def run(dataset_path=DEFAULT_DATASET, dataset_name='timit',
iterator_type=DatasetDTWIterator, batch_size=100,
nframes=13, features="fbank",
init_lr=0.01, max_epochs=500,
network_type="dropout_net", trainer_type="adadelta",
layers_types=[ReLU, ReLU, ReLU, ReLU, LogisticRegression],
layers_sizes=[2400, 2400, 2400, 2400],
dropout_rates=[0.2, 0.5, 0.5, 0.5, 0.5],
recurrent_connections=[],
prefix_fname='',
debug_on_test_only=False,
debug_print=0,
debug_time=False,
debug_plot=0):
"""
FIXME TODO
"""
output_file_name = dataset_name
if prefix_fname != "":
output_file_name = prefix_fname + "_" + dataset_name
output_file_name += "_" + features + str(nframes)
output_file_name += "_" + network_type + "_" + trainer_type
output_file_name += "_emb_" + str(DIM_EMBEDDING)
print "output file name:", output_file_name
n_ins = None
n_outs = None
print "loading dataset from", dataset_path
# TODO DO A FUNCTION
if dataset_path[-7:] != '.joblib':
print >> sys.stderr, "prepare your dataset with align_words.py or lucid.py or buckeye.py"
sys.exit(-1)
### LOADING DATA
data_same = joblib.load(dataset_path)
shuffle(data_same)
has_dev_and_test_set = True
has_test_set_only = False
dev_dataset_path = dataset_path[:-7].replace("train", "") + 'dev.joblib'
test_dataset_path = dataset_path[:-7].replace("train", "") + 'test.joblib'
dev_split_at = len(data_same)
test_split_at = len(data_same)
if not os.path.exists(dev_dataset_path) or not os.path.exists(test_dataset_path):
has_dev_and_test_set = False
if os.path.exists(test_dataset_path):
print >> sys.stderr, "DOESN'T HAVE A SEPARATED DEV SET, WE'LL SPLIT OUT OWN"
has_test_set_only = True
dev_split_at = int(0.9 * dev_split_at)
else:
print >> sys.stderr, "DOESN'T HAVE A SEPARATED DEV AND TEST SET, WE'LL SPLIT OUT OWNS"
dev_split_at = int(0.8 * dev_split_at)
test_split_at = int(0.9 * test_split_at)
print data_same[0]
print data_same[0][3].shape
n_ins = data_same[0][3].shape[1] * nframes
n_outs = DIM_EMBEDDING
normalize = True
min_max_scale = False
marginf = (nframes-1)/2 # TODO
### TRAIN SET
if has_dev_and_test_set:
train_set_iterator = DatasetDTWWrdSpkrIterator(data_same,
normalize=normalize, min_max_scale=min_max_scale,
scale_f1=None, scale_f2=None, nframes=nframes,
batch_size=batch_size, marginf=marginf)
else:
train_set_iterator = DatasetDTWWrdSpkrIterator(
data_same[:dev_split_at], normalize=normalize,
min_max_scale=min_max_scale, scale_f1=None, scale_f2=None,
nframes=nframes, batch_size=batch_size, marginf=marginf)
f1 = train_set_iterator._scale_f1
f2 = train_set_iterator._scale_f2
### DEV SET
if has_dev_and_test_set:
data_same = joblib.load(dev_dataset_path)
valid_set_iterator = DatasetDTWWrdSpkrIterator(data_same,
normalize=normalize, min_max_scale=min_max_scale,
scale_f1=f1, scale_f2=f2,
nframes=nframes, batch_size=batch_size, marginf=marginf)
else:
valid_set_iterator = DatasetDTWWrdSpkrIterator(
data_same[dev_split_at:test_split_at], normalize=normalize,
min_max_scale=min_max_scale, scale_f1=f1, scale_f2=f2,
nframes=nframes, batch_size=batch_size, marginf=marginf)
### TEST SET
if has_dev_and_test_set or has_test_set_only:
data_same = joblib.load(test_dataset_path)
test_set_iterator = DatasetDTWWrdSpkrIterator(data_same,
normalize=normalize, min_max_scale=min_max_scale,
scale_f1=f1, scale_f2=f2, nframes=nframes,
batch_size=batch_size, marginf=marginf)
