def get_concise_model():
param, X_feat, X_seq, y, id_vec = load_example_data(trim_seq_len=1000)
dc = concise_model(pooling_layer="sum",
init_motifs=["TGCGAT", "TATTTAT"],
n_splines=10,
n_covariates=0,
seq_length=X_seq.shape[1],
**param)
dc.fit([X_seq], y, epochs=1, validation_data=([X_seq], y))
return {"model": dc, "out_annotation": np.array(["output_1"])}
def test_serialization_disk(tmpdir):
param, X_feat, X_seq, y, id_vec = load_example_data()
dc = concise_model(pooling_layer="sum",
init_motifs=["TGCGAT", "TATTTAT"],
n_splines=10,
n_covariates=X_feat.shape[1],
seq_length=X_seq.shape[1],
**param)
dc.fit([X_seq, X_feat], y, epochs=1,
validation_data=([X_seq, X_feat], y))
fn = tmpdir.mkdir('data').join('test_keras.h5')
dc.save(str(fn))
dc = keras.models.load_model(str(fn))
assert isinstance(dc, keras.models.Model)
def get_list_input_model():
import keras
from keras.layers.merge import concatenate
from keras.models import Sequential, Model
from keras.layers.core import Dense, Dropout, Flatten
from keras.layers.convolutional import Conv1D, MaxPooling1D
seq_len = 1000
param, X_feat, X_seq, y, id_vec = load_example_data(trim_seq_len=seq_len)
# set hyperparameters
opt = keras.optimizers.Adam(lr=0.0001,
beta_1=0.9,
beta_2=0.999,
epsilon=1e-08,
decay=0.0)
ls = 'binary_crossentropy'
nf = 100
nf_fc = 200
fl = 8
drops = 0.1
pooling_length = 4
cnninit = 'glorot_normal'
winit = "glorot_uniform"
n_output = 1
ls_metrics = ['accuracy']
n1 = Sequential()
n1.add(Dropout(drops, input_shape=(seq_len, 4)))
n1.add(
Conv1D(kernel_size=fl,
kernel_initializer=cnninit,
activation="relu",
filters=nf))
n1.add(MaxPooling1D(pool_size=pooling_length))
n1.add(Dropout(drops))
n1.add(Flatten())
n1.add(Dense(nf_fc, activation='relu', kernel_initializer=winit))
n2 = Sequential()
n2.add(Dropout(drops, input_shape=(seq_len, 4)))
n2.add(
Conv1D(kernel_size=fl,
kernel_initializer=cnninit,
activation="relu",
filters=nf))
n2.add(MaxPooling1D(pool_size=pooling_length))
n2.add(Dropout(drops))
n2.add(Flatten())
n2.add(Dense(nf_fc, activation='relu', kernel_initializer=winit))
merged = concatenate([n1.output, n2.output])
out = Dense(n_output, activation='sigmoid',
kernel_initializer=winit)(merged)
merged_model = Model(inputs=[n1.input, n2.input], outputs=out)
merged_model.compile(optimizer=opt, loss=ls, metrics=ls_metrics)
# this is slow - subset
merged_model.fit([X_seq[:500], X_seq[:500]],
y[:500],
batch_size=128,
epochs=1,
validation_data=([X_seq, X_seq], y))
return {"model": merged_model, "out_annotation": np.array(["output_1"])}
def setup_class(cls):
cls.data = load_example_data(trim_seq_len=1, standardize_features=False)
cls.data[0]["n_motifs"] = 1
cls.data[0]["motif_length"] = 1
cls.data[0]["step_size"] = 0.001
cls.data[0]["early_stop_patience"] = 3
def setup_class(cls):
cls.data = load_example_data(num_tasks=3)
def setup_class(cls):
cls.data = load_example_data()