def testing(args, data, ttns):
rasengan.warn('NOTE: this function presupposes that the parameters were '
'loaded inside the circuit already')
test_result = args.validate_predictions_f(data.test_lex,
data.idx2label,
args,
ttns.test_f_classify,
data.test_y,
data.words_test,
fn='/current.test.txt')
print('Test F1', test_result['f1'])
return test_result['f1']
def transform_train_features_train_labels(train_features, train_labels):
f_list = []
l_list = []
for f, l in zip(train_features, train_labels):
for idx in l.nonzero()[0]:
f_list.append(f)
l_list.append(idx)
rasengan.warn("We are breaking")
break
f_list = np.array(f_list)
l_list = np.array(l_list)
return (f_list, l_list)
def testing(args, data, ttns):
rasengan.warn('NOTE: this function presupposes that the parameters were '
'loaded inside the circuit already')
test_result = args.validate_predictions_f(
data.test_lex,
data.idx2label,
args,
ttns.test_f_classify,
data.test_y,
data.words_test,
fn='/current.test.txt')
print('Test F1', test_result['f1'])
return test_result['f1']
def make_conjunctive_feat(feat, feat_name):
l = []
n = []
base_feat = feat.shape[1]
assert base_feat == len(feat_name)
if base_feat > 200:
rasengan.warn('Making Conjunctive feature with %d base features' %
base_feat)
for i in range(base_feat):
for j in range(i + 1, base_feat):
# Maximummeans OR
l.append(feat[:, i].maximum(feat[:, j]))
# Minimum means AND
# l.append(feat[:, i].minimum(feat[:, j]))
n.append(feat_name[i] + feat_name[j])
return (scipy.sparse.hstack([feat] + l), feat_name + n)
def main():
import transducer_score
args = transducer_score.args
set_dropout_to_zero(args)
data = transducer_score.data
#--------------------------#
# Compile disparate models #
#--------------------------#
models = []
for pkl_fn, changes in pkl_to_combine:
args_clone = rasengan.Namespace(**args)
#--------------------#
# Update args_clone. #
#--------------------#
rasengan.warn('NOTE: Seting pretrained_param_pklfile')
args_clone.pretrained_param_pklfile = pkl_fn
for (k,v) in changes.items():
setattr(args_clone, k, v)
print 'Setting args_clone.%s=%s'%(k,str(v))
#---------------------#
# Compile args_clone. #
#---------------------#
ttns_i = rasengan.Namespace('ttns').update_and_append_prefix(
compile_args(args_clone), 'test_')
load_params_from_pklfile_to_stack_config(
pkl_fn, ttns_i.test_stack_config)
models.append(ttns_i)
#----------------------------#
# Aggregate disparate model. #
#----------------------------#
ttns = Aggregator(models, data)
#-----------------------------------------------#
# Test performance of Aggregated decision rule. #
#-----------------------------------------------#
with rasengan.debug_support():
stats_valid = args.validate_predictions_f(
data.valid_lex,
data.idx2label,
args,
ttns.test_f_classify,
data.valid_y,
data.words_valid,
fn='/combined.valid.txt')
print 'stats_valid', stats_valid
def evaluate_impl(url_mention, TM, E, cat_folds, cat2url,
performance_aggregator, DF, cat_idx, cat, folds):
print >> sys.stderr, 'progress = %.2f\r' % (float(cat_idx) /
len(cat_folds)),
for (train_idx, test_idx) in folds:
for train_set_size in [1]: # 0.5
S = get(cat2url[cat],
train_idx[:int(len(train_idx) * train_set_size)])
EmS = minus(E, S)
Q = get(cat2url[cat], test_idx)
EmSQ = minus(EmS, Q)
# --------------------- #
# Extract Textual Clues #
# --------------------- #
clue_obj = TextualClueObject(S, url_mention, TM)
# ------------------------------------ #
# Hypothesize Recommendation Criterion #
# ------------------------------------ #
rec_obj = NBRecommender(clue_obj, args.ngram_occurrence)
# ------------------------------- #
# Update Recommendation Criterion #
# ------------------------------- #
updated_rec_obj = FunctionWordRemover(rec_obj,
df_obj=DF,
df_lim=args.df_lim)
updated_rec_obj.report()
# ------------------- #
# Apply The Criterion #
# ------------------- #
scores = {}
rasengan.warn('Restricted Entities to 1000')
for e_idx, e in enumerate(EmSQ[:1000] + Q):
try:
scores[e] = updated_rec_obj(url_mention[e], ename=e)
except KeyError as e:
print >> sys.stderr, e
continue
# ------------------- #
# Measure Performance #
# ------------------- #
performance_aggregator(cat, scores, len(S), Q)
def train_transducer_lbfgs(train_lex,
train_y,
args,
ttns,
training_stats,
batch_size=None):
''' This function completes a training epoch by doing one run of LBFGS.
