def main(args):
train_set, valid_set, test_set = prepare_dataset(args.data_dir)
train_x, train_y = train_set
test_x, test_y = test_set
train_y = get_one_hot(train_y, 10)
if args.model_type == "cnn":
train_x = train_x.reshape((-1, 28, 28, 1))
test_x = test_x.reshape((-1, 28, 28, 1))
if args.model_type == "cnn":
net = Net([
Conv2D(kernel=[5, 5, 1, 8], stride=[2, 2], padding="SAME"),
ReLU(),
Conv2D(kernel=[5, 5, 8, 16], stride=[2, 2], padding="SAME"),
ReLU(),
Conv2D(kernel=[5, 5, 16, 32], stride=[2, 2], padding="SAME"),
ReLU(),
Flatten(),
Dense(10)
])
elif args.model_type == "dense":
net = Net([
Dense(200),
ReLU(),
Dense(100),
ReLU(),
Dense(70),
ReLU(),
Dense(30),
ReLU(),
Dense(10)
])
else:
raise ValueError(
"Invalid argument model_type! Must be 'cnn' or 'dense'")
model = Model(net=net,
loss=SoftmaxCrossEntropyLoss(),
optimizer=Adam(lr=args.lr))
iterator = BatchIterator(batch_size=args.batch_size)
evaluator = AccEvaluator()
loss_list = list()
for epoch in range(args.num_ep):
t_start = time.time()
for batch in iterator(train_x, train_y):
pred = model.forward(batch.inputs)
loss, grads = model.backward(pred, batch.targets)
model.apply_grad(grads)
loss_list.append(loss)
print("Epoch %d time cost: %.4f" % (epoch, time.time() - t_start))
# evaluate
model.set_phase("TEST")
test_pred = model.forward(test_x)
test_pred_idx = np.argmax(test_pred, axis=1)
test_y_idx = np.asarray(test_y)
res = evaluator.evaluate(test_pred_idx, test_y_idx)
print(res)
model.set_phase("TRAIN")
def main(args):
train_set, valid_set, test_set = prepare_dataset(args.data_dir)
train_x, train_y = train_set
test_x, test_y = test_set
# train_y = get_one_hot(train_y, 2)
net = Net([Dense(100), ReLU(), Dense(30), ReLU(), Dense(1)])
model = Model(net=net,
loss=SigmoidCrossEntropyLoss(),
optimizer=Adam(lr=args.lr))
iterator = BatchIterator(batch_size=args.batch_size)
evaluator = AccEvaluator()
loss_list = list()
for epoch in range(args.num_ep):
t_start = time.time()
for batch in iterator(train_x, train_y):
pred = model.forward(batch.inputs)
loss, grads = model.backward(pred, batch.targets)
model.apply_grad(grads)
loss_list.append(loss)
print("Epoch %d time cost: %.4f" % (epoch, time.time() - t_start))
for timer in model.timers.values():
timer.report()
# evaluate
model.set_phase("TEST")
test_y_idx = np.asarray(test_y).reshape(-1)
test_pred = model.forward(test_x)
test_pred[test_pred > 0] = 1
test_pred[test_pred <= 0] = 0
test_pred_idx = test_pred.reshape(-1)
res = evaluator.evaluate(test_pred_idx, test_y_idx)
print(res)
model.set_phase("TRAIN")
def main(args):
if args.seed >= 0:
random_seed(args.seed)
train_set, valid_set, test_set = prepare_dataset(args.data_dir)
train_x, train_y = train_set
test_x, test_y = test_set
train_y = get_one_hot(train_y, 10)
train_x = Tensor(train_x)
train_y = Tensor(train_y)
test_x = Tensor(test_x)
test_y = Tensor(test_y)
net = Net([
Dense(200),
ReLU(),
Dense(100),
ReLU(),
Dense(70),
ReLU(),
Dense(30),
ReLU(),
Dense(10)
])
model = Model(net=net,
loss=SoftmaxCrossEntropyLoss(),
optimizer=Adam(lr=args.lr))
loss_layer = SoftmaxCrossEntropyLoss()
iterator = BatchIterator(batch_size=args.batch_size)
evaluator = AccEvaluator()
loss_list = list()
