def main():
Place = paddle.fluid.CUDAPlace(0)
with fluid.dygraph.guard(Place):
model = Transformer(image_size=512,
num_classes=15,
hidden_unit_num=1024,
layer_num=2,
head_num=16,
dropout=0.8,
decoder_name='PUP',
hyber=True,
visualable=False)
preprocess = Transform(512)
dataloader_1 = Dataloader('/home/aistudio/dataset',
'/home/aistudio/dataset/val_list.txt',
transform=preprocess,
shuffle=True)
val_load = fluid.io.DataLoader.from_generator(capacity=1,
use_multiprocess=False)
val_load.set_sample_generator(dataloader_1, batch_size=1, places=Place)
model_dic, optic_dic = load_dygraph(
"./output/SETR-NotZero-Epoch-2-Loss-0.161517-MIOU-0.325002")
model.load_dict(model_dic)
model.eval()
'''result = get_infer_data("/home/aistudio/dataset/infer")
infer_load = Load_infer('/home/aistudio/dataset', result, transform=preprocess, shuffle=False)
loader_infer= fluid.io.DataLoader.from_generator(capacity=1, use_multiprocess=False)
loader_infer.set_sample_generator(infer_load, batch_size=1, places=Place)
process_image(model, loader_infer, result)'''
validation(val_load, model, 15)
for epoch in range(ARGS.n_epochs):
for train_batch in train_iter:
train_batch = utils.rebatch(pad_idx, train_batch)
train_out = model(train_batch.src, train_batch.trg,
train_batch.src_mask, train_batch.trg_mask)
train_loss = criterion(
train_out.contiguous().view(-1, train_out.size(-1)),
train_batch.trg_y.contiguous().view(-1)) / train_batch.ntokens
train_loss.backward()
optimizer.step()
optimizer.optimizer.zero_grad()
print(f'{iter_cnt} - train_loss: {train_loss.item()}')
if iter_cnt % ARGS.eval_steps == 0:
model.eval()
val_total_loss = 0
val_ntokens = 0
with torch.no_grad():
for val_batch in val_iter:
val_batch = utils.rebatch(pad_idx, val_batch)
val_out = model(val_batch.src, val_batch.trg,
val_batch.src_mask, val_batch.trg_mask)
val_loss = criterion(
val_out.contiguous().view(-1, val_out.size(-1)),
val_batch.trg_y.contiguous().view(-1))
val_total_loss += val_loss.item()
val_ntokens += val_batch.ntokens
norm_val_loss = val_total_loss / val_ntokens.float()