def main():
args = get_arguments()
print("initializing device ... ", end="", file=sys.stderr, flush=True)
dev = D.Naive() if args.gpu < 0 else D.CUDA(args.gpu)
Device.set_default(dev)
print("done.", file=sys.stderr)
mode = args.mode
prefix = args.model
if mode == "train":
encdec = EncoderDecoder(args.dropout)
encdec.init(args.src_vocab, args.trg_vocab, args.embed, args.hidden)
optimizer = O.Adam()
optimizer.set_weight_decay(1e-6)
optimizer.set_gradient_clipping(5)
train(encdec, optimizer, args, 1e10)
elif mode == "resume":
print("loading model/optimizer ... ",
end="",
file=sys.stderr,
flush=True)
encdec = EncoderDecoder(args.dropout)
encdec.load(prefix + ".model")
optimizer = O.Adam()
optimizer.load(prefix + ".optimizer")
valid_ppl = load_ppl(prefix + ".valid_ppl")
print("done.", file=sys.stderr)
train(encdec, optimizer, args, valid_ppl)
else:
print("loading model ... ", end="", file=sys.stderr, flush=True)
encdec = EncoderDecoder(args.dropout)
encdec.load(prefix + ".model")
print("done.", file=sys.stderr)
test(encdec, args)
def main():
parser = ArgumentParser()
parser.add_argument("mode", help="(train|resume|test)")
parser.add_argument("model_prefix", help="prefix of the model files.")
args = parser.parse_args()
mode = args.mode
prefix = args.model_prefix
print("mode:", mode, file=sys.stderr)
print("prefix:", prefix, file=sys.stderr)
if mode not in ("train", "resume", "test"):
print("unknown mode:", mode, file=sys.stderr)
return
print("initializing device ... ", end="", file=sys.stderr)
sys.stderr.flush()
dev = D.CUDA(0)
Device.set_default(dev)
print("done.", file=sys.stderr)
if mode == "train":
encdec = AttentionalEncoderDecoder()
encdec.init(SRC_VOCAB_SIZE, TRG_VOCAB_SIZE, NUM_EMBED_UNITS,
NUM_HIDDEN_UNITS)
optimizer = O.Adam()
optimizer.set_weight_decay(1e-6)
optimizer.set_gradient_clipping(5)
train(encdec, optimizer, prefix, 1e10)
elif mode == "resume":
print("loading model/optimizer ... ", end="", file=sys.stderr)
sys.stderr.flush()
encdec = AttentionalEncoderDecoder()
encdec.load(prefix + ".model")
optimizer = O.Adam()
optimizer.load(prefix + ".optimizer")
valid_ppl = load_ppl(prefix + ".valid_ppl")
print("done.", file=sys.stderr)
train(encdec, optimizer, prefix, valid_ppl)
else:
print("loading model ... ", end="", file=sys.stderr)
sys.stderr.flush()
encdec = AttentionalEncoderDecoder()
encdec.load(prefix + ".model")
print("done.", file=sys.stderr)
test(encdec)
def main(config):
mode = config['mode']
if mode == 'preproc':
preproc(config)
return
print('initializing device ...', end='', file=sys.stderr, flush=True)
dev = D.Naive() if config['gpu'] < 0 else D.CUDA(config['gpu'])
Device.set_default(dev)
print("done.", file=sys.stderr, flush=True)
prefix = config['model_prefix']
if mode == 'train':
model = Transformer(config['n_heads'], config['n_stacks'],
config['dropout'], config['generation_limit'])
model.init(config['vocabulary_size'], config['d_model'],
config['d_ff'])
optimizer = O.Adam(alpha=1, beta2=0.98, eps=1e-9)
optimizer.set_gradient_clipping(5)
train(model, optimizer, config, 1e10)
elif mode == 'resume':
print('loading model/optimizer ... ',
end='',
file=sys.stderr,
flush=True)
model = Transformer(config['n_heads'], config['n_stacks'],
config['dropout'], config['generation_limit'])
model.load(prefix + '.model')
optimizer = O.Adam(alpha=1, beta2=0.98, eps=1e-9)
optimizer.set_gradient_clipping(5)
optimizer.load(prefix + '.optimizer')
with Path(prefix).with_suffix('.valid').open() as f:
valid_ppl = float(f.read().strip())
print('done.', file=sys.stderr, flush=True)
train(model, optimizer, config, valid_ppl)
elif mode == 'test':
model = Transformer(config['n_heads'], config['n_stacks'],
config['dropout'], config['generation_limit'])
model.load(prefix + '.model')
test(model, config)
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():
# 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 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)