def run_neural_net(X_train, Y_train, X_test, Y_test):
# Rows are samples, columns are features
INPUT_NODES = X_train.shape[1]
OUTPUT_NODES = 1
# create model
model = Sequential()
model.add(
Dense(512, input_dim=INPUT_NODES, init='uniform', activation='relu'))
model.add(Dropout(0.8))
#model.add(Dense(X.shape[1], init='uniform', activation='relu'))
model.add(Dense(OUTPUT_NODES, init='uniform', activation='sigmoid'))
#model = Sequential([
# Dense(32, input_dim=X.shape[1], init='uniform'),
# Activation('relu'),
# #Dense(10, init='uniform'),
# #Activation('relu'),
# Dense(1, init='uniform'),
# Activation('sigmoid'),
#])
# Compile model
#model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy', recall, precision, window_diff_metric, size1, size2])
model.compile(loss='binary_crossentropy',
optimizer='adam',
metrics=['accuracy', recall, precision])
# Fit the model
model.fit(X_train,
Y_train,
nb_epoch=200,
batch_size=100,
validation_data=(X_test, Y_test))
# evaluate the model
print 'Evaluating...\n'
scores = model.evaluate(X_test, Y_test)
print Y_test.shape
for (name, score) in zip(model.metrics_names, scores):
print " %s: %0.3f" % (name, score)
predictions = model.predict(X_test) # calculate predictions
rounded = np.round(predictions)
rounded = np.expand_dims(rounded, axis=0)
print helper.windiff_metric_NUMPY(np.expand_dims(Y_test, axis=0), rounded)
pdb.set_trace()
def train_LSTM(X, Y, model, train_split=0.8, epochs=10, batch_size=32):
# Clinical
SAMPLE_TYPE_cli, X_cli, Y_cli = get_input(sample_type=4,
shuffle_documents=False,
pad=False)
which_model = 2
if which_model == 2:
custom_fit(X, Y, train_split=train_split, model=model, epochs=epochs)
print "Clinical Data"
custom_fit(X_cli, Y_cli, train_split=1, model=model) # Test clinical
elif which_modle == 1:
# Works for TYPE2 but check for others
# Both these lines work for which_model == 1
X_train, Y_train, X_test, Y_test = split_data(X,
Y,
train_split=train_split)
model.fit(X_train,
Y_train,
shuffle=False,
nb_epoch=epochs,
batch_size=batch_size,
validation_data=(X_test, Y_test))
# WIkipedia
#model.evaluate(X_test, Y_test, batch_size=batch_size)
#pred = model.predict(X_test)
#rounded = np.round(pred)
#result = helper.windiff_metric_NUMPY(Y_test, rounded)
#print result
# Clinical
# Temporary TRUNCATION
TRUNCATE_LEN = X_train.shape[1]
print "NOTE: Truncating the Test dataset(clinical) from %d sentences to %d sentences." % (
X_cli.shape[1], TRUNCATE_LEN)
X_cli, Y_cli = X_cli[:, :TRUNCATE_LEN, :], Y_cli[:, :TRUNCATE_LEN, :]
model.evaluate(X_cli, Y_cli, batch_size=batch_size)
pred = model.predict(X_cli)
rounded = np.round(pred)
result = helper.windiff_metric_NUMPY(Y_cli,
rounded,
win_size=10,
rounded=True)
print result
pdb.set_trace()
def train_LSTM(X,
Y,
model,
embedding_W,
train_split=0.8,
epochs=10,
batch_size=32):
global X_wiki, Y_wiki, X_cli, Y_cli, X_bio, Y_bio
which_model = 2
if which_model == 2:
custom_fit(X,
Y,
model=model,
batch_size=batch_size,
train_split=train_split,
epochs=epochs)
print "############## Clinical Data ###########"
custom_fit(X_cli,
Y_cli,
model=model,
batch_size=batch_size,
train_split=0,
epochs=-1) # Test clinical
print "############## Biography Data ###########"
custom_fit(X_bio,
Y_bio,
model=model,
batch_size=batch_size,
train_split=0,
epochs=-1) # Test biography
elif which_modle == 1:
# Works for TYPE2 but check for others
# Both these lines work for which_model == 1
X_train, Y_train, X_test, Y_test = split_data(X,
Y,
train_split=train_split)
model.fit(X_train,
Y_train,
shuffle=False,
nb_epoch=epochs,
batch_size=batch_size,
validation_data=(X_test, Y_test))
# WIkipedia
#model.evaluate(X_test, Y_test, batch_size=batch_size)
#pred = model.predict(X_test)
#rounded = np.round(pred)
#result = helper.windiff_metric_NUMPY(Y_test, rounded)
#print result
# Clinical
# Temporary TRUNCATION
TRUNCATE_LEN = X_train.shape[1]
print "NOTE: Truncating the Test dataset(clinical) from %d sentences to %d sentences." % (
X_cli.shape[1], TRUNCATE_LEN)
X_cli, Y_cli = X_cli[:, :TRUNCATE_LEN, :], Y_cli[:, :TRUNCATE_LEN, :]
model.evaluate(X_cli, Y_cli, batch_size=batch_size)
pred = model.predict(X_cli)
rounded = np.round(pred)
result = helper.windiff_metric_NUMPY(Y_cli,
rounded,
win_size=10,
rounded=True)
print result
pdb.set_trace()
def custom_fit(X, Y, model, train_split=0.8, epochs=10):
if train_split == 1:
X_test, Y_test = X, Y
else:
# This is only for training! (If train_split =1 then only TEST)
X_train, Y_train, X_test, Y_test = split_data(X,
Y,
train_split=train_split)
print "Batch size = 1"
print 'Train...'