else:
test_set_iterator = DatasetDTWWrdSpkrIterator(
data_same[test_split_at:], normalize=normalize,
min_max_scale=min_max_scale, scale_f1=f1, scale_f2=f2,
nframes=nframes, batch_size=batch_size, marginf=marginf)
assert n_ins != None
assert n_outs != None
# numpy random generator
numpy_rng = numpy.random.RandomState(123)
print '... building the model'
# TODO the proper network type other than just dropout or not
nnet = None
fast_dropout = False
if "dropout" in network_type:
nnet = DropoutABNeuralNet(numpy_rng=numpy_rng, # TODO with 2 Outputs
n_ins=n_ins,
layers_types=layers_types,
layers_sizes=layers_sizes,
n_outs=n_outs,
loss='cos_cos2',
rho=0.95,
eps=1.E-6,
max_norm=4.,
fast_drop=fast_dropout,
debugprint=debug_print)
else:
nnet = ABNeuralNet2Outputs(numpy_rng=numpy_rng,
n_ins=n_ins,
layers_types=layers_types,
layers_sizes=layers_sizes,
n_outs=n_outs,
loss='cos_cos2',
#loss='dot_prod',
rho=0.90,
eps=1.E-6,
max_norm=0.,
debugprint=debug_print)
print "Created a neural net as:",
print str(nnet)
# get the training, validation and testing function for the model
print '... getting the training functions'
print trainer_type
train_fn = None
if debug_plot or debug_print:
if trainer_type == "adadelta":
train_fn = nnet.get_adadelta_trainer(debug=True)
elif trainer_type == "adagrad":
train_fn = nnet.get_adagrad_trainer(debug=True)
else:
train_fn = nnet.get_SGD_trainer(debug=True)
else:
if trainer_type == "adadelta":
train_fn = nnet.get_adadelta_trainer()
elif trainer_type == "adagrad":
train_fn = nnet.get_adagrad_trainer()
else:
train_fn = nnet.get_SGD_trainer()
train_scoref_w = nnet.score_classif_same_diff_word_separated(train_set_iterator)
valid_scoref_w = nnet.score_classif_same_diff_word_separated(valid_set_iterator)
test_scoref_w = nnet.score_classif_same_diff_word_separated(test_set_iterator)
train_scoref_s = nnet.score_classif_same_diff_spkr_separated(train_set_iterator)
valid_scoref_s = nnet.score_classif_same_diff_spkr_separated(valid_set_iterator)
test_scoref_s = nnet.score_classif_same_diff_spkr_separated(test_set_iterator)
data_iterator = train_set_iterator
if debug_on_test_only:
print >> sys.stderr, "NOT IMPLEMENTED"
sys.exit(-1)
data_iterator = test_set_iterator
train_scoref_w = test_scoref_w
train_scoref_s = test_scoref_s
print '... training the model'
# early-stopping parameters
patience = 1000 # look as this many examples regardless TODO
patience_increase = 2. # wait this much longer when a new best is
# found
improvement_threshold = 0.995 # a relative improvement of this much is
# considered significant
best_validation_loss = numpy.inf
test_score = 0.
start_time = time.clock()
done_looping = False
epoch = 0
lr = init_lr
timer = None
if debug_plot:
print_mean_weights_biases(nnet.params)
#with open(output_file_name + 'epoch_0.pickle', 'wb') as f:
# cPickle.dump(nnet, f, protocol=-1)
while (epoch < max_epochs) and (not done_looping):
if REDTW and "ab_net" in network_type and ((epoch + 1) % 20) == 0:
print "recomputing DTW:"
data_iterator.recompute_DTW(nnet.transform_x1())
epoch = epoch + 1
avg_costs = []
avg_params_gradients_updates = []
if debug_time:
timer = time.time()
for iteration, (x, y) in enumerate(data_iterator):
#print "x[0][0]", x[0][0]
#print "x[1][0]", x[1][0]
#print "y[0][0]", y[0][0]
#print "y[1][0]", y[1][0]
avg_cost = 0.
