def train_model(gParameters,
models,
X_train,
Y_train,
X_test,
Y_test,
fold,
verbose=False):
base_run_id = gParameters['run_id']
for epoch in range(gParameters['epochs']):
for k in range(len(models)):
model = models[k]
gParameters['run_id'] = base_run_id + ".{}.{}.{}".format(
fold, epoch, k)
candleRemoteMonitor = candle.CandleRemoteMonitor(
params=gParameters)
timeoutMonitor = candle.TerminateOnTimeOut(gParameters['timeout'])
model.fit({'input': X_train[k]}, {'out_' + str(k): Y_train[k]},
epochs=1,
verbose=verbose,
callbacks=[candleRemoteMonitor, timeoutMonitor],
batch_size=gParameters['batch_size'],
validation_data=(X_test[k], Y_test[k]))
return models
def run(GP):
# set the seed
if GP['seed']:
np.random.seed(GP['seed'])
else:
np.random.seed(np.random.randint(10000))
# Set paths
if not os.path.isdir(GP['home_dir']):
print('Keras home directory not set')
sys.exit(0)
sys.path.append(GP['home_dir'])
# Setup loggin
args = candle.ArgumentStruct(**GP)
# set_seed(args.rng_seed)
# ext = extension_from_parameters(args)
candle.verify_path(args.save_path)
prefix = args.save_path # + ext
logfile = args.logfile if args.logfile else prefix + '.log'
candle.set_up_logger(logfile, logger, False) #args.verbose
logger.info('Params: {}'.format(GP))
import p2b1 as hf
reload(hf)
#import keras_model_utils as KEU
#reload(KEU)
#reload(p2ck)
#reload(p2ck.optimizers)
maps = hf.autoencoder_preprocess()
from keras.optimizers import SGD, RMSprop, Adam
from keras.datasets import mnist
from keras.callbacks import LearningRateScheduler, ModelCheckpoint
from keras import callbacks
from keras.layers.advanced_activations import ELU
from keras.preprocessing.image import ImageDataGenerator
# GP=hf.ReadConfig(opts.config_file)
batch_size = GP['batch_size']
learning_rate = GP['learning_rate']
kerasDefaults = candle.keras_default_config()
##### Read Data ########
import helper
(data_files, fields) = p2b1.get_list_of_data_files(GP)
# Read from local directoy
#(data_files, fields) = helper.get_local_files('/p/gscratchr/brainusr/datasets/cancer/pilot2/3k_run16_10us.35fs-DPPC.20-DIPC.60-CHOL.20.dir/')
#(data_files, fields) = helper.get_local_files('3k_run16', '/p/lscratchf/brainusr/datasets/cancer/pilot2/')
# Define datagenerator
datagen = hf.ImageNoiseDataGenerator(corruption_level=GP['noise_factor'])
# get data dimension ##
num_samples = 0
for f in data_files:
# Seperate different arrays from the data
(X, nbrs, resnums) = helper.get_data_arrays(f)
num_samples += X.shape[0]
(X, nbrs, resnums) = helper.get_data_arrays(data_files[0])
print('\nData chunk shape: ', X.shape)
molecular_hidden_layers = GP['molecular_num_hidden']
if not molecular_hidden_layers:
X_train = hf.get_data(X, case=GP['case'])
input_dim = X_train.shape[1]
else:
# computing input dimension for outer AE
input_dim = X.shape[1] * molecular_hidden_layers[-1]
print('\nState AE input/output dimension: ', input_dim)
# get data dimension for molecular autoencoder
molecular_nbrs = np.int(GP['molecular_nbrs'])
num_molecules = X.shape[1]
num_beads = X.shape[2]
if GP['nbr_type'] == 'relative':
# relative x, y, z positions
num_loc_features = 3
loc_feat_vect = ['rel_x', 'rel_y', 'rel_z']
elif GP['nbr_type'] == 'invariant':
# relative distance and angle
num_loc_features = 2
loc_feat_vect = ['rel_dist', 'rel_angle']
else:
print('Invalid nbr_type!!')
