def start_training(data: TrainingData,
model: Model,
train_params: TrainParams,
evaluators: List[Evaluator],
out: ModelDir,
notes: str = None,
initialize_from=None,
dry_run=False):
""" Train a model from scratch """
if initialize_from is None:
print("Initializing model at: " + out.dir)
model.init(data.get_train_corpus(), data.get_resource_loader())
# Else we assume the model has already completed its first phase of initialization
if not dry_run:
init(out, model, False)
_train(model, data, None, initialize_from, True, train_params, evaluators,
out, notes, dry_run)
def _train_async(model: Model,
data: TrainingData,
checkpoint: Union[str, None],
parameter_checkpoint: Union[str, None],
save_start: bool,
train_params: TrainParams,
evaluators: List[Evaluator],
out: ModelDir,
notes=None,
dry_run=False,
start_eval=False):
""" Train while encoding batches on a seperate thread and storing them in a tensorflow Queue, can
be much faster then using the feed_dict approach """
train = data.get_train()
eval_datasets = data.get_eval()
loader = data.get_resource_loader()
#pdb.set_trace()
print("Training on %d batches" % len(train))
print("Evaluation datasets: " +
" ".join("%s (%d)" % (name, len(data))
for name, data in eval_datasets.items()))
# spec the model for the given datasets
#pdb.set_trace()
model.set_inputs([train] + list(eval_datasets.values()), loader)
placeholders = model.get_placeholders()
train_queue = tf.FIFOQueue(train_params.async_encoding,
[x.dtype for x in placeholders],
name="train_queue")
evaluator_runner = AysncEvaluatorRunner(evaluators, model,
train_params.async_encoding)
#evaluator_runner = EvaluatorRunner(evaluators, model)
train_enqeue = train_queue.enqueue(placeholders)
train_close = train_queue.close(True)
is_train = tf.placeholder(tf.bool, ())
#pdb.set_trace()
input_tensors = tf.cond(is_train, lambda: train_queue.dequeue(),
lambda: evaluator_runner.eval_queue.dequeue())
#new input tensor with no eval
#input_tensors = train_queue.dequeue()
# tensorfow can't infer the shape for an unsized queue, so set it manually
for input_tensor, pl in zip(input_tensors, placeholders):
input_tensor.set_shape(pl.shape)
print("Init model...")
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True))
#pdb.set_trace()
with sess.as_default():
pred = model.get_predictions_for(dict(zip(placeholders,
input_tensors)))
evaluator_runner.set_input(pred) #az need to fix
#pdb.set_trace()
if parameter_checkpoint is not None:
print("Restoring parameters from %s" % parameter_checkpoint)
saver = tf.train.Saver()
saver.restore(sess, checkpoint)
saver = None
print("Setting up model prediction / tf...")
all_vars = tf.global_variables()
loss, summary_tensor, train_opt, global_step, weight_ema = _build_train_ops(
train_params)
#pdb.set_trace()
# Pre-compute tensors we need at evaluations time
eval_tensors = []
for ev in evaluators:
eval_tensors.append(ev.tensors_needed(pred))
#az rmeoved
if train_params.best_weights is not None:
lst = all_vars
if weight_ema is not None:
for x in lst:
v = weight_ema.average(x)
if v is not None:
lst.append(v)
best_weight_saver = tf.train.Saver(var_list=lst, max_to_keep=1)
cur_best = None
else:
best_weight_saver = None
cur_best = None
saver = tf.train.Saver(max_to_keep=train_params.max_checkpoints_to_keep)
summary_writer = tf.summary.FileWriter(out.log_dir)
# Load or initialize the model parameters
if checkpoint is not None:
print("Restoring from checkpoint...")
saver.restore(sess, checkpoint)
print("Loaded checkpoint: " + str(sess.run(global_step)))
else:
print("Initializing parameters...")
sess.run(tf.global_variables_initializer())
# Make sure no bugs occur that add to the graph in the train loop, that can cause (eventuall) OOMs
tf.get_default_graph().finalize()
if dry_run:
return
on_step = sess.run(global_step)
if save_start:
summary_writer.add_graph(sess.graph, global_step=on_step)
save_train_start(out.dir, data, sess.run(global_step), evaluators,
train_params, notes)
def enqueue_train():
try:
# feed data from the dataset iterator -> encoder -> queue
for epoch in range(train_params.num_epochs):
for batch in train.get_epoch():
feed_dict = model.encode(batch, True)
sess.run(train_enqeue, feed_dict)
except tf.errors.CancelledError:
# The queue_close operator has been called, exit gracefully
return
except Exception as e:
# Crashes the main thread with a queue exception
sess.run(train_close)
raise e
train_enqueue_thread = Thread(target=enqueue_train)
train_enqueue_thread.daemon = True # Ensure we exit the program on an excpetion
print("Start training!")
