def train(encoder,
train_senetences,
train_labels,
w2i,
iter=1,
learning_rate=0.001):
optimizer = optim.Adam(encoder.parameters(), lr=learning_rate)
for i in range(iter):
sentence, label = findTrainExample(train_senetences, train_labels)
sentence_tensor = pp.encodeSentence(sentence, w2i)
label_tensor = torch.tensor(label, dtype=torch.float32).view(1, 1)
encoder_hidden = encoder.initHidden()
input_length = sentence_tensor.size(0)
for ei in range(input_length):
output, encoder_hidden = encoder(sentence_tensor[ei],
encoder_hidden)
loss = torch.abs(output - label_tensor)
if i % 100 == 0:
print("loss: ", loss)
optimizer.zero_grad()
loss.backward(retain_graph=True)
optimizer.step()
def batch_train(encoder, train_sentences, train_labels, batch_size, w2i, epochs=5, learning_rate=0.001):
optimizer = optim.Adam(encoder.parameters(), lr=learning_rate)
for i in range(1000):
batch_sentences, batch_labels = findBatch(train_sentences, train_labels, batch_size)
print('found batch')
loss = 0
encoder_hidden = encoder.initHidden()
for j in range(len(batch_sentences)):
sentence = batch_sentences[j]
label = batch_labels[j]
sentence_tensor = pp.encodeSentence(sentence,w2i)
label_tensor = torch.tensor(label,dtype=torch.float32).view(1,1)
input_length = sentence_tensor.size(0)
for ei in range(input_length):
output, encoder_hidden = encoder(sentence_tensor[ei],encoder_hidden)
loss += torch.abs(output - label_tensor)
#print(loss)
loss = loss/len(batch_sentences)
print(loss)
optimizer.zero_grad()
loss.backward(retain_graph=True)
optimizer.step()
def evaluate(encoder, test_sentences, test_labels, w2i):
accuracy = 0.0
#print("encoder weight: ",encoder.out.weight)
for i in range(len(test_sentences)):
sentence = test_sentences[i]
label = test_labels[i]
sentence_tensor = pp.encodeSentence(sentence, w2i)
label_tensor = torch.tensor(label, dtype=torch.float32).view(1, 1)
encoder_hidden = encoder.initHidden()
#encoder_hidden = encoder_hidden.to(device)
#input_length = sentence_tensor.size(0)
#for ei in range(input_length):
sentence_tensor = sentence_tensor.to(device)
label_tensor = label_tensor.to(device)
output = encoder(sentence_tensor, encoder_hidden)
#print("output from encoder: ",output)
#output = torch.abs(output)
output = torch.round(output)
#print("output: ",output)
#print("label: ",label_tensor)
if torch.equal(output, label_tensor):
accuracy += 1
#print ("accuracy: ",accuracy)
return accuracy / len(test_sentences)
def batch_train(encoder, train_sentences, train_labels, batch_size, w2i, epochs=1, learning_rate=0.001):
optimizer = optim.Adam(encoder.parameters(), lr=learning_rate)
for i in range(10):
#print("encoder weight: ",encoder.out.weight)
batch_sentences, batch_labels = findBatch(train_sentences, train_labels, batch_size)
#print('found batch')
loss = 0
#encoder_hidden = encoder.initHidden()
for j in range(len(batch_sentences)):
sentence = batch_sentences[j]
label = batch_labels[j]
sentence_tensor = pp.encodeSentence(sentence,w2i)
label_tensor = torch.tensor(label,dtype=torch.float32).view(1,1)
encoder_hidden = encoder.initHidden()
#encoder_hidden = encoder_hidden.to(device)
#input_length = sentence_tensor.size(0)
#for ei in range(input_length):
sentence_tensor = sentence_tensor.to(device)
label_tensor = label_tensor.to(device)
output = encoder(sentence_tensor,encoder_hidden)
loss += torch.mul((output - label_tensor),(output - label_tensor))
#print(loss)
loss = loss/len(batch_sentences)
print(loss)
optimizer.zero_grad()
loss.backward(retain_graph=True)
optimizer.step()
def evaluate(encoder, test_sentences, test_labels, w2i):
accuracy = 0.0
for i in range(len(test_sentences)):
sentence = test_sentences[i]
label = test_labels[i]
sentence_tensor = pp.encodeSentence(sentence,w2i)
label_tensor = torch.tensor(label,dtype=torch.float32).view(1,1)
encoder_hidden = encoder.initHidden()
input_length = sentence_tensor.size(0)
for ei in range(input_length):
output, encoder_hidden = encoder(sentence_tensor[ei],encoder_hidden)
output = torch.round(output)
if torch.equal(output,label_tensor):
accuracy+=1
return accuracy/len(test_sentences)