`ts` abbreviates `train_stack` in entire function
Params
------
train_lex : A list of input_strings (the strings are represented as np arrays)
train_y : A list of output strings
batch_size : UNUSED : (default None)
'''
assert args.clipping_value < 0
assert args.projection_threshold < 0
ts_param_name = [
str(e) for e in ttns.train_stack_config.updatable_parameters()
]
print('The following params will be trained by lbfgs', ts_param_name)
ts_param_shape_list = [
ttns.train_stack_config[name].get_value().shape
for name in ts_param_name
]
ts_param_shape_map = dict(zip(ts_param_name, ts_param_shape_list))
total_param = sum(
numpy.prod(shape) for shape in ts_param_shape_map.values())
def set_entries_in_ttns(param_vec):
''' Set entries in ttns.train_stack_config
with corresponding values in param_vec.
'''
param_vec = param_vec.astype('float32')
offset = 0
for name in ts_param_name:
shape = ts_param_shape_map[name]
numel = numpy.prod(shape)
ttns.train_stack_config[name].set_value(
param_vec[offset:offset + numel].reshape(shape))
offset += numel
pass
return
def vectorize(param_list, dtype='float32'):
param_vec = numpy.zeros((total_param, ), dtype=dtype)
offset = 0
for idx, param in enumerate(param_list):
shape = param.shape
assert shape == ts_param_shape_list[idx]
numel = numpy.prod(shape)
param_vec[offset:offset + numel] = param.reshape(
(numel, )).astype(dtype)
offset += numel
pass
return param_vec
def get_entries_in_ttns():
''' Set entries in ttns.train_stack_config
with corresponding values in param_vec.
'''
return vectorize([
ttns.train_stack_config[name].get_value() for name in ts_param_name
])
def loss_over_corpus(param_vec):
''' Compute the loss value over the entire corpus.
'''
set_entries_in_ttns(param_vec)
corpus_cost = 0
for idx in range(len(train_lex)):
input_string = train_lex[idx]
output_string = train_y[idx]
corpus_cost += ttns.train_f_cost(input_string, output_string)
return corpus_cost / len(train_lex)
def gradient_over_corpus(param_vec):
set_entries_in_ttns(param_vec)
corpus_grad = numpy.zeros((total_param, ), dtype='float64')
for idx in range(len(train_lex)):
input_string = train_lex[idx]
output_string = train_y[idx]
tmp_grad = ttns.train_f_grad(input_string, output_string)
corpus_grad += vectorize(tmp_grad, 'float64')
return corpus_grad / len(train_lex)
with rasengan.tictoc("Training %d epoch" % training_stats['epoch_id']):
init_param = get_entries_in_ttns()
rasengan.warn('Skipped FD Check')
# print 'Check grad output: Error=', scipy.optimize.check_grad(func=loss_over_corpus, grad=gradient_over_corpus, x0=init_param)
opt_param = scipy.optimize.fmin_l_bfgs_b(loss_over_corpus,
init_param,
fprime=gradient_over_corpus,
disp=2,
maxiter=1000)[0]
set_entries_in_ttns(opt_param)
return
def args_creation_part2(args, data):
if args.penalty_full_decomp_jason:
assert args.use_1bl
assert args.bilstm_stagger_schedule == 'external'
if (args.partition_dev_into_test
and args.partition_dev_into_train):
