def lin_models(lasso=True,
traits=['height', 'BMI', 'WHR', 'BHMD', 'SBP'],
nbsnps=10000,
verbose=0,
hot=False,
unif=False,
reps=1):
alpha = [0.01]
R = {}
for t in traits:
print(t)
x_tr, x_tst, y_tr, y_tst = retrieve_data(t, nbsnps, unif=unif)
x_tr, x_val, y_tr, y_val = train_test_split(x_tr, y_tr, test_size=0.33)
if hot:
x_tr = convert_to_individual_alleles(x_tr)
x_val = convert_to_individual_alleles(x_val)
x_tst = convert_to_individual_alleles(x_tst)
nb_snps = x_tr.shape[1]
res = np.zeros((len(alpha), 3))
n = 0
for a in alpha:
print(a)
for i in range(0, reps):
m = Sequential()
if lasso:
m.add(Dense(1, input_dim=nb_snps,
kernel_regularizer=l1(a)))
else:
m.add(Dense(1, input_dim=nb_snps,
kernel_regularizer=l2(a)))
m.compile(loss='mse', optimizer='adam')
m.fit(x_tr,
y_tr,
epochs=1000,
callbacks=[EarlyStopping()],
validation_data=(x_val, y_val),
verbose=verbose)
if r(m.predict(x_val).ravel(), y_val)[0] > res[n, 0]:
print(r(m.predict(x_val).ravel(), y_val)[0])
print(i)
res[n, 0] = r(m.predict(x_val).ravel(), y_val)[0]
res[n, 1] = r(m.predict(x_tst).ravel(), y_tst)[0]
K.clear_session()
print(res[n, 1])
n += 1
R[t + "val"] = res[:, 0]
R[t + "tst"] = res[:, 1]
R["alpha"] = alpha
print(pd.DataFrame(R).to_csv(float_format='%.3f', index=False))
logging.info(pd.DataFrame(R).to_csv(float_format='%.3f', index=False))
def train_and_score(geneparam, dataset):
"""Train the model, return test loss.
Args:
geneparam (dict): the parameters of the network
dataset (str): Dataset to use for training/evaluating
"""
logging.info("Getting datasets")
x_train, x_test, y_train, y_test = retrieve_data(dataset.trait,
dataset.k,
unif=dataset.unif)
input_shape = x_train.shape[1]
train_data = np.expand_dims(x_train, axis=2)
test_data = np.expand_dims(x_test, axis=2)
input_shape = (input_shape, 1)
logging.info("Compling Keras model")
model = compile_model_cnn(geneparam, input_shape)
model.fit(
train_data,
y_train,
epochs=1200,
# using early stopping so no real limit - don't want to waste time on horrible architectures
verbose=1,
validation_data=(test_data, y_test),
callbacks=[early_stopper])
score = model.evaluate(test_data, y_test, verbose=0)
print('Test mse:', score[0])
print('Test mae:', score[1])
r = pearsonr(model.predict(test_data).ravel(), y_test)[0]
print("Test r:", r)
K.clear_session()
# we do not care about keeping any of this in memory -
# we just need to know the final scores and the architecture
if r != r:
r = -1.0
return r
def CNN(traits=['height', 'BMI', 'WHR', 'BHMD', 'SBP'],
verbose=0,
unif=False,
nbsnps=10000,
p=None,
reps=1):
#cnn1
param = list({
'optimizer': 'nadam',
'size_window': 2,
'activation': 'softplus',
'nb_neurons': 64,
'stride': 'one',
'nb_cnn_layers': 1,
'filters': 16,
'weight_decay': 0.0,
'nb_layers': 3,
'dropout': 0.01,
'batch_norm': True
})
#cnn2
param.append({
'optimizer': 'nadam',
'size_window': 2,
'activation': 'elu',
'nb_neurons': 32,
'stride': 'one',
'nb_cnn_layers': 1,
'filters': 32,
'weight_decay': 0.0,
'nb_layers': 3,
'dropout': 0.01,
'batch_norm': False
})
#cnn3
param.append({
'optimizer': 'rmsprop',
'size_window': 3,
'activation': 'linear',
'nb_neurons': 32,
'stride': 'one',
'nb_cnn_layers': 1,
'filters': 16,
'weight_decay': 0.0,
'nb_layers': 1,
'dropout': 0.01,
'batch_norm': False
})
R = {}
for t in traits:
best = 0
print(t)
x_tr, x_tst, y_tr, y_tst = retrieve_data(t, nbsnps, unif=unif)
x_tr, x_val, y_tr, y_val = train_test_split(x_tr, y_tr, test_size=0.33)
n_snps = x_tr.shape[1]
x_tr = np.expand_dims(x_tr, axis=2)
x_val = np.expand_dims(x_val, axis=2)
x_tst = np.expand_dims(x_tst, axis=2)
f = os.path.join(
