model.apply(init_weights)
print(f'The model has {count_parameters(model):,} train parameters')
optimizer = optim.Adam(model.parameters())
# weight_CE = torch.FloatTensor([0.1,1,1,1,1,1,1,1,1,1]).to(device)
# criterion = nn.CrossEntropyLoss(weight = weight_CE)
criterion = nn.CrossEntropyLoss()
# train data
train_paths = glob.glob("data/train/*.json")
test_paths = glob.glob("data/test/*.json")
train_steps = len(train_paths) // config.s2s_batch_size + 1
test_steps = len(test_paths) // config.s2s_batch_size + 1
train_iterator = gen(train_paths, config.s2s_batch_size,
config.max_box_num, device)
test_iterator = gen(test_paths, config.s2s_batch_size, config.max_box_num,
device)
best_valid_loss = float('inf')
for epoch in range(config.s2s_epoch):
start_time = time.time()
train_loss, train_acc = train(model, train_steps, train_iterator,
optimizer, criterion, config.s2s_clip)
valid_loss, valid_acc = evaluate(model, test_steps, test_iterator,
criterion)
end_time = time.time()
epoch_mins, epoch_secs = epoch_time(start_time, end_time)
attn = Attention(config.s2s_enc_hid, config.s2s_dec_hid)
enc = Encoder(config.s2s_emb_dim, config.s2s_enc_hid, config.s2s_dec_hid,
config.s2s_enc_dropout)
dec = Decoder(len(config.class_char), config.s2s_emb_dim,
config.s2s_enc_hid, config.s2s_dec_hid,
config.s2s_enc_dropout, attn)
model = Seq2Seq(enc, dec, device).to(device)
model.apply(init_weights)
model.load_state_dict(torch.load('weight/s2s.pt'))
model.eval()
data = gen(["data/test/96.json"], 1, config.max_box_num, device)
with torch.no_grad():
src, trg = next(data)
output = model(src)
output = output.permute(1, 0,
2).contiguous().view(-1,
len(config.class_char))
output = torch.max(F.softmax(output, dim=1), 1)
possible, label = output.values, output.indices
acc = np.mean((label == trg).cpu().numpy())
print("trget:", trg.long())
print("label:", label)
print("possible:", possible)
print("acc:", acc)
attn = Attention(config.s2s_enc_hid, config.s2s_dec_hid)
enc = Encoder(config.s2s_emb_dim, config.s2s_enc_hid, config.s2s_dec_hid,
config.s2s_enc_dropout)
dec = Decoder(len(config.class_char), config.s2s_emb_dim,
config.s2s_enc_hid, config.s2s_dec_hid,
config.s2s_enc_dropout, attn)
model = Seq2Seq(enc, dec, device).to(device)
model.apply(init_weights)
model.load_state_dict(torch.load('weight/s2s.pt'))
model.eval()
data = gen(["data/test/OTDR_type2_64.json"], 1, config.max_box_num, device)
with torch.no_grad():
src, trg = next(data)
output = model(src)
output = output.permute(1, 0,
2).contiguous().view(-1,
len(config.class_char))
output = torch.max(F.softmax(output, dim=1), 1)
possible, label = output.values, output.indices
acc = np.mean((label == trg).cpu().numpy())
print("trget:", trg.long())
print("label:", label)
print("possible:", possible)
print("acc:", acc)