def setUp(self):
self.device = Naive()
self.graph = Graph()
Device.set_default(self.device)
Graph.set_default(self.graph)
self.a = np.array([[1, 2], [3, 4]], np.float32)
self.b = np.array([[1, 1], [4, 8]], np.float32)
def test_batch(encdec, src_vocab, trg_vocab, lines):
g = Graph()
Graph.set_default(g)
src_batch = make_batch(lines, list(range(len(lines))), src_vocab)
encdec.encode(src_batch, False)
# Generates target words one-by-one.
trg_ids = [np.array([trg_vocab["<bos>"]] * len(lines))]
eos_id = trg_vocab["<eos>"]
eos_ids = np.array([eos_id] * len(lines))
while (trg_ids[-1] != eos_ids).any():
if len(trg_ids) > GENERATION_LIMIT + 1:
print("Warning: Sentence generation did not finish in",
GENERATION_LIMIT,
"iterations.",
file=sys.stderr)
trg_ids.append(eos_ids)
break
y = encdec.decode_step(trg_ids[-1], False)
trg_ids.append(np.array(y.argmax(0)).T)
return [
hyp[:np.where(hyp == eos_id)[0][0]] for hyp in np.array(trg_ids[1:]).T
]
def setUp(self):
self.device = Naive()
self.graph = Graph()
Device.set_default(self.device)
Graph.set_default(self.graph)
self.ndarray_data = [
np.array([
[1, 2, 3],
[4, 5, 6],
[7, 8, 9],
[10, 11, 12],
], np.float32),
np.array([
[13, 14, 15],
[16, 17, 18],
[19, 20, 21],
[22, 23, 24],
], np.float32),
]
self.list_data = [
1.0,
4.0,
7.0,
10.0,
2.0,
5.0,
8.0,
11.0,
3.0,
6.0,
9.0,
12.0,
13.0,
16.0,
19.0,
22.0,
14.0,
17.0,
20.0,
23.0,
15.0,
18.0,
21.0,
24.0,
]
def train_func(trainer):
dev = D.Naive(12345)
Device.set_default(dev)
g = Graph()
Graph.set_default(g)
pw1 = Parameter([8, 2], I.XavierUniform())
pb1 = Parameter([8], I.Constant(0))
pw2 = Parameter([1, 8], I.XavierUniform())
pb2 = Parameter([1], I.Constant(0))
trainer.add_parameter(pw1)
trainer.add_parameter(pb1)
trainer.add_parameter(pw2)
trainer.add_parameter(pb2)
input_data = [1, 1, 1, -1, -1, 1, -1, -1]
output_data = [1, -1, -1, 1]
for i in range(10):
g.clear()
x = F.input(input_data, Shape([2], 4))
w1 = F.parameter(pw1)
b1 = F.parameter(pb1)
w2 = F.parameter(pw2)
b2 = F.parameter(pb2)
h = F.tanh(w1 @ x + b1)
y = w2 @ h + b2
t = F.input(output_data, Shape([], 4))
diff = t - y
loss = F.batch.mean(diff * diff)
trainer.reset_gradients()
loss.backward()
trainer.update()
return [
pw1.value.to_list(),
pb1.value.to_list(),
pw2.value.to_list(),
pb2.value.to_list()
]
def setUp(self):
self.device = Naive()
self.graph = Graph()
Device.set_default(self.device)
Graph.set_default(self.graph)
self.input_data = [
np.array([
[ 1, 2, 3],
[ 4, 5, 6],
[ 7, 8, 9],
], np.float32),
np.array([
[11,12,13],
[14,15,16],
[17,18,19],
], np.float32),
]
self.list_expected = [
1.0, 4.0, 7.0, 2.0, 5.0, 8.0, 3.0, 6.0, 9.0,
11.0, 14.0, 17.0, 12.0, 15.0, 18.0, 13.0, 16.0, 19.0,
]
def main():
# Loads vocab.
vocab = make_vocab("data/ptb.train.txt")
print("#vocab:", len(vocab)) # maybe 10000
eos_id = vocab["<s>"]
