def run(gParameters):
print('Params:', gParameters)
file_train = gParameters['train_data']
file_test = gParameters['test_data']
url = gParameters['data_url']
train_file = candle.get_file(file_train,
url + file_train,
cache_subdir='Pilot1')
test_file = candle.get_file(file_test,
url + file_test,
cache_subdir='Pilot1')
X_train, Y_train, X_test, Y_test = load_data(train_file, test_file,
gParameters)
print('X_train shape:', X_train.shape)
print('X_test shape:', X_test.shape)
print('Y_train shape:', Y_train.shape)
print('Y_test shape:', Y_test.shape)
x_train_len = X_train.shape[1]
# this reshaping is critical for the Conv1D to work
X_train = np.expand_dims(X_train, axis=2)
X_test = np.expand_dims(X_test, axis=2)
print('X_train shape:', X_train.shape)
print('X_test shape:', X_test.shape)
# Have to add this line or else ALW on 2020-11-15 finds Supervisor jobs using canonically CANDLE-compliant model scripts die as soon as a particular task is used a second time:
# EXCEPTION:
# InvalidArgumentError() ...
# File "<string>", line 23, in <module>
# File "/gpfs/alpine/med106/world-shared/candle/2020-11-11/checkouts/Supervisor/workflows/common/python/model_runner.py", line 241, in run_model
# result, history = run(hyper_parameter_map, obj_return)
# File "/gpfs/alpine/med106/world-shared/candle/2020-11-11/checkouts/Supervisor/workflows/common/python/model_runner.py", line 169, in run
# history = pkg.run(params)
# File "/gpfs/alpine/med106/world-shared/candle/2020-11-11/checkouts/Benchmarks/Pilot1/NT3/nt3_candle_wrappers_baseline_keras2.py", line 211, in run
# callbacks=[csv_logger, reduce_lr, candleRemoteMonitor, timeoutMonitor])
# File "/gpfs/alpine/world-shared/med106/sw/condaenv-200408/lib/python3.6/site-packages/keras/engine/training.py", line 1178, in fit
# validation_freq=validation_freq)
# File "/gpfs/alpine/world-shared/med106/sw/condaenv-200408/lib/python3.6/site-packages/keras/engine/training_arrays.py", line 204, in fit_loop
# outs = fit_function(ins_batch)
# File "/gpfs/alpine/world-shared/med106/sw/condaenv-200408/lib/python3.6/site-packages/keras/backend/tensorflow_backend.py", line 2979, in __call__
# return self._call(inputs)
# File "/gpfs/alpine/world-shared/med106/sw/condaenv-200408/lib/python3.6/site-packages/keras/backend/tensorflow_backend.py", line 2933, in _call
# session)
# File "/gpfs/alpine/world-shared/med106/sw/condaenv-200408/lib/python3.6/site-packages/keras/backend/tensorflow_backend.py", line 2885, in _make_callable
# callable_fn = session._make_callable_from_options(callable_opts)
# File "/gpfs/alpine/world-shared/med106/sw/condaenv-200408/lib/python3.6/site-packages/tensorflow_core/python/client/session.py", line 1505, in _make_callable_from_options
# return BaseSession._Callable(self, callable_options)
# File "/gpfs/alpine/world-shared/med106/sw/condaenv-200408/lib/python3.6/site-packages/tensorflow_core/python/client/session.py", line 1460, in __init__
# session._session, options_ptr)
K.clear_session()
model = Sequential()
layer_list = list(range(0, len(gParameters['conv']), 3))
for _, i in enumerate(layer_list):
filters = gParameters['conv'][i]
filter_len = gParameters['conv'][i + 1]
stride = gParameters['conv'][i + 2]
print(int(i / 3), filters, filter_len, stride)
if gParameters['pool']:
pool_list = gParameters['pool']
if type(pool_list) != list:
pool_list = list(pool_list)
if filters <= 0 or filter_len <= 0 or stride <= 0:
break
if 'locally_connected' in gParameters:
model.add(
LocallyConnected1D(filters,
filter_len,
strides=stride,
padding='valid',
input_shape=(x_train_len, 1)))
else:
# input layer
if i == 0:
model.add(
Conv1D(filters=filters,
kernel_size=filter_len,
strides=stride,
padding='valid',
input_shape=(x_train_len, 1)))
else:
model.add(
Conv1D(filters=filters,
kernel_size=filter_len,
strides=stride,
padding='valid'))
model.add(Activation(gParameters['activation']))
if gParameters['pool']:
model.add(MaxPooling1D(pool_size=pool_list[int(i / 3)]))
model.add(Flatten())
for layer in gParameters['dense']:
if layer:
model.add(Dense(layer))
model.add(Activation(gParameters['activation']))
if gParameters['dropout']:
model.add(Dropout(gParameters['dropout']))
model.add(Dense(gParameters['classes']))
model.add(Activation(gParameters['out_activation']))
# Reference case
# model.add(Conv1D(filters=128, kernel_size=20, strides=1, padding='valid', input_shape=(P, 1)))
# model.add(Activation('relu'))
# model.add(MaxPooling1D(pool_size=1))
# model.add(Conv1D(filters=128, kernel_size=10, strides=1, padding='valid'))
# model.add(Activation('relu'))
# model.add(MaxPooling1D(pool_size=10))
# model.add(Flatten())
# model.add(Dense(200))
# model.add(Activation('relu'))
# model.add(Dropout(0.1))
# model.add(Dense(20))
# model.add(Activation('relu'))
# model.add(Dropout(0.1))
# model.add(Dense(CLASSES))
# model.add(Activation('softmax'))
kerasDefaults = candle.keras_default_config()
# Define optimizer
optimizer = candle.build_optimizer(gParameters['optimizer'],
gParameters['learning_rate'],
kerasDefaults)
model.summary()
model.compile(loss=gParameters['loss'],
optimizer=optimizer,
metrics=[gParameters['metrics']])
output_dir = gParameters['output_dir']
if not os.path.exists(output_dir):
os.makedirs(output_dir)
# calculate trainable and non-trainable params
gParameters.update(candle.compute_trainable_params(model))