rasengan.warn('NOTE: You are pilfering from dev into both train and test')
#------------------------#
# Add Topology Arguments #
#------------------------#
args.in_dim = (data.vocsize + 2)
args.wemb1_win_size = args.win
args.penalty_vocsize = data.vocsize
args.penalty_mid_col = (args.wemb1_win_size - 1)/2
if args.use_0bl:
bilstm_stack = []
elif args.use_1bl:
bilstm_stack = [
(lstm_seqlabel_circuit.BiLSTM, 'bilstm')]
elif args.use_1l:
bilstm_stack = [
(lstm_seqlabel_circuit.BiLSTM, 'bilstm'),]
elif args.use_6bl:
bilstm_stack = [
(lstm_seqlabel_circuit.BiLSTM, 'bilstm'),
(lstm_seqlabel_circuit.BiLSTM, 'bilstmII'),
(lstm_seqlabel_circuit.BiLSTM, 'bilstmIII'),
(lstm_seqlabel_circuit.BiLSTM, 'bilstmIV'),
(lstm_seqlabel_circuit.BiLSTM, 'bilstmV'),
(lstm_seqlabel_circuit.BiLSTM, 'bilstmVI')]
elif args.use_8bl:
bilstm_stack = [
(lstm_seqlabel_circuit.BiLSTM, 'bilstm'),
(lstm_seqlabel_circuit.BiLSTM, 'bilstmII'),
(lstm_seqlabel_circuit.BiLSTM, 'bilstmIII'),
(lstm_seqlabel_circuit.BiLSTM, 'bilstmIV'),
(lstm_seqlabel_circuit.BiLSTM, 'bilstmV'),
(lstm_seqlabel_circuit.BiLSTM, 'bilstmVI'),
(lstm_seqlabel_circuit.BiLSTM, 'bilstmVII'),
(lstm_seqlabel_circuit.BiLSTM, 'bilstmVIII')]
elif args.use_4l:
bilstm_stack = [
(lstm_seqlabel_circuit.BiLSTM, 'bilstm'),
(lstm_seqlabel_circuit.BiLSTM, 'bilstmII'),
(lstm_seqlabel_circuit.BiLSTM, 'bilstmIII'),
(lstm_seqlabel_circuit.BiLSTM, 'bilstmIV')]
else:
bilstm_stack = [
(lstm_seqlabel_circuit.BiLSTM, 'bilstm'),
(lstm_seqlabel_circuit.BiLSTM, 'bilstmII'),
(lstm_seqlabel_circuit.BiLSTM, 'bilstmIII'),
(lstm_seqlabel_circuit.BiLSTM, 'bilstmIV')]
args.chips = (
[(lstm_seqlabel_circuit.Embedding, 'wemb1')]
+ bilstm_stack
+[(transducer_circuit.Penalty, 'penalty')])
#---------------------------------------------#
# Learning Rates, Optimizers, Epoch, EndPoint #
#---------------------------------------------#
args.optimizer = lstm_seqlabel_optimizer.sgd
args.perform_training = 1
args.perform_testing = (0 or args.partition_dev_into_test)
args.lr = 0.4
args.lr_drop = 0.9
args.nepochs = 1000
args.train_f = lstm_seqlabel_training.train_transducer
args.validate_predictions_f = (
lstm_seqlabel_validation.validate_predictions_transducer)
args.verbose = 2
args.skip_validation = 0
INSERTION_LIMIT = 3
args.endpoint = transducer_wrapper.TransducerWrapper(
Transducer(data.vocsize, INSERTION_LIMIT),
sampling_decoding=args.sampling_decoding,
crunching=args.crunching, ryanout=args.ryanout)
args.endpoint.dont_pickle = 1
print args.endpoint
#-------------------------------------------------------#
# Dropout, Gradient Clipping, L2 Projection for *Wemb1* #
#-------------------------------------------------------#
args.wemb1_do_dropout = 1
args.wemb1_dropout_retention_freq = .8
args.wemb1_clip_gradient = 1
args.wemb1_l2_project = 1
#---------------------------------------------------------#
# Dropout, Gradient Clipping, L2 Projection for *Penalty* #
#---------------------------------------------------------#
rasengan.warn('We DONT DO DROPOUT ON PENALTY !!')
