, -1, :] / numpy.float32(config.sample_temperature)
prob = tensor.nnet.softmax(out)
cg = ComputationGraph([prob])
assert(len(cg.inputs) == 1)
assert(cg.inputs[0].name == 'bytes')
state_vars = [theano.shared(v[0:1, :].zeros_like().eval(), v.name+'-gen')
for v, _ in model.states]
givens = [(v, x) for (v, _), x in zip(model.states, state_vars)]
updates= [(x, upd) for x, (_, upd) in zip(state_vars, model.states)]
self.f = theano.function(inputs=cg.inputs, outputs=[prob],
givens=givens, updates=updates)
self.reset_states = theano.function(inputs=[], outputs=[],
updates=[(v, v.zeros_like()) for v in state_vars])
def do(self, which_callback, *args):
print "Sample:"
print "-------"
self.reset_states()
v = numpy.array([ord(i) for i in self.init_text],
dtype='int16')[None, :]
prob, = self.f(v)
sys.stdout.write(self.init_text)
while v.shape[1] < self.max_bytes:
prob = prob / 1.00001
pred = numpy.random.multinomial(1, prob[0, :]).nonzero()[0][0]
v = numpy.concatenate([v, pred[None, None]], axis=1)
sys.stdout.write(chr(int(pred)))
sys.stdout.flush()
prob, = self.f(pred[None, None])
print
print "-------"
print
class ResetStates(SimpleExtension):
def __init__(self, state_vars, **kwargs):
super(ResetStates, self).__init__(**kwargs)
self.f = theano.function(
inputs=[], outputs=[],
updates=[(v, v.zeros_like()) for v in state_vars])
def do(self, which_callback, *args):
self.f()
def train_model(m, train_stream, dump_path=None):
# Define the model
model = Model(m.cost)
cg = ComputationGraph(m.cost_reg)
algorithm = GradientDescent(cost=m.cost_reg,
step_rule=config.step_rule,
params=cg.parameters)
algorithm.add_updates(m.states)
# Load the parameters from a dumped model
if dump_path is not None:
try:
with closing(numpy.load(dump_path)) as source:
logger.info('Loading parameters...')
param_values = {'/' + name.replace(BRICK_DELIMITER, '/'): source[name]
for name in source.keys()
if name != 'pkl' and not 'None' in name}
model.set_param_values(param_values)
except IOError:
pass
main_loop = MainLoop(
model=model,
data_stream=train_stream,
algorithm=algorithm,
extensions=[
Checkpoint(path=dump_path,
after_epoch=False,
use_cpickle=True,
every_n_epochs=config.save_freq),
TrainingDataMonitoring(
[m.cost_reg, m.error_rate_reg, m.cost, m.error_rate],
prefix='train', every_n_epochs=1),
Printing(every_n_epochs=1, after_epoch=False),
Plot(document='text_'+model_name+'_'+config.param_desc,
channels=[['train_cost', 'train_cost_reg'],
['train_error_rate', 'train_error_rate_reg']],
server_url='http://eos21:4201/',
every_n_epochs=1, after_epoch=False),
GenText(m, '\nalex\ttu crois ?\n', config.sample_len,
every_n_epochs=config.sample_freq,
after_epoch=False, before_training=True),
ResetStates([v for v, _ in m.states], after_epoch=True)
]
)
main_loop.run()
if __name__ == "__main__":
# Build datastream
train_stream = datastream.setup_datastream('data/logcompil.txt',
config.num_seqs,
config.seq_len,
config.seq_div_size)
# Build model
m = config.Model()
m.cost.name = 'cost'
m.cost_reg.name = 'cost_reg'
m.error_rate.name = 'error_rate'
m.error_rate_reg.name = 'error_rate_reg'
m.pred.name = 'pred'
# Train the model
saveloc = 'model_data/%s-%s' % (model_name, config.param_desc)
train_model(m, train_stream, dump_path=saveloc)
|
