卷积1D层输出形状错误

huangapple
go评论99阅读模式
英文:

Wrong shape output with conv1D layer

问题

I'm trying some experiments with a small autoencoder defined like this:

  1. input_layer = Input(shape=input_shape)
  3. x = Conv1D(8, kernel_size=5, activation='relu', padding='same')(input_layer)
  4. x = BatchNormalization()(x)
  5. encoded = Conv1D(4, kernel_size=5, activation='relu', padding='same')(x)
  7. x = Conv1D(8, kernel_size=5, activation='relu', padding='same')(encoded)
  8. x = BatchNormalization()(x)
  10. decoded = Conv1D(WINDOW_SIZE, kernel_size=5, activation='relu', padding='same')(x)

input_shape = 8192 and autoencoder.summary() returns:

  1. Layer (type)                Output Shape              Param #
  2. =================================================================
  3. input_1 (InputLayer)        [(None, 8192, 1)]         0
  5. conv1d (Conv1D)             (None, 8192, 8)           48
  7. batch_normalization (BatchNormalization)  (None, 8192, 8)          32
  9. conv1d_1 (Conv1D)           (None, 8192, 4)           164
  11. conv1d_2 (Conv1D)           (None, 8192, 8)           168
  13. batch_normalization_1 (BatchNormalization)  (None, 8192, 8)          32
  15. conv1d_3 (Conv1D)           (None, 8192, 8192)        335,872
  16. =================================================================
  17. Total params: 336,316
  18. Trainable params: 336,284
  19. Non-trainable params: 32

Basically, I want an autoencoder that takes a vector of 8192 values and tries to predict another vector of 8192 values.

But when I do an inference like that:

  1. predictions = autoencoder.predict(batch_data, batch_size=BATCH_SIZE)

I have this shape:

  1. >>> predictions[0].shape
  2. (8192, 8192)

What am I doing wrong? What can I change to get a vector with (8192)?

英文:

I'm trying some experiments with a small autoencoder defined like this:

  1. input_layer = Input(shape=input_shape)
  3. x = Conv1D(8, kernel_size=5, activation='relu', padding='same')(input_layer)
  4. x = BatchNormalization()(x)
  5. encoded = Conv1D(4, kernel_size=5, activation='relu', padding='same')(x)
  7. x = Conv1D(8, kernel_size=5, activation='relu', padding='same')(encoded)
  8. x = BatchNormalization()(x)
  10. decoded = Conv1D(WINDOW_SIZE, kernel_size=5, activation='relu', padding='same')(x)

input_shape = 8192 and autoencoder.summary() returns:

  1.  Layer (type)                Output Shape              Param #
  2. =================================================================
  3.  input_1 (InputLayer)        [(None, 8192, 1)]         0
  5.  conv1d (Conv1D)             (None, 8192, 8)           48
  7.  batch_normalization (BatchN  (None, 8192, 8)          32
  8.  ormalization)
  10.  conv1d_1 (Conv1D)           (None, 8192, 4)           164
  12.  conv1d_2 (Conv1D)           (None, 8192, 8)           168
  14.  batch_normalization_1 (Batc  (None, 8192, 8)          32
  15.  hNormalization)
  17.  conv1d_3 (Conv1D)           (None, 8192, 8192)        335872
  19. =================================================================
  20. Total params: 336,316
  21. Trainable params: 336,284
  22. Non-trainable params: 32

Basically, I want an autoencoder that takes a vector of 8192 values and try to predict an another vector of 8192 values.

but when I do an inference like that:

  1. predictions = autoencoder.predict(batch_data, batch_size=BATCH_SIZE)

I've that shape:

  1. >>> predictions[0].shape
  2. (8192,8192)

What I'm doing wrong ? what can I change to get a vector with (8192) ?