args.penalty_clip_gradient = 1
args.penalty_l2_project = 1
args.penalty_tie_copy_param = 1
args.penalty_vocsize = data.vocsize
#-----------------#
# LSTM parameters #
#-----------------#
# Set the forward LSTM of the first LSTM by hand.
# Forward LSTM
if args.bilstm_externalandcopyatmax:
args.bilstm_stagger_schedule = 'external'
pass
if args.bilstm_runbilstmseparately:
args.bilstm_stagger_schedule = 'external'
pass
args.bilstm_do_backward_pass = not (args.use_1l or args.use_4l)
args.bilstm_forward_do_dropout = 1
args.bilstm_forward_dropout_retention_freq = 0.8
args.bilstm_forward_clip_gradient = 1
args.bilstm_forward_l2_project = 1
args.bilstm_forward_add_bias = 1
for prop_src, prop_dest in zip(
FORWARD_LSTM_PROPERTIES, BACKWARD_LSTM_PROPERTIES):
# Copy Backward LSTM property from the forward part
args.copy_invariant_is_prefix('bilstm', prop_src, prop_dest)
#------------------------------------------------------------#
# Settings for later BiLSTMs : bilstmII, bilstmIII, bilstmIV #
# These settings are simply copied over. #
# There is no need to remove properties, since properties #
# that are not needed would simply not be compiled. #
#------------------------------------------------------------#
for bilstm_height in range(2, len(bilstm_stack)+1):
at_top = (bilstm_height == len(bilstm_stack))
bilstm_height = ARABIC_TO_ROMAN_MAP[bilstm_height]
if args.bilstm_externalandcopyatmax or args.bilstm_runbilstmseparately:
if at_top:
bl_name = ('bilstm%s_forcefully_copy_embedding'
'_to_output'%bilstm_height)
setattr(args, bl_name, 1)
pass
if args.bilstm_runbilstmseparately:
setattr(args, 'bilstm%s_segregate_bilstm_inputs'%bilstm_height,
args.bilstm_runbilstmseparately)
setattr(args, 'bilstm%s_stagger_schedule'%bilstm_height,
args.bilstm_stagger_schedule)
setattr(args, 'bilstm%s_do_backward_pass'%bilstm_height,
args.bilstm_do_backward_pass)
args = util_add_bilstm_prop(args, 'bilstm%s'%bilstm_height)
