def test_train(self):
"""
Perform train on very simple dataset, consisting only with single example
"""
data = ['aa']
encoder = input.InputEncoder(data)
net = model.Model(len(encoder), hidden_size=10)
objective = nn.CrossEntropyLoss()
optimizer = optim.SGD(net.parameters(), lr=0.1)
max_loss = min_loss = None
for _ in range(100):
net.zero_grad()
packed_seq, true_indices = input.batch_to_train(data, encoder)
out, _ = net(packed_seq)
v_loss = objective(out, true_indices)
v_loss.backward()
optimizer.step()
l = v_loss.data[0]
if max_loss is None:
max_loss = min_loss = l
else:
max_loss = max(max_loss, l)
min_loss = min(min_loss, l)
loss_drop = max_loss / min_loss
self.assertGreater(loss_drop, 10.0)
# try to predict letters
out = model.inference_one(encoder.START_TOKEN, encoder, net)
self.assertEquals(out, 'a')
out = model.inference_one('a', encoder, net)
self.assertEquals(out, 'a')
out = model.inference_one('aa', encoder, net)
self.assertEquals(out, encoder.END_TOKEN)
def test_simple(self):
enc = input.InputEncoder(['abc'])
self.assertEqual(5, len(enc)) # 3 + start + end
d = enc.encode("aaa")
self.assertEqual(d.shape, (5, len(enc)))
self.assertEqual(d[0][0], 1.0)
self.assertEqual(d[1][2], 1.0)
d = enc.encode("a", width=4)
self.assertEqual(d.shape, (6, len(enc)))
def test_simple(self):
encoder = input.InputEncoder(['abcde'])
items = ["abc", "ab"]
lens = list(map(len, items))
dat = np.array([
encoder.encode(s, width=max(lens))
for s in items
])
v_dat = Variable(torch.from_numpy(dat))
seq = rnn_utils.pack_padded_sequence(v_dat, lens, batch_first=True)
m = model.Model(len(encoder), 16)
res = m.forward(seq)
def test_generate_batch_2(self):
data = ['aa', 'b']
encoder = input.InputEncoder(data)
net = model.Model(len(encoder), hidden_size=10)
objective = nn.CrossEntropyLoss()
optimizer = optim.SGD(net.parameters(), lr=0.1)
max_loss = min_loss = None
for _ in range(1000):
net.zero_grad()
packed_seq, true_indices = input.batch_to_train(data, encoder)
out, _ = net(packed_seq)
v_loss = objective(out, true_indices)
v_loss.backward()
optimizer.step()
l = v_loss.data[0]
if max_loss is None:
max_loss = min_loss = l
else:
max_loss = max(max_loss, l)
min_loss = min(min_loss, l)
res = [model.generate_name(net, encoder) for _ in range(10)]
print(res)
log = logging.getLogger("train")
if __name__ == "__main__":
logging.basicConfig(
format="%(asctime)-15s %(levelname)s %(name)-14s %(message)s",
level=logging.INFO)
parser = argparse.ArgumentParser()
parser.add_argument("--cuda",
default=False,
action='store_true',
help="Enable cuda")
args = parser.parse_args()
data = input.read_data()
input_encoder = input.InputEncoder(data)
log.info("Read %d train samples, encoder len=%d, first 10: %s", len(data),
len(input_encoder), ', '.join(data[:10]))
net = model.Model(len(input_encoder), hidden_size=HIDDEN_SIZE)
if args.cuda:
net = net.cuda()
objective = nn.CrossEntropyLoss()
lr = 0.004
# decay of LR every 100 epoches
lr_decay = 0.9
optimizer = optim.RMSprop(net.parameters(), lr=lr)
for epoch in range(EPOCHES):
def test_batch_to_train(self):
enc = input.InputEncoder(['abc'])
packed_seq, true_vals = input.batch_to_train(['a', 'b', 'c'], enc)
self.assertEqual(true_vals.data.numpy().tolist(),
[2, 1, 0, 3, 1, 0, 4, 1, 1])
def test_end_token(self):
enc = input.InputEncoder(['abc'])
r = enc.end_token()
self.assertEqual(5, r.shape[0])