batch_time = 0
epoch_best = 0
dev_acc = []
train_acc = []
train_dict = {is_train: True}
eval_dict = {is_train: False}
#pdb.set_trace()
try:
train_enqueue_thread.start()
for epoch in range(train_params.num_epochs):
for batch_ix in range(len(train)):
t0 = time.perf_counter()
on_step = sess.run(global_step) + 1
get_summary = on_step % train_params.log_period == 0
if get_summary:
summary, _, batch_loss = sess.run(
[summary_tensor, train_opt, loss],
feed_dict=train_dict)
print('batch_loss is: ' + str(batch_loss))
else:
summary = None
_, batch_loss = sess.run([train_opt, loss],
feed_dict=train_dict)
#print(batch_loss)
#pdb.set_trace()
#with sess.as_default():
# temp= model.get_predictions_for(dict(zip(placeholders, input_tensors)))
#pdb.set_trace()
if np.isnan(batch_loss):
raise RuntimeError("NaN loss!")
batch_time += time.perf_counter() - t0
if summary is not None:
print("on epoch=%d batch=%d step=%d, time=%.3f" %
(epoch, batch_ix + 1, on_step, batch_time))
summary_writer.add_summary(
tf.Summary(value=[
tf.Summary.Value(tag="time",
simple_value=batch_time)
]), on_step)
summary_writer.add_summary(summary, on_step)
batch_time = 0
# occasional saving
if on_step % train_params.save_period == 0:
print("Checkpointing")
saver.save(sess,
join(out.save_dir,
"checkpoint-" + str(on_step)),
global_step=global_step)
# Occasional evaluation
#pdb.set_trace()
if (on_step % train_params.eval_period == 0) or start_eval:
print("Running evaluation...")
start_eval = False
t0 = time.perf_counter()
for name, data in eval_datasets.items():
n_samples = train_params.eval_samples.get(name)
#pdb.set_trace()
evaluation = evaluator_runner.run_evaluators(
sess, data, name, n_samples, eval_dict)
#pdb.set_trace()
for s in evaluation.to_summaries(name + "-"):
summary_writer.add_summary(s, on_step)
#pdb.set_trace()
# Maybe save as the best weights
if train_params.best_weights is not None and name == train_params.best_weights[
0]:
#pdb.set_trace()
val = evaluation.scalars[
train_params.best_weights[1]]
dev_acc.append(val)
if cur_best is None or val > cur_best:
epoch_best = epoch
send_email(
'epoch: ' + str(epoch_best) + 'acc: ' +
str(val), 'New Best')
print(
"Save weights with current best weights (%s vs %.5f)"
% ("None" if cur_best is None else
("%.5f" % cur_best), val))
best_weight_saver.save(sess,
join(
out.best_weight_dir,
"best"),
global_step=global_step)
cur_best = val
if (cur_best > 0.37):
email_text = 'Best accuracy for dev data: ' + (
'%.3f' %
cur_best) + ' <br> On epoch n: ' + str(
epoch_best) + ' out of: ' + str(
train_params.num_epochs
) + ' <br> Folder: ' + str(
out.save_dir)
email_title = 'Good News EveryOne!'
send_email(email_text, email_title)
print('Current accuracy for dev data: ' +
('%.3f' % val))
print('Best accuracy for dev data: ' +
('%.3f' % cur_best) + 'on epoch n:' +
str(epoch_best))
else:
val_train = evaluation.scalars[
train_params.best_weights[1]]
train_acc.append(val_train)
print('Current accuracy for train data: ' +
('%.3f' % val_train))
print("Evaluation took: %.3f seconds" %
(time.perf_counter() - t0))
finally:
sess.run(
train_close) # terminates the enqueue thread with an exception
train_enqueue_thread.join()
email_text = 'Finished ' + str(
train_params.num_epochs) + ' Best accuracy for dev data: ' + (
'%.3f' % cur_best) + ' <br> On epoch n: ' + str(
epoch_best) + ' <br> Acc for train data last: ' + (
'%.3f' % val_train) + ' <br> Folder: ' + str(out.save_dir)
email_title = 'Test Finished'
image_path = create_train_dev_plot(dev_acc, train_acc, out.save_dir)
send_email(email_text, email_title, image_path)
saver.save(sess,
relpath(join(out.save_dir, "checkpoint-" + str(on_step))),
global_step=global_step)
sess.close()
def _train(model: Model,
data: TrainingData,
checkpoint: Union[str, None],
parameter_checkpoint: Union[str, None],
save_start: bool,
train_params: TrainParams,
evaluators: List[Evaluator],
out: ModelDir,
notes=None,
dry_run=False,
start_eval=False):
if train_params.async_encoding:
_train_async(model, data, checkpoint, parameter_checkpoint, save_start,
train_params, evaluators, out, notes, dry_run, start_eval)
return
if train_params.best_weights is not None:
raise NotImplementedError
# spec the model for the current voc/input/batching
train = data.get_train()
eval_datasets = data.get_eval()
loader = data.get_resource_loader()
evaluator_runner = EvaluatorRunner(evaluators, model)
print("Training on %d batches" % len(train))
print("Evaluation datasets: " +
" ".join("%s (%d)" % (name, len(data))
for name, data in eval_datasets.items()))
print("Init model...")