#----------------------------------------------#
# The clipping Value and Projection Threshold. #
#----------------------------------------------#
args.clipping_value = 10
args.projection_threshold = 7
#------------------------------------------#
# Settings for blocking updates to layers. #
#------------------------------------------#
args.wemb1_block_update = 0
args.bilstm_forward_block_update = 0
args.bilstm_backward_block_update = 0
args.bilstmII_forward_block_update = 0
args.bilstmII_backward_block_update = 0
args.penalty_block_update = 0
#----------------------------#
# Learning Rate Controllers. #
#----------------------------#
args.decay = 0
args.decay_epochs = 0
args.minimum_lr = 1e-5
# The learning rate decay exponent.
args.lr_decay_exponent = 0
#-------------------------#
# Loading Pretrained PKL. #
#-------------------------#
rasengan.warn('NOTE: I have set pretrained_param_pklfile to None')
args.pretrained_param_pklfile = None
return args
def args_creation_part2(args, data):
if args.penalty_full_decomp_jason:
assert args.use_1bl
assert args.bilstm_stagger_schedule == 'external'
if (args.partition_dev_into_test and args.partition_dev_into_train):
rasengan.warn(
'NOTE: You are pilfering from dev into both train and test')
#------------------------#
# Add Topology Arguments #
#------------------------#
args.in_dim = (data.vocsize + 2)
args.wemb1_win_size = args.win
args.penalty_vocsize = data.vocsize
args.penalty_mid_col = (args.wemb1_win_size - 1) / 2
if args.use_0bl:
bilstm_stack = []
elif args.use_1bl:
bilstm_stack = [(lstm_seqlabel_circuit.BiLSTM, 'bilstm')]
elif args.use_1l:
bilstm_stack = [
(lstm_seqlabel_circuit.BiLSTM, 'bilstm'),
]
elif args.use_6bl:
bilstm_stack = [(lstm_seqlabel_circuit.BiLSTM, 'bilstm'),
(lstm_seqlabel_circuit.BiLSTM, 'bilstmII'),
(lstm_seqlabel_circuit.BiLSTM, 'bilstmIII'),
(lstm_seqlabel_circuit.BiLSTM, 'bilstmIV'),
(lstm_seqlabel_circuit.BiLSTM, 'bilstmV'),
(lstm_seqlabel_circuit.BiLSTM, 'bilstmVI')]
elif args.use_8bl:
bilstm_stack = [(lstm_seqlabel_circuit.BiLSTM, 'bilstm'),
(lstm_seqlabel_circuit.BiLSTM, 'bilstmII'),
(lstm_seqlabel_circuit.BiLSTM, 'bilstmIII'),
(lstm_seqlabel_circuit.BiLSTM, 'bilstmIV'),
(lstm_seqlabel_circuit.BiLSTM, 'bilstmV'),
(lstm_seqlabel_circuit.BiLSTM, 'bilstmVI'),
(lstm_seqlabel_circuit.BiLSTM, 'bilstmVII'),
(lstm_seqlabel_circuit.BiLSTM, 'bilstmVIII')]
elif args.use_4l:
bilstm_stack = [(lstm_seqlabel_circuit.BiLSTM, 'bilstm'),
(lstm_seqlabel_circuit.BiLSTM, 'bilstmII'),
(lstm_seqlabel_circuit.BiLSTM, 'bilstmIII'),
(lstm_seqlabel_circuit.BiLSTM, 'bilstmIV')]
else:
bilstm_stack = [(lstm_seqlabel_circuit.BiLSTM, 'bilstm'),
(lstm_seqlabel_circuit.BiLSTM, 'bilstmII'),
(lstm_seqlabel_circuit.BiLSTM, 'bilstmIII'),
(lstm_seqlabel_circuit.BiLSTM, 'bilstmIV')]
args.chips = ([(lstm_seqlabel_circuit.Embedding, 'wemb1')] + bilstm_stack +
[(transducer_circuit.Penalty, 'penalty')])
#---------------------------------------------#
# Learning Rates, Optimizers, Epoch, EndPoint #
#---------------------------------------------#
args.optimizer = lstm_seqlabel_optimizer.sgd
args.perform_training = 1
args.perform_testing = (0 or args.partition_dev_into_test)
args.lr = 0.4
args.lr_drop = 0.9
args.nepochs = 1000
args.train_f = lstm_seqlabel_training.train_transducer
args.validate_predictions_f = (
lstm_seqlabel_validation.validate_predictions_transducer)
args.verbose = 2
args.skip_validation = 0
INSERTION_LIMIT = 3
args.endpoint = transducer_wrapper.TransducerWrapper(
Transducer(data.vocsize, INSERTION_LIMIT),
sampling_decoding=args.sampling_decoding,
crunching=args.crunching)
args.endpoint.dont_pickle = 1
print args.endpoint
#-------------------------------------------------------#
# Dropout, Gradient Clipping, L2 Projection for *Wemb1* #
#-------------------------------------------------------#
args.wemb1_do_dropout = 1
args.wemb1_dropout_retention_freq = .8
args.wemb1_clip_gradient = 1
args.wemb1_l2_project = 1
#---------------------------------------------------------#
# Dropout, Gradient Clipping, L2 Projection for *Penalty* #
#---------------------------------------------------------#
rasengan.warn('We DONT DO DROPOUT ON PENALTY !!')