os.path.expanduser("~"), 'Code/genomic_cnn/models',
"Model_" + t + "_cnn_" + str(n_snps / 1000) + "k" +
("_unif" if unif else "_best") + ".h5")
n = 0
if p is None:
res = np.zeros((len(param), 2))
for g in param:
print(g)
for x in range(0, reps):
m = compile_model_cnn(g, (n_snps, 1))
m.fit(x_tr,
y_tr,
epochs=1200,
verbose=verbose,
validation_data=(x_val, y_val),
callbacks=[early_stopper])
if r(m.predict(x_val).ravel(), y_val)[0] > res[n, 0]:
print(r(m.predict(x_val).ravel(), y_val)[0])
print(x)
res[n, 0] = r(m.predict(x_val).ravel(), y_val)[0]
res[n, 1] = r(m.predict(x_tst).ravel(), y_tst)[0]
if res[n, 0] > best:
print("A better network was found with r: %.3f" %
res[n, 0])
print(g)
m.save(f)
best = res[n, 0]
n = n + 1
else:
res = np.zeros((reps, 2))
g = param[p]
for i in range(0, reps):
m = compile_model_cnn(g, (n_snps, 1))
m.fit(x_tr,
y_tr,
epochs=1200,
verbose=verbose,
validation_data=(x_val, y_val),
callbacks=[early_stopper])
res[i, :] = (r(m.predict(x_val).ravel(),
y_val)[0], r(m.predict(x_tst).ravel(),
y_tst)[0])
R[t + "_tr"] = res[:, 0]
R[t + "_tst"] = res[:, 1]
print(pd.DataFrame(R).to_csv(float_format='%.3f', index=False))
logging.info(pd.DataFrame(R).to_csv(float_format='%.3f', index=False))
def MLP(traits=['height', 'BMI', 'WHR', 'BHMD', 'SBP'],
verbose=0,
unif=False,
nbsnps=10000,
p=None,
reps=1,
hot=False):
#mlp1
geneparam = list({
'optimizer': 'rmsprop',
'activation': 'elu',
'nb_neurons': 32,
'weight_decay': 0.01,
'nb_layers': 1,
'dropout': 0.02
})
# mlp2
geneparam.append({
'optimizer': 'adagrad',
'activation': 'elu',
'nb_neurons': 64,
'weight_decay': 0.01,
'nb_layers': 2,
'dropout': 0.03
})
# mlp3
geneparam.append({
'optimizer': 'adam',
'activation': 'softplus',
'nb_neurons': 32,
'weight_decay': 0.01,
'nb_layers': 5,
'dropout': 0.02
})
R = {}
for t in traits:
print(t)
best = 0
x_tr, x_tst, y_tr, y_tst = retrieve_data(t, nbsnps, unif=unif)
x_tr, x_val, y_tr, y_val = train_test_split(x_tr, y_tr, test_size=0.33)
if hot:
x_tr = convert_to_individual_alleles(x_tr)
x_val = convert_to_individual_alleles(x_val)
x_tst = convert_to_individual_alleles(x_tst)
n_snps = x_tr.shape[1]
f = os.path.join(os.path.expanduser("~"), 'Code/genomic_cnn/models',
"Model_" + t + "_mlp_" + str(n_snps / 1000) \
+ "kHot" + ("_unif" if unif else "_best") + ".h5")
else:
n_snps = x_tr.shape[1]
f = os.path.join(
os.path.expanduser("~"), 'Code/genomic_cnn/models',
"Model_" + t + "_mlp_" + str(n_snps / 1000) + "k" +
("_unif" if unif else "_best") + ".h5")
n = 0
if p is None:
res = np.zeros((len(geneparam), 2))
for g in geneparam:
print(g)
for x in range(0, reps):
m = compile_model_mlp(g, n_snps)
m.fit(x_tr,
y_tr,
epochs=1200,
validation_data=(x_val, y_val),
callbacks=[early_stopper],
verbose=verbose)
if r(m.predict(x_val).ravel(), y_val)[0] > res[n, 0]:
print(r(m.predict(x_val).ravel(), y_val)[0])
print(x)
res[n, 0] = r(m.predict(x_val).ravel(), y_val)[0]
res[n, 1] = r(m.predict(x_tst).ravel(), y_tst)[0]
if res[n, 0] > best:
print("A better network was found with r: %.3f" %
res[n, 0])
print(g)
m.save(f)
best = res[n, 0]
K.clear_session()
n = n + 1
else:
res = np.zeros((reps, 2))
g = geneparam[p]
for i in range(0, reps):
m = compile_model_mlp(g, n_snps)
m.fit(x_tr,
y_tr,
epochs=1200,
verbose=verbose,
validation_data=(x_val, y_val),
callbacks=[early_stopper])
res[i, :] = (r(m.predict(x_val).ravel(),
y_val)[0], r(m.predict(x_tst).ravel(),
y_tst)[0])
R[t + "_tr"] = res[:, 0]
R[t + "_tst"] = res[:, 1]
print(pd.DataFrame(R).to_csv(float_format='%.3f', index=False))
logging.info(pd.DataFrame(R).to_csv(float_format='%.3f', index=False))