model.set_inputs([train] + list(eval_datasets.values()), loader)
print("Setting up model prediction / tf...")
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True))
with sess.as_default():
pred = model.get_prediction()
evaluator_runner.set_input(pred)
if parameter_checkpoint is not None:
print("Restoring parameters from %s" % parameter_checkpoint)
saver = tf.train.Saver(tf.trainable_variables())
saver.restore(sess, parameter_checkpoint)
saver = None
loss, summary_tensor, train_opt, global_step, _ = _build_train_ops(
train_params)
# Pre-compute tensors we need at evaluations time
eval_tensors = []
for ev in evaluators:
eval_tensors.append(ev.tensors_needed(pred))
saver = tf.train.Saver(max_to_keep=train_params.max_checkpoints_to_keep)
summary_writer = tf.summary.FileWriter(out.log_dir)
# Load or initialize the model parameters
if checkpoint is not None:
print("Restoring training from checkpoint...")
saver.restore(sess, checkpoint)
print("Loaded checkpoint: " + str(sess.run(global_step)))
return
else:
if parameter_checkpoint is not None:
print("Initializing training variables...")
vars = [
x for x in tf.global_variables()
if x not in tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)
]
sess.run(tf.variables_initializer(vars))
else:
print("Initializing parameters...")
sess.run(tf.global_variables_initializer())
# Make sure no bugs occur that add to the graph in the train loop, that can cause (eventuall) OOMs
tf.get_default_graph().finalize()
print("Start training!")
on_step = sess.run(global_step)
if save_start:
summary_writer.add_graph(sess.graph, global_step=on_step)
save_train_start(out.dir, data, on_step, evaluators, train_params,
notes)
if train_params.eval_at_zero:
print("Running evaluation...")
start_eval = False
for name, data in eval_datasets.items():
n_samples = train_params.eval_samples.get(name)
evaluation = evaluator_runner.run_evaluators(
sess, data, name, n_samples)
for s in evaluation.to_summaries(name + "-"):
summary_writer.add_summary(s, on_step)
batch_time = 0
for epoch in range(train_params.num_epochs):
for batch_ix, batch in enumerate(train.get_epoch()):
t0 = time.perf_counter()
on_step = sess.run(
global_step
) + 1 # +1 because all calculations are done after step
get_summary = on_step % train_params.log_period == 0
encoded = model.encode(batch, True)
if get_summary:
summary, _, batch_loss = sess.run(
[summary_tensor, train_opt, loss], feed_dict=encoded)
else:
summary = None
_, batch_loss = sess.run([train_opt, loss], feed_dict=encoded)
if np.isnan(batch_loss):
raise RuntimeError("NaN loss!")
batch_time += time.perf_counter() - t0
if get_summary:
print("on epoch=%d batch=%d step=%d time=%.3f" %
(epoch, batch_ix + 1, on_step, batch_time))
summary_writer.add_summary(
tf.Summary(value=[
tf.Summary.Value(tag="time", simple_value=batch_time)
]), on_step)
summary_writer.add_summary(summary, on_step)
batch_time = 0
# occasional saving
if on_step % train_params.save_period == 0:
print("Checkpointing")
saver.save(sess,
join(out.save_dir, "checkpoint-" + str(on_step)),
global_step=global_step)
# Occasional evaluation
if (on_step % train_params.eval_period == 0) or start_eval:
print("Running evaluation...")
start_eval = False
t0 = time.perf_counter()
for name, data in eval_datasets.items():
n_samples = train_params.eval_samples.get(name)
evaluation = evaluator_runner.run_evaluators(
sess, data, name, n_samples)
for s in evaluation.to_summaries(name + "-"):
summary_writer.add_summary(s, on_step)
print("Evaluation took: %.3f seconds" %
(time.perf_counter() - t0))
saver.save(sess,
relpath(join(out.save_dir, "checkpoint-" + str(on_step))),
global_step=global_step)
sess.close()