# Loads all corpus.
train_corpus = load_corpus("data/ptb.train.txt", vocab)
valid_corpus = load_corpus("data/ptb.valid.txt", vocab)
num_train_sents = len(train_corpus)
num_valid_sents = len(valid_corpus)
num_train_labels = count_labels(train_corpus)
num_valid_labels = count_labels(valid_corpus)
print("train:", num_train_sents, "sentences,", num_train_labels, "labels")
print("valid:", num_valid_sents, "sentences,", num_valid_labels, "labels")
# Device and computation graph.
dev = D.CUDA(0)
Device.set_default(dev)
g = Graph()
Graph.set_default(g)
# Our LM.
lm = RNNLM(len(vocab), eos_id)
# Optimizer.
optimizer = O.SGD(1)
#optimizer.set_weight_decay(1e-6)
optimizer.set_gradient_clipping(5)
optimizer.add(lm)
# Sentence IDs.
train_ids = list(range(num_train_sents))
valid_ids = list(range(num_valid_sents))
best_valid_ppl = 1e10
# Train/valid loop.
for epoch in range(MAX_EPOCH):
print("epoch", epoch + 1, "/", MAX_EPOCH, ":")
# Shuffles train sentence IDs.
random.shuffle(train_ids)
# Training.
train_loss = 0
for ofs in range(0, num_train_sents, BATCH_SIZE):
batch_ids = train_ids[ofs:min(ofs + BATCH_SIZE, num_train_sents)]
batch = make_batch(train_corpus, batch_ids, eos_id)
g.clear()
outputs = lm.forward(batch, True)
loss = lm.loss(outputs, batch)
train_loss += loss.to_float() * len(batch_ids)
optimizer.reset_gradients()
loss.backward()
optimizer.update()
print("%d" % ofs, end="\r")
sys.stdout.flush()
train_ppl = math.exp(train_loss / num_train_labels)
print(" train ppl =", train_ppl)
# Validation.
valid_loss = 0
for ofs in range(0, num_valid_sents, BATCH_SIZE):
batch_ids = valid_ids[ofs:min(ofs + BATCH_SIZE, num_valid_sents)]
batch = make_batch(valid_corpus, batch_ids, eos_id)
g.clear()
outputs = lm.forward(batch, False)
loss = lm.loss(outputs, batch)
valid_loss += loss.to_float() * len(batch_ids)
print("%d" % ofs, end="\r")
sys.stdout.flush()
valid_ppl = math.exp(valid_loss / num_valid_labels)
print(" valid ppl =", valid_ppl)
if valid_ppl < best_valid_ppl:
best_valid_ppl = valid_ppl
print(" BEST")
else:
old_lr = optimizer.get_learning_rate_scaling()
new_lr = 0.5 * old_lr
optimizer.set_learning_rate_scaling(new_lr)
print(" learning rate scaled:", old_lr, "->", new_lr)
def main():
# Loads data
train_inputs = load_images("data/train-images-idx3-ubyte", NUM_TRAIN_SAMPLES)
train_labels = load_labels("data/train-labels-idx1-ubyte", NUM_TRAIN_SAMPLES)
test_inputs = load_images("data/t10k-images-idx3-ubyte", NUM_TEST_SAMPLES)
test_labels = load_labels("data/t10k-labels-idx1-ubyte", NUM_TEST_SAMPLES)
# Initializes 2 device objects which manage different GPUs.
dev0 = D.CUDA(0)
dev1 = D.CUDA(1)
# Parameters on GPU 0.
pw1 = Parameter([NUM_HIDDEN_UNITS, NUM_INPUT_UNITS], I.XavierUniform(), dev0)
pb1 = Parameter([NUM_HIDDEN_UNITS], I.Constant(0), dev0)
# Parameters on GPU 1.
pw2 = Parameter([NUM_OUTPUT_UNITS, NUM_HIDDEN_UNITS], I.XavierUniform(), dev1)
pb2 = Parameter([NUM_OUTPUT_UNITS], I.Constant(0), dev1)
trainer = T.SGD(.1)
trainer.add_parameter(pw1)
trainer.add_parameter(pb1)
trainer.add_parameter(pw2)
trainer.add_parameter(pb2)
def make_graph(inputs):
