def print_output_stats(self, output_stats):
table = OrderedDict()
for name, stats in six.iteritems(output_stats):
table.setdefault('name', []).append(name)
for key in stats:
table.setdefault(key, []).append(stats[key])
print('Output statistics:')
print(format_table(table))
print()
def print_output_stats(self, output_stats):
table = OrderedDict()
for name, stats in six.iteritems(output_stats):
table.setdefault('name', []).append(name)
for key in stats:
table.setdefault(key, []).append(stats[key])
print('Output statistics:')
print(format_table(table))
print()
def print_class_weights(self, class_weights):
table = OrderedDict()
for name, class_weight in six.iteritems(class_weights):
if not class_weight:
continue
column = []
for cla, weight in six.iteritems(class_weight):
column.append('%s=%.2f' % (cla, weight))
table[name] = column
if table:
print('Class weights:')
print(format_table(table))
print()
def print_class_weights(self, class_weights):
table = OrderedDict()
for name, class_weight in six.iteritems(class_weights):
if not class_weight:
continue
column = []
for cla, weight in six.iteritems(class_weight):
column.append('%s=%.2f' % (cla, weight))
table[name] = column
if table:
print('Class weights:')
print(format_table(table))
print()
def set_trainability(self, model):
opts = self.opts
trainable = []
not_trainable = []
if opts.fine_tune:
not_trainable.append('.*')
elif opts.train_models:
not_trainable.append('.*')
for name in opts.train_models:
trainable.append('%s/' % name)
if opts.freeze_filter:
not_trainable.append(mod.get_first_conv_layer(model.layers).name)
if not trainable and opts.trainable:
trainable = opts.trainable
if not not_trainable and opts.not_trainable:
not_trainable = opts.not_trainable
if not trainable and not not_trainable:
return
table = OrderedDict()
table['layer'] = []
table['trainable'] = []
for layer in model.layers:
if is_input_layer(layer) or is_output_layer(layer, model):
continue
if not hasattr(layer, 'trainable'):
continue
for regex in not_trainable:
if re.match(regex, layer.name):
layer.trainable = False
for regex in trainable:
if re.match(regex, layer.name):
layer.trainable = True
table['layer'].append(layer.name)
table['trainable'].append(layer.trainable)
print('Layer trainability:')
print(format_table(table))
print()
def set_trainability(self, model):
opts = self.opts
trainable = []
not_trainable = []
if opts.fine_tune:
not_trainable.append('.*')
elif opts.train_models:
not_trainable.append('.*')
for name in opts.train_models:
trainable.append('%s/' % name)
if opts.freeze_filter:
not_trainable.append(mod.get_first_conv_layer(model.layers).name)
if not trainable and opts.trainable:
trainable = opts.trainable
if not not_trainable and opts.not_trainable:
not_trainable = opts.not_trainable
if not trainable and not not_trainable:
return
table = OrderedDict()
table['layer'] = []
table['trainable'] = []
for layer in model.layers:
if is_input_layer(layer) or is_output_layer(layer, model):
continue
if not hasattr(layer, 'trainable'):
continue
for regex in not_trainable:
if re.match(regex, layer.name):
layer.trainable = False
for regex in trainable:
if re.match(regex, layer.name):
layer.trainable = True
table['layer'].append(layer.name)
table['trainable'].append(layer.trainable)
print('Layer trainability:')
print(format_table(table))
print()
def set_trainability(self, model):
opts = self.opts
trainable = [] #create a list
not_trainable = [] #create a list
if opts.fine_tune: #only train output layers
not_trainable.append('.*')
elif opts.train_models: #only train the specified model, including dna, cpg, and joint
not_trainable.append('.*')
for name in opts.train_models:
trainable.append('%s/' % name)
if opts.freeze_filter: #Exclude filter weights of first convolutional layer from training
not_trainable.append(mod.get_first_conv_layer(model.layers).name)
if not trainable and opts.trainable:
trainable = opts.trainable
if not not_trainable and opts.not_trainable:
not_trainable = opts.not_trainable
if not trainable and not not_trainable:
return
table = OrderedDict() #dictionary which remember the order
table['layer'] = []
table['trainable'] = []
for layer in model.layers:
if layer not in model.input_layers + model.output_layers:
if not hasattr(layer, 'trainable'):
continue
for regex in not_trainable:
if re.match(regex, layer.name):
layer.trainable = False
for regex in trainable:
if re.match(regex, layer.name):
layer.trainable = True
table['layer'].append(layer.name)
table['trainable'].append(layer.trainable)
print('Layer trainability:')
print(format_table(table))
print()
def main(self, name, opts):
logging.basicConfig(filename=opts.log_file,
format='%(levelname)s (%(asctime)s): %(message)s')
log = logging.getLogger(name)
if opts.verbose:
log.setLevel(logging.DEBUG)
else:
log.setLevel(logging.INFO)
if opts.seed is not None:
np.random.seed(opts.seed)
random.seed(opts.seed)
self.log = log
self.opts = opts
make_dir(opts.out_dir)
log.info('Building model ...')
