def March(self, objects, Collisions):
counter = 0
current_position = self.position
pygame.draw.circle(self.screen, (255, 41, 120), self.position, 5)
while counter < 100:
record = 2000
closest = None
for object in objects:
distance = SignedDistance(current_position, object,
object.radius)
if distance < record:
record = distance
closest = object
if record < 1:
Collisions.insert(
0, (int(current_position[0]), int(current_position[1])))
break
self.x_position = current_position[0] + cos(self.angle) * record
self.y_position = current_position[1] + sin(self.angle) * record
a_X = current_position[0] + cos(self.angle) * record
a_Y = current_position[1] + sin(self.angle) * record
pygame.draw.circle(
self.screen, (180, 244, 187),
(int(current_position[0]), int(current_position[1])),
abs(int(record)), 1)
pygame.draw.line(self.screen, self.color,
(self.position[0], self.position[1]),
(normalize(self.x_position) * a_X,
normalize(self.y_position) * a_Y), 4)
current_position[0] = normalize(self.x_position) * a_X
current_position[1] = normalize(self.x_position) * a_Y
pygame.draw.circle(self.screen, self.color,
(int(self.x_position), int(self.y_position)), 4)
closest.display(self.screen, (249, 129, 58))
if offScreen([self.x_position, self.y_position], 1920, 1080):
break
if offScreen(current_position, 1920, 1080):
break
counter += 1
pygame.display.update()
def March(self, objects):
counter = 0
current_position = self.position
pygame.draw.circle(self.screen, (255, 41, 120), self.position, 5)
while counter < 100:
record = 2000
closest = None
for obj in objects:
distance = SignedDistance(current_position, obj, obj.radius)
if distance < record:
record = distance
closest = obj
if record < 1:
break
self.x_position = current_position[0] + cos(self.angle) * record
self.y_position = current_position[1] + sin(self.angle) * record
a_X = current_position[0] + cos(self.angle) * record
a_Y = current_position[1] + sin(self.angle) * record
pygame.draw.circle(self.screen, (0, 204, 37), (int(current_position[0]), int(current_position[1])), abs(int(record)), 1)
pygame.draw.line(self.screen, self.color, (self.position[0], self.position[1]), (normalize(self.x_position) * a_X, normalize(self.y_position) * a_Y))
current_position[0] = normalize(self.x_position) * a_X
current_position[1] = normalize(self.x_position) * a_Y
pygame.draw.circle(self.screen, self.color, (int(self.x_position), int(self.y_position)), 5)
closest.display(self.screen, (130, 190, 41))
if offScreen([self.x_position, self.y_position], 1280, 720):
break
if offScreen(current_position, 1280, 720):
break
counter += 1
pygame.display.update()
def main():
@tf.function
def train_cutmix_image(image_cutmix, target_a, target_b, lam):
with tf.GradientTape() as tape:
output = model(image_cutmix, training=True)
loss = criterion(target_a, output) * lam + criterion(
target_b, output) * (1. - lam)
gradients = tape.gradient(loss, model.trainable_variables)
optimizer.apply_gradients(zip(gradients, model.trainable_variables))
return loss, output
@tf.function
def train_original_image(image, target):
with tf.GradientTape() as tape:
output = model(image, training=True)
loss = criterion(target, output)
gradients = tape.gradient(loss, model.trainable_variables)
optimizer.apply_gradients(zip(gradients, model.trainable_variables))
return loss, output
args = parser.parse_args()
assert len(args.lr_boundaries) + 1 == len(
args.lr_values), "Epochs and Learning rates are not matched"
if args.dataset == 'cifar10':
(train_images,
train_labels), (test_images,
test_labels) = datasets.cifar10.load_data()
num_classes = 10
elif args.dataset == 'cifar100':
(train_images,
train_labels), (test_images,
test_labels) = datasets.cifar100.load_data()