exit()
if not GP['type_bool']:
# only consider molecular location coordinates
num_type_features = 0
type_feat_vect = []
else:
num_type_features = 5
type_feat_vect = list(fields.keys())[3:8]
num_features = num_loc_features + num_type_features + num_beads
dim = np.prod([num_beads, num_features, molecular_nbrs + 1])
bead_kernel_size = num_features
molecular_input_dim = dim
mol_kernel_size = num_beads
feature_vector = loc_feat_vect + type_feat_vect + list(fields.keys())[8:]
print('\nMolecular AE input/output dimension: ', molecular_input_dim)
print(
'\nData Format:\n[Frames (%s), Molecules (%s), Beads (%s), %s (%s)]' %
(num_samples, num_molecules, num_beads, feature_vector, num_features))
### Define Model, Solver and Compile ##########
print('\nDefine the model and compile')
opt = candle.build_optimizer(GP['optimizer'], learning_rate, kerasDefaults)
model_type = 'mlp'
memo = '%s_%s' % (GP['base_memo'], model_type)
######## Define Molecular Model, Solver and Compile #########
molecular_nonlinearity = GP['molecular_nonlinearity']
len_molecular_hidden_layers = len(molecular_hidden_layers)
conv_bool = GP['conv_bool']
full_conv_bool = GP['full_conv_bool']
if conv_bool:
molecular_model, molecular_encoder = AE_models.conv_dense_mol_auto(
bead_k_size=bead_kernel_size,
mol_k_size=mol_kernel_size,
weights_path=None,
input_shape=(1, molecular_input_dim, 1),
nonlinearity=molecular_nonlinearity,
hidden_layers=molecular_hidden_layers,
l2_reg=GP['l2_reg'],
drop=float(GP['drop_prob']))
elif full_conv_bool:
molecular_model, molecular_encoder = AE_models.full_conv_mol_auto(
bead_k_size=bead_kernel_size,
mol_k_size=mol_kernel_size,
weights_path=None,
input_shape=(1, molecular_input_dim, 1),
nonlinearity=molecular_nonlinearity,
hidden_layers=molecular_hidden_layers,
l2_reg=GP['l2_reg'],
drop=float(GP['drop_prob']))
else:
molecular_model, molecular_encoder = AE_models.dense_auto(
weights_path=None,
input_shape=(molecular_input_dim, ),
nonlinearity=molecular_nonlinearity,
hidden_layers=molecular_hidden_layers,
l2_reg=GP['l2_reg'],
drop=float(GP['drop_prob']))
if GP['loss'] == 'mse':
loss_func = 'mse'
elif GP['loss'] == 'custom':
loss_func = helper.combined_loss
molecular_model.compile(
optimizer=opt,
loss=loss_func,
metrics=['mean_squared_error', 'mean_absolute_error'])
print('\nModel Summary: \n')
molecular_model.summary()
##### set up callbacks and cooling for the molecular_model ##########
drop = 0.5
mb_epochs = GP['epochs']
initial_lrate = GP['learning_rate']
epochs_drop = 1 + int(np.floor(mb_epochs / 3))
def step_decay(epoch):
global initial_lrate, epochs_drop, drop
lrate = initial_lrate * np.power(drop,
np.floor((1 + epoch) / epochs_drop))
return lrate
lr_scheduler = LearningRateScheduler(step_decay)
history = callbacks.History()
# callbacks=[history,lr_scheduler]
history_logger = candle.LoggingCallback(logger.debug)
candleRemoteMonitor = candle.CandleRemoteMonitor(params=GP)
timeoutMonitor = candle.TerminateOnTimeOut(TIMEOUT)
callbacks = [history, history_logger, candleRemoteMonitor, timeoutMonitor]
loss = 0.
#### Save the Model to disk
if GP['save_path'] != None:
save_path = GP['save_path']
if not os.path.exists(save_path):
os.makedirs(save_path)
else:
save_path = '.'
model_json = molecular_model.to_json()
with open(save_path + '/model.json', "w") as json_file:
json_file.write(model_json)
encoder_json = molecular_encoder.to_json()
with open(save_path + '/encoder.json', "w") as json_file:
json_file.write(encoder_json)
print('Saved model to disk')
#### Train the Model
if GP['train_bool']:
ct = hf.Candle_Molecular_Train(
molecular_model,
molecular_encoder,
data_files,
mb_epochs,
callbacks,
batch_size=batch_size,
nbr_type=GP['nbr_type'],
save_path=GP['save_path'],
len_molecular_hidden_layers=len_molecular_hidden_layers,
molecular_nbrs=molecular_nbrs,
conv_bool=conv_bool,
full_conv_bool=full_conv_bool,
type_bool=GP['type_bool'],
sampling_density=GP['sampling_density'])
frame_loss, frame_mse = ct.train_ac()
else:
frame_mse = []
frame_loss = []
return frame_loss, frame_mse
def run(params):
args = Struct(**params)
set_seed(args.rng_seed)
ext = extension_from_parameters(args)
verify_path(args.save)
prefix = args.save + 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,
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()
# 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])
plot_history(prefix, history, 'loss')
plot_history(prefix, history, 'r2')
if K.backend() == 'tensorflow':
K.clear_session()
if not args.gen:
pred_fname = prefix + '.predicted.growth.tsv'
if args.use_combo_score:
pred_fname = prefix + '.predicted.score.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)
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'])
prefix = '{}{}'.format(params['save'], 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['drop']
dense_layers = params['dense']
