def get_function(self, func_name):
if func_name == 'tanh':
return T.tanh
elif func_name == 'hardtanh':
L.warning('Current hardTanh implementation is slow!')
return lambda x: ((abs(x) <= 1) * x) + ((1 < abs(x)) * T.sgn(x))
elif func_name == 'xtanh':
return lambda x: T.tanh(x) + 0.1 * x
elif func_name == 'sigmoid':
return T.nnet.sigmoid
elif func_name == 'fastsigmoid':
L.error('T.nnet.ultra_fast_sigmoid function has some problems')
elif func_name == 'hardsigmoid':
return T.nnet.hard_sigmoid
elif func_name == 'xsigmoid':
return lambda x: T.nnet.sigmoid(x) + 0.1 * x
elif func_name == 'softplus':
return T.nnet.softplus
elif func_name == 'relu':
#return lambda x: T.maximum(x, 0)
return lambda x: x * (x > 0)
#return T.nnet.relu # Update theano and then use this one instead
elif func_name == 'leakyrelu':
return lambda x: T.maximum(x, 0.01 * x)
elif func_name == 'cappedrelu':
return lambda x: T.minimum(x * (x > 0), 6)
elif func_name == 'softmax':
return T.nnet.softmax
elif func_name == 'norm1':
return lambda x: x / T.nlinalg.norm(x, 1)
elif func_name == 'norm2':
#return lambda x: x / T.nlinalg.norm(x, 2)
return lambda x: x / T.dot(x, x)**0.5
else:
L.error('Invalid function name given: ' + func_name)
def curr_version():
import dlm.io.logging as L
info_path = os.path.dirname(sys.argv[0]) + '/.git/refs/heads/master'
if os.path.exists(info_path):
with open(info_path, 'r') as info_file:
return info_file.next().strip()
L.warning('Unable to read current version.')
return None
def curr_version():
import dlm.io.logging as L
info_path = os.path.dirname(sys.argv[0]) + '/.git/refs/heads/master'
if os.path.exists(info_path):
with open(info_path, 'r') as info_file:
return info_file.next().strip()
L.warning('Unable to read current version.')
return None
def set_theano_device(device, threads):
import sys
import dlm.io.logging as L
xassert(device == "cpu" or device.startswith("gpu"), "The device can only be 'cpu', 'gpu' or 'gpu<id>'")
if device.startswith("gpu") and len(device) > 3:
try:
gpu_id = int(device[3:])
if not is_gpu_free(gpu_id):
L.warning('The selected GPU (GPU' + str(gpu_id) + ') is apparently busy.')
except ValueError:
L.error("Unknown GPU device format: " + device)
if device.startswith("gpu"):
L.warning('Running on GPU yields non-deterministic results.')
xassert(sys.modules.has_key('theano') == False, "dlm.utils.set_theano_device() function cannot be called after importing theano")
os.environ['OMP_NUM_THREADS'] = str(threads)
os.environ['THEANO_FLAGS'] = 'device=' + device
os.environ['THEANO_FLAGS'] += ',force_device=True'
os.environ['THEANO_FLAGS'] += ',floatX=float32'
os.environ['THEANO_FLAGS'] += ',warn_float64=warn'
os.environ['THEANO_FLAGS'] += ',cast_policy=numpy+floatX'
#os.environ['THEANO_FLAGS'] += ',allow_gc=True'
os.environ['THEANO_FLAGS'] += ',print_active_device=False'
os.environ['THEANO_FLAGS'] += ',exception_verbosity=high' # Highly verbose debugging
os.environ['THEANO_FLAGS'] += ',mode=FAST_RUN'
os.environ['THEANO_FLAGS'] += ',nvcc.fastmath=False' # True: makes div and sqrt faster at the cost of precision, and possible bugs
#os.environ['THEANO_FLAGS'] += ',optimizer_including=cudnn' # Comment out if CUDNN is not available
# change theano to wrapper
try:
#import theano
import backend.nn_wrapper as K
except EnvironmentError:
L.exception()
global logger
#if theano.config.device == "gpu":
# L.info(
# "Device: " + theano.config.device.upper() + " "
# + str(theano.sandbox.cuda.active_device_number())
# + " (" + str(theano.sandbox.cuda.active_device_name()) + ")"
# )
#else:
# L.info("Device: " + theano.config.device.upper())
#global K
try:
K.set_platform('tensorflow') # theano is working
L.info("Creating a variable inside utils")
import numpy as np
val = np.random.random((4, 2))
tmp = K.variable(val)
except:
print >> sys.stderr, "Unexpected error:", sys.exc_info()
raise TypeError("Cannot set the platform")
def set_theano_device(device, threads):
import sys
import dlm.io.logging as L
xassert(device == "cpu" or device.startswith("gpu"),
"The device can only be 'cpu', 'gpu' or 'gpu<id>'")
if device.startswith("gpu") and len(device) > 3:
try:
gpu_id = int(device[3:])
if not is_gpu_free(gpu_id):
L.warning('The selected GPU (GPU' + str(gpu_id) +
') is apparently busy.')
