def train_test_dnn_vary_parameter(prefix,
model_parameters,
param_name,
param_values,
X_train=None, y_train=None,
X_valid=None, y_valid=None,
X_test=None, y_test=None):
X_train = np.concatenate((X_train, reverse_complement(X_train)))
y_train = np.concatenate((y_train, y_train))
dnn_results = []
for param_value in param_values:
model_parameters[param_name] = param_value
ofname_infix = dict2string(model_parameters)
ofname_prefix = "%s.%s" % (prefix, ofname_infix)
model_fname = "%s.arch.json" % (ofname_prefix)
weights_fname = "%s.weights.hd5" % (ofname_prefix)
try:
logger.debug("Checking for model files {} and {}...".format(model_fname, weights_fname))
dnn = SequenceDNN.load(model_fname, weights_fname)
logger.debug("Model files found. Loaded model successfully!")
except:
logger.debug("Model files not found. Training model...")
dnn = SequenceDNN(**model_parameters)
logger.info("training with %s %s .." % (param_name, param_value))
dnn.train(X_train, y_train, (X_valid, y_valid))
dnn.save(model_fname, weights_fname)
dnn_results.append(dnn.test(X_test, y_test))
return dnn_results
def run(use_deep_CNN, use_RNN, label, golden_first_sequence, golden_results):
seq_length = 100
num_sequences = 200
num_positives = 100
num_negatives = num_sequences - num_positives
GC_fraction = 0.4
test_fraction = 0.2
num_epochs = 1
sequences, labels, embeddings = simulate_single_motif_detection(
'SPI1_disc1', seq_length, num_positives, num_negatives, GC_fraction)
assert sequences[0] == golden_first_sequence, 'first sequence = {}, golden = {}'.format(
sequences[0], golden_first_sequence)
encoded_sequences = one_hot_encode(sequences)
X_train, X_test, y_train, y_test = train_test_split(
encoded_sequences, labels, test_size=test_fraction)
X_train = np.concatenate((X_train, reverse_complement(X_train)))
y_train = np.concatenate((y_train, y_train))
random_order = np.arange(len(X_train))
np.random.shuffle(random_order)
X_train = X_train[random_order]
y_train = y_train[random_order]
hyperparameters = {'seq_length': seq_length, 'use_RNN': use_RNN,
'num_filters': (45,), 'pool_width': 25, 'conv_width': (10,),
'L1': 0, 'dropout': 0.2, 'num_epochs': num_epochs}
if use_deep_CNN:
hyperparameters.update({'num_filters': (45, 50, 50), 'conv_width': (10, 8, 5)})
if use_RNN:
hyperparameters.update({'GRU_size': 35, 'TDD_size': 45})
model = SequenceDNN(**hyperparameters)
model.train(X_train, y_train, validation_data=(X_test, y_test))
results = model.test(X_test, y_test).results[0]
assert np.allclose(tuple(results.values()), tuple(golden_results.values())), \
'{}: result = {}, golden = {}'.format(label, results, golden_results)
def run(use_deep_CNN, use_RNN, label, golden_results):
seq_length = 100
num_sequences = 200
test_fraction = 0.2
num_epochs = 1
sequences = np.array([''.join(random.choice('ACGT') for base in range(seq_length)) for sequence in range(num_sequences)])
labels = np.random.choice((True, False), size=num_sequences)[:, None]
encoded_sequences = one_hot_encode(sequences)
X_train, X_test, y_train, y_test = train_test_split(
encoded_sequences, labels, test_size=test_fraction)