for epoch in range(args.num_ep):
t_start = time.time()
for batch in iterator(train_x, train_y):
model.zero_grad()
pred = model.forward(batch.inputs)
loss = loss_layer.loss(pred, batch.targets)
loss.backward()
model.step()
loss_list.append(loss.values)
print("Epoch %d tim cost: %.4f" % (epoch, time.time() - t_start))
# evaluate
model.set_phase("TEST")
test_pred = model.forward(test_x)
test_pred_idx = np.argmax(test_pred, axis=1)
test_y_idx = test_y.values
res = evaluator.evaluate(test_pred_idx, test_y_idx)
print(res)
model.set_phase("TRAIN")
def test_batch_iterator():
batch_size = 10
n_data = 10 * batch_size # 10 batches
iterator = BatchIterator(batch_size=batch_size)
x_dim, y_dim = 10, 5
fake_x = np.random.randint(0, 100, size=(n_data, x_dim))
fake_y = np.random.randint(0, 100, size=(n_data, y_dim))
n_batches = 0
for batch_x, batch_y in iterator(fake_x, fake_y):
assert batch_x.shape == (batch_size, x_dim)
assert batch_y.shape == (batch_size, y_dim)
n_batches += 1
assert n_batches == 10
def main(args):
if args.seed >= 0:
random_seed(args.seed)
train_set, valid_set, test_set = prepare_dataset(args.data_dir)
train_x, train_y = train_set
rbm = RBM(500, k=args.k)
iterator = BatchIterator(batch_size=args.batch_size)
for epoch in range(args.num_ep):
print('epoch', epoch)
for batch in iterator(train_x, train_y):
visible = rbm.gibs_sampling(batch.inputs)
rbm.step(args.lr)
filename = './out/epoch-%d.png' % epoch
save_batch_as_images(filename, visible, title='epoch %d' % epoch)
def main(args):
if args.seed >= 0:
random_seed(args.seed)
# data preparing
data_path = os.path.join(args.data_dir, args.file_name)
train_x, train_y, img_shape = prepare_dataset(data_path)
net = Net([
Dense(30),
ReLU(),
Dense(60),
ReLU(),
Dense(60),
ReLU(),
Dense(30),
ReLU(),
Dense(3),
Sigmoid()
])
model = Model(net=net, loss=MSELoss(), optimizer=Adam())
mse_evaluator = MSEEvaluator()
iterator = BatchIterator(batch_size=args.batch_size)
for epoch in range(args.num_ep):
t_start = time.time()
for batch in iterator(train_x, train_y):
preds = model.forward(batch.inputs)
loss, grads = model.backward(preds, batch.targets)
model.apply_grad(grads)
# evaluate
preds = net.forward(train_x)
mse = mse_evaluator.evaluate(preds, train_y)
print(mse)
if args.paint:
# generate painting
preds = preds.reshape(img_shape[0], img_shape[1], -1)
preds = (preds * 255.0).astype("uint8")
filename, ext = os.path.splitext(args.file_name)
output_filename = "output" + ext
output_path = os.path.join(args.data_dir, output_filename)
Image.fromarray(preds).save(output_path)
print("Epoch %d time cost: %.2f" % (epoch, time.time() - t_start))
def main(args):
if args.seed >= 0:
random_seed(args.seed)
# data preparing
train_x, train_y, img_shape = prepare_dataset(args.img)
net = Net([
Dense(30),
ReLU(),
Dense(100),
ReLU(),
Dense(100),
ReLU(),
Dense(30),
ReLU(),
Dense(3),
Sigmoid()
])
model = Model(net=net, loss=MSE(), optimizer=Adam())
iterator = BatchIterator(batch_size=args.batch_size)
for epoch in range(args.num_ep):
for batch in iterator(train_x, train_y):
preds = model.forward(batch.inputs)
loss, grads = model.backward(preds, batch.targets)
model.apply_grad(grads)
# evaluate
preds = net.forward(train_x)
