def main(args):
#Initialise config vars
NGRAMS = args.ngrams
BATCH_SIZE = args.batch
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
if (args.backend == 'gpu' and not torch.cuda.is_available()):
logger.error('Backend device: %s not available', args.backend)
if args.backend != 'auto':
device = torch.device('cpu' if args.backend == 'cpu' else 'cuda')
EMBED_DIM = args.model_dim
N_EPOCHS = args.epochs
MAX_SEQ_LEN = args.max_seq_len
TRAIN_SPLIT = args.train_split
EVOLVED = args.evolved
#logging config vars
logger.info(
'Device:%s|Batch size:%s|EmbedDim:%s|Epochs:%s|Ngrams:%s|MAX_LEN:%s|Split:%s',
device.type, BATCH_SIZE, EMBED_DIM, N_EPOCHS, NGRAMS, MAX_SEQ_LEN,
TRAIN_SPLIT)
import os
#download data
logger.info('Loading Data')
# train_dataset,test_dataset,x_vocab,y_vocab=read_data('./.data/ag_news_csv',ngrams=NGRAMS)
data_reader = DataReader('./.data/ag_news_csv')
#train_dataset, test_dataset = text_classification.DATASETS['AG_NEWS'](root='./.data', ngrams=NGRAMS, vocab=None)
logger.info('Data Loaded')
VOCAB_SIZE = data_reader.get_vocab_size()
NUM_CLASS = data_reader.get_num_classes()
train_dataset = data_reader.get_training_data()
test_dataset = data_reader.get_testing_data()
model = ClassificationTransformer(EMBED_DIM,
VOCAB_SIZE,
NUM_CLASS,
max_seq_len=MAX_SEQ_LEN,
evolved=EVOLVED)
model.to(device)
from torch.utils.data.dataset import random_split
min_valid_loss = float('inf')
criterion = torch.nn.CrossEntropyLoss().to(device)
optimizer = torch.optim.SGD(model.parameters(), lr=0.01)
scheduler = torch.optim.lr_scheduler.StepLR(optimizer, 1, gamma=0.9)
train_len = int(len(train_dataset) * TRAIN_SPLIT)
sub_train_, sub_valid_ = random_split(
train_dataset, [train_len, len(train_dataset) - train_len])
for epoch in range(N_EPOCHS):
train_loss, train_acc = train_func(sub_train_, data_reader, model,
BATCH_SIZE, device, optimizer,
scheduler, criterion, MAX_SEQ_LEN)
logger.info('Trained for epoch %s', str(epoch))
valid_loss, valid_acc = test(sub_valid_, data_reader, model,
BATCH_SIZE, device, optimizer, criterion,
MAX_SEQ_LEN)
TrainingLossStr = f'\tLoss: {train_loss:.4f}(train)\t|\tAcc: {train_acc * 100:.1f}%(train)'
ValidationLossStr = f'\tLoss: {valid_loss:.4f}(valid)\t|\tAcc: {valid_acc * 100:.1f}%(valid)'
logger.info(TrainingLossStr)
logger.info(ValidationLossStr)
