def train_distill_model(dataset, args):
# load dataset
train_x, train_y, test_x, test_y = dataset
# load or train a teacher model
teacher = Model(net=teacher_net,
loss=SoftmaxCrossEntropy(),
optimizer=Adam(lr=args.lr))
teacher_model_path = os.path.join(args.model_dir, "teacher.model")
if not os.path.isfile(teacher_model_path):
print("No teacher model founded. Training a new one...")
train_single_model(teacher, dataset, args, name="teacher")
teacher.load(teacher_model_path)
teacher.set_phase("TEST")
print("training distill model")
# define a student model
student = Model(net=student_net,
loss=DistillationLoss(alpha=args.alpha, T=args.T),
optimizer=Adam(lr=args.lr))
# run training
iterator = BatchIterator(batch_size=args.batch_size)
for epoch in range(args.num_ep):
t_start = time.time()
for i, batch in enumerate(iterator(train_x, train_y)):
pred = student.forward(batch.inputs)
teacher_out = teacher.forward(batch.inputs)
teacher_out_prob = softmax(teacher_out, t=args.T)
loss = student.loss.loss(pred, batch.targets, teacher_out_prob)
grad_from_loss = student.loss.grad(pred, batch.targets,
teacher_out_prob)
grads = student.net.backward(grad_from_loss)
student.apply_grads(grads)
print("Epoch %d time cost: %.4f" % (epoch, time.time() - t_start))
# evaluate
student.set_phase("TEST")
hit, total = 0, 0
for i, batch in enumerate(iterator(test_x, test_y)):
pred = student.forward(batch.inputs)
res = accuracy(np.argmax(pred, 1), np.argmax(batch.targets, 1))
hit += res["hit_num"]
total += res["total_num"]
print("accuracy: %.4f" % (1.0 * hit / total))
student.set_phase("TRAIN")
# save the distilled model
if not os.path.isdir(args.model_dir):
os.makedirs(args.model_dir)
model_path = os.path.join(args.model_dir, "distill-%d.model" % args.T)
student.save(model_path)
print("model saved in %s" % model_path)
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_grads(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 test_parameters_change(fake_dataset):
# make sure the parameters does change after apply gradients
# fake dataset
X, y = fake_dataset
# simple model
net = Net([Dense(10), Dense(1)])
loss = MSE()
opt = SGD(lr=1.0)
model = Model(net, loss, opt)
# forward and backward
pred = model.forward(X)
loss, grads = model.backward(pred, y)
# parameters change test
params_before = model.net.params.values
model.apply_grads(grads)
params_after = model.net.params.values
for p1, p2 in zip(params_before, params_after):
assert np.all(p1 != p2)
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]])
if args.model_type == "cnn":
X = X.reshape((-1, 28, 28, 1))
G_net, D_net = G_cnn(), D_cnn()
elif args.model_type == "mlp":
G_net, D_net = G_mlp(), D_mlp()
else:
raise ValueError("Invalid argument: model_type")
fix_noise = get_noise(size=(args.batch_size, args.nz))
loss = SigmoidCrossEntropy()
G = Model(net=G_net,
loss=loss,
optimizer=Adam(args.lr_g, beta1=args.beta1))
D = Model(net=D_net,
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_grads(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_grads(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)
train_set, _, test_set = mnist(args.data_dir, one_hot=True)
train_x, train_y = train_set
test_x, test_y = test_set
if args.model_type == "mlp":
# A multilayer perceptron model
net = Net([
Dense(200),
ReLU(),
Dense(100),
ReLU(),
Dense(70),
ReLU(),
Dense(30),
ReLU(),
Dense(10)
])
elif args.model_type == "cnn":
# A LeNet-5 model with activation function changed to ReLU
train_x = train_x.reshape((-1, 28, 28, 1))
test_x = test_x.reshape((-1, 28, 28, 1))
net = Net([
Conv2D(kernel=[5, 5, 1, 6], stride=[1, 1]),
ReLU(),
MaxPool2D(pool_size=[2, 2], stride=[2, 2]),
Conv2D(kernel=[5, 5, 6, 16], stride=[1, 1]),
ReLU(),
MaxPool2D(pool_size=[2, 2], stride=[2, 2]),
Flatten(),
Dense(120),
ReLU(),
Dense(84),
ReLU(),
Dense(10)
])
elif args.model_type == "rnn":
# A simple recurrent neural net to classify images.
train_x = train_x.reshape((-1, 28, 28))
test_x = test_x.reshape((-1, 28, 28))
net = Net([RNN(num_hidden=50, activation=Tanh()), 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_grads(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.argmax(test_y, axis=1)
res = accuracy(test_pred_idx, test_y_idx)
print(res)
model.set_phase("TRAIN")
# save model
if not os.path.isdir(args.model_dir):
os.makedirs(args.model_dir)
model_name = "mnist-%s-epoch%d.pkl" % (args.model_type, args.num_ep)
model_path = os.path.join(args.model_dir, model_name)
model.save(model_path)
print("model saved in %s" % model_path)