dropout_layer = keras.layers.noise.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'] + 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')
plot_history(prefix, history, 'loss')
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)
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)
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
def run(params):
args = Struct(**params)
set_seed(args.rng_seed)
ext = extension_from_parameters(args)
verify_path(args.save)
prefix = args.save + ext
logfile = args.logfile if args.logfile else prefix + '.log'
set_up_logger(logfile, args.verbose)
logger.info('Params: {}'.format(params))
loader = CombinedDataLoader(seed=args.rng_seed)
loader.load(
cache=args.cache,
ncols=args.feature_subsample,
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,
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,
)
val_split = args.validation_split
train_split = 1 - val_split
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)
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
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)
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)
model = build_model(loader, args, silent=True)
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)
train_gen = CombinedDataGenerator(loader,
fold=fold,
batch_size=args.batch_size,
shuffle=args.shuffle)
val_gen = 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['Growth'].values
y_shuf = np.random.permutation(y_val)
log_evaluation(evaluate_prediction(y_val, y_shuf),
description='Between random pairs in y_val:')
candleRemoteMonitor = CandleRemoteMonitor(params=params)
callbacks.append(candleRemoteMonitor)
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.flow(single=args.single),
train_gen.steps,
epochs=args.epochs,
callbacks=callbacks,
validation_data=val_gen.flow(single=args.single),
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.flow(single=args.single), val_gen.steps)
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_pred_list.append(df_val)
plot_history(prefix, history, 'loss')
plot_history(prefix, history, 'r2')
pred_fname = prefix + '.predicted.tsv'
df_pred = pd.concat(df_pred_list)
df_pred.sort_values(
['Source', '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['Growth'],
df_pred['PredictedGrowth'])
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['Growth'].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
hyperparams.update(candle.compute_trainable_params(model))
# set up a bunch of callbacks to do work during model training..
model_name = hyperparams['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=hyperparams)
timeoutMonitor = candle.TerminateOnTimeOut(hyperparams['timeout'])
history2 = model.fit(
X_train,
Y_train,
batch_size=hyperparams['batch_size'],
epochs=hyperparams['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)
val_to_return = score[0]
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)
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['drop']:
model.add(Dropout(gParameters['drop']))
model.add(Dense(gParameters['classes']))
model.add(Activation(gParameters['out_act']))
#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['save']
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 run_cnn(GP,
train_x,
train_y,
test_x,
test_y,
learning_rate=0.01,
batch_size=10,
epochs=10,
dropout=0.5,
optimizer='adam',
wv_len=300,
filter_sizes=[3, 4, 5],
num_filters=[300, 300, 300],
emb_l2=0.001,
w_l2=0.01):
max_vocab = np.max(train_x)
max_vocab2 = np.max(test_x)
if max_vocab2 > max_vocab:
max_vocab = max_vocab2
wv_mat = np.random.randn(max_vocab + 1, wv_len).astype('float32') * 0.1
num_classes = []
num_classes.append(np.max(train_y[:, 0]) + 1)
num_classes.append(np.max(train_y[:, 1]) + 1)
num_classes.append(np.max(train_y[:, 2]) + 1)
num_classes.append(np.max(train_y[:, 3]) + 1)
kerasDefaults = candle.keras_default_config()
optimizer_run = candle.build_optimizer(optimizer, learning_rate,
kerasDefaults)
cnn = keras_mt_shared_cnn.init_export_network(num_classes=num_classes,
in_seq_len=1500,
vocab_size=len(wv_mat),
wv_space=wv_len,
filter_sizes=filter_sizes,
num_filters=num_filters,
concat_dropout_prob=dropout,
emb_l2=emb_l2,
w_l2=w_l2,
optimizer=optimizer_run)
print(cnn.summary())
validation_data = ({
'Input': test_x
}, {
'Dense0': test_y[:, 0],
'Dense1': test_y[:, 1],
'Dense2': test_y[:, 2],
'Dense3': test_y[:, 3]
})
# candleRemoteMonitor = CandleRemoteMonitor(params= GP)
# timeoutMonitor = TerminateOnTimeOut(TIMEOUT)
candleRemoteMonitor = candle.CandleRemoteMonitor(params=GP)
timeoutMonitor = candle.TerminateOnTimeOut(GP['timeout'])
history = cnn.fit(x=np.array(train_x),
y=[
np.array(train_y[:, 0]),
np.array(train_y[:, 1]),
np.array(train_y[:, 2]),
np.array(train_y[:, 3])
],
batch_size=batch_size,
epochs=epochs,
verbose=2,
validation_data=validation_data,
callbacks=[candleRemoteMonitor, timeoutMonitor])
return history