if "delta" in trainer_type: # TODO remove need for this if
avg_cost = train_fn(x[0], x[1], y[0], y[1])
else:
avg_cost = train_fn(x[0], x[1], y[0], y[1], lr)
if debug_print >= 3:
print "cost:", avg_cost[0]
if debug_plot >= 2:
plot_costs(avg_cost[0])
if not len(avg_params_gradients_updates):
avg_params_gradients_updates = map(numpy.asarray, avg_cost[1:])
else:
avg_params_gradients_updates = rolling_avg_pgu(
iteration, avg_params_gradients_updates,
map(numpy.asarray, avg_cost[1:]))
if debug_plot >= 3:
plot_params_gradients_updates(iteration, avg_cost[1:])
if type(avg_cost) == list:
avg_costs.append(avg_cost[0])
else:
avg_costs.append(avg_cost)
if debug_print >= 2:
print_mean_weights_biases(nnet.params)
if debug_plot >= 2:
plot_params_gradients_updates(epoch, avg_params_gradients_updates)
if debug_time:
print(' epoch %i took %f seconds' % (epoch, time.time() - timer))
avg_cost = numpy.mean(avg_costs)
if numpy.isnan(avg_cost):
print("avg costs is NaN so we're stopping here!")
break
print(' epoch %i, avg costs %f' % \
(epoch, avg_cost))
tmp_train = zip(*train_scoref_w())
print(' epoch %i, training sim same words %f, diff words %f' % \
(epoch, numpy.mean(tmp_train[0]), numpy.mean(tmp_train[1])))
tmp_train = zip(*train_scoref_s())
print(' epoch %i, training sim same spkrs %f, diff spkrs %f' % \
(epoch, numpy.mean(tmp_train[0]), numpy.mean(tmp_train[1])))
# TODO update lr(t) = lr(0) / (1 + lr(0) * lambda * t)
lr = numpy.float32(init_lr / (numpy.sqrt(iteration) + 1.)) ### TODO
#lr = numpy.float32(init_lr / (iteration + 1.)) ### TODO
# or another scheme for learning rate decay
#with open(output_file_name + 'epoch_' +str(epoch) + '.pickle', 'wb') as f:
# cPickle.dump(nnet, f, protocol=-1)
if debug_on_test_only:
continue
# we check the validation loss on every epoch
validation_losses_w = zip(*valid_scoref_w())
validation_losses_s = zip(*valid_scoref_s())
this_validation_loss = 0.25*(1.-numpy.mean(validation_losses_w[0])) +\
0.25*numpy.mean(validation_losses_w[1]) +\
0.25*(1.-numpy.mean(validation_losses_s[0])) +\
0.25*numpy.mean(validation_losses_s[1])
print(' epoch %i, valid sim same words %f, diff words %f' % \
(epoch, numpy.mean(validation_losses_w[0]), numpy.mean(validation_losses_w[1])))
print(' epoch %i, valid sim same spkrs %f, diff spkrs %f' % \
(epoch, numpy.mean(validation_losses_s[0]), numpy.mean(validation_losses_s[1])))
# if we got the best validation score until now
if this_validation_loss < best_validation_loss:
with open(output_file_name + '.pickle', 'wb') as f:
cPickle.dump(nnet, f, protocol=-1)
# improve patience if loss improvement is good enough
if (this_validation_loss < best_validation_loss *
improvement_threshold):
patience = max(patience, iteration * patience_increase)
# save best validation score and iteration number
best_validation_loss = this_validation_loss
# test it on the test set
test_losses_w = zip(*test_scoref_w())
test_losses_s = zip(*test_scoref_s())
print(' epoch %i, test sim same words %f, diff words %f' % \
(epoch, numpy.mean(test_losses_w[0]), numpy.mean(test_losses_w[1])))
print(' epoch %i, test sim same spkrs %f, diff spkrs %f' % \
(epoch, numpy.mean(test_losses_s[0]), numpy.mean(test_losses_s[1])))
if patience <= iteration: # TODO correct that
done_looping = True
break
end_time = time.clock()
print(('Optimization complete with best validation score of %f, '
'with test performance %f') %
(best_validation_loss, test_score))
print >> sys.stderr, ('The fine tuning code for file ' +
os.path.split(__file__)[1] +
' ran for %.2fm' % ((end_time - start_time)
/ 60.))