num_classes = 100
# dataset
train_dataset = tf.data.Dataset.from_tensor_slices(
(train_images, train_labels))
train_dataset = train_dataset.shuffle(int(len(train_labels) / 4))
train_dataset = train_dataset.map(train_augment, num_parallel_calls=AUTO)
train_dataset = train_dataset.batch(args.batch_size)
train_dataset = train_dataset.prefetch(AUTO)
test_dataset = tf.data.Dataset.from_tensor_slices(
(test_images, test_labels))
test_dataset = test_dataset.map(test_augment, num_parallel_calls=AUTO)
test_dataset = test_dataset.batch(args.batch_size)
# model building (ResNet50)
model = ResNet(num_blocks=args.layers,
num_classes=num_classes,
bottleneck=args.bottleneck)
# loss & optimizer
criterion = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
lr_schedule = learning_rate_schedule(args.lr_boundaries, args.lr_values,
train_labels, args.batch_size,
args.epochs)
optimizer = tf.keras.optimizers.SGD(lr_schedule,
momentum=args.momentum,
nesterov=True)
# ckpt manager
current_time = datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
ckpt_dir = os.path.join(args.ckpt_dir, current_time)
ckpt = tf.train.Checkpoint(model=model)
ckpt_manager = tf.train.CheckpointManager(ckpt, ckpt_dir, max_to_keep=1)
# metrics
train_loss = tf.keras.metrics.Mean(name='train_loss')
train_accuracy_top1 = tf.keras.metrics.SparseTopKCategoricalAccuracy(
k=1, name='train_accuracy_top1')
train_accuracy_top5 = tf.keras.metrics.SparseTopKCategoricalAccuracy(
k=5, name='train_accuracy_top5')
test_loss = tf.keras.metrics.Mean(name='test_loss')
test_accuracy_top1 = tf.keras.metrics.SparseTopKCategoricalAccuracy(
k=1, name='test_accuracy_top1')
test_accuracy_top5 = tf.keras.metrics.SparseTopKCategoricalAccuracy(
k=5, name='test_accuracy_top5')
# tensorboard writer
train_log_dir = os.path.join(args.tensorboard_dir, current_time, 'train')
train_summary_writer = tf.summary.create_file_writer(train_log_dir)
print(f'******* Train result log written to {train_log_dir} ******')
test_log_dir = os.path.join(args.tensorboard_dir, current_time, 'test')
test_summary_writer = tf.summary.create_file_writer(test_log_dir)
print(f'******* Test result log written to {test_log_dir} ******')
template = '******* Epoch: {:03d} / {:03d}:, LR: {:.3f}, Train loss: {:.3f}, Train top1 err: {:.3f}, Train top5 err: {:.3f}, Test loss: {:.3f}, Test top1 err: {:.3f}, Test top5 err: {:.3f}, Time: {:.3f}'
# custom training
for epoch in range(args.epochs):
s_time = time.time()
train_loss.reset_states()
train_accuracy_top1.reset_states()
train_accuracy_top5.reset_states()
test_loss.reset_states()
test_accuracy_top1.reset_states()
test_accuracy_top5.reset_states()
for image, target in train_dataset:
r = np.random.rand(1)
# training on cutmix image
if args.beta > 0 and r < args.cutmix_prob:
lam = np.random.beta(args.beta, args.beta)
rand_index = tf.random.shuffle(tf.range(len(target)))
target_a = target
target_b = tf.gather(target, rand_index)
bbx1, bby1, bbx2, bby2 = rand_bbox(image.shape, lam)
image_a = image
image_b = tf.gather(image, rand_index)
mask = np.ones_like(image)
mask[:, bbx1:bbx2, bby1:bby2, :] = 0
image_cutmix = normalize(
tf.math.multiply(image_a, mask) +
tf.math.multiply(image_b, (abs(1. - mask))))
image = normalize(image)
lam = tf.convert_to_tensor(
1 - ((bbx2 - bbx1) * (bby2 - bby1) /
(image_cutmix.shape[1] * image_cutmix.shape[2])),
dtype=tf.float32)
loss, output = train_cutmix_image(image_cutmix, target_a,
target_b, lam)
# training on original image
else:
image = normalize(image)
loss, output = train_original_image(image, target)
train_loss(loss)
train_accuracy_top1(target, output)
train_accuracy_top5(target, output)
for image, target in test_dataset:
image = normalize(image)