# We first store input values explicitly on GPU 0.
x = F.input(inputs, device=dev0)
w1 = F.parameter(pw1)
b1 = F.parameter(pb1)
w2 = F.parameter(pw2)
b2 = F.parameter(pb2)
# The hidden layer is calculated and implicitly stored on GPU 0.
h_on_gpu0 = F.relu(w1 @ x + b1)
# `copy()` transfers the hiddne layer to GPU 1.
h_on_gpu1 = F.copy(h_on_gpu0, dev1)
# The output layer is calculated and implicitly stored on GPU 1.
return w2 @ h_on_gpu1 + b2
ids = list(range(NUM_TRAIN_SAMPLES))
g = Graph()
Graph.set_default(g)
for epoch in range(MAX_EPOCH):
random.shuffle(ids)
# Training loop
for batch in range(NUM_TRAIN_BATCHES):
print("\rTraining... %d / %d" % (batch + 1, NUM_TRAIN_BATCHES), end="")
inputs = [train_inputs[ids[batch * BATCH_SIZE + i]] for i in range(BATCH_SIZE)]
labels = [train_labels[ids[batch * BATCH_SIZE + i]] for i in range(BATCH_SIZE)]
g.clear()
y = make_graph(inputs)
loss = F.softmax_cross_entropy(y, labels, 0)
avg_loss = F.batch.mean(loss)
trainer.reset_gradients()
avg_loss.backward()
trainer.update()
print()
match = 0
# Test loop
for batch in range(NUM_TEST_BATCHES):
print("\rTesting... %d / %d" % (batch + 1, NUM_TEST_BATCHES), end="")
inputs = [test_inputs[batch * BATCH_SIZE + i] for i in range(BATCH_SIZE)]
g.clear()
y = make_graph(inputs)
y_val = y.to_list()
for i in range(BATCH_SIZE):
maxval = -1e10
argmax = -1
for j in range(NUM_OUTPUT_UNITS):
v = y_val[j + i * NUM_OUTPUT_UNITS]
if (v > maxval):
maxval = v
argmax = j
if argmax == test_labels[i + batch * BATCH_SIZE]:
match += 1
accuracy = 100.0 * match / NUM_TEST_SAMPLES
print("\nepoch %d: accuracy: %.2f%%\n" % (epoch, accuracy))
def main():
dev = D.Naive() # or D.CUDA(gpuid)
Device.set_default(dev)
# Parameters
pw1 = Parameter([8, 2], I.XavierUniform())
pb1 = Parameter([8], I.Constant(0))
pw2 = Parameter([1, 8], I.XavierUniform())
pb2 = Parameter([], I.Constant(0))
# Optimizer
optimizer = O.SGD(0.1)
# Registers parameters.
optimizer.add_parameter(pw1)
optimizer.add_parameter(pb1)
optimizer.add_parameter(pw2)
optimizer.add_parameter(pb2)
# Training data
input_data = [
np.array([1, 1], dtype=np.float32), # Sample 1
np.array([1, -1], dtype=np.float32), # Sample 2
np.array([-1, 1], dtype=np.float32), # Sample 3
np.array([-1, -1], dtype=np.float32), # Sample 4
]
output_data = [
np.array([1], dtype=np.float32), # Label 1
np.array([-1], dtype=np.float32), # Label 2
np.array([-1], dtype=np.float32), # Label 3
np.array([1], dtype=np.float32), # Label 4
]
g = Graph()
Graph.set_default(g)
for i in range(10):
g.clear()
# Builds a computation graph.
x = F.input(input_data)
w1 = F.parameter(pw1)
b1 = F.parameter(pb1)
w2 = F.parameter(pw2)
b2 = F.parameter(pb2)
h = F.tanh(w1 @ x + b1)
y = w2 @ h + b2
# Obtains values.
y_val = y.to_list()
print("epoch ", i, ":")
for j in range(4):
print(" [", j, "]: ", y_val[j])
# Extends the computation graph to calculate loss values.
t = F.input(output_data)
diff = t - y
loss = F.batch.mean(diff * diff)
# Obtains the loss.
loss_val = loss.to_float()
print(" loss: ", loss_val)
