def make_model():
model = Sequential()
model.add(Input(shape=input_shape))
model.add(Conv2D(32, kernel_size=(3, 3), padding='same'))
model.add(ReLU())
model.add(Conv2D(32, kernel_size=(3, 3), padding='same'))
model.add(ReLU())
model.add(Conv2D(32, kernel_size=(3, 3), padding='same'))
model.add(ReLU())
model.add(Conv2D(32, kernel_size=(3, 3), padding='same'))
model.add(ReLU())
model.add(Conv2D(32, kernel_size=(3, 3), padding='same'))
model.add(ReLU())
model.add(Flatten())
model.add(Dense(2500, kernel_initializer='He'))
model.add(ReLU())
model.add(Dense(1500, kernel_initializer='He'))
model.add(ReLU())
model.add(Dense(10, kernel_initializer='He'))
model.add(Softmax())
model.summary()
model.compile(Adam(), 'categorical_crossentropy', 'accuracy')
return model
def make_model():
model = Sequential()
model.add(Input(shape=input_shape))
model.add(
Conv2D(16,
kernel_size=3,
strides=1,
padding='same',
kernel_regularizer=l2(1e-4)))
model.add(BatchNormalization_v2())
model.add(ReLU())
add_residual_block(model, num_filters=16)
add_residual_block(model, num_filters=16)
add_residual_block(model, num_filters=16)
add_residual_block(model, num_filters=32, strides=2, cnn_shortcut=True)
add_residual_block(model, num_filters=32)
add_residual_block(model, num_filters=32)
add_residual_block(model, num_filters=64, strides=2, cnn_shortcut=True)
add_residual_block(model, num_filters=64)
add_residual_block(model, num_filters=64)
model.add(AveragePooling2DAll())
model.add(Flatten())
model.add(Dense(10, kernel_initializer='He'))
model.add(Softmax())
model.summary()
model.compile(Adam(lr=0.001, decay=1e-4), 'categorical_crossentropy',
'accuracy')
return model
def make_model():
model = Sequential()
model.add(Input(shape=input_shape))
model.add(Conv2D(32, kernel_size=(3, 3), padding='same'))
model.add(BN_LAYER())
model.add(ReLU())
model.add(Conv2D(32, kernel_size=(3, 3), padding='same'))
model.add(BN_LAYER())
model.add(ReLU())
model.add(MaxPooling2D(2, 2, stride=2))
model.add(Conv2D(64, kernel_size=(3, 3), padding='same'))
model.add(BN_LAYER())
model.add(ReLU())
model.add(Conv2D(64, kernel_size=(3, 3), padding='same'))
model.add(BN_LAYER())
model.add(ReLU())
model.add(MaxPooling2D(2, 2, stride=2))
model.add(Flatten())
model.add(Dense(512, kernel_initializer='He'))
model.add(BN_LAYER())
model.add(ReLU())
model.add(Dense(10, kernel_initializer='He'))
model.add(Softmax())
model.summary()
model.compile(Adam(), loss='categorical_crossentropy', metric='accuracy')
return model
def main():
(X_train, _), (_, _), (X_test, _) = make_mnist_data(valid_ratio=0)
X_train = np.concatenate([X_train, X_test])
adam = Adam(lr=0.0002, beta_1=0.5)
generator, discriminator = make_generator(adam), make_discriminator(adam)
gan_model = make_gan_model(generator, discriminator, adam)
train(X_train, generator, discriminator,
gan_model, epochs=epochs, batchSize=batchSize, plot_freq=plot_freq)
def main():
transform_train = torchvision.transforms.Compose([
torchvision.transforms.RandomCrop(32, padding=4, padding_mode='edge'),
torchvision.transforms.RandomHorizontalFlip(),
# torchvision.transforms.RandomRotation(15),
torchvision.transforms.ToTensor(),
torchvision.transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
])
transform_test = torchvision.transforms.Compose([
torchvision.transforms.ToTensor(),
torchvision.transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
])
