## Append sample to training list
x_test.append(line)
## Convert list to array
x_test = np.asarray(x_test)
## Correct classification (first half positive, second half negative
y_test = np.zeros((25000, ))
y_test[0:12500] = 1
## Allow 8000 words + 1 unknown
vocab_size += 1
## Define model with 500 hidden units
model = BOW_model(vocab_size, 500)
model.cuda()
## Define optimizer
# opt = 'sgd'
# LR = 0.01
opt = 'adam'
LR = 0.01
if (opt == 'adam'):
optimizer = optim.Adam(model.parameters(), lr=LR)
elif (opt == 'sgd'):
optimizer = optim.SGD(model.parameters(), lr=LR, momentum=0.9)
## Hyper-parameters
batch_size = 200
no_of_epochs = 40
def main(input_optimizer, input_batch_size, input_hidden_units, input_epochs):
glove_embeddings = np.load('../preprocessed_data/glove_embeddings.npy')
vocab_size = 100000
x_train = []
with io.open('../preprocessed_data/imdb_train_glove.txt',
'r',
encoding='utf-8') as f:
lines = f.readlines()
for line in lines:
line = line.strip()
line = line.split(' ')
line = np.asarray(line, dtype=np.int)
line[line > vocab_size] = 0
line = line[line != 0]
line = np.mean(glove_embeddings[line], axis=0)
x_train.append(line)
x_train = np.asarray(x_train)
x_train = x_train[0:25000]
y_train = np.zeros((25000, ))
y_train[0:12500] = 1
x_test = []
with io.open('../preprocessed_data/imdb_test_glove.txt',
'r',
encoding='utf-8') as f:
lines = f.readlines()
for line in lines:
line = line.strip()
line = line.split(' ')
line = np.asarray(line, dtype=np.int)
line[line > vocab_size] = 0
line = line[line != 0]
line = np.mean(glove_embeddings[line], axis=0)
x_test.append(line)
x_test = np.asarray(x_test)
y_test = np.zeros((25000, ))
y_test[0:12500] = 1
vocab_size += 1
# no_hidden_units = 500 # try 300 as well
no_hidden_units = input_hidden_units
model = BOW_model(no_hidden_units)
model.cuda()
# opt = 'sgd'
# LR = 0.01
# opt = 'adam'
opt = input_optimizer
LR = 0.001
if (opt == 'adam'):
optimizer = optim.Adam(model.parameters(), lr=LR)
elif (opt == 'sgd'):
optimizer = optim.SGD(model.parameters(), lr=LR, momentum=0.9)
# batch_size = 200
batch_size = input_batch_size
# no_of_epochs = 6
no_of_epochs = input_epochs
L_Y_train = len(y_train)
L_Y_test = len(y_test)
model.train()
train_loss = []
train_accu = []
test_accu = []
print("Optmizer: %s" % opt, "LR: %.6f" % LR,
"EpochSize: %d" % no_of_epochs, "BatchSize: %d" % batch_size,
"VocalSize: %d" % (vocab_size - 1),
"HidenSize: %d" % no_hidden_units)
for epoch in range(no_of_epochs):
# training
model.train()
epoch_acc = 0.0
epoch_loss = 0.0
epoch_counter = 0
time1 = time.time()
I_permutation = np.random.permutation(L_Y_train)
for i in range(0, L_Y_train, batch_size):
x_input = x_train[I_permutation[i:i + batch_size]]
y_input = y_train[I_permutation[i:i + batch_size]]
data = Variable(torch.FloatTensor(x_input)).cuda()
target = Variable(torch.FloatTensor(y_input)).cuda()
optimizer.zero_grad()
loss, pred = model(data, target)
loss.backward()
optimizer.step() # update weights
prediction = pred >= 0.0
truth = target >= 0.5
acc = prediction.eq(truth).sum().cpu().data.numpy()
epoch_acc += acc
epoch_loss += loss.data.item()
epoch_counter += batch_size
epoch_acc /= epoch_counter
epoch_loss /= (epoch_counter / batch_size)
train_loss.append(epoch_loss)
train_accu.append(epoch_acc)
print(epoch, "%.2f" % (epoch_acc * 100.0), "%.4f" % epoch_loss,
"%.4f" % float(time.time() - time1))
# ## test
model.eval()
epoch_acc = 0.0
epoch_loss = 0.0
epoch_counter = 0
time1 = time.time()
I_permutation = np.random.permutation(L_Y_test)
for i in range(0, L_Y_test, batch_size):
x_input = x_train[I_permutation[i:i + batch_size]]
y_input = y_train[I_permutation[i:i + batch_size]]
data = Variable(torch.FloatTensor(x_input)).cuda()
target = Variable(torch.FloatTensor(y_input)).cuda()
with torch.no_grad():
loss, pred = model(data, target)
prediction = pred >= 0.0
truth = target >= 0.5
acc = prediction.eq(truth).sum().cpu().data.numpy()
epoch_acc += acc
epoch_loss += loss.data.item()
epoch_counter += batch_size
epoch_acc /= epoch_counter
epoch_loss /= (epoch_counter / batch_size)
test_accu.append(epoch_acc)
time2 = time.time()
time_elapsed = time2 - time1
print(" ", "%.2f" % (epoch_acc * 100.0), "%.4f" % epoch_loss)
torch.save(model, 'BOW.model')
data = [train_loss, train_accu, test_accu]
data = np.asarray(data)
np.save('data.npy', data)
line = line.split(' ')