with open(output_file_name + '_final.pickle', 'wb') as f:
cPickle.dump(nnet, f, protocol=-1)
def run(dataset_path=DEFAULT_DATASET,
dataset_name='timit',
batch_size=100,
nframes=13,
features="fbank",
init_lr=0.01,
max_epochs=500,
network_type="AB",
trainer_type="adadelta",
layers_types=[ReLU, ReLU, ReLU, ReLU, LogisticRegression],
layers_sizes=[2400, 2400, 2400, 2400],
dropout_rates=[0.2, 0.5, 0.5, 0.5, 0.5],
recurrent_connections=[],
prefix_fname='',
debug_print=0,
debug_time=False,
debug_plot=0):
"""
Configures and run the neural net on the given dataset.
"""
output_file_name = dataset_name
if prefix_fname != "":
output_file_name = prefix_fname + "_" + dataset_name
output_file_name += "_" + features + str(nframes)
output_file_name += "_" + network_type + "_" + trainer_type
output_file_name += "_emb_" + str(DIM_EMBEDDING)
print "output file name:", output_file_name
n_ins = None
n_outs = None
print "loading dataset from", dataset_path
# TODO DO A FUNCTION FOR DATASET LOADING CRAP
if dataset_path[-7:] != '.joblib':
print >> sys.stderr, "prepare your dataset with align_words.py or lucid.py or buckeye.py"
sys.exit(-1)
### LOADING DATA
data_same = joblib.load(dataset_path)
shuffle(data_same)
has_dev_set = True
test_dataset_path = dataset_path[:-7].replace("train", "") + 'test.joblib'
dev_split_at = int(0.9 * len(data_same))
test_split_at = len(data_same)
if not os.path.exists(test_dataset_path):
has_dev_set = False
test_split_at = int(0.95 * test_split_at)
print data_same[0]
print data_same[0][3].shape
n_ins = data_same[0][3].shape[1] * nframes
n_outs = DIM_EMBEDDING
normalize = True
min_max_scale = False
marginf = (nframes - 1) / 2 # TODO
### TRAIN SET
train_set_iterator = DatasetDTWWrdSpkrIterator(data_same[:dev_split_at],
normalize=normalize,
min_max_scale=min_max_scale,
scale_f1=None,
scale_f2=None,
nframes=nframes,
batch_size=batch_size,
marginf=marginf)
f1 = train_set_iterator._scale_f1
f2 = train_set_iterator._scale_f2
### DEV SET
valid_set_iterator = DatasetDTWWrdSpkrIterator(
data_same[dev_split_at:test_split_at],
normalize=normalize,
min_max_scale=min_max_scale,
scale_f1=f1,
scale_f2=f2,
nframes=nframes,
batch_size=batch_size,
marginf=marginf)
### TEST SET
if has_dev_set:
data_same = joblib.load(test_dataset_path)
test_set_iterator = DatasetDTWWrdSpkrIterator(
data_same,
normalize=normalize,
min_max_scale=min_max_scale,
scale_f1=f1,
scale_f2=f2,
nframes=nframes,
batch_size=batch_size,
marginf=marginf)
else:
test_set_iterator = DatasetDTWWrdSpkrIterator(
data_same[test_split_at:],
normalize=normalize,
min_max_scale=min_max_scale,
scale_f1=f1,
scale_f2=f2,
nframes=nframes,
batch_size=batch_size,
marginf=marginf)
assert n_ins != None
assert n_outs != None
# numpy random generator
numpy_rng = numpy.random.RandomState(123)
print '... building the model'
nnet = None
fast_dropout = False
if "dropout" in network_type:
print >> sys.stderr, "Dropout is not implemented for ABnets with 2 Outputs"
nnet = DropoutABNeuralNet(
numpy_rng=numpy_rng, # TODO with 2 Outputs
n_ins=n_ins,
layers_types=layers_types,