mse = mean_square_error(preds, train_y)
print("Epoch %d %s" % (epoch, mse))
# generate painting
if epoch % 5 == 0:
preds = preds.reshape(img_shape[0], img_shape[1], -1)
preds = (preds * 255.0).astype("uint8")
name, ext = os.path.splitext(args.img)
filename = os.path.basename(name)
out_filename = filename + "-paint-epoch" + str(epoch) + ext
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
out_path = os.path.join(args.output_dir, out_filename)
Image.fromarray(preds).save(out_path)
print("save painting to %s" % out_path)
def main(args):
train_set, valid_set, test_set = prepare_dataset(args.data_dir)
train_x, train_y = train_set
test_x, test_y = test_set
train_y = get_one_hot(train_y, 10)
net = Net([
Dense(784, 200),
ReLU(),
Dense(200, 100),
ReLU(),
Dense(100, 70),
ReLU(),
Dense(70, 30),
ReLU(),
Dense(30, 10)
])
model = Model(net=net,
loss=SoftmaxCrossEntropyLoss(),
optimizer=Adam(lr=args.lr))
iterator = BatchIterator(batch_size=args.batch_size)
evaluator = AccEvaluator()
loss_list = list()
for epoch in range(args.num_ep):
t_start = time.time()
for batch in iterator(train_x, train_y):
pred = model.forward(batch.inputs)
loss, grads = model.backward(pred, batch.targets)
model.apply_grad(grads)
loss_list.append(loss)
t_end = time.time()
# evaluate
test_pred = model.forward(test_x)
test_pred_idx = np.argmax(test_pred, axis=1)
test_y_idx = np.asarray(test_y)
res = evaluator.evaluate(test_pred_idx, test_y_idx)
print("Epoch %d time cost: %.4f\t %s" % (epoch, t_end - t_start, res))
def main(args):
if args.seed >= 0:
random_seed(args.seed)
train_set, valid_set, test_set = mnist(args.data_dir)
train_x, train_y = train_set
test_x, test_y = test_set
train_y = get_one_hot(train_y, 10)
if args.model_type == "cnn":
train_x = train_x.reshape((-1, 28, 28, 1))
test_x = test_x.reshape((-1, 28, 28, 1))
if args.model_type == "cnn":
# a LeNet-5 model with activation function changed to ReLU
net = Net([
Conv2D(kernel=[5, 5, 1, 6], stride=[1, 1], padding="SAME"),
ReLU(),
MaxPool2D(pool_size=[2, 2], stride=[2, 2]),
Conv2D(kernel=[5, 5, 6, 16], stride=[1, 1], padding="SAME"),
ReLU(),
MaxPool2D(pool_size=[2, 2], stride=[2, 2]),
Flatten(),
Dense(120),
ReLU(),
Dense(84),
ReLU(),
Dense(10)
])
elif args.model_type == "dense":
net = Net([
Dense(200),
ReLU(),
Dense(100),
ReLU(),
Dense(70),
ReLU(),
Dense(30),
ReLU(),
Dense(10)
])
else:
raise ValueError("Invalid argument: model_type")
model = Model(net=net,
loss=SoftmaxCrossEntropy(),
optimizer=Adam(lr=args.lr))
iterator = BatchIterator(batch_size=args.batch_size)
loss_list = list()
for epoch in range(args.num_ep):
t_start = time.time()
for batch in iterator(train_x, train_y):
pred = model.forward(batch.inputs)
loss, grads = model.backward(pred, batch.targets)
model.apply_grad(grads)
loss_list.append(loss)
print("Epoch %d time cost: %.4f" % (epoch, time.time() - t_start))
# evaluate
model.set_phase("TEST")
test_pred = model.forward(test_x)
test_pred_idx = np.argmax(test_pred, axis=1)
test_y_idx = np.asarray(test_y)
res = accuracy(test_pred_idx, test_y_idx)
print(res)
model.set_phase("TRAIN")
def train(args):
# prepare dataset
train_, valid, test = mnist(args.data_dir)
X = np.concatenate([train_[0], valid[0], test[0]])
