def discriminator_loss(real_imgs, fake_imgs):
real_loss = cross_entropy(tf.ones_like(real_imgs), real_imgs)
fake_loss = cross_entropy(tf.zeros_like(fake_imgs), fake_imgs)
total_loss = real_loss + fake_loss
return total_loss
def discriminator_loss(real_output, fake_output):
real_loss = cross_entropy(from_logits=True)(tf.ones_like(real_output),
real_output)
fake_loss = cross_entropy(from_logits=True)(tf.zeros_like(fake_output),
fake_output)
total_loss = real_loss + fake_loss
return total_loss
def distribution_focal_loss(y_true, y_pred):
r"""Distribution Focal Loss (DFL) is from `Generalized Focal Loss: Learning
Qualified and Distributed Bounding Boxes for Dense Object Detection
<https://arxiv.org/abs/2006.04388>`_.
Args:
pred (tf.Tensor): Predicted general distribution of bounding boxes
(before softmax) with shape (N, n+1), n is the max value of the
integral set `{0, ..., n}` in paper.
label (tf.Tensor): Target distance label for bounding boxes with
shape (N,).
Returns:
torch.Tensor: Loss tensor with shape (N,).
"""
dis_left = tf.cast(y_true, tf.int32)
dis_right = dis_left + 1
weight_left = tf.cast(dis_right, tf.float32) - y_true
weight_right = y_true - tf.cast(dis_left, tf.float32)
loss = cross_entropy(dis_left, y_pred, from_logits=True) * weight_left \
+ cross_entropy(dis_right, y_pred, from_logits=True) * weight_right
return loss
def generator_loss(fake_imgs):
return cross_entropy(tf.ones_like(fake_imgs), fake_imgs)
## Define Constans
GENERATE_SQUARE = 96
IMG_CHANNEL = 3
NOISE = 100
BATCH_SIZE = 32
BUFFER_DATA = 20000
EPOCHS = 200
IMG_SHAPE = (GENERATE_SQUARE, GENERATE_SQUARE, IMG_CHANNEL)
PATH = "Anime"
train = []
PREVIEW_ROWS = 4
PREVIEW_COLS = 7
PREVIEW_MARGIN = 16
cross_entropy = cross_entropy()
generator_optimizer = tf.keras.optimizers.Adam(1.5e-4, 0.5)
discriminator_optimizer = tf.keras.optimizers.Adam(1.5e-4, 0.5)
## Prepare Data
for filename in tqdm(os.listdir(PATH)):
img_ = Image.open(PATH + "/" + filename).resize(
(GENERATE_SQUARE, GENERATE_SQUARE), Image.ANTIALIAS)
train.append(np.asarray(img_))
train = np.reshape(train, (-1, GENERATE_SQUARE, GENERATE_SQUARE, IMG_CHANNEL))
train = train.astype(np.float32)
train = train / 127.5 - 1.
train_ = tf.data.Dataset.from_tensor_slices(train).shuffle(BUFFER_DATA).batch(
BATCH_SIZE)
def generator_loss(fake_output):
return cross_entropy(from_logits=True)(tf.ones_like(fake_output),
fake_output)