def train_and_test(s2smodel, train_reader, test_reader, block_size,
num_quantization_bits, max_epochs, epoch_size,
minibatch_size, progress_printer, warm_up):
from Sequence2Sequence import create_criterion_function, create_model_train
model_train = create_model_train(s2smodel)
criterion = create_criterion_function(model_train)
# Create learner
if block_size is not None and num_quantization_bits != default_quantization_bits:
raise RuntimeError(
"Block momentum cannot be used with quantization, please remove quantized_bits option."
)
lr = 0.001 if use_attention else 0.005 # TODO: can we use the same value for both?
local_learner = fsadagrad(
model_train.parameters,
lr=learning_rate_schedule([lr] * 2 + [lr / 2] * 3 + [lr / 4],
UnitType.sample, epoch_size),
momentum=momentum_as_time_constant_schedule(1100),
gradient_clipping_threshold_per_sample=2.3,
gradient_clipping_with_truncation=True)
if block_size != None:
learner = block_momentum_distributed_learner(local_learner,
block_size=block_size)
else:
learner = data_parallel_distributed_learner(
local_learner,
num_quantization_bits=num_quantization_bits,
distributed_after=warm_up)
trainer = Trainer(None, criterion, learner, progress_printer)
train_bind = {
criterion.arguments[0]: train_reader.streams.features,
criterion.arguments[1]: train_reader.streams.labels
}
training_session(
mb_source=train_reader,
trainer=trainer,
model_inputs_to_streams=train_bind,
mb_size=minibatch_size,
progress_frequency=epoch_size,
checkpoint_config=CheckpointConfig(frequency=epoch_size,
filename=os.path.join(
model_path,
"SequenceToSequence"),
restore=False),
cv_config=CrossValidationConfig(source=test_reader,
mb_size=minibatch_size)).train()
def train(reader, model, max_epochs):
# declare the model's input dimension, so that the saved model is usable
model.update_signature(Sequence[SparseTensor[vocab_size]])
#model.declare_args(vocab_size)
# criterion: (model args, labels) -> (loss, metric)
# here (query, slot_labels) -> (ce, errs)
criterion = create_criterion_function(model)
labels = reader.streams.slot_labels
#labels = reader.streams.intent_labels # for intent classification
#from cntk.logging.graph import plot
#plot(criterion, filename=data_dir + "/model.pdf")
# iteration parameters --needed here because learner schedule needs it
epoch_size = 36000
minibatch_size = 70
#epoch_size = 1000 ; max_epochs = 1 # uncomment for faster testing
# SGD parameters
learner = fsadagrad(criterion.parameters,
lr = learning_rate_schedule([0.003]*2+[0.0015]*12+[0.0003], UnitType.sample, epoch_size),
momentum = momentum_as_time_constant_schedule(minibatch_size / -math.log(0.9)),
gradient_clipping_threshold_per_sample = 15,
gradient_clipping_with_truncation = True)
# trainer
trainer = Trainer(None, criterion, learner)
# process minibatches and perform model training
log_number_of_parameters(model) ; print()
progress_printer = ProgressPrinter(freq=100, first=10, tag='Training') # more detailed logging
#progress_printer = ProgressPrinter(tag='Training')
t = 0
for epoch in range(max_epochs): # loop over epochs
peek(model, epoch) # log some interesting info
epoch_end = (epoch+1) * epoch_size
while t < epoch_end: # loop over minibatches on the epoch
# BUGBUG: The change of minibatch_size parameter vv has no effect.
# TODO: change all examples to this pattern; then remove this comment
data = reader.next_minibatch(min(minibatch_size, epoch_end-t)) # fetch minibatch
#trainer.train_minibatch(data[reader.streams.query], data[labels]) # update model with it
trainer.train_minibatch({criterion.arguments[0]: data[reader.streams.query], criterion.arguments[1]: data[labels]}) # update model with it
t += data[labels].num_samples # count samples processed so far
progress_printer.update_with_trainer(trainer, with_metric=True) # log progress
loss, metric, actual_samples = progress_printer.epoch_summary(with_metric=True)
return loss, metric # return values from last epoch
def train_and_test(s2smodel, train_reader, test_reader, block_size, num_quantization_bits, max_epochs, epoch_size, minibatch_size, progress_printer, warm_up):
from Sequence2Sequence import create_criterion_function, create_model_train
model_train = create_model_train(s2smodel)
criterion = create_criterion_function(model_train)
# Create learner
if block_size is not None and num_quantization_bits != default_quantization_bits:
raise RuntimeError("Block momentum cannot be used with quantization, please remove quantized_bits option.")
lr = 0.001 if use_attention else 0.005 # TODO: can we use the same value for both?