# set up a bunch of callbacks to do work during model training..
model_name = gParameters['model_name']
path = '{}/{}.autosave.model.h5'.format(output_dir, model_name)
# checkpointer = ModelCheckpoint(filepath=path, verbose=1, save_weights_only=False, save_best_only=True)
csv_logger = CSVLogger('{}/training.log'.format(output_dir))
reduce_lr = ReduceLROnPlateau(monitor='val_loss',
factor=0.1,
patience=10,
verbose=1,
mode='auto',
epsilon=0.0001,
cooldown=0,
min_lr=0)
candleRemoteMonitor = candle.CandleRemoteMonitor(params=gParameters)
timeoutMonitor = candle.TerminateOnTimeOut(gParameters['timeout'])
history = model.fit(
X_train,
Y_train,
batch_size=gParameters['batch_size'],
epochs=gParameters['epochs'],
verbose=1,
validation_data=(X_test, Y_test),
callbacks=[csv_logger, reduce_lr, candleRemoteMonitor, timeoutMonitor])
score = model.evaluate(X_test, Y_test, verbose=0)
if False:
print('Test score:', score[0])
print('Test accuracy:', score[1])
# serialize model to JSON
model_json = model.to_json()
with open("{}/{}.model.json".format(output_dir, model_name),
"w") as json_file:
json_file.write(model_json)
# serialize model to YAML
model_yaml = model.to_yaml()
with open("{}/{}.model.yaml".format(output_dir, model_name),
"w") as yaml_file:
yaml_file.write(model_yaml)
# serialize weights to HDF5
model.save_weights("{}/{}.weights.h5".format(output_dir, model_name))
print("Saved model to disk")
# load json and create model
json_file = open('{}/{}.model.json'.format(output_dir, model_name),
'r')
loaded_model_json = json_file.read()
json_file.close()
loaded_model_json = model_from_json(loaded_model_json)
# load yaml and create model
yaml_file = open('{}/{}.model.yaml'.format(output_dir, model_name),
'r')
loaded_model_yaml = yaml_file.read()
yaml_file.close()
loaded_model_yaml = model_from_yaml(loaded_model_yaml)
# load weights into new model
loaded_model_json.load_weights('{}/{}.weights.h5'.format(
output_dir, model_name))
print("Loaded json model from disk")
# evaluate json loaded model on test data
loaded_model_json.compile(loss=gParameters['loss'],
optimizer=gParameters['optimizer'],
metrics=[gParameters['metrics']])
score_json = loaded_model_json.evaluate(X_test, Y_test, verbose=0)
print('json Test score:', score_json[0])
print('json Test accuracy:', score_json[1])
print("json %s: %.2f%%" %
(loaded_model_json.metrics_names[1], score_json[1] * 100))
# load weights into new model
loaded_model_yaml.load_weights('{}/{}.weights.h5'.format(
output_dir, model_name))
print("Loaded yaml model from disk")
# evaluate loaded model on test data
loaded_model_yaml.compile(loss=gParameters['loss'],
optimizer=gParameters['optimizer'],
metrics=[gParameters['metrics']])
score_yaml = loaded_model_yaml.evaluate(X_test, Y_test, verbose=0)
print('yaml Test score:', score_yaml[0])
print('yaml Test accuracy:', score_yaml[1])
print("yaml %s: %.2f%%" %
(loaded_model_yaml.metrics_names[1], score_yaml[1] * 100))
return history
def build_model(gParameters, kerasDefaults,
shared_nnet_spec, individual_nnet_spec,
input_dim, Y_train, Y_test,
verbose=False):
labels_train = []
labels_test = []
n_out_nodes = []
for l in range( len( Y_train ) ):
truth_train = np.array( Y_train[l], dtype='int32' )
truth_test = np.array( Y_test[l], dtype='int32' )
mv = int( np.max( truth_train ) )
label_train = np.zeros( ( len( truth_train ), mv + 1 ) )
for i in range( len( truth_train ) ):
label_train[ i, truth_train[ i ] ] = 1
label_test = np.zeros( ( len( truth_test ), mv + 1 ) )
for i in range( len(truth_test) ):
label_test[ i, truth_test[ i ] ] = 1
labels_train.append( label_train )
labels_test.append( label_test )
n_out_nodes.append( mv + 1 )
shared_layers = []
# input layer
layer = Input( shape = ( input_dim, ), name= 'input' )
shared_layers.append( layer )
# shared layers
for k in range( len( shared_nnet_spec ) ):
layer = Dense( shared_nnet_spec[ k ], activation=gParameters['activation'],
name= 'shared_layer_' + str( k ) )( shared_layers[ -1 ] )
if gParameters['drop'] > 0:
layer = Dropout( gParameters['drop'] )( shared_layers[ -1 ] )
shared_layers.append( layer )
# individual layers
indiv_layers_arr = []
models = []
trainable_count = 0
non_trainable_count = 0
for l in range( len( individual_nnet_spec ) ):
indiv_layers = [ shared_layers[-1] ]
for k in range( len( individual_nnet_spec[l] ) + 1 ):
if k < len( individual_nnet_spec[l] ):
layer = Dense( individual_nnet_spec[l][k], activation=gParameters['activation'],
name= 'indiv_layer_' + str( l ) + '_' + str( k ) )( indiv_layers[-1] )
indiv_layers.append( layer )
if gParameters['drop'] > 0:
layer = Dropout( gParameters['drop'] )( indiv_layers[-1] )
indiv_layers.append( layer )
else:
layer = Dense( n_out_nodes[l], activation=gParameters['out_activation'],
name= 'out_' + str( l ) )( indiv_layers[-1] )
indiv_layers.append( layer )
indiv_layers_arr.append( indiv_layers )
model = Model( inputs=[shared_layers[0]], outputs=[indiv_layers[-1]] )
# calculate trainable/non-trainable param count for each model
param_counts = candle.compute_trainable_params(model)
trainable_count += param_counts['trainable_params']
non_trainable_count += param_counts['non_trainable_params']
models.append( model )
# capture total param counts
gParameters['trainable_params'] = trainable_count
gParameters['non_trainable_params'] = non_trainable_count
gParameters['total_params'] = trainable_count + non_trainable_count
# Define optimizer
optimizer = candle.build_optimizer(gParameters['optimizer'],
gParameters['learning_rate'],
kerasDefaults)
# DEBUG - verify
if verbose:
for k in range( len( models ) ):
model = models[k]
print('Model: ', k)
model.summary()
for k in range( len( models ) ):
model = models[ k ]
model.compile( loss=gParameters['loss'], optimizer=optimizer, metrics=[gParameters['metrics']] )
return models, labels_train, labels_test
def run(params):
args = Struct(**params)
set_seed(args.rng_seed)
ext = extension_from_parameters(args)
verify_path(args.save_path)
prefix = args.save_path + ext
logfile = args.logfile if args.logfile else prefix + '.log'
set_up_logger(logfile, args.verbose)
logger.info('Params: {}'.format(params))
if (len(args.gpus) > 0):
import tensorflow as tf
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.gpu_options.visible_device_list = ",".join(map(str, args.gpus))
K.set_session(tf.Session(config=config))
loader = CombinedDataLoader(seed=args.rng_seed)
loader.load(
cache=args.cache,
ncols=args.feature_subsample,
agg_dose=args.agg_dose,
cell_features=args.cell_features,
drug_features=args.drug_features,
drug_median_response_min=args.drug_median_response_min,
drug_median_response_max=args.drug_median_response_max,
use_landmark_genes=args.use_landmark_genes,
use_filtered_genes=args.use_filtered_genes,
cell_feature_subset_path=args.cell_feature_subset_path
or args.feature_subset_path,
drug_feature_subset_path=args.drug_feature_subset_path
or args.feature_subset_path,
preprocess_rnaseq=args.preprocess_rnaseq,
single=args.single,
train_sources=args.train_sources,
test_sources=args.test_sources,
embed_feature_source=not args.no_feature_source,
encode_response_source=not args.no_response_source,
)
target = args.agg_dose or 'Growth'
val_split = args.validation_split
train_split = 1 - val_split
if args.export_csv:
fname = args.export_csv
loader.partition_data(cv_folds=args.cv,
train_split=train_split,
val_split=val_split,
cell_types=args.cell_types,
by_cell=args.by_cell,
by_drug=args.by_drug,
cell_subset_path=args.cell_subset_path,
drug_subset_path=args.drug_subset_path)
train_gen = CombinedDataGenerator(loader,
batch_size=args.batch_size,
shuffle=args.shuffle)
val_gen = CombinedDataGenerator(loader,
partition='val',
batch_size=args.batch_size,
shuffle=args.shuffle)
x_train_list, y_train = train_gen.get_slice(size=train_gen.size,
dataframe=True,
single=args.single)
x_val_list, y_val = val_gen.get_slice(size=val_gen.size,
dataframe=True,
single=args.single)
df_train = pd.concat([y_train] + x_train_list, axis=1)
df_val = pd.concat([y_val] + x_val_list, axis=1)
df = pd.concat([df_train, df_val]).reset_index(drop=True)
if args.growth_bins > 1:
df = uno_data.discretize(df, 'Growth', bins=args.growth_bins)
