@alan0324 对 看报错是你输入的原始图像和label图像尺寸不一致导致的 可以检查下;)

from attentive-gan-derainnet.

@alan0324 检查src_image和label_image的尺寸是不是对应:)

from attentive-gan-derainnet.

感謝您的回覆!! 請問是以下嗎if name == 'main':
input_image = tf.placeholder(dtype=tf.float32, shape=[1, 256, 256, 3])
auto_label_image = tf.placeholder(dtype=tf.float32, shape=[1, 256, 256, 3])
rnn_label_image = tf.placeholder(dtype=tf.float32, shape=[1, 256, 256, 1])
net = GenerativeNet(phase=tf.constant('train', tf.string))
rnn_loss = net.compute_attentive_rnn_loss(input_image, rnn_label_image, name='rnn_loss')
autoencoder_loss = net.compute_autoencoder_loss(input_image, auto_label_image, name='autoencoder_loss')
for vv in tf.trainable_variables():
print(vv.name)
不好意思 對這塊還不太熟QQ

from attentive-gan-derainnet.

另外補充我在for迴圈print出來的結果lm_loss shape: ()
lm_loss size: tf.Tensor(1, shape=(), dtype=int32)
mse_loss shape: (1, 60, 90)
mse_loss size: tf.Tensor(5400, shape=(), dtype=int32)
lm_loss shape: (1, 60, 90)
lm_loss size: tf.Tensor(5400, shape=(), dtype=int32)
mse_loss shape: (1, 120, 180)
mse_loss size: tf.Tensor(21600, shape=(), dtype=int32)

from attentive-gan-derainnet.

@MaybeShewill-CV 您好，經過了一番嘗試，即使更改了原始圖像與label圖像的尺寸也沒有改變我的報錯信息，我的mse_loss還是會在for迴圈的計算中增加一倍，請問除了尺寸不一致以外，還有什麼可能會導致錯誤的原因嗎?

from attentive-gan-derainnet.

附上我的尺寸信息
__C.TRAIN.IMG_HEIGHT = 480

Set train image width

__C.TRAIN.IMG_WIDTH = 720

Set train image height

__C.TRAIN.CROP_IMG_HEIGHT = 240

Set train image width

__C.TRAIN.CROP_IMG_WIDTH = 360

if name == 'main':
input_image = tf.keras.Input(dtype=tf.float32, shape=[1, 240, 360, 3])
auto_label_image = tf.keras.Input(dtype=tf.float32, shape=[1, 240, 360, 3])
rnn_label_image = tf.keras.Input(dtype=tf.float32, shape=[1, 240, 360, 1])

if name == 'main':
"""
test
"""
input_tensor = tf.keras.Input(dtype=tf.float32, shape=[5, 480, 720, 3])
label_tensor = tf.keras.Input(dtype=tf.float32, shape=[5, 480, 720, 3])
mask_tensor = tf.keras.Input(dtype=tf.float32, shape=[5, 480, 720, 1])

from attentive-gan-derainnet.