except ValueError:
L.error("Unknown GPU device format: " + device)
if device.startswith("gpu"):
L.warning('Running on GPU yields non-deterministic results.')
xassert(
sys.modules.has_key('theano') == False,
"dlm.utils.set_theano_device() function cannot be called after importing theano"
)
os.environ['OMP_NUM_THREADS'] = str(threads)
os.environ['THEANO_FLAGS'] = 'device=' + device
os.environ['THEANO_FLAGS'] += ',force_device=True'
os.environ['THEANO_FLAGS'] += ',floatX=float32'
os.environ['THEANO_FLAGS'] += ',warn_float64=warn'
os.environ['THEANO_FLAGS'] += ',cast_policy=numpy+floatX'
#os.environ['THEANO_FLAGS'] += ',allow_gc=True'
os.environ['THEANO_FLAGS'] += ',print_active_device=False'
os.environ[
'THEANO_FLAGS'] += ',exception_verbosity=high' # Highly verbose debugging
os.environ['THEANO_FLAGS'] += ',mode=FAST_RUN'
os.environ[
'THEANO_FLAGS'] += ',nvcc.fastmath=False' # True: makes div and sqrt faster at the cost of precision, and possible bugs
#os.environ['THEANO_FLAGS'] += ',optimizer_including=cudnn' # Comment out if CUDNN is not available
try:
import theano
except EnvironmentError:
L.exception()
global logger
if theano.config.device == "gpu":
L.info("Device: " + theano.config.device.upper() + " " +
str(theano.sandbox.cuda.active_device_number()) + " (" +
str(theano.sandbox.cuda.active_device_name()) + ")")
else:
L.info("Device: " + theano.config.device.upper())
def set_theano_device(device, threads):
import sys
import dlm.io.logging as L
xassert(device == "cpu" or device.startswith("gpu"), "The device can only be 'cpu', 'gpu' or 'gpu<id>'")
if device.startswith("gpu") and len(device) > 3:
try:
gpu_id = int(device[3:])
if not is_gpu_free(gpu_id):
L.warning('The selected GPU (GPU' + str(gpu_id) + ') is apparently busy.')
except ValueError:
L.error("Unknown GPU device format: " + device)
if device.startswith("gpu"):
L.warning('Running on GPU yields non-deterministic results.')
xassert(sys.modules.has_key('theano') == False, "dlm.utils.set_theano_device() function cannot be called after importing theano")
os.environ['OMP_NUM_THREADS'] = str(threads)
os.environ['THEANO_FLAGS'] = 'device=' + device
os.environ['THEANO_FLAGS'] += ',force_device=True'
os.environ['THEANO_FLAGS'] += ',floatX=float32'
os.environ['THEANO_FLAGS'] += ',warn_float64=warn'
os.environ['THEANO_FLAGS'] += ',cast_policy=numpy+floatX'
# os.environ['THEANO_FLAGS'] += ',cuda.root=/usr/local/cuda'
#os.environ['THEANO_FLAGS'] += ',allow_gc=True'
os.environ['THEANO_FLAGS'] += ',print_active_device=False'
os.environ['THEANO_FLAGS'] += ',exception_verbosity=high' # Highly verbose debugging
os.environ['THEANO_FLAGS'] += ',mode=FAST_RUN'
os.environ['THEANO_FLAGS'] += ',nvcc.fastmath=False' # True: makes div and sqrt faster at the cost of precision, and possible bugs
#os.environ['THEANO_FLAGS'] += ',optimizer_including=cudnn' # Comment out if CUDNN is not available
try:
import theano
except EnvironmentError:
L.exception()
global logger
if theano.config.device == "gpu":
L.info(
"Device: " + theano.config.device.upper() + " "
+ str(theano.sandbox.cuda.active_device_number())
+ " (" + str(theano.sandbox.cuda.active_device_name()) + ")"
)
else:
L.info("Device: " + theano.config.device.upper())
from dlm.reranker import augmenter
output_nbest_path = args.out_dir + '/augmented.nbest'
if args.no_aug:
shutil.copy(args.input_nbest, output_nbest_path)
else:
augmenter.augment(args.model_path, args.input_nbest, args.vocab_path,
output_nbest_path)
with open(args.weights, 'r') as input_weights:
lines = input_weights.readlines()
if len(lines) > 1:
L.warning(
"Weights file has more than one line. I'll read the 1st and ignore the rest."