local_learner = fsadagrad(model_train.parameters,
lr = learning_rate_schedule([lr]*2+[lr/2]*3+[lr/4], UnitType.sample, epoch_size),
momentum = momentum_as_time_constant_schedule(1100),
gradient_clipping_threshold_per_sample=2.3,
gradient_clipping_with_truncation=True)
if block_size != None:
learner = block_momentum_distributed_learner(local_learner, block_size=block_size)
else:
learner = data_parallel_distributed_learner(local_learner, num_quantization_bits=num_quantization_bits, distributed_after=warm_up)
trainer = Trainer(None, criterion, learner, progress_printer)
train_bind = {criterion.arguments[0]: train_reader.streams.features,
criterion.arguments[1]: train_reader.streams.labels}
training_session(
mb_source = train_reader,
trainer=trainer,
model_inputs_to_streams=train_bind,
mb_size=minibatch_size,
progress_frequency=epoch_size,
checkpoint_config=CheckpointConfig(frequency = epoch_size,
filename = os.path.join(model_path, "SequenceToSequence"),
restore = False),
cv_config=CrossValidationConfig(source=test_reader, mb_size=minibatch_size)
).train()
def train(train_reader, valid_reader, vocab, i2w, s2smodel, max_epochs,
epoch_size):
# Note: We would like to set the signature of 's2smodel' (s2smodel.update_signature()), but that will cause
# an error since the training criterion uses a reduced sequence axis for the labels.
# This is because it removes the initial <s> symbol. Hence, we must leave the model
# with unspecified input shapes and axes.
# create the training wrapper for the s2smodel, as well as the criterion function
model_train = create_model_train(s2smodel)
criterion = create_criterion_function(model_train)
# also wire in a greedy decoder so that we can properly log progress on a validation example
# This is not used for the actual training process.
model_greedy = create_model_greedy(s2smodel)
# This does not need to be done in training generally though
# Instantiate the trainer object to drive the model training
minibatch_size = 72
lr = 0.001 if use_attention else 0.005 # TODO: can we use the same value for both?
learner = fsadagrad(
model_train.parameters,
lr=learning_rate_schedule([lr] * 2 + [lr / 2] * 3 + [lr / 4],
UnitType.sample, epoch_size),
momentum=momentum_as_time_constant_schedule(1100),
gradient_clipping_threshold_per_sample=2.3,
gradient_clipping_with_truncation=True)
trainer = Trainer(None, criterion, learner)
# Get minibatches of sequences to train with and perform model training
total_samples = 0
mbs = 0
eval_freq = 100
# print out some useful training information
log_number_of_parameters(model_train)
print()
progress_printer = ProgressPrinter(freq=30, tag='Training')
#progress_printer = ProgressPrinter(freq=30, tag='Training', log_to_file=model_path_stem + ".log") # use this to log to file
sparse_to_dense = create_sparse_to_dense(input_vocab_dim)
for epoch in range(max_epochs):
print("Saving model to '%s'" % model_path(epoch))
s2smodel.save(model_path(epoch))
while total_samples < (epoch + 1) * epoch_size:
# get next minibatch of training data
mb_train = train_reader.next_minibatch(minibatch_size)
#trainer.train_minibatch(mb_train[train_reader.streams.features], mb_train[train_reader.streams.labels])
trainer.train_minibatch({
criterion.arguments[0]:
mb_train[train_reader.streams.features],
criterion.arguments[1]:
mb_train[train_reader.streams.labels]
})
progress_printer.update_with_trainer(
trainer, with_metric=True) # log progress
# every N MBs evaluate on a test sequence to visually show how we're doing
if mbs % eval_freq == 0:
mb_valid = valid_reader.next_minibatch(1)
# run an eval on the decoder output model (i.e. don't use the groundtruth)
e = model_greedy(mb_valid[valid_reader.streams.features])
print(
format_sequences(
sparse_to_dense(
mb_valid[valid_reader.streams.features]), i2w))
print("->")
print(format_sequences(e, i2w))
# debugging attention
if use_attention:
debug_attention(model_greedy,
mb_valid[valid_reader.streams.features])
total_samples += mb_train[train_reader.streams.labels].num_samples
mbs += 1
# log a summary of the stats for the epoch
progress_printer.epoch_summary(with_metric=True)
# done: save the final model
print("Saving final model to '%s'" % model_path(max_epochs))
s2smodel.save(model_path(max_epochs))
print("%d epochs complete." % max_epochs)
def train(reader, model, max_epochs):
# declare the model's input dimension, so that the saved model is usable
model.update_signature(Sequence[SparseTensor[vocab_size]])
#model.declare_args(vocab_size)
# criterion: (model args, labels) -> (loss, metric)
# here (query, slot_labels) -> (ce, errs)
criterion = create_criterion_function(model)
labels = reader.streams.slot_labels
#labels = reader.streams.intent_labels # for intent classification
#from cntk.logging.graph import plot
#plot(criterion, filename=data_dir + "/model.pdf")
# iteration parameters --needed here because learner schedule needs it
epoch_size = 36000
minibatch_size = 70
#epoch_size = 1000 ; max_epochs = 1 # uncomment for faster testing
# SGD parameters
learner = fsadagrad(
criterion.parameters,
lr=learning_rate_schedule([0.003] * 2 + [0.0015] * 12 + [0.0003],
UnitType.sample, epoch_size),
momentum=momentum_as_time_constant_schedule(minibatch_size /
-math.log(0.9)),
gradient_clipping_threshold_per_sample=15,
gradient_clipping_with_truncation=True)
# trainer
trainer = Trainer(None, criterion, learner)
# process minibatches and perform model training
log_number_of_parameters(model)
print()
progress_printer = ProgressPrinter(freq=100, first=10,
tag='Training') # more detailed logging
#progress_printer = ProgressPrinter(tag='Training')