X_train = np.concatenate((X_train, reverse_complement(X_train)))
y_train = np.concatenate((y_train, y_train))
random_order = np.arange(len(X_train))
np.random.shuffle(random_order)
X_train = X_train[random_order]
y_train = y_train[random_order]
hyperparameters = {'seq_length': seq_length, 'use_RNN': use_RNN,
'num_filters': (45,), 'pool_width': 25, 'conv_width': (10,),
'L1': 0, 'dropout': 0.2, 'num_epochs': num_epochs}
if use_deep_CNN:
hyperparameters.update({'num_filters': (45, 50, 50), 'conv_width': (10, 8, 5)})
if use_RNN:
hyperparameters.update({'GRU_size': 35, 'TDD_size': 45})
model = SequenceDNN(**hyperparameters)
model.train(X_train, y_train, validation_data=(X_test, y_test))
results = model.test(X_test, y_test).results[0]
assert np.allclose(tuple(results.values()), tuple(golden_results.values())), \
'{}: result = {}, golden = {}'.format(label, results, golden_results)
def train_test_dnn_vary_parameter(prefix,
model_parameters,
param_name,
param_values,
X_train=None,
y_train=None,
X_valid=None,
y_valid=None,
X_test=None,
y_test=None):
X_train = np.concatenate((X_train, reverse_complement(X_train)))
y_train = np.concatenate((y_train, y_train))
dnn_results = []
for param_value in param_values:
model_parameters[param_name] = param_value
ofname_infix = dict2string(model_parameters)
ofname_prefix = "%s.%s" % (prefix, ofname_infix)
model_fname = "%s.arch.json" % (ofname_prefix)
weights_fname = "%s.weights.h5" % (ofname_prefix)
try:
logger.debug("Checking for model files {} and {}...".format(
model_fname, weights_fname))
dnn = SequenceDNN.load(model_fname, weights_fname)
logger.debug("Model files found. Loaded model successfully!")
except:
logger.debug("Model files not found. Training model...")
dnn = SequenceDNN(**model_parameters)
logger.info("training with %s %s .." % (param_name, param_value))
dnn.train(X_train, y_train, (X_valid, y_valid))
dnn.save(ofname_prefix)
dnn_results.append(dnn.test(X_test, y_test))
return dnn_results
def train_test_dnn_vary_data_size(prefix,
model_parameters=None,
X_train=None,
y_train=None,
X_valid=None,
y_valid=None,
X_test=None,
y_test=None,
train_set_sizes=None):
dnn_results = []
for train_set_size in train_set_sizes:
ofname_infix = dict2string(model_parameters)
ofname_infix = "%s.train_set_size_%s" % (ofname_infix,
str(train_set_size))
ofname_prefix = "%s.%s" % (prefix, ofname_infix)
model_fname = "%s.arch.json" % (ofname_prefix)
weights_fname = "%s.weights.h5" % (ofname_prefix)
try:
logger.debug("Checking for model files {} and {}...".format(
model_fname, weights_fname))
best_dnn = SequenceDNN.load(model_fname, weights_fname)
logger.debug("Model files found. Loaded model successfully!")
except:
logger.debug("Model files not found. Training model...")
# try 3 attempts, take best auROC, save that model
X_train_subset = X_train[:train_set_size]
X_train_subset = np.concatenate(
(X_train_subset, reverse_complement(X_train_subset)))
y_train_subset = np.concatenate(
(y_train[:train_set_size], y_train[:train_set_size]))
best_auROC = 0
best_dnn = None
for random_seed in [1, 2, 3]:
np.random.seed(random_seed)
random.seed(random_seed)