output = model(image, training=False)
loss = criterion(target, output)
test_loss(loss)
test_accuracy_top1(target, output)
test_accuracy_top5(target, output)
train_err1 = (1 - train_accuracy_top1.result()) * 100
train_err5 = (1 - train_accuracy_top5.result()) * 100
test_err1 = (1 - test_accuracy_top1.result()) * 100
test_err5 = (1 - test_accuracy_top5.result()) * 100
# writing to tensorboard
with train_summary_writer.as_default():
tf.summary.scalar('train_loss', train_loss.result(), step=epoch)
tf.summary.scalar('train_err_top1', train_err1, step=epoch)
tf.summary.scalar('train_err_top5', train_err5, step=epoch)
with test_summary_writer.as_default():
tf.summary.scalar('test_loss', test_loss.result(), step=epoch)
tf.summary.scalar('test_err_top1', test_err1, step=epoch)
tf.summary.scalar('test_err_top5', test_err5, step=epoch)
# print metrics to screen
if ((epoch + 1) % args.print_freq == 0) & (args.verbose == 1):
print(
template.format(
epoch + 1, args.epochs,
optimizer.learning_rate(optimizer.iterations).numpy(),
train_loss.result(), train_err1, train_err5,
test_loss.result(), test_err1, test_err5,
time.time() - s_time))
# # saving only the best model
# is_best = test_err1 <= best_err1
# best_err1 = min(test_err1, best_err1)
# if is_best:
# print('Model being saved...')
# manager.save(checkpoint_number=optimizer.iterations)
# print('Best accuracy (top-1 and 5 error):', best_err1, best_err5)
# save only the last epoch model
ckpt_manager.save(checkpoint_number=optimizer.iterations)
print(f'******* Model saved to {ckpt_dir} ******')
print(f'******* Test top-1 err: {test_err1} top-5 err: {test_err5}')
def main():
args = parser.parse_args()
if args.dataset == 'cifar10':
_, (test_images, test_labels) = datasets.cifar10.load_data()
num_classes = 10
height, width, channels = 32, 32, 3
elif args.dataset == 'cifar100':
_, (test_images, test_labels) = datasets.cifar100.load_data()
num_classes = 100
height, width, channels = 32, 32, 3
test_dataset = tf.data.Dataset.from_tensor_slices(
(test_images, test_labels))
test_dataset = test_dataset.map(test_augment, num_parallel_calls=AUTO)
test_dataset = test_dataset.batch(args.batch_size)
test_accuracy_top1 = tf.keras.metrics.SparseTopKCategoricalAccuracy(
k=1, name='test_accuracy_top1')
test_accuracy_top5 = tf.keras.metrics.SparseTopKCategoricalAccuracy(
k=5, name='test_accuracy_top5')
# model initialization
model = ResNet(dataset=args.dataset,
depth=args.depth,
num_classes=num_classes,
bottleneck=args.bottleneck)
# load the latest checkpoint from the given ckpt folder
ckpt = tf.train.Checkpoint(model=model)
manager = tf.train.CheckpointManager(ckpt, args.ckpt_dir, max_to_keep=1)
if manager.latest_checkpoint:
ckpt.restore(
manager.latest_checkpoint).expect_partial() # optimizer not loaded
print('checkpoint restored')
else:
raise Except
print('checkpoint not restored')
# test_summary_writer
current_time = datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
test_log_dir = os.path.join(args.tensorboard_dir, current_time, 'test')
print(f'******* Test result log written to {test_log_dir} ******')
test_summary_writer = tf.summary.create_file_writer(test_log_dir)
# training
s_time = time.time()
for image, target in test_dataset:
image = normalize(image)
output = model(image, training=False)
test_accuracy_top1(target, output)
test_accuracy_top5(target, output)
test_err1 = (1 - test_accuracy_top1.result()) * 100
test_err5 = (1 - test_accuracy_top5.result()) * 100
with test_summary_writer.as_default():
tf.summary.scalar('test_err_top1', test_err1, step=0)
tf.summary.scalar('test_err_top5', test_err5, step=0)
if args.verbose == 1:
print(template.format(test_err1, test_err5, time.time() - s_time))