df.to_csv(fname, sep='\t', index=False, float_format="%.3g")
return
if args.export_data:
fname = args.export_data
loader.partition_data(cv_folds=args.cv,
train_split=train_split,
val_split=val_split,
cell_types=args.cell_types,
by_cell=args.by_cell,
by_drug=args.by_drug,
cell_subset_path=args.cell_subset_path,
drug_subset_path=args.drug_subset_path)
train_gen = CombinedDataGenerator(loader,
batch_size=args.batch_size,
shuffle=args.shuffle)
val_gen = CombinedDataGenerator(loader,
partition='val',
batch_size=args.batch_size,
shuffle=args.shuffle)
store = pd.HDFStore(fname, complevel=9, complib='blosc:snappy')
config_min_itemsize = {'Sample': 30, 'Drug1': 10}
if not args.single:
config_min_itemsize['Drug2'] = 10
for partition in ['train', 'val']:
gen = train_gen if partition == 'train' else val_gen
for i in range(gen.steps):
x_list, y = gen.get_slice(size=args.batch_size,
dataframe=True,
single=args.single)
for j, input_feature in enumerate(x_list):
input_feature.columns = [''] * len(input_feature.columns)
store.append('x_{}_{}'.format(partition, j),
input_feature.astype('float32'),
format='table',
data_column=True)
store.append('y_{}'.format(partition),
y.astype({target: 'float32'}),
format='table',
data_column=True,
min_itemsize=config_min_itemsize)
logger.info('Generating {} dataset. {} / {}'.format(
partition, i, gen.steps))
store.close()
logger.info('Completed generating {}'.format(fname))
return
loader.partition_data(cv_folds=args.cv,
train_split=train_split,
val_split=val_split,
cell_types=args.cell_types,
by_cell=args.by_cell,
by_drug=args.by_drug,
cell_subset_path=args.cell_subset_path,
drug_subset_path=args.drug_subset_path)
model = build_model(loader, args)
logger.info('Combined model:')
model.summary(print_fn=logger.info)
# plot_model(model, to_file=prefix+'.model.png', show_shapes=True)
if args.cp:
model_json = model.to_json()
with open(prefix + '.model.json', 'w') as f:
print(model_json, file=f)
def warmup_scheduler(epoch):
lr = args.learning_rate or base_lr * args.batch_size / 100
if epoch <= 5:
K.set_value(model.optimizer.lr,
(base_lr * (5 - epoch) + lr * epoch) / 5)
logger.debug('Epoch {}: lr={:.5g}'.format(
epoch, K.get_value(model.optimizer.lr)))
return K.get_value(model.optimizer.lr)
df_pred_list = []
cv_ext = ''
cv = args.cv if args.cv > 1 else 1
for fold in range(cv):
if args.cv > 1:
logger.info('Cross validation fold {}/{}:'.format(fold + 1, cv))
cv_ext = '.cv{}'.format(fold + 1)
template_model = build_model(loader, args, silent=True)
if args.initial_weights:
logger.info("Loading weights from {}".format(args.initial_weights))
template_model.load_weights(args.initial_weights)
if len(args.gpus) > 1:
from keras.utils import multi_gpu_model
gpu_count = len(args.gpus)
logger.info("Multi GPU with {} gpus".format(gpu_count))
model = multi_gpu_model(template_model,
cpu_merge=False,
gpus=gpu_count)
else:
model = template_model
optimizer = optimizers.deserialize({
'class_name': args.optimizer,
'config': {}
})
base_lr = args.base_lr or K.get_value(optimizer.lr)
if args.learning_rate:
K.set_value(optimizer.lr, args.learning_rate)
model.compile(loss=args.loss, optimizer=optimizer, metrics=[mae, r2])
# calculate trainable and non-trainable params
params.update(candle.compute_trainable_params(model))
candle_monitor = candle.CandleRemoteMonitor(params=params)
timeout_monitor = candle.TerminateOnTimeOut(params['timeout'])
reduce_lr = ReduceLROnPlateau(monitor='val_loss',
factor=0.5,
patience=5,
min_lr=0.00001)
warmup_lr = LearningRateScheduler(warmup_scheduler)
checkpointer = MultiGPUCheckpoint(prefix + cv_ext + '.model.h5',
save_best_only=True)
tensorboard = TensorBoard(
log_dir="tb/{}{}{}".format(args.tb_prefix, ext, cv_ext))
history_logger = LoggingCallback(logger.debug)
callbacks = [candle_monitor, timeout_monitor, history_logger]
if args.reduce_lr:
callbacks.append(reduce_lr)
if args.warmup_lr:
callbacks.append(warmup_lr)
if args.cp:
callbacks.append(checkpointer)
if args.tb:
callbacks.append(tensorboard)
if args.save_weights:
callbacks.append(
SimpleWeightSaver(args.save_path + '/' + args.save_weights))
if args.use_exported_data is not None:
train_gen = DataFeeder(filename=args.use_exported_data,
batch_size=args.batch_size,
shuffle=args.shuffle,
single=args.single,
agg_dose=args.agg_dose)
val_gen = DataFeeder(partition='val',
filename=args.use_exported_data,
batch_size=args.batch_size,
shuffle=args.shuffle,
single=args.single,
agg_dose=args.agg_dose)
else:
train_gen = CombinedDataGenerator(loader,
fold=fold,
batch_size=args.batch_size,
shuffle=args.shuffle,
single=args.single)
val_gen = CombinedDataGenerator(loader,
partition='val',
fold=fold,
batch_size=args.batch_size,
shuffle=args.shuffle,
single=args.single)
df_val = val_gen.get_response(copy=True)
y_val = df_val[target].values
y_shuf = np.random.permutation(y_val)
log_evaluation(evaluate_prediction(y_val, y_shuf),
description='Between random pairs in y_val:')
if args.no_gen:
x_train_list, y_train = train_gen.get_slice(size=train_gen.size,
single=args.single)
x_val_list, y_val = val_gen.get_slice(size=val_gen.size,
single=args.single)
history = model.fit(x_train_list,
y_train,
batch_size=args.batch_size,
epochs=args.epochs,
callbacks=callbacks,
validation_data=(x_val_list, y_val))
else:
logger.info('Data points per epoch: train = %d, val = %d',
train_gen.size, val_gen.size)
logger.info('Steps per epoch: train = %d, val = %d',
train_gen.steps, val_gen.steps)
history = model.fit_generator(train_gen,
train_gen.steps,
epochs=args.epochs,
callbacks=callbacks,
validation_data=val_gen,
validation_steps=val_gen.steps)
if args.no_gen:
y_val_pred = model.predict(x_val_list, batch_size=args.batch_size)
else:
val_gen.reset()
y_val_pred = model.predict_generator(val_gen, val_gen.steps + 1)
y_val_pred = y_val_pred[:val_gen.size]
y_val_pred = y_val_pred.flatten()
scores = evaluate_prediction(y_val, y_val_pred)
log_evaluation(scores)
# df_val = df_val.assign(PredictedGrowth=y_val_pred, GrowthError=y_val_pred - y_val)
df_val['Predicted' + target] = y_val_pred
df_val[target + 'Error'] = y_val_pred - y_val
df_pred_list.append(df_val)
if hasattr(history, 'loss'):
plot_history(prefix, history, 'loss')
if hasattr(history, 'r2'):
plot_history(prefix, history, 'r2')
pred_fname = prefix + '.predicted.tsv'
df_pred = pd.concat(df_pred_list)
if args.agg_dose:
if args.single:
df_pred.sort_values(['Sample', 'Drug1', target], inplace=True)
else:
df_pred.sort_values(['Source', 'Sample', 'Drug1', 'Drug2', target],
inplace=True)
else:
if args.single:
df_pred.sort_values(['Sample', 'Drug1', 'Dose1', 'Growth'],
inplace=True)
else:
df_pred.sort_values(
['Sample', 'Drug1', 'Drug2', 'Dose1', 'Dose2', 'Growth'],
inplace=True)
df_pred.to_csv(pred_fname, sep='\t', index=False, float_format='%.4g')
if args.cv > 1:
scores = evaluate_prediction(df_pred[target],
df_pred['Predicted' + target])
log_evaluation(scores, description='Combining cross validation folds:')
for test_source in loader.test_sep_sources:
test_gen = CombinedDataGenerator(loader,
partition='test',
batch_size=args.batch_size,
source=test_source)
df_test = test_gen.get_response(copy=True)
y_test = df_test[target].values
n_test = len(y_test)
if n_test == 0:
continue
if args.no_gen:
x_test_list, y_test = test_gen.get_slice(size=test_gen.size,
single=args.single)
y_test_pred = model.predict(x_test_list,
batch_size=args.batch_size)
else:
y_test_pred = model.predict_generator(
test_gen.flow(single=args.single), test_gen.steps)
y_test_pred = y_test_pred[:test_gen.size]
y_test_pred = y_test_pred.flatten()
scores = evaluate_prediction(y_test, y_test_pred)
log_evaluation(scores,
description='Testing on data from {} ({})'.format(
test_source, n_test))
if K.backend() == 'tensorflow':
K.clear_session()
logger.handlers = []
return history
optimizer = candle.build_optimizer(candle_params['optimizer'],
candle_params['learning_rate'],
kerasDefaults)
model.summary()
model.compile(loss=candle_params['loss'],
optimizer=optimizer,
metrics=[candle_params['metrics']])
output_dir = candle_params['save']
if not os.path.exists(output_dir):
os.makedirs(output_dir)
# calculate trainable and non-trainable params
candle_params.update(candle.compute_trainable_params(model))