dnn = SequenceDNN(**model_parameters)
logger.info("training with %i examples.." % (train_set_size))
dnn.train(X_train_subset, y_train_subset, (X_valid, y_valid))
result = dnn.test(X_test, y_test)
auROCs = [
result.results[i]["auROC"]
for i in range(y_valid.shape[-1])
]
# get average auROC across tasks
mean_auROC = sum(auROCs) / len(auROCs)
if mean_auROC > best_auROC:
best_auROC = mean_auROC
dnn.save(ofname_prefix)
best_dnn = dnn
dnn_results.append(best_dnn.test(X_test, y_test))
# reset to original random seed
np.random.seed(1)
random.seed(1)
return dnn_results
def run(use_deep_CNN, use_RNN, label, golden_results):
import random
np.random.seed(1)
random.seed(1)
from dragonn.models import SequenceDNN
from simdna.simulations import simulate_single_motif_detection
from dragonn.utils import one_hot_encode, reverse_complement
from sklearn.cross_validation import train_test_split
seq_length = 50
num_sequences = 100
num_positives = 50
num_negatives = num_sequences - num_positives
GC_fraction = 0.4
test_fraction = 0.2
validation_fraction = 0.2
num_epochs = 1
sequences, labels = simulate_single_motif_detection(
'SPI1_disc1', seq_length, num_positives, num_negatives, GC_fraction)
encoded_sequences = one_hot_encode(sequences)
X_train, X_test, y_train, y_test = train_test_split(
encoded_sequences, labels, test_size=test_fraction)
X_train, X_valid, y_train, y_valid = train_test_split(
X_train, y_train, test_size=validation_fraction)
X_train = np.concatenate((X_train, reverse_complement(X_train)))
y_train = np.concatenate((y_train, y_train))
random_order = np.arange(len(X_train))
np.random.shuffle(random_order)
X_train = X_train[random_order]
y_train = y_train[random_order]
hyperparameters = {
'seq_length': seq_length,
'use_RNN': use_RNN,
'num_filters': (45, ),
'pool_width': 25,
'conv_width': (10, ),
'L1': 0,
'dropout': 0.2,
'num_epochs': num_epochs
}
if use_deep_CNN:
hyperparameters.update({
'num_filters': (45, 50, 50),
'conv_width': (10, 8, 5)
})
if use_RNN:
hyperparameters.update({'GRU_size': 35, 'TDD_size': 45})
model = SequenceDNN(**hyperparameters)
model.train(X_train, y_train, validation_data=(X_valid, y_valid))
results = model.test(X_test, y_test).results[0]
assert np.allclose(tuple(results.values()), tuple(golden_results.values())), \
'{}: result = {}, golden = {}'.format(label, results, golden_results)
def train_test_dnn_vary_data_size(prefix, model_parameters=None,
X_train=None, y_train=None,
X_valid=None, y_valid=None,
X_test=None, y_test=None,
train_set_sizes=None):
dnn_results = []
for train_set_size in train_set_sizes:
ofname_infix = dict2string(model_parameters)
ofname_infix = "%s.train_set_size_%s" % (ofname_infix, str(train_set_size))
ofname_prefix = "%s.%s" % (prefix, ofname_infix)
model_fname = "%s.arch.json" % (ofname_prefix)
weights_fname = "%s.weights.hd5" % (ofname_prefix)
try:
logger.debug("Checking for model files {} and {}...".format(model_fname, weights_fname))
best_dnn = SequenceDNN.load(model_fname, weights_fname)
logger.debug("Model files found. Loaded model successfully!")
except:
logger.debug("Model files not found. Training model...")
