loss function problem

[alan0324]

您好，我在執行train_model.py時遇到了loss張量不合的問題，以下是我遇到的問題: Traceback (most recent call last): File "F:\temp\demo\00\pythonProject\attentive-gan-derainnet\tools\train_model.py", line 324, in train_model(args.dataset_dir, weights_path=args.weights_path) File "F:\temp\demo\00\pythonProject\attentive-gan-derainnet\tools\train_model.py", line 116, in train_model train_gan_loss, train_discriminative_loss, train_net_output = derain_net.compute_loss( File "F:\temp\demo\00\pythonProject\attentive-gan-derainnet\attentive_gan_model\derain_drop_net.py", line 52, in compute_loss auto_encoder_loss, auto_encoder_output = self._attentive_gan.compute_autoencoder_loss( File "F:\temp\demo\00\pythonProject\attentive-gan-derainnet\attentive_gan_model\attentive_gan_net.py", line 333, in compute_autoencoder_loss lm_loss = tf.add(lm_loss, mse_loss) File "F:\temp\demo\00\pythonProject\venv\lib\site-packages\tensorflow\python\util\traceback_utils.py", line 153, in error_handler raise e.with_traceback(filtered_tb) from None File "F:\temp\demo\00\pythonProject\venv\lib\site-packages\tensorflow\python\framework\ops.py", line 7262, in raise_from_not_ok_status raise core._status_to_exception(e) from None # pylint: disable=protected-access tensorflow.python.framework.errors_impl.InvalidArgumentError: {{function_node wrappedAddV2device/job:localhost/replica:0/task:0/device:CPU:0}} Incompatible shapes: [1,60,90] vs. [1,120,180] 請問要如何解決此問題 非常感謝您提供的代碼!

[MaybeShewill-CV]

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

[alan0324]

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

[alan0324]

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)

[MaybeShewill-CV]

@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;)

[alan0324]

@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?

[alan0324]

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])