def main():
device = torch.device('cuda')
embedding_vectors = torch.load(f'{EMBEDDINGS_DIR}/vectors.pkl')
text_processor = TextProcessor(
wti=pickle.load(open(f'{EMBEDDINGS_DIR}/wti.pkl', 'rb')),
tokenizer=get_tokenizer('basic_english'),
standardize=True,
min_len=3,
)
dataset = TextDataset(CORPUS_DIR, text_processor)
# split into training and test set
# TODO: fix this splitting sometimes failing when corpus size changes
train_set, test_set = torch.utils.data.random_split(
dataset, [
int(len(dataset) * DATA_SPLIT),
int(len(dataset) * (1.0 - DATA_SPLIT))
])
# count number of samples in each class
class_count = [0, 0]
for data, label in dataset:
class_count[int(label.item())] += 1
# get relative weights for classes
_sum = sum(class_count)
class_count[0] /= _sum
class_count[1] /= _sum
# reverse the weights since we're getting the inverse for the sampler
class_count = list(reversed(class_count))
# set weight for every sample
weights = [class_count[int(x[1].item())] for x in train_set]
# weighted sampler
sampler = torch.utils.data.WeightedRandomSampler(
weights=weights, num_samples=len(train_set), replacement=True)
train_loader = DataLoader(dataset=train_set,
batch_size=32,
collate_fn=Sequencer(SEQUENCE_LEN),
sampler=sampler)
test_loader = DataLoader(dataset=test_set,
batch_size=32,
collate_fn=Sequencer(SEQUENCE_LEN))
# number of filters in each convolutional filter
N_FILTERS = 64
# sizes and number of convolutional layers
FILTER_SIZES = [2, 3]
# dropout for between conv and dense layers
DROPOUT = 0.5
model = TextCNN(
embeddings=embedding_vectors,
n_filters=N_FILTERS,
filter_sizes=FILTER_SIZES,
dropout=DROPOUT,
).to(device)
print(model)
print('Trainable params:',
sum(p.numel() for p in model.parameters() if p.requires_grad))
criterion = nn.BCELoss()
optimizer = torch.optim.Adam(model.parameters(), lr=1e-3)
EPOCHS = 12
best_acc = 0.0
# training loop
for epoch in range(EPOCHS):
print('Epoch', epoch + 1)
for i, data in tqdm(enumerate(train_loader), total=len(train_loader)):
# get word indices vector and corresponding labels
x, labels = data
# send to device
x = x.to(device)
labels = labels.to(device)
# make predictions
predictions = model(x).squeeze()
# calculate loss
loss = criterion(predictions, labels)
# learning stuff...
optimizer.zero_grad()
loss.backward()
optimizer.step()
# evaluate
with torch.no_grad():
model.eval()
correct = 0
wrong = 0
m = [[0, 0], [0, 0]]
for data in test_loader:
x, label = data
x = x.to(device)
predictions = model(x).squeeze()
for truth, prediction in zip(label, predictions):
y = int(truth.item())
y_pred = 1 if prediction.item() > 0.5 else 0
m[y][y_pred] += 1
if y == y_pred:
correct += 1
else:
wrong += 1
model.train()
acc = correct / (correct + wrong)
if acc > best_acc:
best_acc = acc
for file in glob.glob('models/model_*.pth'):
os.remove(file)
torch.save(model.state_dict(), f'models/state_{epoch}.pth')
print()
print('Correct:', f'{correct}/{correct + wrong}', 'Accuracy:', acc)
print('[[TN, FP], [FN, TP]]')
print(m)
print()
# put into evaluation mode
model.eval()
text_processor.do_standardize = True
with torch.no_grad():
while True:
text = input('Prompt: ')
x = text_processor.process(text)
x = torch.tensor(x).unsqueeze(dim=0)
print(model(x.to(device)).squeeze())