# try 3 attempts, take best auROC, save that model
X_train_subset = X_train[:train_set_size]
X_train_subset = np.concatenate((X_train_subset, reverse_complement(X_train_subset)))
y_train_subset = np.concatenate((y_train[:train_set_size], y_train[:train_set_size]))
best_auROC = 0
best_dnn = None
for random_seed in [1, 2, 3]:
np.random.seed(random_seed)
random.seed(random_seed)
dnn = SequenceDNN(**model_parameters)
logger.info("training with %i examples.." % (train_set_size))
dnn.train(X_train_subset, y_train_subset, (X_valid, y_valid))
result = dnn.test(X_test, y_test)
auROCs = [result.results[i]["auROC"] for i in range(y_valid.shape[-1])]
# get average auROC across tasks
mean_auROC = sum(auROCs) / len(auROCs)
if mean_auROC > best_auROC:
best_auROC = mean_auROC
dnn.save(model_fname, weights_fname)
best_dnn = dnn
dnn_results.append(best_dnn.test(X_test, y_test))
# reset to original random seed
np.random.seed(1)
random.seed(1)
return dnn_results
def run(use_deep_CNN, use_RNN, label, golden_results):
seq_length = 100
num_sequences = 200
test_fraction = 0.2
num_epochs = 1
sequences = np.array([
''.join(random.choice('ACGT') for base in range(seq_length))
for sequence in range(num_sequences)
])
labels = np.random.choice((True, False), size=num_sequences)[:, None]
encoded_sequences = one_hot_encode(sequences)
X_train, X_test, y_train, y_test = train_test_split(
encoded_sequences, labels, test_size=test_fraction)
X_train = np.concatenate((X_train, reverse_complement(X_train)))
y_train = np.concatenate((y_train, y_train))
random_order = np.arange(len(X_train))
np.random.shuffle(random_order)
X_train = X_train[random_order]
y_train = y_train[random_order]
hyperparameters = {
'seq_length': seq_length,
'use_RNN': use_RNN,
'num_filters': (45, ),
'pool_width': 25,
'conv_width': (10, ),
'L1': 0,
'dropout': 0.2,
'num_epochs': num_epochs
}
if use_deep_CNN:
hyperparameters.update({
'num_filters': (45, 50, 50),
'conv_width': (10, 8, 5)
})
if use_RNN:
hyperparameters.update({'GRU_size': 35, 'TDD_size': 45})
model = SequenceDNN(**hyperparameters)
model.train(X_train, y_train, validation_data=(X_test, y_test))
results = model.test(X_test, y_test).results[0]
assert np.allclose(tuple(results.values()), tuple(golden_results.values())), \
'{}: result = {}, golden = {}'.format(label, results, golden_results)
# DATASET MANIPULATION & CREATION =============================================
# Create dataset
X, y = simulate_single_motif_detection('CTCF_known2', 1000, 5000, 5000, 0.4)
X = one_hot_encode(X)
random_order = np.arange(len(X))
np.random.shuffle(random_order)
X, y = X[random_order], y[random_order]
# Split dataset into a training and test set
test_fraction = validation_fraction = 0.2
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=test_fraction)
# Concatenate reverse complements
X_train = np.concatenate((X_train, reverse_complement(X_train)))
y_train = np.concatenate((y_train, y_train))
X_test = np.concatenate((X_test, reverse_complement(X_test)))
y_test = np.concatenate((y_test, y_test))
# Reshape dataset per Tensorflow convention: [# examples, height, width, channels]
X_train = X_train.reshape((16000, 4, 1000, 1))
X_test = X_test.reshape((4000, 4, 1000, 1))
# TF data norm is Height, Width, num_channels
input_shape = (None, ) + np.shape(X_train)[1:]
seq_length = input_shape[-1]
# BUILDING MODEL ==================================================================
hyperparameters = {'seq_length': seq_length, 'use_deep_CNN': True, 'use_RNN': False,
print('One-hot encoding sequences...')
encoded_sequences = one_hot_encode(sequences)
print('Getting motif scores...')
motif_scores = get_motif_scores(encoded_sequences, motif_names=['SPI1_disc1'])
print('Partitioning data into training, validation and test sets...')
X_train, X_test, y_train, y_test = train_test_split(encoded_sequences, labels, test_size=test_fraction)
X_train, X_valid, y_train, y_valid = train_test_split(X_train, y_train, test_size=validation_fraction)
print('Adding reverse complements...')
X_train = np.concatenate((X_train, reverse_complement(X_train)))
y_train = np.concatenate((y_train, y_train))
print('Randomly splitting data into training and test sets...')
random_order = np.arange(len(X_train))
np.random.shuffle(random_order)
X_train = X_train[random_order]
y_train = y_train[random_order]
# Build and train model
if not do_hyperparameter_search:
hyperparameters = {'seq_length': seq_length, 'use_RNN': use_RNN,
'num_filters': (45,), 'pool_width': 25, 'conv_width': (10,),
'L1': 0, 'dropout': 0.2, 'num_epochs': num_epochs}