for epoch in range(epochs):
mean_tr_acc = []
mean_tr_loss = []
for i in range(len(X_train)):
#y_true = Y_train[i]
for sequence, truth in zip(X_train[i],
Y_train[i]): # Sequence in document
sequence = sequence.reshape((1, sequence.shape[0]))
sequence = np.expand_dims(sequence, axis=0)
tr_loss, tr_acc = model.train_on_batch(
[sequence, sequence, sequence, sequence], truth)
mean_tr_acc.append(tr_acc)
mean_tr_loss.append(tr_loss)
model.reset_states()
print('accuracy training = {}'.format(np.mean(mean_tr_acc)))
print('loss training = {}'.format(np.mean(mean_tr_loss)))
print('___________________________________')
mean_te_acc = []
mean_te_loss = []
predictions = []
for i in range(len(X_test)):
for sequence, truth in zip(X_test[i], Y_test[i]):
sequence = sequence.reshape((1, sequence.shape[0]))
sequence = np.expand_dims(sequence, axis=0)
te_loss, te_acc = model.test_on_batch(
[sequence, sequence, sequence, sequence], truth)
mean_te_acc.append(te_acc)
mean_te_loss.append(te_loss)
model.reset_states()
predictions.append([])
for sequence, truth in zip(X_test[i], Y_test[i]):
sequence = sequence.reshape((1, sequence.shape[0]))
sequence = np.expand_dims(sequence, axis=0)
y_pred = model.predict_on_batch(
[sequence, sequence, sequence, sequence])
predictions[i].append(y_pred)
model.reset_states()
print('accuracy testing = {}'.format(np.mean(mean_te_acc)))
print('loss testing = {}'.format(np.mean(mean_te_loss)))
print "Check windiff value"
#rounded = np.round(predictions)
result = helper.windiff_metric_NUMPY(Y_test,
predictions,
win_size=-1,
rounded=False)
print result
print('___________________________________')
def custom_fit(X, Y, model, train_split=0.8, epochs=10):
if train_split == 0:
X_test, Y_test = X, Y
else:
# This is only for training! (If train_split =1 then only TEST)
X_train, Y_train, X_test, Y_test = split_data(X, Y, train_split=train_split)
print "Batch size = 1"
print 'Train...'
_total_docs = len(X_train)
_total_sentences = sum([sequence.shape[0] for sequence in X_train])
for epoch in range(epochs):
mean_tr_acc = []
mean_tr_loss = []
_sentence_no = 0
for i in range(len(X_train)):
#y_true = Y_train[i]
for sequence, truth in zip(X_train[i], Y_train[i]): # Sequence in document
sequence = sequence.reshape((1, sequence.shape[0]))
#sequence = np.expand_dims(sequence, axis=0)
tr_loss, tr_acc = model.train_on_batch([sequence], truth)
mean_tr_acc.append(tr_acc)
mean_tr_loss.append(tr_loss)
_sentence_no += 1
print ">> Epoch: %d/%d | Doc: %d/%d | Sent: %d/%d" %(epoch+1, epochs, i+1, _total_docs, _sentence_no+1, _total_sentences)
model.reset_states()
print('accuracy training = {}'.format(np.mean(mean_tr_acc)))
print('loss training = {}'.format(np.mean(mean_tr_loss)))
print('___________________________________')
# Testing
mean_te_acc = []
mean_te_loss = []
predictions = []
_total_docs = len(X_test)
_total_sentences = sum([sequence.shape[0] for sequence in X_test])
_sentence_no = 0
for i in range(len(X_test)):
for sequence, truth in zip(X_test[i], Y_test[i]):
sequence = sequence.reshape((1, sequence.shape[0]))
#sequence = np.expand_dims(sequence, axis=0)
te_loss, te_acc = model.test_on_batch([sequence], truth)
mean_te_acc.append(te_acc)
mean_te_loss.append(te_loss)
_sentence_no += 1
print ">> TEST >> Doc: %d/%d | Sent: %d/%d" %(i+1, _total_docs, _sentence_no+1, _total_sentences)
model.reset_states()
print('accuracy testing = {}'.format(np.mean(mean_te_acc)))
print('loss testing = {}'.format(np.mean(mean_te_loss)))
print("Predicting...")
for i in range(len(X_test)):
predictions.append([])
for sequence, truth in zip(X_test[i], Y_test[i]):
sequence = sequence.reshape((1, sequence.shape[0]))
#sequence = np.expand_dims(sequence, axis=0)
y_pred = model.predict_on_batch([sequence])
predictions[i].append(y_pred)
model.reset_states()
print "Check windiff value"
#rounded = np.round(predictions)
result = helper.windiff_metric_NUMPY(Y_test, predictions, win_size=-1, rounded=False)
print result
print('___________________________________')