# Updates parameters.
optimizer.reset_gradients()
loss.backward()
optimizer.update()
def train(model, optimizer, config, best_valid):
max_epoch = config.get("max_epoch", int(1e9))
max_iteration = config.get("max_iteration", int(1e9))
max_sentences = config.get("max_sentences", 1e9)
max_tokens = config.get("max_tokens", 1e9)
update_freq = config.get('update_freq', 1)
optimizer.add(model)
corpus_prefix = Path(config['corpus_prefix']) / "subword"
model_path = corpus_prefix / "spm.model"
tokenizer = spm.SentencePieceProcessor()
tokenizer.Load(str(model_path))
train_src = load_corpus(corpus_prefix / Path(config["train_source"]).name,
tokenizer)
train_trg = load_corpus(corpus_prefix / Path(config["train_target"]).name,
tokenizer)
train_src, train_trg = clean_corpus(train_src, train_trg, config)
dev_src = load_corpus(corpus_prefix / Path(config["dev_source"]).name,
tokenizer)
dev_trg = load_corpus(corpus_prefix / Path(config["dev_target"]).name,
tokenizer)
dev_src, dev_trg = clean_corpus(dev_src, dev_trg, config)
num_train_sents = len(train_src)
num_dev_sents = len(dev_src)
eos_id = tokenizer.eos_id()
epoch = 0
iteration = 0
while epoch < max_epoch and iteration < max_iteration:
epoch += 1
g = Graph()
Graph.set_default(g)
train_itr = create_batch_itr(train_src,
train_trg,
max_tokens,
max_sentences,
shuffle=True)
train_itr = tqdm(train_itr, desc='train epoch {}'.format(epoch))
train_loss = 0.
itr_loss = 0.
itr_tokens = 0
itr_sentences = 0
optimizer.reset_gradients()
for step, batch_ids in enumerate(train_itr):
src_batch = make_batch(train_src, batch_ids, eos_id)
trg_batch = make_batch(train_trg, batch_ids, eos_id)
src_mask = padding_mask(src_batch, eos_id)
trg_mask = [
x | subsequent_mask(len(trg_batch) - 1)
for x in padding_mask(trg_batch[:-1], eos_id)
]
itr_tokens += len(src_batch) * len(src_batch[0])
itr_sentences += len(batch_ids)
g.clear()
loss = model.loss(src_batch, trg_batch, src_mask, trg_mask)
loss /= update_freq
loss.backward()
loss_val = loss.to_float()
train_loss += loss_val * update_freq * len(batch_ids)
itr_loss += loss_val
# with open('graph.dot', 'w') as f:
# print(g.dump("dot"), end="", file=f)
if (step + 1) % update_freq == 0:
step_num = optimizer.get_epoch() + 1
new_scale = config['d_model'] ** (-0.5) * \
min(step_num ** (-0.5), step_num * config['warmup_steps'] ** (-1.5))
optimizer.set_learning_rate_scaling(new_scale)
optimizer.update()
optimizer.reset_gradients()
iteration += 1
train_itr.set_postfix(itr=("%d" % (iteration)),
loss=("%.3lf" % (itr_loss)),
wpb=("%d" % (itr_tokens)),
spb=("%d" % (itr_sentences)),
lr=optimizer.get_learning_rate_scaling())
itr_loss = 0.
itr_tokens = 0
itr_sentences = 0
if iteration >= max_iteration:
break
print("\ttrain loss = %.4f" % (train_loss / num_train_sents))
g.clear()
valid_loss = 0.
valid_itr = create_batch_itr(dev_src,
dev_trg,
max_tokens,
max_sentences,
shuffle=False)
valid_itr = tqdm(valid_itr, desc='valid epoch {}'.format(epoch))
for batch_ids in valid_itr:
src_batch = make_batch(dev_src, batch_ids, eos_id)
trg_batch = make_batch(dev_trg, batch_ids, eos_id)
src_mask = padding_mask(src_batch, eos_id)
trg_mask = [
x | subsequent_mask(len(trg_batch) - 1)
for x in padding_mask(trg_batch[:-1], eos_id)
]
loss = model.loss(src_batch,
trg_batch,
src_mask,
trg_mask,
train=False)
valid_loss += loss.to_float() * len(batch_ids)
valid_itr.set_postfix(loss=loss.to_float())
print("\tvalid loss = %.4f" % (valid_loss / num_dev_sents))
if valid_loss < best_valid:
best_valid = valid_loss
print('\tsaving model/optimizer ... ', end="", flush=True)
prefix = config['model_prefix']
model.save(prefix + '.model')
optimizer.save(prefix + '.optimizer')
with Path(prefix).with_suffix('.valid').open('w') as f:
f.write(str(best_valid))
print('done.')