layers_sizes=layers_sizes,
n_outs=n_outs,
loss='cos_cos2',
rho=0.95,
eps=1.E-6,
max_norm=4.,
fast_drop=fast_dropout,
debugprint=debug_print)
else:
nnet = ABNeuralNet2Outputs(numpy_rng=numpy_rng,
n_ins=n_ins,
layers_types=layers_types,
layers_sizes=layers_sizes,
n_outs=n_outs,
loss='cos_cos2',
rho=0.90,
eps=1.E-6,
max_norm=0.,
debugprint=debug_print)
print "Created a neural net as:",
print str(nnet)
# get the training, validation and testing function for the model
print '... getting the training functions'
print trainer_type
train_fn = None
if debug_plot or debug_print:
if trainer_type == "adadelta":
train_fn = nnet.get_adadelta_trainer(debug=True)
elif trainer_type == "adagrad":
train_fn = nnet.get_adagrad_trainer(debug=True)
else:
train_fn = nnet.get_SGD_trainer(debug=True)
else:
if trainer_type == "adadelta":
train_fn = nnet.get_adadelta_trainer()
elif trainer_type == "adagrad":
train_fn = nnet.get_adagrad_trainer()
else:
train_fn = nnet.get_SGD_trainer()
train_scoref_w = nnet.score_classif_same_diff_word_separated(
train_set_iterator)
valid_scoref_w = nnet.score_classif_same_diff_word_separated(
valid_set_iterator)
test_scoref_w = nnet.score_classif_same_diff_word_separated(
test_set_iterator)
train_scoref_s = nnet.score_classif_same_diff_spkr_separated(
train_set_iterator)
valid_scoref_s = nnet.score_classif_same_diff_spkr_separated(
valid_set_iterator)
test_scoref_s = nnet.score_classif_same_diff_spkr_separated(
test_set_iterator)
data_iterator = train_set_iterator
print '... training the model'
best_validation_loss = numpy.inf
test_score = 0.
start_time = time.clock()
epoch = 0
lr = init_lr
timer = None
if debug_plot:
print_mean_weights_biases(nnet.params)
#with open(output_file_name + 'epoch_0.pickle', 'wb') as f:
# cPickle.dump(nnet, f, protocol=-1)
while (epoch < max_epochs):
epoch = epoch + 1
avg_costs = []
avg_params_gradients_updates = []
if debug_time:
timer = time.time()
for iteration, (x, y) in enumerate(data_iterator):
#print "x[0][0]", x[0][0]
#print "x[1][0]", x[1][0]
#print "y[0][0]", y[0][0]
#print "y[1][0]", y[1][0]
avg_cost = 0.
if "delta" in trainer_type: # TODO remove need for this if
avg_cost = train_fn(x[0], x[1], y[0], y[1])
else:
avg_cost = train_fn(x[0], x[1], y[0], y[1], lr)
if debug_print >= 3:
print "cost:", avg_cost[0]
if debug_plot >= 2:
plot_costs(avg_cost[0])
if not len(avg_params_gradients_updates):
avg_params_gradients_updates = map(numpy.asarray,
avg_cost[1:])
else:
avg_params_gradients_updates = rolling_avg_pgu(
iteration, avg_params_gradients_updates,
map(numpy.asarray, avg_cost[1:]))
if debug_plot >= 3:
plot_params_gradients_updates(iteration, avg_cost[1:])
if type(avg_cost) == list:
avg_costs.append(avg_cost[0])
else:
avg_costs.append(avg_cost)
if debug_print >= 2:
print_mean_weights_biases(nnet.params)
if debug_plot >= 2:
plot_params_gradients_updates(epoch, avg_params_gradients_updates)
if debug_time:
print(' epoch %i took %f seconds' % (epoch, time.time() - timer))
avg_cost = numpy.mean(avg_costs)
if numpy.isnan(avg_cost):
print("avg costs is NaN so we're stopping here!")