args.penalty_clip_gradient = 1
args.penalty_l2_project = 1
args.penalty_tie_copy_param = 1
args.penalty_vocsize = data.vocsize
#-----------------#
# LSTM parameters #
#-----------------#
# Set the forward LSTM of the first LSTM by hand.
# Forward LSTM
if args.bilstm_externalandcopyatmax:
args.bilstm_stagger_schedule = 'external'
pass
if args.bilstm_runbilstmseparately:
args.bilstm_stagger_schedule = 'external'
pass
args.bilstm_do_backward_pass = not (args.use_1l or args.use_4l)
args.bilstm_forward_do_dropout = 1
args.bilstm_forward_dropout_retention_freq = 0.8
args.bilstm_forward_clip_gradient = 1
args.bilstm_forward_l2_project = 1
args.bilstm_forward_add_bias = 1
for prop_src, prop_dest in zip(FORWARD_LSTM_PROPERTIES,
BACKWARD_LSTM_PROPERTIES):
# Copy Backward LSTM property from the forward part
args.copy_invariant_is_prefix('bilstm', prop_src, prop_dest)
#------------------------------------------------------------#
# Settings for later BiLSTMs : bilstmII, bilstmIII, bilstmIV #
# These settings are simply copied over. #
# There is no need to remove properties, since properties #
# that are not needed would simply not be compiled. #
#------------------------------------------------------------#
for bilstm_height in range(2, len(bilstm_stack) + 1):
at_top = (bilstm_height == len(bilstm_stack))
bilstm_height = ARABIC_TO_ROMAN_MAP[bilstm_height]
if args.bilstm_externalandcopyatmax or args.bilstm_runbilstmseparately:
if at_top:
bl_name = ('bilstm%s_forcefully_copy_embedding'
'_to_output' % bilstm_height)
setattr(args, bl_name, 1)
pass
if args.bilstm_runbilstmseparately:
setattr(args, 'bilstm%s_segregate_bilstm_inputs' % bilstm_height,
args.bilstm_runbilstmseparately)
setattr(args, 'bilstm%s_stagger_schedule' % bilstm_height,
args.bilstm_stagger_schedule)
setattr(args, 'bilstm%s_do_backward_pass' % bilstm_height,
args.bilstm_do_backward_pass)
args = util_add_bilstm_prop(args, 'bilstm%s' % bilstm_height)
#----------------------------------------------#
# The clipping Value and Projection Threshold. #
#----------------------------------------------#
args.clipping_value = 10
args.projection_threshold = 7
#------------------------------------------#
# Settings for blocking updates to layers. #
#------------------------------------------#
args.wemb1_block_update = 0
args.bilstm_forward_block_update = 0
args.bilstm_backward_block_update = 0
args.bilstmII_forward_block_update = 0
args.bilstmII_backward_block_update = 0
args.penalty_block_update = 0
#----------------------------#
# Learning Rate Controllers. #
#----------------------------#
args.decay = 0
args.decay_epochs = 0
args.minimum_lr = 1e-5
# The learning rate decay exponent.
args.lr_decay_exponent = 0
#-------------------------#
# Loading Pretrained PKL. #
#-------------------------#
rasengan.warn('NOTE: I have set pretrained_param_pklfile to None')
args.pretrained_param_pklfile = None
return args
def train_transducer_lbfgs(
train_lex, train_y, args, ttns, training_stats, batch_size=None):
''' This function completes a training epoch by doing one run of LBFGS.
`ts` abbreviates `train_stack` in entire function
Params
------
train_lex : A list of input_strings (the strings are represented as np arrays)
train_y : A list of output strings
batch_size : UNUSED : (default None)
'''
assert args.clipping_value < 0
assert args.projection_threshold < 0
ts_param_name = [
str(e) for e in ttns.train_stack_config.updatable_parameters()]
print 'The following params will be trained by lbfgs', ts_param_name
ts_param_shape_list = [ttns.train_stack_config[name].get_value().shape
for name in ts_param_name]
ts_param_shape_map = dict(zip(ts_param_name, ts_param_shape_list))
total_param = sum(numpy.prod(shape)
for shape
in ts_param_shape_map.values())
def set_entries_in_ttns(param_vec):
''' Set entries in ttns.train_stack_config
with corresponding values in param_vec.