model = self.build_model()
model.summary()
self.set_trainability(model)
if opts.filter_weights:
conv_layer = mod.get_first_conv_layer(model.layers)
log.info('Initializing filters of %s ...' % conv_layer.name)
self.init_filter_weights(opts.filter_weights, conv_layer)
mod.save_model(model, os.path.join(opts.out_dir, 'model.json'))
log.info('Computing output statistics ...')
output_names = []
for output_layer in model.output_layers:
output_names.append(output_layer.name)
output_stats = OrderedDict()
if opts.no_class_weights:
class_weights = None
else:
class_weights = OrderedDict()
for name in output_names:
output = hdf.read(opts.train_files,
'outputs/%s' % name,
nb_sample=opts.nb_train_sample)
output = list(output.values())[0]
output_stats[name] = get_output_stats(output)
if class_weights is not None:
class_weights[name] = get_output_class_weights(name, output)
self.print_output_stats(output_stats)
if class_weights:
self.print_class_weights(class_weights)
output_weights = None
if opts.output_weights:
log.info('Initializing output weights ...')
output_weights = get_output_weights(output_names,
opts.output_weights)
print('Output weights:')
for output_name in output_names:
if output_name in output_weights:
print('%s: %.2f' %
(output_name, output_weights[output_name]))
print()
self.metrics = dict()
for output_name in output_names:
self.metrics[output_name] = get_metrics(output_name)
optimizer = Adam(lr=opts.learning_rate)
model.compile(optimizer=optimizer,
loss=mod.get_objectives(output_names),
loss_weights=output_weights,
metrics=self.metrics)
log.info('Loading data ...')
replicate_names = dat.get_replicate_names(opts.train_files[0],
regex=opts.replicate_names,
nb_key=opts.nb_replicate)
data_reader = mod.data_reader_from_model(
model, replicate_names=replicate_names)
nb_train_sample = dat.get_nb_sample(opts.train_files,
opts.nb_train_sample)
train_data = data_reader(opts.train_files,
class_weights=class_weights,
batch_size=opts.batch_size,
nb_sample=nb_train_sample,
shuffle=True,
loop=True)
if opts.val_files:
nb_val_sample = dat.get_nb_sample(opts.val_files,
opts.nb_val_sample)
val_data = data_reader(opts.val_files,
batch_size=opts.batch_size,
nb_sample=nb_val_sample,
shuffle=False,
loop=True)
else:
val_data = None
nb_val_sample = None
log.info('Initializing callbacks ...')
callbacks = self.get_callbacks()
log.info('Training model ...')
print()
print('Training samples: %d' % nb_train_sample)
if nb_val_sample:
print('Validation samples: %d' % nb_val_sample)
model.fit_generator(train_data,
nb_train_sample,
opts.nb_epoch,
callbacks=callbacks,
validation_data=val_data,
nb_val_samples=nb_val_sample,
max_q_size=opts.data_q_size,
nb_worker=opts.data_nb_worker,
verbose=0)
print('\nTraining set performance:')
print(
format_table(self.perf_logger.epoch_logs, precision=LOG_PRECISION))
if self.perf_logger.val_epoch_logs:
print('\nValidation set performance:')
print(
format_table(self.perf_logger.val_epoch_logs,
precision=LOG_PRECISION))
# Restore model with highest validation performance
filename = os.path.join(opts.out_dir, 'model_weights_val.h5')
if os.path.isfile(filename):
model.load_weights(filename)
# Delete metrics since they cause problems when loading the model
# from HDF5 file. Metrics can be loaded from json + weights file.
model.metrics = None
model.metrics_names = None
model.metrics_tensors = None
model.save(os.path.join(opts.out_dir, 'model.h5'))
log.info('Done!')