# set up a bunch of callbacks to do work during model training..
model_name = candle_params['model_name']
path = '{}/{}.autosave.model.h5'.format(output_dir, model_name)
csv_logger = CSVLogger('{}/training.log'.format(output_dir))
reduce_lr = ReduceLROnPlateau(monitor='val_loss', factor=0.1, patience=10, verbose=1, mode='auto', epsilon=0.0001, cooldown=0, min_lr=0)
candleRemoteMonitor = candle.CandleRemoteMonitor(params=candle_params)
timeoutMonitor = candle.TerminateOnTimeOut(candle_params['timeout'])
history2 = model.fit(X_train, Y_train,
batch_size=candle_params['batch_size'],
epochs=candle_params['epochs'],
verbose=1,
validation_data=(X_test, Y_test),
callbacks = [csv_logger, reduce_lr, candleRemoteMonitor, timeoutMonitor])
def run(params):
args = Struct(**params)
set_seed(args.rng_seed)
ext = extension_from_parameters(args)
verify_path(args.save_path)
prefix = args.save_path + ext
logfile = args.logfile if args.logfile else prefix + '.log'
set_up_logger(logfile, args.verbose)
logger.info('Params: {}'.format(params))
loader = ComboDataLoader(seed=args.rng_seed,
val_split=args.validation_split,
cell_features=args.cell_features,
drug_features=args.drug_features,
use_mean_growth=args.use_mean_growth,
response_url=args.response_url,
use_landmark_genes=args.use_landmark_genes,
preprocess_rnaseq=args.preprocess_rnaseq,
exclude_cells=args.exclude_cells,
exclude_drugs=args.exclude_drugs,
use_combo_score=args.use_combo_score,
cv_partition=args.cv_partition,
cv=args.cv)
# test_loader(loader)
# test_generator(loader)
train_gen = ComboDataGenerator(loader, batch_size=args.batch_size).flow()
val_gen = ComboDataGenerator(loader,
partition='val',
batch_size=args.batch_size).flow()
train_steps = int(loader.n_train / args.batch_size)
val_steps = int(loader.n_val / args.batch_size)
model = build_model(loader, args, verbose=True)
model.summary()
# candle.plot_model(model, to_file=prefix+'.model.png', show_shapes=True)
if args.cp:
model_json = model.to_json()
with open(prefix + '.model.json', 'w') as f:
print(model_json, file=f)
def warmup_scheduler(epoch):
lr = args.learning_rate or base_lr * args.batch_size / 100
if epoch <= 5:
K.set_value(model.optimizer.lr,
(base_lr * (5 - epoch) + lr * epoch) / 5)
logger.debug('Epoch {}: lr={}'.format(epoch,
K.get_value(model.optimizer.lr)))
return K.get_value(model.optimizer.lr)
df_pred_list = []
cv_ext = ''
cv = args.cv if args.cv > 1 else 1
fold = 0
while fold < cv:
if args.cv > 1:
logger.info('Cross validation fold {}/{}:'.format(fold + 1, cv))
cv_ext = '.cv{}'.format(fold + 1)
model = build_model(loader, args)
optimizer = optimizers.deserialize({
'class_name': args.optimizer,
'config': {}
})
base_lr = args.base_lr or K.get_value(optimizer.lr)
if args.learning_rate:
K.set_value(optimizer.lr, args.learning_rate)
model.compile(loss=args.loss, optimizer=optimizer, metrics=[mae, r2])
# calculate trainable and non-trainable params
params.update(candle.compute_trainable_params(model))
candle_monitor = candle.CandleRemoteMonitor(params=params)
timeout_monitor = candle.TerminateOnTimeOut(params['timeout'])
reduce_lr = ReduceLROnPlateau(monitor='val_loss',
factor=0.5,
patience=5,
min_lr=0.00001)
warmup_lr = LearningRateScheduler(warmup_scheduler)
checkpointer = ModelCheckpoint(prefix + cv_ext + '.weights.h5',
save_best_only=True,
save_weights_only=True)
tensorboard = TensorBoard(log_dir="tb/tb{}{}".format(ext, cv_ext))
history_logger = LoggingCallback(logger.debug)
model_recorder = ModelRecorder()
# callbacks = [history_logger, model_recorder]
callbacks = [
candle_monitor, timeout_monitor, history_logger, model_recorder
]
if args.reduce_lr:
callbacks.append(reduce_lr)
if args.warmup_lr:
callbacks.append(warmup_lr)
if args.cp:
callbacks.append(checkpointer)
if args.tb:
callbacks.append(tensorboard)
if args.gen:
history = model.fit_generator(train_gen,
train_steps,
epochs=args.epochs,
callbacks=callbacks,
validation_data=val_gen,
validation_steps=val_steps)
fold += 1
else:
if args.cv > 1:
x_train_list, y_train, x_val_list, y_val, df_train, df_val = loader.load_data_cv(
fold)
else:
x_train_list, y_train, x_val_list, y_val, df_train, df_val = loader.load_data(
)
y_shuf = np.random.permutation(y_val)
log_evaluation(evaluate_prediction(y_val, y_shuf),
description='Between random pairs in y_val:')
history = model.fit(x_train_list,
y_train,
batch_size=args.batch_size,
shuffle=args.shuffle,
epochs=args.epochs,
callbacks=callbacks,
validation_data=(x_val_list, y_val))
if args.cp:
model.load_weights(prefix + cv_ext + '.weights.h5')
if not args.gen:
y_val_pred = model.predict(x_val_list,
batch_size=args.batch_size).flatten()
scores = evaluate_prediction(y_val, y_val_pred)
if args.cv > 1 and scores[args.loss] > args.max_val_loss:
logger.warn(
'Best val_loss {} is greater than {}; retrain the model...'