def main():
# Loads data
train_inputs = load_images("data/train-images-idx3-ubyte", NUM_TRAIN_SAMPLES)
train_labels = load_labels("data/train-labels-idx1-ubyte", NUM_TRAIN_SAMPLES)
test_inputs = load_images("data/t10k-images-idx3-ubyte", NUM_TEST_SAMPLES)
test_labels = load_labels("data/t10k-labels-idx1-ubyte", NUM_TEST_SAMPLES)
dev = D.Naive() # or D.CUDA(gpuid)
Device.set_default(dev)
pw1 = Parameter([NUM_HIDDEN_UNITS, NUM_INPUT_UNITS], I.XavierUniform())
pb1 = Parameter([NUM_HIDDEN_UNITS], I.Constant(0))
pw2 = Parameter([NUM_OUTPUT_UNITS, NUM_HIDDEN_UNITS], I.XavierUniform())
pb2 = Parameter([NUM_OUTPUT_UNITS], I.Constant(0))
optimizer = O.SGD(.5)
optimizer.add(pw1, pb1, pw2, pb2)
def make_graph(inputs, train):
x = F.input(inputs)
w1 = F.parameter(pw1)
b1 = F.parameter(pb1)
h = F.relu(w1 @ x + b1)
h = F.dropout(h, .5, train)
w2 = F.parameter(pw2)
b2 = F.parameter(pb2)
return w2 @ h + b2
ids = list(range(NUM_TRAIN_SAMPLES))
g = Graph()
Graph.set_default(g)
for epoch in range(MAX_EPOCH):
random.shuffle(ids)
# Training loop
for batch in range(NUM_TRAIN_BATCHES):
print("\rTraining... %d / %d" % (batch + 1, NUM_TRAIN_BATCHES), end="")
inputs = [train_inputs[ids[batch * BATCH_SIZE + i]] for i in range(BATCH_SIZE)]
labels = [train_labels[ids[batch * BATCH_SIZE + i]] for i in range(BATCH_SIZE)]
g.clear()
y = make_graph(inputs, True)
loss = F.softmax_cross_entropy(y, labels, 0)
avg_loss = F.batch.mean(loss)
optimizer.reset_gradients()
avg_loss.backward()
optimizer.update()
print()
match = 0
# Test loop
for batch in range(NUM_TEST_BATCHES):
print("\rTesting... %d / %d" % (batch + 1, NUM_TEST_BATCHES), end="")
inputs = [test_inputs[batch * BATCH_SIZE + i] for i in range(BATCH_SIZE)]
g.clear()
y = make_graph(inputs, False)
y_val = y.to_list()
for i in range(BATCH_SIZE):
maxval = -1e10
argmax = -1
for j in range(NUM_OUTPUT_UNITS):
v = y_val[j + i * NUM_OUTPUT_UNITS]
if (v > maxval):
maxval = v
argmax = j
if argmax == test_labels[i + batch * BATCH_SIZE]:
match += 1
accuracy = 100.0 * match / NUM_TEST_SAMPLES
print("\nepoch %d: accuracy: %.2f%%\n" % (epoch, accuracy))
def train(encdec, optimizer, prefix, best_valid_ppl):
# Registers all parameters to the optimizer.
optimizer.add_model(encdec)