return 0
def main(self, name, opts):
logging.basicConfig(filename=opts.log_file,
format='%(levelname)s (%(asctime)s): %(message)s')
log = logging.getLogger(name)
if opts.verbose:
log.setLevel(logging.DEBUG)
else:
log.setLevel(logging.INFO)
if opts.seed is not None:
np.random.seed(opts.seed)
random.seed(opts.seed)
self.log = log
self.opts = opts
make_dir(opts.out_dir)
log.info('Building model ...')
model = self.build_model()
model.summary()
self.set_trainability(model)
if opts.filter_weights:
conv_layer = mod.get_first_conv_layer(model.layers)
log.info('Initializing filters of %s ...' % conv_layer.name)
self.init_filter_weights(opts.filter_weights, conv_layer)
mod.save_model(model, os.path.join(opts.out_dir, 'model.json'))
log.info('Computing output statistics ...')
output_names = model.output_names
output_stats = OrderedDict()
if opts.no_class_weights:
class_weights = None
else:
class_weights = OrderedDict()
for name in output_names:
output = hdf.read(opts.train_files, 'outputs/%s' % name,
nb_sample=opts.nb_train_sample)
output = list(output.values())[0]
output_stats[name] = get_output_stats(output)
if class_weights is not None:
class_weights[name] = get_output_class_weights(name, output)
self.print_output_stats(output_stats)
if class_weights:
self.print_class_weights(class_weights)
output_weights = None
if opts.output_weights:
log.info('Initializing output weights ...')
output_weights = get_output_weights(output_names,
opts.output_weights)
print('Output weights:')
for output_name in output_names:
if output_name in output_weights:
print('%s: %.2f' % (output_name,
output_weights[output_name]))
print()
self.metrics = dict()
for output_name in output_names:
self.metrics[output_name] = get_metrics(output_name)
optimizer = Adam(lr=opts.learning_rate)
model.compile(optimizer=optimizer,
loss=mod.get_objectives(output_names),
loss_weights=output_weights,
metrics=self.metrics)
log.info('Loading data ...')
replicate_names = dat.get_replicate_names(
opts.train_files[0],
regex=opts.replicate_names,
nb_key=opts.nb_replicate)
data_reader = mod.data_reader_from_model(
model, replicate_names=replicate_names)
nb_train_sample = dat.get_nb_sample(opts.train_files,
opts.nb_train_sample)
train_data = data_reader(opts.train_files,
class_weights=class_weights,
batch_size=opts.batch_size,
nb_sample=nb_train_sample,
shuffle=True,
loop=True)
if opts.val_files:
nb_val_sample = dat.get_nb_sample(opts.val_files,
opts.nb_val_sample)
val_data = data_reader(opts.val_files,
batch_size=opts.batch_size,
nb_sample=nb_val_sample,
shuffle=False,
loop=True)
else:
val_data = None
nb_val_sample = None
log.info('Initializing callbacks ...')
callbacks = self.get_callbacks()
log.info('Training model ...')
print()
print('Training samples: %d' % nb_train_sample)
if nb_val_sample:
print('Validation samples: %d' % nb_val_sample)
model.fit_generator(
train_data,
steps_per_epoch=nb_train_sample // opts.batch_size,
epochs=opts.nb_epoch,
callbacks=callbacks,
validation_data=val_data,
validation_steps=nb_val_sample // opts.batch_size,
max_queue_size=opts.data_q_size,
workers=opts.data_nb_worker,
verbose=0)
print('\nTraining set performance:')
print(format_table(self.perf_logger.epoch_logs,
precision=LOG_PRECISION))
if self.perf_logger.val_epoch_logs:
print('\nValidation set performance:')
print(format_table(self.perf_logger.val_epoch_logs,
precision=LOG_PRECISION))
# Restore model with highest validation performance
filename = os.path.join(opts.out_dir, 'model_weights_val.h5')
if os.path.isfile(filename):
model.load_weights(filename)
# Delete metrics since they cause problems when loading the model
# from HDF5 file. Metrics can be loaded from json + weights file.
model.metrics = None
model.metrics_names = None
model.metrics_tensors = None
model.save(os.path.join(opts.out_dir, 'model.h5'))
log.info('Done!')
return 0