.format(scores[args.loss], args.max_val_loss))
continue
else:
fold += 1
log_evaluation(scores)
df_val.is_copy = False
df_val['GROWTH_PRED'] = y_val_pred
df_val['GROWTH_ERROR'] = y_val_pred - y_val
df_pred_list.append(df_val)
if args.cp:
# model.save(prefix+'.model.h5')
model_recorder.best_model.save(prefix + '.model.h5')
# test reloadded model prediction
# new_model = keras.models.load_model(prefix+'.model.h5')
# new_model.load_weights(prefix+cv_ext+'.weights.h5')
# new_pred = new_model.predict(x_val_list, batch_size=args.batch_size).flatten()
# print('y_val:', y_val[:10])
# print('old_pred:', y_val_pred[:10])
# print('new_pred:', new_pred[:10])
candle.plot_history(prefix, history, 'loss')
candle.plot_history(prefix, history, 'r2')
if K.backend() == 'tensorflow':
K.clear_session()
if not args.gen:
if args.use_combo_score:
pred_fname = prefix + '.predicted.score.tsv'
elif args.use_mean_growth:
pred_fname = prefix + '.predicted.mean.growth.tsv'
else:
pred_fname = prefix + '.predicted.growth.tsv'
df_pred = pd.concat(df_pred_list)
df_pred.to_csv(pred_fname, sep='\t', index=False, float_format='%.4g')
logger.handlers = []
return history
def run(params):
args = candle.ArgumentStruct(**params)
candle.set_seed(args.rng_seed)
ext = uno.extension_from_parameters(args)
candle.verify_path(args.save_path)
prefix = args.save_path + 'uno' + ext
logfile = args.logfile if args.logfile else prefix+'.log'
uno.set_up_logger(logfile, logger, uno.loggerUno, args.verbose)
logger.info('Params: {}'.format(params))
# Exclude drugs / cells for UQ
if 'uq_exclude_drugs_file' in params.keys():
args.exclude_drugs = uno.read_IDs_file(args.uq_exclude_drugs_file)
logger.info('Drugs to exclude: {}'.format(args.exclude_drugs))
else:
args.exclude_drugs = []
if 'uq_exclude_cells_file' in params.keys():
args.exclude_cells = uno.read_IDs_file(args.uq_exclude_cells_file)
logger.info('Cells to exclude: {}'.format(args.exclude_cells))
else:
args.exclude_cells = []
if 'uq_exclude_indices_file' in params.keys():
exclude_indices_ = uno.read_IDs_file(args.uq_exclude_indices_file)
args.exclude_indices = [int(x) for x in exclude_indices_]
logger.info('Indices to exclude: {}'.format(args.exclude_indices))
else:
args.exclude_indices = []
if (len(args.gpus) > 0):
import tensorflow as tf
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.gpu_options.visible_device_list = ",".join(map(str, args.gpus))
K.set_session(tf.Session(config=config))
loader = uno_combined_data_loader.CombinedDataLoader(seed=args.rng_seed)
loader.load(cache=args.cache,
ncols=args.feature_subsample,
agg_dose=args.agg_dose,
cell_features=args.cell_features,
drug_features=args.drug_features,
drug_median_response_min=args.drug_median_response_min,
drug_median_response_max=args.drug_median_response_max,
use_landmark_genes=args.use_landmark_genes,
use_filtered_genes=args.use_filtered_genes,
cell_feature_subset_path=args.cell_feature_subset_path or args.feature_subset_path,
drug_feature_subset_path=args.drug_feature_subset_path or args.feature_subset_path,
preprocess_rnaseq=args.preprocess_rnaseq,
single=args.single,
train_sources=args.train_sources,
test_sources=args.test_sources,
embed_feature_source=not args.no_feature_source,
encode_response_source=not args.no_response_source,
)
target = args.agg_dose or 'Growth'
val_split = args.val_split
train_split = 1 - val_split
loader.partition_data(partition_by=args.partition_by,
cv_folds=args.cv, train_split=train_split, val_split=val_split,
cell_types=args.cell_types, by_cell=args.by_cell, by_drug=args.by_drug,
cell_subset_path=args.cell_subset_path,
drug_subset_path=args.drug_subset_path,
exclude_cells=args.exclude_cells,
exclude_drugs=args.exclude_drugs,
exclude_indices=args.exclude_indices
)
model = uno_model_utils.build_model(loader, args, logger)
logger.info('Combined model:')
model.summary(print_fn=logger.info)
# plot_model(model, to_file=prefix+'.model.png', show_shapes=True)
if args.cp:
model_json = model.to_json()
with open(prefix+'.model.json', 'w') as f:
print(model_json, file=f)
def warmup_scheduler(epoch):
lr = args.learning_rate or base_lr * args.batch_size/100
if epoch <= 5:
K.set_value(model.optimizer.lr, (base_lr * (5-epoch) + lr * epoch) / 5)
logger.debug('Epoch {}: lr={:.5g}'.format(epoch, K.get_value(model.optimizer.lr)))
return K.get_value(model.optimizer.lr)
df_pred_list = []
cv_ext = ''
cv = args.cv if args.cv > 1 else 1
for fold in range(cv):
if args.cv > 1:
logger.info('Cross validation fold {}/{}:'.format(fold+1, cv))
cv_ext = '.cv{}'.format(fold+1)
# model = uno_model_utils.build_model(loader, args, logger, silent=True)
template_model = uno_model_utils.build_model(loader, args, logger, silent=True)
if args.initial_weights:
logger.info("Loading weights from {}".format(args.initial_weights))
template_model.load_weights(args.initial_weights)
if len(args.gpus) > 1:
from keras.utils import multi_gpu_model
gpu_count = len(args.gpus)
logger.info("Multi GPU with {} gpus".format(gpu_count))
model = multi_gpu_model(template_model, cpu_merge=False, gpus=gpu_count)
else:
model = template_model
optimizer = optimizers.deserialize({'class_name': args.optimizer, 'config': {}})
base_lr = args.base_lr or K.get_value(optimizer.lr)
if args.learning_rate:
K.set_value(optimizer.lr, args.learning_rate)
if args.loss == 'heteroscedastic':
logger.info('Training heteroscedastic model:')
model.compile(loss=heteroscedastic_loss, optimizer=optimizer, metrics=[uno_model_utils.mae_heteroscedastic, uno_model_utils.r2_heteroscedastic, uno_model_utils.meanS_heteroscesdastic])
elif args.loss == 'quantile':
logger.info('Training quantile model:')
model.compile(loss=triple_quantile_loss, optimizer=optimizer, metrics=[uno_model_utils.quantile50, uno_model_utils.quantile10, uno_model_utils.quantile90])
else:
logger.info('Training homoscedastic model:')
model.compile(loss=args.loss, optimizer=optimizer, metrics=[candle.mae, candle.r2])
# calculate trainable and non-trainable params
params.update(candle.compute_trainable_params(model))
candle_monitor = candle.CandleRemoteMonitor(params=params)
timeout_monitor = candle.TerminateOnTimeOut(params['timeout'])
reduce_lr = ReduceLROnPlateau(monitor='val_loss', factor=0.5, patience=5, min_lr=0.00001)
warmup_lr = LearningRateScheduler(warmup_scheduler)
#checkpointer = ModelCheckpoint(prefix+cv_ext+'.weights.h5', save_best_only=True, save_weights_only=True)
checkpointer = candle.MultiGPUCheckpoint(prefix + cv_ext + '.model.h5', save_best_only=True)
tensorboard = TensorBoard(log_dir="tb/{}{}{}".format(args.tb_prefix, ext, cv_ext))
history_logger = candle.LoggingCallback(logger.debug)
# model_recorder = uno_model_utils.ModelRecorder()
# callbacks = [history_logger, model_recorder]
callbacks = [candle_monitor, timeout_monitor, history_logger]#, model_recorder]
if args.reduce_lr:
callbacks.append(reduce_lr)
if args.warmup_lr:
callbacks.append(warmup_lr)
if args.cp:
callbacks.append(checkpointer)
if args.tb:
callbacks.append(tensorboard)
if args.save_weights:
callbacks.append(uno_model_utils.SimpleWeightSaver(args.save_path + '/' + args.save_weights))
train_gen = uno_combined_data_generator.CombinedDataGenerator(loader, fold=fold, batch_size=args.batch_size, shuffle=args.shuffle)
val_gen = uno_combined_data_generator.CombinedDataGenerator(loader, partition='val', fold=fold, batch_size=args.batch_size, shuffle=args.shuffle)
df_val = val_gen.get_response(copy=True)
y_val = df_val[target].values
y_shuf = np.random.permutation(y_val)
uno.log_evaluation(uno.evaluate_prediction(y_val, y_shuf), logger,
description='Between random pairs in y_val:')
if args.no_gen:
x_train_list, y_train = train_gen.get_slice(size=train_gen.size, single=args.single)
x_val_list, y_val = val_gen.get_slice(size=val_gen.size, single=args.single)
history = model.fit(x_train_list, y_train,
batch_size=args.batch_size,
epochs=args.epochs,
callbacks=callbacks,
validation_data=(x_val_list, y_val))
else:
logger.info('Data points per epoch: train = %d, val = %d',train_gen.size, val_gen.size)
logger.info('Steps per epoch: train = %d, val = %d',train_gen.steps, val_gen.steps)
history = model.fit_generator(train_gen, train_gen.steps,
epochs=args.epochs,
callbacks=callbacks,
validation_data=val_gen,
validation_steps=val_gen.steps)
# if args.cp:
# model.load_weights(prefix+cv_ext+'.weights.h5')
# model = model_recorder.best_model
if args.no_gen:
y_val_pred = model.predict(x_val_list, batch_size=args.batch_size)
else:
val_gen.reset()
y_val_pred = model.predict_generator(val_gen, val_gen.steps + 1)
y_val_pred = y_val_pred[:val_gen.size]
if args.loss == 'heteroscedastic':
y_val_pred_ = y_val_pred[:,0]
s_val_pred = y_val_pred[:,1]
y_val_pred = y_val_pred_.flatten()
df_val['Predicted_'+target] = y_val_pred
df_val[target+'_Error'] = y_val_pred-y_val
df_val['Pred_S_'+target] = s_val_pred
elif args.loss == 'quantile':
y_val_pred_50q = y_val_pred[:,0]
y_val_pred_10q = y_val_pred[:,1]
y_val_pred_90q = y_val_pred[:,2]
y_val_pred = y_val_pred_50q.flatten() # 50th quantile prediction
df_val['Predicted_50q_'+target] = y_val_pred
df_val[target+'_Error_50q'] = y_val_pred-y_val
df_val['Predicted_10q_'+target] = y_val_pred_10q.flatten()
df_val['Predicted_90q_'+target] = y_val_pred_90q.flatten()
else:
y_val_pred = y_val_pred.flatten()
# df_val = df_val.assign(PredictedGrowth=y_val_pred, GrowthError=y_val_pred-y_val)
df_val['Predicted'+target] = y_val_pred
df_val[target+'Error'] = y_val_pred-y_val
scores = uno.evaluate_prediction(y_val, y_val_pred)
uno.log_evaluation(scores, logger)
df_pred_list.append(df_val)
# if args.cp:
# model_recorder.best_model.save(prefix+'.model.h5')
if hasattr(history, 'loss'):
candle.plot_history(prefix, history, 'loss')
if args.loss == 'heteroscedastic':
if hasattr(history, 'r2_heteroscedastic'):
candle.plot_history(prefix, history, 'r2_heteroscedastic')
if hasattr(history, 'meanS_heteroscedastic'):
candle.plot_history(prefix, history, 'meanS_heteroscesdastic')
elif args.loss == 'quantile':
if hasattr(history, 'quantile50'):
candle.plot_history(prefix, history, 'quantile50')
if hasattr(history, 'quantile10'):
candle.plot_history(prefix, history, 'quantile10')
if hasattr(history, 'quantile90'):
candle.plot_history(prefix, history, 'quantile90')
else:
if hasattr(history, 'r2'):
candle.plot_history(prefix, history, 'r2')
pred_fname = prefix + '.predicted.tsv'
df_pred = pd.concat(df_pred_list)
if args.agg_dose:
if args.single:
# df_pred.sort_values(['Source', 'Sample', 'Drug1', target], inplace=True)
df_pred.sort_values(['Sample', 'Drug1', target], inplace=True)
else:
df_pred.sort_values(['Source', 'Sample', 'Drug1', 'Drug2', target], inplace=True)
else:
if args.single:
# df_pred.sort_values(['Source', 'Sample', 'Drug1', 'Dose1', 'Growth'], inplace=True)
df_pred.sort_values(['Sample', 'Drug1', 'Dose1', 'Growth'], inplace=True)
else:
# df_pred.sort_values(['Source', 'Sample', 'Drug1', 'Drug2', 'Dose1', 'Dose2', 'Growth'], inplace=True)
df_pred.sort_values(['Sample', 'Drug1', 'Drug2', 'Dose1', 'Dose2', 'Growth'], inplace=True)
df_pred.to_csv(pred_fname, sep='\t', index=False, float_format='%.4g')
logger.info('Testing predictions stored in file: {}'.format(pred_fname))
if args.cp:
logger.info('Model stored in file: {}'.format(prefix+'.model.h5'))
# logger.info('Model weights stored in file: {}'.format(prefix+cv_ext+'.weights.h5'))
logger.info('Model weights stored in file: {}'.format(args.save_path + '/' + args.save_weights))
if args.cv > 1:
scores = uno.evaluate_prediction(df_pred[target], df_pred['Predicted'+target])
uno.log_evaluation(scores, logger, description='Combining cross validation folds:')
for test_source in loader.test_sep_sources:
test_gen = uno_combined_data_generator.CombinedDataGenerator(loader, partition='test', batch_size=args.batch_size, source=test_source)
df_test = test_gen.get_response(copy=True)
y_test = df_test[target].values
n_test = len(y_test)
if n_test == 0:
continue
if args.no_gen:
x_test_list, y_test = test_gen.get_slice(size=test_gen.size, single=args.single)
y_test_pred = model.predict(x_test_list, batch_size=args.batch_size)
if args.loss == 'heteroscedastic':
y_test_pred = y_test_pred[:,0]
elif args.loss == 'quantile':
y_test_pred = y_test_pred[:,0] # 50th quantile prediction
else:
y_test_pred = model.predict_generator(test_gen.flow(single=args.single), test_gen.steps)
if args.loss == 'heteroscedastic':
y_test_pred = y_test_pred[:test_gen.size,0]
elif args.loss == 'quantile':
y_test_pred = y_test_pred[:test_gen.size,0] # 50th quantile prediction
else:
y_test_pred = y_test_pred[:test_gen.size]
y_test_pred = y_test_pred.flatten()
scores = uno.evaluate_prediction(y_test, y_test_pred)
uno.log_evaluation(scores, logger, description='Testing on data from {} ({})'.format(test_source, n_test))
if K.backend() == 'tensorflow':
K.clear_session()
logger.handlers = []
return history
def run(gParameters):
print ('Params:', gParameters)
file_train = gParameters['train_data']
file_test = gParameters['test_data']
url = gParameters['data_url']
train_file = candle.get_file(file_train, url+file_train, cache_subdir='Pilot1')
test_file = candle.get_file(file_test, url+file_test, cache_subdir='Pilot1')
X_train, Y_train, X_test, Y_test = bmk.load_data(train_file, test_file, gParameters)
print('X_train shape:', X_train.shape)
print('X_test shape:', X_test.shape)
print('Y_train shape:', Y_train.shape)
print('Y_test shape:', Y_test.shape)
x_train_len = X_train.shape[1]
# this reshaping is critical for the Conv1D to work
X_train = np.expand_dims(X_train, axis=2)
X_test = np.expand_dims(X_test, axis=2)
print('X_train shape:', X_train.shape)
print('X_test shape:', X_test.shape)
model = Sequential()
layer_list = list(range(0, len(gParameters['conv']), 3))
for l, i in enumerate(layer_list):
filters = gParameters['conv'][i]
filter_len = gParameters['conv'][i+1]
stride = gParameters['conv'][i+2]
print(int(i/3), filters, filter_len, stride)
if gParameters['pool']:
pool_list=gParameters['pool']
if type(pool_list) != list:
pool_list=list(pool_list)
if filters <= 0 or filter_len <= 0 or stride <= 0:
break
if 'locally_connected' in gParameters:
model.add(LocallyConnected1D(filters, filter_len, strides=stride, padding='valid', input_shape=(x_train_len, 1)))
else:
#input layer
if i == 0:
model.add(Conv1D(filters=filters, kernel_size=filter_len, strides=stride, padding='valid', input_shape=(x_train_len, 1)))
else:
model.add(Conv1D(filters=filters, kernel_size=filter_len, strides=stride, padding='valid'))
model.add(Activation(gParameters['activation']))
if gParameters['pool']:
model.add(MaxPooling1D(pool_size=pool_list[int(i/3)]))
model.add(Flatten())
for layer in gParameters['dense']:
if layer:
model.add(Dense(layer))
model.add(Activation(gParameters['activation']))
if gParameters['dropout']:
model.add(Dropout(gParameters['dropout']))
model.add(Dense(gParameters['classes']))
model.add(Activation(gParameters['out_activation']))
#Reference case
#model.add(Conv1D(filters=128, kernel_size=20, strides=1, padding='valid', input_shape=(P, 1)))
#model.add(Activation('relu'))
#model.add(MaxPooling1D(pool_size=1))
#model.add(Conv1D(filters=128, kernel_size=10, strides=1, padding='valid'))
#model.add(Activation('relu'))
#model.add(MaxPooling1D(pool_size=10))
#model.add(Flatten())
#model.add(Dense(200))
#model.add(Activation('relu'))
#model.add(Dropout(0.1))
#model.add(Dense(20))
#model.add(Activation('relu'))
#model.add(Dropout(0.1))
#model.add(Dense(CLASSES))
#model.add(Activation('softmax'))
kerasDefaults = candle.keras_default_config()
# Define optimizer
optimizer = candle.build_optimizer(gParameters['optimizer'],
gParameters['learning_rate'],
kerasDefaults)
model.summary()
model.compile(loss=gParameters['loss'],
optimizer=optimizer,
metrics=[gParameters['metrics']])
output_dir = gParameters['output_dir']
if not os.path.exists(output_dir):
os.makedirs(output_dir)
# calculate trainable and non-trainable params
gParameters.update(candle.compute_trainable_params(model))
# set up a bunch of callbacks to do work during model training..
model_name = gParameters['model_name']
path = '{}/{}.autosave.model.h5'.format(output_dir, model_name)
checkpointer = ModelCheckpoint(filepath=path, verbose=1, save_weights_only=False, save_best_only=True)
csv_logger = CSVLogger('{}/training.log'.format(output_dir))
reduce_lr = ReduceLROnPlateau(monitor='val_loss', factor=0.1, patience=10, verbose=1, mode='auto', epsilon=0.0001, cooldown=0, min_lr=0)
candleRemoteMonitor = candle.CandleRemoteMonitor(params=gParameters)
timeoutMonitor = candle.TerminateOnTimeOut(gParameters['timeout'])
history = model.fit(X_train, Y_train,
batch_size=gParameters['batch_size'],
epochs=gParameters['epochs'],
verbose=1,
validation_data=(X_test, Y_test),
callbacks = [csv_logger, reduce_lr, candleRemoteMonitor, timeoutMonitor])
score = model.evaluate(X_test, Y_test, verbose=0)
print('Test score:', score[0])
print('Test accuracy:', score[1])
# serialize model to JSON
model_json = model.to_json()
with open("{}/{}.model.json".format(output_dir, model_name), "w") as json_file:
json_file.write(model_json)
# serialize model to YAML
model_yaml = model.to_yaml()
with open("{}/{}.model.yaml".format(output_dir, model_name), "w") as yaml_file:
yaml_file.write(model_yaml)
# serialize weights to HDF5
model.save_weights("{}/{}.weights.h5".format(output_dir, model_name))
print("Saved model to disk")
# load json and create model
json_file = open('{}/{}.model.json'.format(output_dir, model_name), 'r')
loaded_model_json = json_file.read()
json_file.close()
loaded_model_json = model_from_json(loaded_model_json)
# load yaml and create model
yaml_file = open('{}/{}.model.yaml'.format(output_dir, model_name), 'r')
loaded_model_yaml = yaml_file.read()
yaml_file.close()
loaded_model_yaml = model_from_yaml(loaded_model_yaml)
# load weights into new model
loaded_model_json.load_weights('{}/{}.weights.h5'.format(output_dir, model_name))
print("Loaded json model from disk")
# evaluate json loaded model on test data
loaded_model_json.compile(loss=gParameters['loss'],
optimizer=gParameters['optimizer'],
metrics=[gParameters['metrics']])
score_json = loaded_model_json.evaluate(X_test, Y_test, verbose=0)
print('json Test score:', score_json[0])
print('json Test accuracy:', score_json[1])
print("json %s: %.2f%%" % (loaded_model_json.metrics_names[1], score_json[1]*100))
# load weights into new model
loaded_model_yaml.load_weights('{}/{}.weights.h5'.format(output_dir, model_name))
print("Loaded yaml model from disk")
# evaluate loaded model on test data
loaded_model_yaml.compile(loss=gParameters['loss'],
optimizer=gParameters['optimizer'],
metrics=[gParameters['metrics']])
score_yaml = loaded_model_yaml.evaluate(X_test, Y_test, verbose=0)
print('yaml Test score:', score_yaml[0])
print('yaml Test accuracy:', score_yaml[1])
print("yaml %s: %.2f%%" % (loaded_model_yaml.metrics_names[1], score_yaml[1]*100))
return history
def run(params):
args = candle.ArgumentStruct(**params)
seed = args.rng_seed
candle.set_seed(seed)
# Construct extension to save model
ext = p1b1.extension_from_parameters(params, '.keras')
candle.verify_path(params['save_path'])
prefix = '{}{}'.format(params['save_path'], ext)
logfile = params['logfile'] if params['logfile'] else prefix+'.log'
candle.set_up_logger(logfile, p1b1.logger, params['verbose'])
p1b1.logger.info('Params: {}'.format(params))
# Get default parameters for initialization and optimizer functions
keras_defaults = candle.keras_default_config()
# Load dataset
x_train, y_train, x_val, y_val, x_test, y_test, x_labels, y_labels = p1b1.load_data(params, seed)
# cache_file = 'data_l1000_cache.h5'
# save_cache(cache_file, x_train, y_train, x_val, y_val, x_test, y_test, x_labels, y_labels)
# x_train, y_train, x_val, y_val, x_test, y_test, x_labels, y_labels = load_cache(cache_file)
p1b1.logger.info("Shape x_train: {}".format(x_train.shape))
p1b1.logger.info("Shape x_val: {}".format(x_val.shape))
p1b1.logger.info("Shape x_test: {}".format(x_test.shape))
p1b1.logger.info("Range x_train: [{:.3g}, {:.3g}]".format(np.min(x_train), np.max(x_train)))
p1b1.logger.info("Range x_val: [{:.3g}, {:.3g}]".format(np.min(x_val), np.max(x_val)))
p1b1.logger.info("Range x_test: [{:.3g}, {:.3g}]".format(np.min(x_test), np.max(x_test)))
p1b1.logger.debug('Class labels')
for i, label in enumerate(y_labels):
p1b1.logger.debug(' {}: {}'.format(i, label))
# clf = build_type_classifier(x_train, y_train, x_val, y_val)
n_classes = len(y_labels)
cond_train = y_train
cond_val = y_val
cond_test = y_test
input_dim = x_train.shape[1]
cond_dim = cond_train.shape[1]
latent_dim = params['latent_dim']
activation = params['activation']
dropout = params['dropout']
dense_layers = params['dense']
dropout_layer = AlphaDropout if params['alpha_dropout'] else Dropout
# Initialize weights and learning rule
initializer_weights = candle.build_initializer(params['initialization'], keras_defaults, seed)
initializer_bias = candle.build_initializer('constant', keras_defaults, 0.)