)
weights = np.asarray(lines[0].strip().split(" "), dtype=float)
prefix = os.path.basename(args.input_nbest)
input_aug_nbest = NBestList(output_nbest_path, mode='r')
output_nbest = NBestList(args.out_dir + '/' + prefix + '.reranked.nbest',
mode='w')
output_1best = codecs.open(args.out_dir + '/' + prefix + '.reranked.1best',
mode='w',
encoding='UTF-8')
def is_number(s):
try:
float(s)
def set_theano_device(device, threads):
import sys
import dlm.io.logging as L
xassert(device == "cpu" or device.startswith("gpu"),
"The device can only be 'cpu', 'gpu' or 'gpu<id>'")
if device.startswith("gpu") and len(device) > 3:
try:
gpu_id = int(device[3:])
if not is_gpu_free(gpu_id):
L.warning('The selected GPU (GPU' + str(gpu_id) +
') is apparently busy.')
except ValueError:
L.error("Unknown GPU device format: " + device)
if device.startswith("gpu"):
L.warning('Running on GPU yields non-deterministic results.')
xassert(
sys.modules.has_key('theano') == False,
"dlm.utils.set_theano_device() function cannot be called after importing theano"
)
os.environ['OMP_NUM_THREADS'] = str(threads)
os.environ['THEANO_FLAGS'] = 'device=' + device
os.environ['THEANO_FLAGS'] += ',force_device=True'
os.environ['THEANO_FLAGS'] += ',floatX=float32'
os.environ['THEANO_FLAGS'] += ',warn_float64=warn'
os.environ['THEANO_FLAGS'] += ',cast_policy=numpy+floatX'
#os.environ['THEANO_FLAGS'] += ',allow_gc=True'
os.environ['THEANO_FLAGS'] += ',print_active_device=False'
os.environ[
'THEANO_FLAGS'] += ',exception_verbosity=high' # Highly verbose debugging
os.environ['THEANO_FLAGS'] += ',mode=FAST_RUN'
os.environ[
'THEANO_FLAGS'] += ',nvcc.fastmath=False' # True: makes div and sqrt faster at the cost of precision, and possible bugs
#os.environ['THEANO_FLAGS'] += ',optimizer_including=cudnn' # Comment out if CUDNN is not available
# change theano to wrapper
try:
#import theano
import backend.nn_wrapper as K
except EnvironmentError:
L.exception()
global logger
#if theano.config.device == "gpu":
# L.info(
# "Device: " + theano.config.device.upper() + " "
# + str(theano.sandbox.cuda.active_device_number())
# + " (" + str(theano.sandbox.cuda.active_device_name()) + ")"
# )
#else:
# L.info("Device: " + theano.config.device.upper())
#global K
try:
K.set_platform('tensorflow') # theano is working
L.info("Creating a variable inside utils")
import numpy as np
val = np.random.random((4, 2))
tmp = K.variable(val)
except:
print >> sys.stderr, "Unexpected error:", sys.exc_info()
raise TypeError("Cannot set the platform")
U.mkdir_p(args.out_dir)
from dlm.reranker import augmenter
output_nbest_path = args.out_dir + '/augmented.nbest'
if args.no_aug:
shutil.copy(args.input_nbest, output_nbest_path)
else:
augmenter.augment(args.model_path, args.input_nbest, args.vocab_path, output_nbest_path)
with open(args.weights, 'r') as input_weights:
lines = input_weights.readlines()
if len(lines) > 1:
L.warning("Weights file has more than one line. I'll read the 1st and ignore the rest.")
weights = np.asarray(lines[0].strip().split(" "), dtype=float)
prefix = os.path.basename(args.input_nbest)
input_aug_nbest = NBestList(output_nbest_path, mode='r')
output_nbest = NBestList(args.out_dir + '/' + prefix + '.reranked.nbest', mode='w')
output_1best = codecs.open(args.out_dir + '/' + prefix + '.reranked.1best', mode='w', encoding='UTF-8')
def is_number(s):
try:
float(s)
return True
except ValueError:
return False
counter = 0