'''
param_vec = param_vec.astype('float32')
offset = 0
for name in ts_param_name:
shape = ts_param_shape_map[name]
numel = numpy.prod(shape)
ttns.train_stack_config[name].set_value(
param_vec[offset:offset + numel].reshape(shape))
offset += numel
pass
return
def vectorize(param_list, dtype='float32'):
param_vec = numpy.zeros((total_param,), dtype=dtype)
offset = 0
for idx, param in enumerate(param_list):
shape = param.shape
assert shape == ts_param_shape_list[idx]
numel = numpy.prod(shape)
param_vec[offset:offset + numel] = param.reshape((numel,)).astype(dtype)
offset += numel
pass
return param_vec
def get_entries_in_ttns():
''' Set entries in ttns.train_stack_config
with corresponding values in param_vec.
'''
return vectorize(
[ttns.train_stack_config[name].get_value()
for name
in ts_param_name])
def loss_over_corpus(param_vec):
''' Compute the loss value over the entire corpus.
'''
set_entries_in_ttns(param_vec)
corpus_cost = 0
for idx in range(len(train_lex)):
input_string = train_lex[idx]
output_string = train_y[idx]
corpus_cost += ttns.train_f_cost(input_string, output_string)
return corpus_cost / len(train_lex)
def gradient_over_corpus(param_vec):
set_entries_in_ttns(param_vec)
corpus_grad = numpy.zeros((total_param,), dtype='float64')
for idx in range(len(train_lex)):
input_string = train_lex[idx]
output_string = train_y[idx]
tmp_grad = ttns.train_f_grad(input_string, output_string)
corpus_grad += vectorize(tmp_grad, 'float64')
return corpus_grad / len(train_lex)
with rasengan.tictoc("Training %d epoch"%training_stats['epoch_id']):
init_param = get_entries_in_ttns()
rasengan.warn('Skipped FD Check')
# print 'Check grad output: Error=', scipy.optimize.check_grad(func=loss_over_corpus, grad=gradient_over_corpus, x0=init_param)
opt_param = scipy.optimize.fmin_l_bfgs_b(
loss_over_corpus, init_param,
fprime=gradient_over_corpus, disp=2, maxiter=1000)[0]
set_entries_in_ttns(opt_param)
return
| Last-Updated: Sun May 1 09:07:58 2016 (-0400)
| By: Pushpendre Rastogi
| Update #: 64
'''
import util_lstm_seqlabel
import numpy
import time
import rasengan
import lstm_seqlabel_load_save_model
import functools
import codecs
try:
from dependency_parser.RectangleDependencyParser import DependencyParser
dp_viterbi_parse = DependencyParser().viterbi_parse
except:
rasengan.warn(
'You dont have the depenedency parser. Dont worry if you just want the transducer')
def get_conlleval_for_task(args):
if args.task == 'slu':
from data.atis import conlleval
elif args.task == 'chunking':
from data.conll2003_ner import conlleval
elif args.task == 'ner':
from data.conll2003_ner import conlleval
elif args.task == 'postag':
from data.conll_postag import eval_pwa as conlleval
else:
raise NotImplementedError
return conlleval
def config_overide(msg, args):
assert ' ' not in msg
args.folder = args.folder + '_' + msg
rasengan.warn('NOTE: I set args.folder to ' + args.folder)
yield
pass
| Last-Updated: Sun May 1 09:07:58 2016 (-0400)
| By: Pushpendre Rastogi
| Update #: 64
'''
import util_lstm_seqlabel
import numpy
import time
import rasengan
import lstm_seqlabel_load_save_model
import functools
import codecs
try:
from dependency_parser.RectangleDependencyParser import DependencyParser
dp_viterbi_parse = DependencyParser().viterbi_parse
except:
rasengan.warn(
'You dont have the depenedency parser. Dont worry if you just want the transducer')
def get_conlleval_for_task(args):
if args.task == 'slu':
from data.atis import conlleval
elif args.task == 'chunking':
from data.conll2003_ner import conlleval
elif args.task == 'ner':
from data.conll2003_ner import conlleval
elif args.task == 'postag':
from data.conll_postag import eval_pwa as conlleval
else:
raise NotImplementedError
return conlleval
def config_overide(msg, args):
assert ' ' not in msg
args.folder = args.folder + '_' + msg
rasengan.warn('NOTE: I set args.folder to ' + args.folder)
yield
pass