def train(data_fpath, index_fpath):
dataset = TextData(data_fpath, index_fpath)
dataloader = DataLoader(
dataset,
batch_size=16,
shuffle=True,
num_workers=4,
collate_fn=collate_text
)
criterion = nn.BCEWithLogitsLoss()
print('number of tokens {}'.format(dataset.ntoks))
m = TextClassifier(dataset.ntoks)
optimizer = optim.Adam(
m.parameters(),
lr=0.0001
)
for eidx in range(10):
for bidx, batch in enumerate(dataloader):
optimizer.zero_grad()
#print(bidx, batch['transcriptions_one_hot'])
out = m(batch['transcriptions_one_hot'])
loss = criterion(out, batch['labels'])
print('epoch: {}\tstep: {}\tloss: {:.4f}'.format(eidx, bidx, loss.item()))
print('pred', (F.sigmoid(out).data > 0.5).numpy().astype(np.int))
print('true', batch['labels'].numpy().astype(np.int))
loss.backward()
optimizer.step()
def main(args):
data = TextData(args.source)
if args.checkpoint:
rnn = load_crnn(args.checkpoint)
else:
rnn = CharRNN(in_out_size=data.num_classes, state_size=args.state)
opt = Adagrad(rnn, 0.1, stateful=True, clip=5)
setup_plot()
sequence_pairs = list(data.get_seqs(25))
print('Training on {}:\n'
'- {} total chars\n'
'- {} unique chars\n'
'- {} sequences of length 25'.format(args.source, data.tot_chars,
data.num_classes,
len(sequence_pairs)))
opt.train(sequence_pairs,
epochs=40,
callback=partial(callback, data=data, start=time.time()),
callback_every=4321,
epoch_callback=epoch_callback)
plt.savefig('plots/{}.png'.format(basename(args.source)))
TextCNN新闻文本分类 ''' import numpy as np from data import TextData from public import TCNNConfig, plt_model from keras import Input from keras.preprocessing.sequence import pad_sequences from keras.utils import to_categorical, plot_model from keras.models import Sequential, load_model, Model from keras.layers import Dense, Dropout from keras.layers import Embedding, Conv1D, MaxPool1D, concatenate, Flatten from keras.optimizers import Adam # 读入训练数据 td = TextData() (x_data, y_data, z_data), (x_labels, y_labels, z_labels) = td.load_idata() word2id = td.word2id cat2id = td.cat2id num_classes = len(set(x_labels)) vocab_size = len(word2id) # 将每条文本固定为相同长度 x_data = pad_sequences(x_data, TCNNConfig.seq_length) x_labels = to_categorical(x_labels, num_classes=num_classes) y_data = pad_sequences(y_data, TCNNConfig.seq_length) y_labels = to_categorical(y_labels, num_classes=num_classes) def text_cnn(CNNConfig):
def setUp(cls):
cls.goblet = TextData('data/goblet_book.txt')
cls.rnn = CharRNN(in_out_size=cls.goblet.num_classes, state_size=100)
def train(self, train, test, num_epochs, path):
## Putting data into loaders
train_loader = DataLoader(
TextData(train),
batch_size=self.params['batch_size'],
num_workers=1,
shuffle=True)
test_loader = DataLoader(
TextData(test),
batch_size=self.params['batch_size'],
num_workers=1,
shuffle=True)
## Begin training
for epoch in range(num_epochs):
## Train cycle
self.model.train()
train_loss = 0
train_batches = 0
for X, y in tqdm.tqdm(train_loader):
X, y = X.cuda(), y.cuda()
self.model.opt.zero_grad()
## Forward
logits = self.model(X)
loss = self.criterion(logits, y)
## Backward
loss.backward()
self.model.opt.step()
## Measure stats
train_loss += loss.item()
train_batches += 1
## Log train loss for each batch so we get a good graph
wandb.log({'train_loss':loss.item()})
## Testing cycle
self.model.eval()
test_loss = 0
test_acc = 0
test_batches = 0
for X, y in tqdm.tqdm(test_loader):
X, y = X.cuda(), y.cuda()
## Forward
with torch.no_grad():
logits = self.model(X)
loss = self.criterion(logits, y)
## Measure stats
test_loss += loss.item()
probs = torch.softmax(logits, dim=-1)
test_acc += self.accuracy(probs, y)
test_batches += 1
## Report statistics
avg_train_loss = train_loss / test_batches
avg_test_loss = test_loss / test_batches
avg_test_acc = test_acc / test_batches
print(f'\
Train loss: {train_loss / train_batches}, Test loss: {avg_test_loss}, Test acc: {avg_test_acc}')
wandb.log({
'test_loss':avg_test_loss,
'test_acc':avg_test_acc})
## Save model each epoch
torch.save(self.model.state_dict(), path / f'model{epoch}.pt')
## Save model parameters at the end
dictToJson(self.params, path / 'model_params.json')