# Loads vocab.
src_vocab = make_vocab(SRC_TRAIN_FILE, SRC_VOCAB_SIZE)
trg_vocab = make_vocab(TRG_TRAIN_FILE, TRG_VOCAB_SIZE)
inv_trg_vocab = make_inv_vocab(trg_vocab)
print("#src_vocab:", len(src_vocab))
print("#trg_vocab:", len(trg_vocab))
# Loads all corpus
train_src_corpus = load_corpus(SRC_TRAIN_FILE, src_vocab)
train_trg_corpus = load_corpus(TRG_TRAIN_FILE, trg_vocab)
valid_src_corpus = load_corpus(SRC_VALID_FILE, src_vocab)
valid_trg_corpus = load_corpus(TRG_VALID_FILE, trg_vocab)
test_src_corpus = load_corpus(SRC_TEST_FILE, src_vocab)
test_ref_corpus = load_corpus_ref(REF_TEST_FILE, trg_vocab)
num_train_sents = len(train_trg_corpus)
num_valid_sents = len(valid_trg_corpus)
num_test_sents = len(test_ref_corpus)
num_train_labels = count_labels(train_trg_corpus)
num_valid_labels = count_labels(valid_trg_corpus)
print("train:", num_train_sents, "sentences,", num_train_labels, "labels")
print("valid:", num_valid_sents, "sentences,", num_valid_labels, "labels")
# Sentence IDs
train_ids = list(range(num_train_sents))
valid_ids = list(range(num_valid_sents))
# Train/valid loop.
for epoch in range(MAX_EPOCH):
# Computation graph.
g = Graph()
Graph.set_default(g)
print("epoch %d/%d:" % (epoch + 1, MAX_EPOCH))
print(" learning rate scale = %.4e" %
optimizer.get_learning_rate_scaling())
# Shuffles train sentence IDs.
random.shuffle(train_ids)
# Training.
train_loss = 0.
for ofs in range(0, num_train_sents, BATCH_SIZE):
print("%d" % ofs, end="\r")
sys.stdout.flush()
batch_ids = train_ids[ofs:min(ofs + BATCH_SIZE, num_train_sents)]
src_batch = make_batch(train_src_corpus, batch_ids, src_vocab)
trg_batch = make_batch(train_trg_corpus, batch_ids, trg_vocab)
g.clear()
encdec.encode(src_batch, True)
loss = encdec.loss(trg_batch, True)
train_loss += loss.to_float() * len(batch_ids)
optimizer.reset_gradients()
loss.backward()
optimizer.update()
train_ppl = math.exp(train_loss / num_train_labels)
print(" train PPL = %.4f" % train_ppl)
# Validation.
valid_loss = 0.
for ofs in range(0, num_valid_sents, BATCH_SIZE):
print("%d" % ofs, end="\r")
sys.stdout.flush()
batch_ids = valid_ids[ofs:min(ofs + BATCH_SIZE, num_valid_sents)]
src_batch = make_batch(valid_src_corpus, batch_ids, src_vocab)
trg_batch = make_batch(valid_trg_corpus, batch_ids, trg_vocab)
g.clear()
encdec.encode(src_batch, False)
loss = encdec.loss(trg_batch, False)
valid_loss += loss.to_float() * len(batch_ids)
valid_ppl = math.exp(valid_loss / num_valid_labels)
print(" valid PPL = %.4f" % valid_ppl)
# Calculates test BLEU.
stats = defaultdict(int)
for ofs in range(0, num_test_sents, BATCH_SIZE):
print("%d" % ofs, end="\r")
sys.stdout.flush()
src_batch = test_src_corpus[ofs:min(ofs +
BATCH_SIZE, num_test_sents)]
ref_batch = test_ref_corpus[ofs:min(ofs +
BATCH_SIZE, num_test_sents)]
hyp_ids = test_batch(encdec, src_vocab, trg_vocab, src_batch)
for hyp_line, ref_line in zip(hyp_ids, ref_batch):
for k, v in get_bleu_stats(ref_line[1:-1], hyp_line).items():
stats[k] += v
bleu = calculate_bleu(stats)
print(" test BLEU = %.2f" % (100 * bleu))
# Saves best model/optimizer.
if valid_ppl < best_valid_ppl:
best_valid_ppl = valid_ppl
print(" saving model/optimizer ... ", end="")
sys.stdout.flush()
encdec.save(prefix + ".model")
optimizer.save(prefix + ".optimizer")
save_ppl(prefix + ".valid_ppl", best_valid_ppl)
print("done.")