y = np.concatenate([train_[1], valid[1], test[1]])
fix_noise = get_noise(size=(args.batch_size, args.nz))
loss = SigmoidCrossEntropy()
# TODO: replace mlp with cnn
G = Model(net=mlp_G(),
loss=loss,
optimizer=Adam(args.lr_g, beta1=args.beta1))
D = Model(net=mlp_D(),
loss=loss,
optimizer=Adam(args.lr_d, beta1=args.beta1))
running_g_err, running_d_err = 0, 0
iterator = BatchIterator(batch_size=args.batch_size)
for epoch in range(args.num_ep):
for i, batch in enumerate(iterator(X, y)):
# --- Train Discriminator ---
# feed with real data (maximize log(D(x)))
d_pred_real = D.forward(batch.inputs)
label_real = np.ones_like(d_pred_real)
d_real_err, d_real_grad = D.backward(d_pred_real, label_real)
# feed with fake data (maximize log(1 - D(G(z))))
noise = get_noise(size=(len(batch.inputs), args.nz))
g_out = G.forward(noise)
d_pred_fake = D.forward(g_out)
label_fake = np.zeros_like(d_pred_fake)
d_fake_err, d_fake_grad = D.backward(d_pred_fake, label_fake)
# train D
d_err = d_real_err + d_fake_err
d_grads = d_real_grad + d_fake_grad
D.apply_grad(d_grads)
# ---- Train Generator ---
# maximize log(D(G(z)))
d_pred_fake = D.forward(g_out)
g_err, d_grad = D.backward(d_pred_fake, label_real)
g_grads = G.net.backward(d_grad.wrt_input)
G.apply_grad(g_grads)
running_d_err = 0.9 * running_d_err + 0.1 * d_err
running_g_err = 0.9 * running_g_err + 0.1 * g_err
if i % 100 == 0:
print("epoch-%d iter-%d d_err: %.4f g_err: %.4f" %
(epoch + 1, i + 1, running_d_err, running_g_err))
# sampling
print("epoch: %d/%d d_err: %.4f g_err: %.4f" %
(epoch + 1, args.num_ep, running_d_err, running_g_err))
samples = G.forward(fix_noise)
img_name = "ep%d.png" % (epoch + 1)
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
save_path = os.path.join(args.output_dir, img_name)
save_batch_as_images(save_path, samples)
# save generator
model_path = os.path.join(args.output_dir, args.model_name)
G.save(model_path)
print("Saving generator ", model_path)
def main(args):
if args.seed >= 0:
random_seed(args.seed)
# create output directory for saving result images
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
# prepare and read dataset
train_set, _, test_set = mnist(args.data_dir)
train_x, train_y = train_set
test_x, test_y = test_set
# specify the encoder and decoder net structure
encoder_net = Net([Dense(256), ReLU(), Dense(64)])
decoder_net = Net([ReLU(), Dense(256), Tanh(), Dense(784), Tanh()])
nets = (encoder_net, decoder_net)
optimizers = (Adam(args.lr), Adam(args.lr))
model = AutoEncoder(nets, loss=MSE(), optimizer=optimizers)
# for pre-trained model, test generated images from latent space
if args.load_model is not None:
# load pre-trained model
model.load(os.path.join(args.output_dir, args.load_model))
print("Loaded model fom %s" % args.load_model)
# transition from test[from_idx] to test[to_idx] in n steps
idx_arr, n = [2, 4, 32, 12, 82], 160
print("Transition in numbers", [test_y[i] for i in idx_arr],
"in %d steps ..." % n)
stops = [model.en_net.forward(test_x[i]) for i in idx_arr]
k = int(n / (len(idx_arr) - 1)) # number of code per transition
# generate all transition codes
code_arr = []
for i in range(len(stops) - 1):
t = [c.copy() for c in transition(stops[i], stops[i + 1], k)]