break
print(' epoch %i, avg costs %f' % \
(epoch, avg_cost))
tmp_train = zip(*train_scoref_w())
print(' epoch %i, training sim same words %f, diff words %f' % \
(epoch, numpy.mean(tmp_train[0]), numpy.mean(tmp_train[1])))
tmp_train = zip(*train_scoref_s())
print(' epoch %i, training sim same spkrs %f, diff spkrs %f' % \
(epoch, numpy.mean(tmp_train[0]), numpy.mean(tmp_train[1])))
# TODO update lr(t) = lr(0) / (1 + lr(0) * lambda * t)
lr = numpy.float32(init_lr / (numpy.sqrt(iteration) + 1.)) ### TODO
# or another scheme for learning rate decay
#with open(output_file_name + 'epoch_' +str(epoch) + '.pickle', 'wb') as f:
# cPickle.dump(nnet, f, protocol=-1)
# we check the validation loss on every epoch
validation_losses_w = zip(*valid_scoref_w())
validation_losses_s = zip(*valid_scoref_s())
this_validation_loss = 0.25*(1.-numpy.mean(validation_losses_w[0])) +\
0.25*numpy.mean(validation_losses_w[1]) +\
0.25*(1.-numpy.mean(validation_losses_s[0])) +\
0.25*numpy.mean(validation_losses_s[1])
print(' epoch %i, valid sim same words %f, diff words %f' % \
(epoch, numpy.mean(validation_losses_w[0]), numpy.mean(validation_losses_w[1])))
print(' epoch %i, valid sim same spkrs %f, diff spkrs %f' % \
(epoch, numpy.mean(validation_losses_s[0]), numpy.mean(validation_losses_s[1])))
# if we got the best validation score until now
if this_validation_loss < best_validation_loss:
with open(output_file_name + '.pickle', 'wb') as f:
cPickle.dump(nnet, f, protocol=-1)
# save best validation score and iteration number
best_validation_loss = this_validation_loss
# test it on the test set
test_losses_w = zip(*test_scoref_w())
test_losses_s = zip(*test_scoref_s())
print(' epoch %i, test sim same words %f, diff words %f' % \
(epoch, numpy.mean(test_losses_w[0]), numpy.mean(test_losses_w[1])))
print(' epoch %i, test sim same spkrs %f, diff spkrs %f' % \
(epoch, numpy.mean(test_losses_s[0]), numpy.mean(test_losses_s[1])))
end_time = time.clock()
print(('Optimization complete with best validation score of %f, '
'with test performance %f') % (best_validation_loss, test_score))
print >> sys.stderr, ('The fine tuning code for file ' +
os.path.split(__file__)[1] + ' ran for %.2fm' %
((end_time - start_time) / 60.))
with open(output_file_name + '_final.pickle', 'wb') as f:
cPickle.dump(nnet, f, protocol=-1)
def run(dataset_path="from_aren.joblib", dataset_name='timit',
batch_size=100,
nframes=13, features="fbank",
init_lr=0.01, max_epochs=500,
network_type="AB", trainer_type="adadelta",
layers_types=[ReLU, ReLU, ReLU, ReLU, LogisticRegression],
layers_sizes=[2400, 2400, 2400, 2400],
dropout_rates=[0.2, 0.5, 0.5, 0.5, 0.5],
loss='cos_cos2',
recurrent_connections=[],
prefix_fname='',
debug_print=0,
debug_time=False,
debug_plot=0):
"""
Configures and run the neural net on the given dataset.