答案1

得分: 1

可能是这样的吧？

  1. import tensorflow as tf
  3. input_layer = tf.keras.layers.Input(shape=(8192, 1))
  5. x = tf.keras.layers.Conv1D(8, kernel_size=5, activation='relu', padding='same')(input_layer)
  6. x = tf.keras.layers.BatchNormalization()(x)
  7. encoded = tf.keras.layers.Conv1D(4, kernel_size=5, activation='relu', padding='same')(x)
  9. x = tf.keras.layers.Conv1D(8, kernel_size=5, activation='relu', padding='same')(encoded)
  10. x = tf.keras.layers.BatchNormalization()(x)
  12. decoded = tf.keras.layers.Conv1D(1, kernel_size=5, activation='relu', padding='same')(x)
  13. model = tf.keras.Model(input_layer, decoded)
  14. model.summary()
  1. Model: "model_1"
  2. _________________________________________________________________
  3.  Layer (type)                Output Shape              Param #   
  4. =================================================================
  5.  input_2 (InputLayer)        [(None, 8192, 1)]         0         
  6.                                                                  
  7.  conv1d_4 (Conv1D)           (None, 8192, 8)           48        
  8.                                                                  
  9.  batch_normalization_2 (Batc  (None, 8192, 8)          32        
  10.  hNormalization)                                                 
  11.                                                                  
  12.  conv1d_5 (Conv1D)           (None, 8192, 4)           164       
  13.                                                                  
  14.  conv1d_6 (Conv1D)           (None, 8192, 8)           168       
  15.                                                                  
  16.  batch_normalization_3 (Batc  (None, 8192, 8)          32        
  17.  hNormalization)                                                 
  18.                                                                  
  19.  conv1d_7 (Conv1D)           (None, 8192, 1)           41        
  20.                                                                  
  21. =================================================================
  22. Total params: 485
  23. Trainable params: 453
  24. Non-trainable params: 32
  25. _________________________________________________________________
英文:

Maybe something like this?

  1. import tensorflow as tf
  3. input_layer = tf.keras.layers.Input(shape=(8192, 1))
  5. x = tf.keras.layers.Conv1D(8, kernel_size=5, activation='relu', padding='same')(input_layer)
  6. x = tf.keras.layers.BatchNormalization()(x)
  7. encoded = tf.keras.layers.Conv1D(4, kernel_size=5, activation='relu', padding='same')(x)
  9. x = tf.keras.layers.Conv1D(8, kernel_size=5, activation='relu', padding='same')(encoded)
  10. x = tf.keras.layers.BatchNormalization()(x)
  12. decoded = tf.keras.layers.Conv1D(1, kernel_size=5, activation='relu', padding='same')(x)
  13. model = tf.keras.Model(input_layer, decoded)
  14. model.summary()
  1. Model: "model_1"
  2. _________________________________________________________________
  3.  Layer (type)                Output Shape              Param #   
  4. =================================================================
  5.  input_2 (InputLayer)        [(None, 8192, 1)]         0         
  6.                                                                  
  7.  conv1d_4 (Conv1D)           (None, 8192, 8)           48        
  8.                                                                  
  9.  batch_normalization_2 (Batc  (None, 8192, 8)          32        
  10.  hNormalization)                                                 
  11.                                                                  
  12.  conv1d_5 (Conv1D)           (None, 8192, 4)           164       
  13.                                                                  
  14.  conv1d_6 (Conv1D)           (None, 8192, 8)           168       
  15.                                                                  
  16.  batch_normalization_3 (Batc  (None, 8192, 8)          32        
  17.  hNormalization)                                                 
  18.                                                                  
  19.  conv1d_7 (Conv1D)           (None, 8192, 1)           41        
  20.                                                                  
  21. =================================================================
  22. Total params: 485
  23. Trainable params: 453
  24. Non-trainable params: 32
  25. _________________________________________________________________

huangapple
  • 本文由 发表于 2023年3月9日 19:30:36
  • 转载请务必保留本文链接：https://go.coder-hub.com/75684005.html
  • autoencoder
  • keras
  • python
  • tensorflow
匿名

发表评论

匿名网友

:?: :razz: :sad: :evil: :!: :smile: :oops: :grin: :eek: :shock: :???: :cool: :lol: :mad: :twisted: :roll: :wink: :idea: :arrow: :neutral: :cry: :mrgreen:

确定