code_arr += t
# apply decoding all n "code" from latent space...
batch = None
for code in code_arr:
# translate latent space to image
genn = model.de_net.forward(code)
# save decoded results in a batch
if batch is None:
batch = np.array(genn)
else:
batch = np.concatenate((batch, genn))
output_path = os.path.join(args.output_dir, "genn-latent.png")
save_batch_as_images(output_path, batch)
quit()
# train the auto-encoder
iterator = BatchIterator(batch_size=args.batch_size)
for epoch in range(args.num_ep):
for batch in iterator(train_x, train_y):
origin_in = batch.inputs
# make noisy inputs
m = origin_in.shape[0] # batch size
mu = args.gaussian_mean # mean
sigma = args.gaussian_std # standard deviation
noises = np.random.normal(mu, sigma, (m, 784))
noises_in = origin_in + noises # noisy inputs
# forward
genn = model.forward(noises_in)
# back-propagate
loss, grads = model.backward(genn, origin_in)
# apply gradients
model.apply_grad(grads)
print("Epoch: %d Loss: %.3f" % (epoch, loss))
# save all the generated images and original inputs for this batch
noises_in_path = os.path.join(args.output_dir,
"ep%d-input.png" % epoch)
genn_path = os.path.join(args.output_dir, "ep%d-genn.png" % epoch)
save_batch_as_images(noises_in_path, noises_in, titles=batch.targets)
save_batch_as_images(genn_path, genn, titles=batch.targets)
# save the model after training
model.save(os.path.join(args.output_dir, args.save_model))
def main(args):
if args.seed >= 0:
random_seed(args.seed)
# create output directory for saving result images
if not os.path.exists('./output'):
os.mkdir('./output')
# prepare and read dataset
train_set, valid_set, test_set = prepare_dataset(args.data_dir)
train_x, train_y = train_set
test_x, test_y = test_set
# batch iterator
iterator = BatchIterator(batch_size=args.batch_size)
# specify the encoder and decoder net structure
encoder = Net([Dense(256), ReLU(), Dense(64)])
decoder = Net([ReLU(), Dense(256), Tanh(), Dense(784), Tanh()])
# create AutoEncoder model
model = AutoEncoder(encoder=encoder,
decoder=decoder,
loss=MSELoss(),
optimizer=Adam(args.lr))
# for pretrained model, test generated images from latent space
if args.load_model is not None:
# load pretrained model
model.load(args.load_model)
print('Loaded model from %s' % args.load_model)
# transition from test[from_idx] to test[to_idx] in n steps
idx_arr, n = [2, 4, 32, 12, 82], 160
print("Transition in numbers", [test_y[i] for i in idx_arr],
"in %d steps ..." % n)
stops = [model.encoder.forward(test_x[i]) for i in idx_arr]
k = int(n / (len(idx_arr) - 1)) # number of code per transition
# generate all transition codes
code_arr = []
for i in range(len(stops) - 1):
t = [c.copy() for c in transition(stops[i], stops[i + 1], k)]
code_arr += t
# apply decoding all n "code" from latent space...
batch = None
for code in code_arr:
# translate latent space to image
genn = model.decoder.forward(code)
# save decoded results in a batch
if batch is None:
batch = np.array(genn)
else:
batch = np.concatenate((batch, genn))
save_batch_as_images('output/genn-latent.png', batch)
quit()
# train the autoencoder
for epoch in range(args.num_ep):
print('epoch %d ...' % epoch)
for batch in iterator(train_x, train_y):
origin_in = batch.inputs
# make noisy inputs
m = origin_in.shape[0] # batch size
mu = args.guassian_mean # mean
sigma = args.guassian_std # standard deviation
noises = np.random.normal(mu, sigma, (m, 784))
noises_in = origin_in + noises # noisy inputs
# train the representation
genn = model.forward(noises_in)
loss, grads = model.backward(genn, origin_in)
model.apply_grad(grads)
print('Loss: %.3f' % loss)
# save all the generated images and original inputs for this batch
save_batch_as_images('output/ep%d-input.png' % epoch,
noises_in,
titles=batch.targets)
save_batch_as_images('output/ep%d-genn.png' % epoch,
genn,
titles=batch.targets)
# save the model after training
model.save('output/model.pkl')