training_data = torchvision.datasets.CIFAR10('./data/cifar10', train=True, transform=transform_train, download=True)
testing_data = torchvision.datasets.CIFAR10('./data/cifar10', train=False, transform=transform_test, download=True)
model = resnet_v2(input_shape, depth, n_classes)
model.summary()
model.compile(Adam(lr=0.001), 'categorical_crossentropy', 'accuracy')
def lr_schduler(model):
lr = 1e-3
if model.n_epoch > 180:
lr *= 0.5e-3
elif model.n_epoch > 160:
lr *= 1e-3
elif model.n_epoch > 120:
lr *= 1e-2
elif model.n_epoch > 80:
lr *= 1e-1
model.optimizer.lr = lr
print('Learning rate: ', lr)
return lr
lr_reducer = ReduceLROnPlateau(factor=np.sqrt(0.1),
cooldown=0,
patience=6,
min_lr=0.5e-6)
history = model.fit_torchvision(training_data, batch_size=batch_size, epochs=epochs,
validation_data=testing_data, callbacks=[lr_schduler])
plot_history(history, 'cifar10_resnet56_v2.jpg')
model.save('resnet56_v2.h8')
def main():
(X_train,
_), (_, _), (X_test,
_) = make_cifar10_data(valid_ratio=0,
image_data_format='channels_last')
X_train = np.concatenate([X_train, X_test])
X_train = X_train.reshape(X_train.shape[0], -1)
adam = Adam(lr=0.0002, beta_1=0.5)
generator, discriminator = make_generator(adam), make_discriminator(adam)
gan_model = make_gan_model(generator, discriminator, adam)
train(X_train,
generator,
discriminator,
gan_model,
epochs=epochs,
batchSize=batchSize,
plot_freq=plot_freq)
def make_model():
model = Sequential()
model.add(Input(shape=input_shape))
model.add(Dense(4096))
model.add(ReLU())
model.add(Dense(4096))
model.add(ReLU())
model.add(Dense(4096))
model.add(ReLU())
model.add(Dense(4096))
model.add(ReLU())
model.add(Dense(4096))
model.add(ReLU())
model.add(Dense(10))
model.add(Softmax())
model.summary()
model.compile(Adam(), 'categorical_crossentropy', 'accuracy')
return model
def main():
class bbox_transform:
def __init__(self, img_size):
self.img_size = img_size
def __call__(self, x):
bbox_num = len(x)
target = np.zeros((bbox_num, 6))
for index, bbox in enumerate(x):
target[index, 1] = bbox['category_id']
target[index, 2:] = bbox['bbox']
target[:, 2:] /= self.img_size
return target
transform_train = torchvision.transforms.Compose([
torchvision.transforms.Resize((112, 112)),
torchvision.transforms.ToTensor(),
torchvision.transforms.Normalize((0.4914, 0.4822, 0.4465),
(0.2023, 0.1994, 0.2010)),
])
training_data = torchvision.datasets.CocoDetection(
'../datasets/shape/train',
'../datasets/shape/annotations/instances_train.json',
transform=transform_train,
target_transform=bbox_transform(512))
testing_data = torchvision.datasets.CocoDetection(
'../datasets/shape/val',
'../datasets/shape/annotations/instances_val.json',
transform=transform_train,
target_transform=bbox_transform(512))
def collate_fn(batch):
imgs, targets = list(zip(*batch))
# Remove empty boxes
targets = [boxes for boxes in targets if boxes is not None]
# set the sample index
for b_i, boxes in enumerate(targets):
boxes[:, 0] = b_i
targets = np.concatenate(targets, 0)
imgs = np.stack([img for img in imgs])
return imgs, targets
training_laoder = DataLoader(training_data,
batch_size=8,
collate_fn=collate_fn)
loss_params = {
"scaled_anchors": anchors,
"ignore_th": 0.5,
"obj_scale": 1,
"noobj_scale": 100,
}
model = make_my_yolo(input_shape, len(anchors), n_classes)
model.summary()
model.compile(Adam(lr=0.001), yolo_loss(**loss_params), None)
history = model.fit_dataloader(training_laoder,
batch_size=batch_size,
epochs=epochs)