"""
output_file_name = dataset_name
if prefix_fname != "":
output_file_name = prefix_fname + "_" + dataset_name
output_file_name += "_" + features + str(nframes)
output_file_name += "_" + network_type + "_" + trainer_type
output_file_name += "_emb_" + str(DIM_EMBEDDING)
print "output file name:", output_file_name
n_ins = None
n_outs = None
print "loading dataset from", dataset_path
# TODO DO A FUNCTION FOR DATASET LOADING CRAP
if dataset_path[-7:] != '.joblib':
print >> sys.stderr, "prepare your dataset with align_words.py or lucid.py or buckeye.py"
sys.exit(-1)
### LOADING DATA
data_same = joblib.load(dataset_path)
shuffle(data_same)
dev_split_at = int(0.9 * len(data_same))
print data_same[0]
print data_same[0][3].shape
n_ins = data_same[0][3].shape[1] * nframes
n_outs = DIM_EMBEDDING
normalize = True
min_max_scale = False
marginf = (nframes-1)/2 # TODO
### TRAIN SET
train_set_iterator = DatasetDTWWrdSpkrIterator(
data_same[:dev_split_at], normalize=normalize,
min_max_scale=min_max_scale, scale_f1=None, scale_f2=None,
nframes=nframes, batch_size=batch_size, marginf=marginf)
f1 = train_set_iterator._scale_f1
f2 = train_set_iterator._scale_f2
### DEV SET
valid_set_iterator = DatasetDTWWrdSpkrIterator(
data_same[dev_split_at:], normalize=normalize,
min_max_scale=min_max_scale, scale_f1=f1, scale_f2=f2,
nframes=nframes, batch_size=batch_size, marginf=marginf)
assert n_ins != None
assert n_outs != None
# numpy random generator
numpy_rng = numpy.random.RandomState(123)
print '... building the model'
nnet = None
fast_dropout = False
if "dropout" in network_type:
print >> sys.stderr, "Dropout is not implemented for ABnets with 2 Outputs"
nnet = DropoutABNeuralNet(numpy_rng=numpy_rng, # TODO with 2 Outputs
n_ins=n_ins,
layers_types=layers_types,
layers_sizes=layers_sizes,
n_outs=n_outs,
loss=loss,
rho=0.95,
eps=1.E-6,
max_norm=4.,
fast_drop=fast_dropout,
debugprint=debug_print)
else:
nnet = ABNeuralNet2Outputs(numpy_rng=numpy_rng,
n_ins=n_ins,
layers_types=layers_types,
layers_sizes=layers_sizes,
n_outs=n_outs,
loss=loss,
rho=0.90,
eps=1.E-6,
max_norm=0.,
debugprint=debug_print)
print "Created a neural net as:",
print str(nnet)
# get the training, validation and testing function for the model
print '... getting the training functions'
print trainer_type
train_fn = None
if debug_plot or debug_print:
if trainer_type == "adadelta":
train_fn = nnet.get_adadelta_trainer(debug=True)
elif trainer_type == "adagrad":
train_fn = nnet.get_adagrad_trainer(debug=True)
else:
train_fn = nnet.get_SGD_trainer(debug=True)
else:
if trainer_type == "adadelta":
train_fn = nnet.get_adadelta_trainer()
elif trainer_type == "adagrad":
train_fn = nnet.get_adagrad_trainer()
else:
train_fn = nnet.get_SGD_trainer()
train_scoref_w = nnet.score_classif_same_diff_word_separated(train_set_iterator)
valid_scoref_w = nnet.score_classif_same_diff_word_separated(valid_set_iterator)
train_scoref_s = nnet.score_classif_same_diff_spkr_separated(train_set_iterator)
valid_scoref_s = nnet.score_classif_same_diff_spkr_separated(valid_set_iterator)
data_iterator = train_set_iterator
print '... training the model'
best_validation_loss = numpy.inf
test_score = 0.
