def on_epoch_begin(self, epoch, logs={}):
if epoch > 0 and epoch % self.eval_frequency == 0:
# Unhappy hack to work around h5py not being able to write to GCS.
# Force snapshots and saves to local filesystem, then copy them over to GCS.
model_path_glob = 'checkpoint.*'
if not self.job_dir.startswith("gs://"):
model_path_glob = os.path.join(self.job_dir, model_path_glob)
checkpoints = glob.glob(model_path_glob)
if len(checkpoints) > 0:
checkpoints.sort()
census_model = load_model(checkpoints[-1])
census_model = model.compile_model(census_model,
self.learning_rate)
loss, acc = census_model.evaluate_generator(
model.generator_input(self.eval_files,
chunk_size=CHUNK_SIZE),
steps=self.steps)
print '\nEvaluation epoch[{}] metrics[{:.2f}, {:.2f}] {}'.format(
epoch, loss, acc, census_model.metrics_names)
if self.job_dir.startswith("gs://"):
copy_file_to_gcs(self.job_dir, checkpoints[-1])
else:
print '\nEvaluation epoch[{}] (no checkpoints found)'.format(
epoch)
def dispatch(train_files, eval_files, job_dir, train_steps, eval_steps,
train_batch_size, eval_batch_size, learning_rate, eval_frequency,
first_layer_size, num_layers, scale_factor, eval_num_epochs,
num_epochs, checkpoint_epochs):
census_model = model.model_fn(INPUT_SIZE, CLASS_SIZE)
try:
os.makedirs(job_dir)
except:
pass
# Unhappy hack to work around h5py not being able to write to GCS.
# Force snapshots and saves to local filesystem, then copy them over to GCS.
checkpoint_path = FILE_PATH
if not job_dir.startswith("gs://"):
checkpoint_path = os.path.join(job_dir, checkpoint_path)
# Model checkpoint callback
checkpoint = keras.callbacks.ModelCheckpoint(checkpoint_path,
monitor='val_loss',
verbose=1,
period=checkpoint_epochs,
mode='max')
# Continuous eval callback
evaluation = ContinuousEval(eval_frequency,
eval_files,
learning_rate,
job_dir,
steps=train_steps)
# Tensorboard logs callback
tblog = keras.callbacks.TensorBoard(log_dir=os.path.join(job_dir, 'logs'),
histogram_freq=0,
write_graph=True,
embeddings_freq=0)
callbacks = [checkpoint, evaluation, tblog]
census_model.fit_generator(model.generator_input(train_files,
chunk_size=CHUNK_SIZE),
steps_per_epoch=train_steps,
epochs=num_epochs,
callbacks=callbacks)
# Unhappy hack to work around h5py not being able to write to GCS.
# Force snapshots and saves to local filesystem, then copy them over to GCS.
if job_dir.startswith("gs://"):
census_model.save(CENSUS_MODEL)
copy_file_to_gcs(job_dir, CENSUS_MODEL)
else:
census_model.save(os.path.join(job_dir, CENSUS_MODEL))
# Convert the Keras model to TensorFlow SavedModel
model.to_savedmodel(census_model, os.path.join(job_dir, 'export'))
def on_epoch_begin(self, epoch, logs={}):
if epoch > 0 and epoch % self.eval_frequency == 0:
checkpoints = glob.glob(os.path.join(self.job_dir, 'checkpoint.*'))
checkpoints.sort()
census_model = load_model(checkpoints[-1])
census_model = model.compile_model(census_model, self.learning_rate)
loss, acc = census_model.evaluate_generator(
model.generator_input(self.eval_files, chunk_size=CHUNK_SIZE),
steps=self.steps)
print('\nEvaluation epoch[{}] metrics[{:.2f}, {:.2f}] {}'.format(
epoch, loss, acc, census_model.metrics_names))
def dispatch(train_files, eval_files, job_dir, train_steps, eval_steps,
train_batch_size, eval_batch_size, learning_rate, eval_frequency,
first_layer_size, num_layers, scale_factor, eval_num_epochs,
num_epochs, checkpoint_epochs):
census_model = model.model_fn(INPUT_SIZE, CLASS_SIZE)
try:
os.makedirs(job_dir)
except:
pass
# Model checkpoint callback
checkpoint = keras.callbacks.ModelCheckpoint(os.path.join(
job_dir, FILE_PATH),
monitor='val_loss',
verbose=1,
period=checkpoint_epochs,
mode='max')
# Continuous eval callback
evaluation = ContinuousEval(eval_frequency, eval_files, learning_rate,
job_dir)
# Tensorboard logs callback
tblog = keras.callbacks.TensorBoard(log_dir=os.path.join(job_dir, 'logs'),
histogram_freq=0,
write_graph=True,
embeddings_freq=0)
# TODO: This needs to be fixed in h5py so that writes to GCS are possible
# Don't attempt to create checkpoints on Cloud ML Engine for now because
# h5py doesn't come with native GCS write capability
if job_dir.startswith('gs://'):
callbacks = [evaluation, tblog]
else:
callbacks = [checkpoint, evaluation, tblog]
start_time = time.time()
census_model.fit_generator(model.generator_input(train_files,
chunk_size=CHUNK_SIZE),
steps_per_epoch=train_steps,
epochs=num_epochs,
callbacks=callbacks)
print "\nTime used.", time.time() - start_time
census_model.save(os.path.join(job_dir, CENSUS_MODEL))