if dense_layers is not None:
if type(dense_layers) != list:
dense_layers = list(dense_layers)
else:
dense_layers = []
# Encoder Part
x_input = Input(shape=(input_dim,))
cond_input = Input(shape=(cond_dim,))
h = x_input
if params['model'] == 'cvae':
h = keras.layers.concatenate([x_input, cond_input])
for i, layer in enumerate(dense_layers):
if layer > 0:
x = h
h = Dense(layer, activation=activation,
kernel_initializer=initializer_weights,
bias_initializer=initializer_bias)(h)
if params['residual']:
try:
h = keras.layers.add([h, x])
except ValueError:
pass
if params['batch_normalization']:
h = BatchNormalization()(h)
if dropout > 0:
h = dropout_layer(dropout)(h)
if params['model'] == 'ae':
encoded = Dense(latent_dim, activation=activation,
kernel_initializer=initializer_weights,
bias_initializer=initializer_bias)(h)
else:
epsilon_std = params['epsilon_std']
z_mean = Dense(latent_dim, name='z_mean')(h)
z_log_var = Dense(latent_dim, name='z_log_var')(h)
encoded = z_mean
def vae_loss(x, x_decoded_mean):
xent_loss = binary_crossentropy(x, x_decoded_mean)
kl_loss = - 0.5 * K.sum(1 + z_log_var - K.square(z_mean) - K.exp(z_log_var), axis=-1)
return K.mean(xent_loss + kl_loss/input_dim)
def sampling(params):
z_mean_, z_log_var_ = params
batch_size = K.shape(z_mean_)[0]
epsilon = K.random_normal(shape=(batch_size, latent_dim),
mean=0., stddev=epsilon_std)
return z_mean_ + K.exp(z_log_var_ / 2) * epsilon
z = Lambda(sampling, output_shape=(latent_dim,))([z_mean, z_log_var])
if params['model'] == 'cvae':
z_cond = keras.layers.concatenate([z, cond_input])
# Decoder Part
decoder_input = Input(shape=(latent_dim,))
h = decoder_input
if params['model'] == 'cvae':
h = keras.layers.concatenate([decoder_input, cond_input])
for i, layer in reversed(list(enumerate(dense_layers))):
if layer > 0:
x = h
h = Dense(layer, activation=activation,
kernel_initializer=initializer_weights,
bias_initializer=initializer_bias)(h)
if params['residual']:
try:
h = keras.layers.add([h, x])
except ValueError:
pass
if params['batch_normalization']:
h = BatchNormalization()(h)
if dropout > 0:
h = dropout_layer(dropout)(h)
decoded = Dense(input_dim, activation='sigmoid',
kernel_initializer=initializer_weights,
bias_initializer=initializer_bias)(h)
# Build autoencoder model
if params['model'] == 'cvae':
encoder = Model([x_input, cond_input], encoded)
decoder = Model([decoder_input, cond_input], decoded)
model = Model([x_input, cond_input], decoder([z, cond_input]))
loss = vae_loss
metrics = [xent, corr, mse]
elif params['model'] == 'vae':
encoder = Model(x_input, encoded)
decoder = Model(decoder_input, decoded)
model = Model(x_input, decoder(z))
loss = vae_loss
metrics = [xent, corr, mse]
else:
encoder = Model(x_input, encoded)
decoder = Model(decoder_input, decoded)
model = Model(x_input, decoder(encoded))
loss = params['loss']
metrics = [xent, corr]
model.summary()
decoder.summary()
if params['cp']:
model_json = model.to_json()
with open(prefix+'.model.json', 'w') as f:
print(model_json, file=f)
# Define optimizer
# optimizer = candle.build_optimizer(params['optimizer'],
# params['learning_rate'],
# keras_defaults)
optimizer = optimizers.deserialize({'class_name': params['optimizer'], 'config': {}})
base_lr = params['base_lr'] or K.get_value(optimizer.lr)
if params['learning_rate']:
K.set_value(optimizer.lr, params['learning_rate'])
model.compile(loss=loss, optimizer=optimizer, metrics=metrics)
# calculate trainable and non-trainable params
params.update(candle.compute_trainable_params(model))
def warmup_scheduler(epoch):
lr = params['learning_rate'] or base_lr * params['batch_size']/100
if epoch <= 5:
K.set_value(model.optimizer.lr, (base_lr * (5-epoch) + lr * epoch) / 5)
p1b1.logger.debug('Epoch {}: lr={}'.format(epoch, K.get_value(model.optimizer.lr)))
return K.get_value(model.optimizer.lr)
reduce_lr = ReduceLROnPlateau(monitor='val_loss', factor=0.5, patience=5, min_lr=0.00001)
warmup_lr = LearningRateScheduler(warmup_scheduler)
checkpointer = ModelCheckpoint(params['save_path']+ext+'.weights.h5', save_best_only=True, save_weights_only=True)
tensorboard = TensorBoard(log_dir="tb/tb{}".format(ext))
candle_monitor = candle.CandleRemoteMonitor(params=params)
timeout_monitor = candle.TerminateOnTimeOut(params['timeout'])
history_logger = LoggingCallback(p1b1.logger.debug)
callbacks = [candle_monitor, timeout_monitor, history_logger]
if params['reduce_lr']:
callbacks.append(reduce_lr)
if params['warmup_lr']:
callbacks.append(warmup_lr)
if params['cp']:
callbacks.append(checkpointer)
if params['tb']:
callbacks.append(tensorboard)
x_val2 = np.copy(x_val)
np.random.shuffle(x_val2)
start_scores = p1b1.evaluate_autoencoder(x_val, x_val2)
p1b1.logger.info('\nBetween random pairs of validation samples: {}'.format(start_scores))
if params['model'] == 'cvae':
inputs = [x_train, cond_train]
val_inputs = [x_val, cond_val]
test_inputs = [x_test, cond_test]
else:
inputs = x_train
val_inputs = x_val
test_inputs = x_test
outputs = x_train
val_outputs = x_val
test_outputs = x_test
history = model.fit(inputs, outputs,
verbose=2,
batch_size=params['batch_size'],
epochs=params['epochs'],
callbacks=callbacks,
validation_data=(val_inputs, val_outputs))
if params['cp']:
encoder.save(prefix+'.encoder.h5')
decoder.save(prefix+'.decoder.h5')
candle.plot_history(prefix, history, 'loss')
candle.plot_history(prefix, history, 'corr', 'streaming pearson correlation')
# Evalute model on test set
x_pred = model.predict(test_inputs)
scores = p1b1.evaluate_autoencoder(x_pred, x_test)
p1b1.logger.info('\nEvaluation on test data: {}'.format(scores))
x_test_encoded = encoder.predict(test_inputs, batch_size=params['batch_size'])
y_test_classes = np.argmax(y_test, axis=1)
candle.plot_scatter(x_test_encoded, y_test_classes, prefix+'.latent')
if params['tsne']:
tsne = TSNE(n_components=2, random_state=seed)
x_test_encoded_tsne = tsne.fit_transform(x_test_encoded)
candle.plot_scatter(x_test_encoded_tsne, y_test_classes, prefix+'.latent.tsne')
# diff = x_pred - x_test
# plt.hist(diff.ravel(), bins='auto')
# plt.title("Histogram of Errors with 'auto' bins")
# plt.savefig('histogram_keras.png')
# generate synthetic data
# epsilon_std = 1.0
# for i in range(1000):
# z_sample = np.random.normal(size=(1, 2)) * epsilon_std
# x_decoded = decoder.predict(z_sample)
p1b1.logger.handlers = []
return history
