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
|
#!/usr/bin/env python
import importlib
import logging
import operator
import os
import sys
from functools import reduce
from theano import tensor
from blocks import roles
from blocks.algorithms import AdaDelta, CompositeRule, GradientDescent, RemoveNotFinite, StepRule
from blocks.extensions import Printing, FinishAfter
from blocks.extensions.monitoring import DataStreamMonitoring, TrainingDataMonitoring
from blocks.extensions.plot import Plot
from blocks.extensions.saveload import LoadFromDump, Dump
from blocks.dump import MainLoopDumpManager
from blocks.filter import VariableFilter
from blocks.graph import ComputationGraph, apply_dropout, apply_noise
from blocks.main_loop import MainLoop
from blocks.model import Model
logger = logging.getLogger(__name__)
class ElementwiseRemoveNotFinite(StepRule):
"""A step rule that replaces non-finite coefficients by zeros.
Replaces non-finite elements (such as ``inf`` or ``NaN``) in a step
(the parameter update of a single shared variable)
with a scaled version of the parameters being updated instead.
Parameters
----------
scaler : float, optional
The scaling applied to the parameter in case the step contains
non-finite elements. Defaults to 0.1.
Notes
-----
This trick was originally used in the GroundHog_ framework.
.. _GroundHog: https://github.com/lisa-groundhog/GroundHog
"""
def __init__(self, scaler=0.1):
self.scaler = scaler
def compute_step(self, param, previous_step):
not_finite = tensor.isnan(previous_step) + tensor.isinf(previous_step)
step = tensor.switch(not_finite, self.scaler * param, previous_step)
return step, []
class CustomDumpManager(MainLoopDumpManager):
def dump(self, main_loop):
"""Dumps the main loop to the root folder.
See :mod:`blocks.dump`.
Overwrites the old data if present.
"""
if not os.path.exists(self.folder):
os.mkdir(self.folder)
self.dump_parameters(main_loop)
self.dump_log(main_loop)
def load(self):
return (self.load_parameters(), self.load_log())
def load_to(self, main_loop):
"""Loads the dump from the root folder into the main loop."""
parameters, log = self.load()
main_loop.model.set_param_values(parameters)
main_loop.log = log
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, 'config')
logger.info('# Configuration: %s' % config.__name__)
for key in dir(config):
if not key.startswith('__') and isinstance(getattr(config, key), (int, str, list, tuple)):
logger.info(' %20s %s' % (key, str(getattr(config, key))))
model = config.Model(config)
model.initialize()
stream = config.Stream(config)
inputs = stream.inputs()
req_vars = model.cost.inputs
train_stream = stream.train(req_vars)
valid_stream = stream.valid(req_vars)
cost = model.cost(**inputs)
cg = ComputationGraph(cost)
monitored = set([cost] + VariableFilter(roles=[roles.COST])(cg.variables))
if hasattr(config, 'dropout') and config.dropout < 1.0:
cg = apply_dropout(cg, config.dropout_inputs(cg), config.dropout)
if hasattr(config, 'noise') and config.noise > 0.0:
cg = apply_noise(cg, config.noise_inputs(cg), config.noise)
cost = cg.outputs[0]
cg = Model(cost)
logger.info('# Parameter shapes:')
parameters_size = 0
for key, value in cg.get_params().iteritems():
logger.info(' %20s %s' % (value.get_value().shape, key))
parameters_size += reduce(operator.mul, value.get_value().shape, 1)
logger.info('Total number of parameters: %d in %d matrices' % (parameters_size, len(cg.get_params())))
params = cg.parameters
algorithm = GradientDescent(
cost=cost,
step_rule=CompositeRule([
ElementwiseRemoveNotFinite(),
AdaDelta(),
#Momentum(learning_rate=config.learning_rate, momentum=config.momentum),
]),
params=params)
plot_vars = [['valid_' + x.name for x in monitored]]
logger.info('Plotted variables: %s' % str(plot_vars))
dump_path = os.path.join('model_data', model_name)
logger.info('Dump path: %s' % dump_path)
load_dump_ext = LoadFromDump(dump_path)
dump_ext = Dump(dump_path, every_n_batches=1000)
load_dump_ext.manager = CustomDumpManager(dump_path)
dump_ext.manager = CustomDumpManager(dump_path)
extensions=[TrainingDataMonitoring(monitored, prefix='train', every_n_batches=1000),
DataStreamMonitoring(monitored, valid_stream,
prefix='valid',
every_n_batches=1000),
Printing(every_n_batches=1000),
Plot(model_name, channels=plot_vars, every_n_batches=500),
load_dump_ext,
dump_ext
]
main_loop = MainLoop(
model=cg,
data_stream=train_stream,
algorithm=algorithm,
extensions=extensions)
main_loop.run()
main_loop.profile.report()
|