t = 0
for epoch in range(max_epochs): # loop over epochs
peek(model, epoch) # log some interesting info
epoch_end = (epoch + 1) * epoch_size
while t < epoch_end: # loop over minibatches on the epoch
# BUGBUG: The change of minibatch_size parameter vv has no effect.
# TODO: change all examples to this pattern; then remove this comment
data = reader.next_minibatch(min(minibatch_size,
epoch_end - t)) # fetch minibatch
#trainer.train_minibatch(data[reader.streams.query], data[labels]) # update model with it
trainer.train_minibatch({
criterion.arguments[0]:
data[reader.streams.query],
criterion.arguments[1]:
data[labels]
}) # update model with it
t += data[labels].num_samples # count samples processed so far
progress_printer.update_with_trainer(
trainer, with_metric=True) # log progress
loss, metric, actual_samples = progress_printer.epoch_summary(
with_metric=True)
return loss, metric # return values from last epoch
def train(train_reader, valid_reader, vocab, i2w, s2smodel, max_epochs, epoch_size):
# Note: We would like to set the signature of 's2smodel' (s2smodel.update_signature()), but that will cause
# an error since the training criterion uses a reduced sequence axis for the labels.
# This is because it removes the initial <s> symbol. Hence, we must leave the model
# with unspecified input shapes and axes.
# create the training wrapper for the s2smodel, as well as the criterion function
model_train = create_model_train(s2smodel)
criterion = create_criterion_function(model_train)
# also wire in a greedy decoder so that we can properly log progress on a validation example
# This is not used for the actual training process.
model_greedy = create_model_greedy(s2smodel)
# This does not need to be done in training generally though
# Instantiate the trainer object to drive the model training
minibatch_size = 72
lr = 0.001 if use_attention else 0.005 # TODO: can we use the same value for both?
learner = fsadagrad(model_train.parameters,
lr = learning_rate_schedule([lr]*2+[lr/2]*3+[lr/4], UnitType.sample, epoch_size),
momentum = momentum_as_time_constant_schedule(1100),
gradient_clipping_threshold_per_sample=2.3,
gradient_clipping_with_truncation=True)
trainer = Trainer(None, criterion, learner)
# Get minibatches of sequences to train with and perform model training
total_samples = 0
mbs = 0
eval_freq = 100
# print out some useful training information
log_number_of_parameters(model_train) ; print()
progress_printer = ProgressPrinter(freq=30, tag='Training')
#progress_printer = ProgressPrinter(freq=30, tag='Training', log_to_file=model_path_stem + ".log") # use this to log to file
sparse_to_dense = create_sparse_to_dense(input_vocab_dim)
for epoch in range(max_epochs):
print("Saving model to '%s'" % model_path(epoch))
s2smodel.save(model_path(epoch))
while total_samples < (epoch+1) * epoch_size:
# get next minibatch of training data
mb_train = train_reader.next_minibatch(minibatch_size)
#trainer.train_minibatch(mb_train[train_reader.streams.features], mb_train[train_reader.streams.labels])
trainer.train_minibatch({criterion.arguments[0]: mb_train[train_reader.streams.features], criterion.arguments[1]: mb_train[train_reader.streams.labels]})
progress_printer.update_with_trainer(trainer, with_metric=True) # log progress
# every N MBs evaluate on a test sequence to visually show how we're doing
if mbs % eval_freq == 0:
mb_valid = valid_reader.next_minibatch(1)
# run an eval on the decoder output model (i.e. don't use the groundtruth)
e = model_greedy(mb_valid[valid_reader.streams.features])
print(format_sequences(sparse_to_dense(mb_valid[valid_reader.streams.features]), i2w))
print("->")
print(format_sequences(e, i2w))
# debugging attention
if use_attention:
debug_attention(model_greedy, mb_valid[valid_reader.streams.features])
total_samples += mb_train[train_reader.streams.labels].num_samples
mbs += 1
# log a summary of the stats for the epoch
progress_printer.epoch_summary(with_metric=True)
# done: save the final model
print("Saving final model to '%s'" % model_path(max_epochs))
s2smodel.save(model_path(max_epochs))
print("%d epochs complete." % max_epochs)