start_time = time.clock()
epoch = 0
lr = init_lr
timer = None
if debug_plot:
print_mean_weights_biases(nnet.params)
#with open(output_file_name + 'epoch_0.pickle', 'wb') as f:
# cPickle.dump(nnet, f, protocol=-1)
while (epoch < max_epochs):
epoch = epoch + 1
avg_costs = []
avg_params_gradients_updates = []
if debug_time:
timer = time.time()
for iteration, (x, y) in enumerate(data_iterator):
#print "x[0][0]", x[0][0]
#print "x[1][0]", x[1][0]
#print "y[0][0]", y[0][0]
#print "y[1][0]", y[1][0]
avg_cost = 0.
if "delta" in trainer_type: # TODO remove need for this if
avg_cost = train_fn(x[0], x[1], y[0], y[1])
else:
avg_cost = train_fn(x[0], x[1], y[0], y[1], lr)
if debug_print >= 3:
print "cost:", avg_cost[0]
if debug_plot >= 2:
plot_costs(avg_cost[0])
if not len(avg_params_gradients_updates):
avg_params_gradients_updates = map(numpy.asarray, avg_cost[1:])
else:
avg_params_gradients_updates = rolling_avg_pgu(
iteration, avg_params_gradients_updates,
map(numpy.asarray, avg_cost[1:]))
if debug_plot >= 3:
plot_params_gradients_updates(iteration, avg_cost[1:])
if type(avg_cost) == list:
avg_costs.append(avg_cost[0])
else:
avg_costs.append(avg_cost)
if debug_print >= 2:
print_mean_weights_biases(nnet.params)
if debug_plot >= 2:
plot_params_gradients_updates(epoch, avg_params_gradients_updates)
if debug_time:
print(' epoch %i took %f seconds' % (epoch, time.time() - timer))
avg_cost = numpy.mean(avg_costs)
if numpy.isnan(avg_cost):
print("avg costs is NaN so we're stopping here!")
break
print(' epoch %i, avg costs %f' % \
(epoch, avg_cost))
tmp_train = zip(*train_scoref_w())
print(' epoch %i, training sim same words %f, diff words %f' % \
(epoch, numpy.mean(tmp_train[0]), numpy.mean(tmp_train[1])))
tmp_train = zip(*train_scoref_s())
print(' epoch %i, training sim same spkrs %f, diff spkrs %f' % \
(epoch, numpy.mean(tmp_train[0]), numpy.mean(tmp_train[1])))
# TODO update lr(t) = lr(0) / (1 + lr(0) * lambda * t)
lr = numpy.float32(init_lr / (numpy.sqrt(iteration) + 1.)) ### TODO
# or another scheme for learning rate decay
#with open(output_file_name + 'epoch_' +str(epoch) + '.pickle', 'wb') as f:
# cPickle.dump(nnet, f, protocol=-1)
# we check the validation loss on every epoch
validation_losses_w = zip(*valid_scoref_w())
validation_losses_s = zip(*valid_scoref_s())
this_validation_loss = 0.25*(1.-numpy.mean(validation_losses_w[0])) +\
0.25*numpy.mean(validation_losses_w[1]) +\
0.25*(1.-numpy.mean(validation_losses_s[0])) +\
0.25*numpy.mean(validation_losses_s[1])
print(' epoch %i, valid sim same words %f, diff words %f' % \
(epoch, numpy.mean(validation_losses_w[0]), numpy.mean(validation_losses_w[1])))
print(' epoch %i, valid sim same spkrs %f, diff spkrs %f' % \
(epoch, numpy.mean(validation_losses_s[0]), numpy.mean(validation_losses_s[1])))
# if we got the best validation score until now
if this_validation_loss < best_validation_loss:
with open(output_file_name + '.pickle', 'wb') as f:
cPickle.dump(nnet, f, protocol=-1)
# save best validation score and iteration number
best_validation_loss = this_validation_loss
end_time = time.clock()
print(('Optimization complete with best validation score of %f, ') %
(best_validation_loss))
print >> sys.stderr, ('The fine tuning code for file ' +
os.path.split(__file__)[1] +
' ran for %.2fm' % ((end_time - start_time)
/ 60.))
with open(output_file_name + '_final.pickle', 'wb') as f:
cPickle.dump(nnet, f, protocol=-1)