line = np.asarray(line, dtype=np.int)
line[line > vocab_size] = 0
line = line[line != 0]
line = np.mean(glove_embeddings[line], axis=0)
x_test.append(line)
x_test = np.asarray(x_test)
y_test = np.zeros((25000, ))
y_test[0:12500] = 1
vocab_size += 1
model = BOW_model(1000) # try 300 as well
model.cuda()
####################################################################
# opt = 'sgd'
# LR = 0.01
opt = 'adam'
LR = 0.001
if (opt == 'adam'):
optimizer = optim.Adam(model.parameters(), lr=LR)
elif (opt == 'sgd'):
optimizer = optim.SGD(model.parameters(), lr=LR, momentum=0.9)
batch_size = 200
line = np.asarray(line, dtype = np.int) line[line > vocab_size] = 0 line = line[line != 0] line = np.mean(glove_embeddings[line], axis = 0) x_test.append(line) x_test = np.asarray(x_test) y_test = np.zeros((25000, )) y_test[0:12500] = 1 #----------------------------------------------------------------- # Load model vocab_size += 1 model = BOW_model(no_of_hidden_units = 100) model.cuda() # Optimizer and learning rate # opt = "SGD" # LR = 0.01 opt = "Adam" LR = 0.001 if opt == "SGD": optimizer = optim.SGD(model.parameters(), lr = LR, momentum = 0.9) elif opt == "Adam": optimizer = optim.Adam(model.parameters(),
encoding='utf-8') as f:
lines = f.readlines()
for line in lines:
line = line.strip()
line = line.split(' ')
line = np.asarray(line, dtype=np.int)
line[line > vocab_size] = 0
x_test.append(line)
y_test = np.zeros((25000, ))
y_test[0:12500] = 1
vocab_size += 1
model = BOW_model(vocab_size, 1000)
model.cuda()
# opt = 'sgd'
# LR = 0.01
opt = 'adam'
LR = 0.001
if (opt == 'adam'):
optimizer = optim.Adam(model.parameters(), lr=LR, weight_decay=10**-6)
elif (opt == 'sgd'):
optimizer = optim.SGD(model.parameters(), lr=LR, momentum=0.9)
batch_size = 200
no_of_epochs = 100
L_Y_train = len(y_train)
L_Y_test = len(y_test)
for line in lines:
line = line.strip()
line = line.split(' ')
line = np.asarray(line, dtype=np.int)
line[line > vocab_size] = 0
line = line[line != 0]
line = np.mean(glove_embeddings[line], axis=0)
x_test.append(line)
x_test = np.asarray(x_test)
y_test = np.zeros((25000, ))
y_test[0:12500] = 1
model = BOW_model(500)
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model.to(device)
# opt = 'sgd'
# LR = 0.01
opt = 'adam'
LR = 0.001
if (opt == 'adam'):
optimizer = optim.Adam(model.parameters(), lr=LR)
elif (opt == 'sgd'):
optimizer = optim.SGD(model.parameters(), lr=LR, momentum=0.9)
batch_size = 200
no_of_epochs = 6
L_Y_train = len(y_train)
line = line.split(' ')
line = np.asarray(line, dtype=np.int)
line[line > vocab_size] = 0
line = line[line != 0]
line = np.mean(glove_embeddings[line], axis=0)
x_test.append(line)
x_test = np.asarray(x_test)
y_test = np.zeros((25000, ))
y_test[0:12500] = 1
vocab_size += 1
model = BOW_model(300) # try 300 as well
model.cuda()
####################################################################
# opt = 'sgd'
# LR = 0.01
opt = 'adam'
LR = 0.001
if (opt == 'adam'):
optimizer = optim.Adam(model.parameters(), lr=LR, weight_decay=10**-4)
elif (opt == 'sgd'):
optimizer = optim.SGD(model.parameters(), lr=LR, momentum=0.9)
batch_size = 200
lines = f.readlines()
for line in lines:
line = line.strip()
line = line.split(' ')
line = np.asarray(line, dtype=np.int)
line[line > vocab_size] = 0
x_test.append(line)
y_test = np.zeros((25000, ))
y_test[0:12500] = 1
vocab_size += 1
for idx, no_of_hidden_units in enumerate(no_of_hidden_units_list):
model = BOW_model(vocab_size, no_of_hidden_units)
model.cuda()
optimizer = optim.Adam(model.parameters(), lr=LR)
L_Y_train = len(y_train)
L_Y_test = len(y_test)
model.train()
train_loss = []
train_accu = []
test_accu = []
for epoch in range(no_of_epochs[idx]):
with io.open('../preprocessed_data/imdb_test.txt', 'r', encoding='utf-8') as f:
lines = f.readlines()
for line in lines:
line = line.strip()
line = line.split(' ')
line = np.asarray(line, dtype=np.int)
line[line > vocab_size] = 0
x_test.append(line)
y_test = np.zeros((25000, ))
y_test[0:12500] = 1
vocab_size += 1
model = BOW_model(vocab_size, 500)
model.cuda()
# opt = 'sgd'
# LR = 0.01
opt = 'adam'
LR = 0.001
if (opt == 'adam'):
optimizer = optim.Adam(model.parameters(), lr=LR)
elif (opt == 'sgd'):
optimizer = optim.SGD(model.parameters(), lr=LR, momentum=0.9)
batch_size = 200
no_of_epochs = 6
L_Y_train = len(y_train)
L_Y_test = len(y_test)