else:
# Learning rate decay by 1/sqrt(2)
new_scale = .7071 * optimizer.get_learning_rate_scaling()
optimizer.set_learning_rate_scaling(new_scale)
def main():
# Loads data
train_inputs = load_images("data/train-images-idx3-ubyte",
NUM_TRAIN_SAMPLES)
train_labels = load_labels("data/train-labels-idx1-ubyte",
NUM_TRAIN_SAMPLES)
test_inputs = load_images("data/t10k-images-idx3-ubyte", NUM_TEST_SAMPLES)
test_labels = load_labels("data/t10k-labels-idx1-ubyte", NUM_TEST_SAMPLES)
dev = D.CUDA(0)
Device.set_default(dev)
g = Graph()
Graph.set_default(g)
# Parameters of CNNs
# Shape: {kernel_height, kernel_width, in_channels, out_channels}
pw_cnn1 = Parameter(Shape([KERNEL_SIZE1, KERNEL_SIZE1, 1, NUM_CHANNELS1]),
I.XavierUniformConv2D())
pw_cnn2 = Parameter(
Shape([KERNEL_SIZE2, KERNEL_SIZE2, NUM_CHANNELS1, NUM_CHANNELS2]),
I.XavierUniformConv2D())
# Parameters of FC layers
pw_fc1 = Parameter(Shape([NUM_HIDDEN_UNITS, NUM_INPUT_UNITS]),
I.XavierUniform())
pw_fc2 = Parameter(Shape([NUM_OUTPUT_UNITS, NUM_HIDDEN_UNITS]),
I.XavierUniform())
pb_fc1 = Parameter(Shape([NUM_HIDDEN_UNITS]), I.Constant(0))
pb_fc2 = Parameter(Shape([NUM_OUTPUT_UNITS]), I.Constant(0))
# Optimizer
optimizer = O.SGD(.1)
optimizer.add(pw_cnn1, pw_cnn2, pw_fc1, pw_fc2, pb_fc1, pb_fc2)
# Helper lambda to construct the predictor network.
def make_graph(inputs, train):
# Input and parameters.
#x = F.input(Shape([IMAGE_HEIGHT, IMAGE_WIDTH], BATCH_SIZE), inputs)
x = F.input(inputs)
w_cnn1 = F.parameter(pw_cnn1)
w_cnn2 = F.parameter(pw_cnn2)
w_fc1 = F.parameter(pw_fc1)
w_fc2 = F.parameter(pw_fc2)
b_fc1 = F.parameter(pb_fc1)
b_fc2 = F.parameter(pb_fc2)
# CNNs
h_cnn1 = F.relu(F.conv2d(x, w_cnn1, PADDING1, PADDING1, 1, 1, 1, 1))
h_pool1 = F.max_pool2d(h_cnn1, 2, 2, 0, 0, 2, 2)
h_cnn2 = F.relu(
F.conv2d(h_pool1, w_cnn2, PADDING2, PADDING2, 1, 1, 1, 1))
h_pool2 = F.max_pool2d(h_cnn2, 2, 2, 0, 0, 2, 2)
# FC layers
x_fc = F.dropout(F.flatten(h_pool2), .5, train)
h_fc = F.dropout(F.relu(F.matmul(w_fc1, x_fc) + b_fc1), .5, train)
return F.matmul(w_fc2, h_fc) + b_fc2
# Batch randomizer
ids = list(range(NUM_TRAIN_SAMPLES))
for epoch in range(MAX_EPOCH):
# Shuffles sample IDs.
random.shuffle(ids)
# Training loop
for batch in range(NUM_TRAIN_BATCHES):
print("\rTraining... %d / %d" % (batch + 1, NUM_TRAIN_BATCHES),
end="")
# Makes a minibatch for training.
inputs = [
train_inputs[ids[batch * BATCH_SIZE + i]]
for i in range(BATCH_SIZE)
]
labels = [
train_labels[ids[batch * BATCH_SIZE + i]]
for i in range(BATCH_SIZE)
]
# Constructs the graph.
g.clear()
y = make_graph(inputs, True)
loss = F.softmax_cross_entropy(y, labels, 0)
avg_loss = F.batch.mean(loss)
# Dump computation graph at the first time.
# if epoch == 0 and batch == 0:
# print(g.dump("dot"))
# Implicit forward, backward, and updates parameters.
optimizer.reset_gradients()
avg_loss.backward()
optimizer.update()
print()
match = 0
# Test loop
for batch in range(NUM_TEST_BATCHES):
print("\rTesting... %d / %d" % (batch + 1, NUM_TEST_BATCHES),
end="")
# Makes a test minibatch.
inputs = [
test_inputs[batch * BATCH_SIZE + i] for i in range(BATCH_SIZE)
]
# Constructs the graph.
g.clear()
y = make_graph(inputs, False)
# Gets outputs, argmax, and compares them with the label.
y_val = y.to_list()
for i in range(BATCH_SIZE):
maxval = -1e10
argmax = -1
for j in range(NUM_OUTPUT_UNITS):
v = y_val[j + i * NUM_OUTPUT_UNITS]
if v > maxval:
maxval = v
argmax = j
if argmax == test_labels[i + batch * BATCH_SIZE]:
match += 1
accuracy = 100.0 * match / NUM_TEST_SAMPLES
print("epoch %d: accuracy: %.2f%%" % (epoch, accuracy))
return 0
def setUp(self):
self.device = Naive()
Device.set_default(self.device)
self.graph = Graph()
Graph.set_default(self.graph)
def main():
# Loads vocab.
vocab = make_vocab("data/ptb.train.txt")
print("#vocab:", len(vocab)) # maybe 10000
eos_id = vocab["<s>"]
# Loads all corpus.
train_corpus = load_corpus("data/ptb.train.txt", vocab)
valid_corpus = load_corpus("data/ptb.valid.txt", vocab)
num_train_sents = len(train_corpus)
num_valid_sents = len(valid_corpus)
num_train_labels = count_labels(train_corpus)
num_valid_labels = count_labels(valid_corpus)
print("train:", num_train_sents, "sentences,", num_train_labels, "labels")
print("valid:", num_valid_sents, "sentences,", num_valid_labels, "labels")
dev = D.CUDA(0)
Device.set_default(dev)
# Trainer.
trainer = T.Adam()
trainer.set_weight_decay(1e-6)
trainer.set_gradient_clipping(5)
# Our LM.
lm = RNNLM(len(vocab), eos_id, trainer)
# Sentence IDs.
train_ids = list(range(num_train_sents))
valid_ids = list(range(num_valid_sents))
g = Graph()
Graph.set_default(g)
# Train/valid loop.
for epoch in range(MAX_EPOCH):
print("epoch", (epoch + 1), "/", MAX_EPOCH, ":")
# Shuffles train sentence IDs.
random.shuffle(train_ids)
# Training.
train_loss = 0
for ofs in range(0, num_train_sents, BATCH_SIZE):
batch_ids = train_ids[ofs:min(ofs + BATCH_SIZE, num_train_sents)]
batch = make_batch(train_corpus, batch_ids, eos_id)
g.clear()
outputs = lm.forward(batch)
loss = lm.forward_loss(outputs, batch)
train_loss += loss.to_float() * len(batch_ids)
trainer.reset_gradients()
loss.backward()
trainer.update()
print("\r%d" % ofs, end="")
sys.stdout.flush()
print()
train_ppl = math.exp(train_loss / num_train_labels)
print(" train ppl =", train_ppl)
# Validation.
valid_loss = 0
for ofs in range(0, num_valid_sents, BATCH_SIZE):
batch_ids = valid_ids[ofs:min(ofs + BATCH_SIZE, num_valid_sents)]
batch = make_batch(valid_corpus, batch_ids, eos_id)
g.clear()
outputs = lm.forward(batch)
loss = lm.forward_loss(outputs, batch)
valid_loss += loss.to_float() * len(batch_ids)
print("\r%d" % ofs, end="")
sys.stdout.flush()
print()
valid_ppl = math.exp(valid_loss / num_valid_labels)
print(" valid ppl =", valid_ppl)
