UFCNN值错误：检查模型输入时

1 回答

• 2

  Theano 1.0.0，Keras 2.1.3，Python 2.7.10

  Shape of Input and Output: 对于我对keras和CNN的理解，你需要在形状样本上训练网络 (n_samples, sequence_length, features) . 这意味着在rolling window view中扩展您的时间序列，将形状 (6400, 1, 1) 广播到 (6400, 64, 1) . 你这样做是通过as_strided：

  ## reshape data
  cos_view = np.lib.stride_tricks.as_strided(
      cos,
      shape=[cos.shape[0] - sequence_length, sequence_length, 1],
      strides=cos.strides
  )
  expected_output_view = np.lib.stride_tricks.as_strided(
      expected_output,
      shape=[expected_output.shape[0] - sequence_length, sequence_length, 1],
      strides=expected_output.strides
  )
  

  但是 expected_output 也必须具有更高的维度 . 在 gen_cosine_amp ：

  def gen_cosine_amp(amp=100, period=25, x0=0, xn=500, step=1, k=0.0001):
      ...
      expected_output = np.zeros((len(cos), 1, 1))
      for i in range(len(cos) - lahead):
          expected_output[i, 0, 0] = np.mean(cos[i + 1:i + lahead + 1])
  

  现在可以在这些广播视图上训练您的模型：

  UFCNN_1.fit(cos_view, expected_output_view, verbose=1,epochs=1,shuffle=False, batch_size=batch_size)
  

  Fixing Model Bug: 但请注意，您的代码中还有其他错误 . 图层 conv_7 是您的输出图层 . 其输出/过滤器维度不应为 nb_filter ，而是 output_dim

  conv_7 = Conv1D(filters=output_dim, kernel_size=filter_length, padding='same')(relu_6)
  

  Concerns: 即使有了这些修改，对数据的训练效果也很好，我想这个模型将会延伸到时间序列的未来（假设你的信号是lukovkin的ufcnn-keras中的时间序列） . 也许Conv1D -layer与 padding="causal" 和 dilation_rate=x （x> 1）会更好 . （我还在尝试时间序列预测）

  此外，请注意，模型通过add合并层，而不是通过concatenate合并层，这类似于Roni Mittelman中描述的UFCNN模型 . The code below doesn't reflect these concerns.

  Putting it all together:

  from __future__ import absolute_import
  from __future__ import print_function
  import numpy as np
  import keras
  from keras.models import Model
  from keras.models import Sequential
  from keras.layers import Input, merge
  from keras.layers.core import Activation
  from keras.layers.convolutional import Conv1D
  import matplotlib.pyplot as plt
  from keras.preprocessing import sequence
  
  def ufcnn_regression_model(sequence_length=5000,
                             features=1,
                             nb_filter=150,
                             filter_length=5,
                             output_dim=1,
                             optimizer='adagrad',
                             loss='mse'):
      inputs = Input(shape=(sequence_length, features), name = 'input')
  
      #########################################################
      conv_1 = Conv1D(filters=nb_filter, kernel_size=filter_length, padding='same')(inputs)
      relu_1 = Activation('relu')(conv_1)
      #########################################################
      conv_2 = Conv1D(filters=nb_filter, kernel_size=filter_length, padding='same')(relu_1)
      relu_2 = Activation('relu')(conv_2)
      #########################################################
      conv_3 = Conv1D(filters=nb_filter, kernel_size=filter_length, padding='same')(relu_2)
      relu_3 = Activation('relu')(conv_3)
      #########################################################
      conv_4 = Conv1D(filters=nb_filter, kernel_size=filter_length, padding='same')(relu_3)
      relu_4 = Activation('relu')(conv_4)
      #########################################################
      merge_1 = keras.layers.add([relu_2, relu_4])
      conv_5 =Conv1D(filters=nb_filter, kernel_size=filter_length, padding='same')(merge_1)
      relu_5 = Activation('relu')(conv_5)
      #########################################################
      merge_2 = keras.layers.add([relu_1, relu_5])
      conv_6 = Conv1D(filters=nb_filter, kernel_size=filter_length, padding='same')(merge_2)
      relu_6 = Activation('relu')(conv_6)
      #########################################################
      conv_7 = Conv1D(filters=output_dim, kernel_size=filter_length, padding='same')(relu_6)
      #########################################################
      model = Model(inputs = inputs, outputs = conv_7)
      model.compile(optimizer=optimizer, loss=loss)
      print(model.summary())
      return model
  
  ## Input & Output function
  def gen_cosine_amp(amp=100, period=25, x0=0, xn=500, step=1, k=0.0001):
      cos = np.zeros(((xn - x0) * step,  1, 1))
      print("Cos. Shape",cos.shape)
      for i in range(len(cos)):
          idx = x0 + i * step
          cos[i, 0, 0] = amp * np.cos(idx / (2 * np.pi * period))
          cos[i, 0, 0] = cos[i, 0, 0] * np.exp(-k * idx)
  
      lahead = 1
      expected_output = np.zeros((len(cos), 1, 1))
      for i in range(len(cos) - lahead):
          expected_output[i, 0, 0] = np.mean(cos[i + 1:i + lahead + 1])
      return cos, expected_output
  
  ##  Parameter
  sequence_length = 64
  features = 1
  nb_filter = 150
  filter_length = 5
  output_dim = 1
  epochs = 5
  batch_size = 128
  
  ##  UFCNN_1 model summary
  UFCNN_1 = ufcnn_regression_model(sequence_length=sequence_length, nb_filter=nb_filter)
  
  ## Inputs and ouputs to be trained
  cos, expected_output = gen_cosine_amp(xn = sequence_length * 100)
  
  ## reshape data
  cos_view = np.lib.stride_tricks.as_strided(
      cos,
      shape=[cos.shape[0] - sequence_length, sequence_length, 1],
      strides=cos.strides
  )
  expected_output_view = np.lib.stride_tricks.as_strided(
      expected_output,
      shape=[expected_output.shape[0] - sequence_length, sequence_length, 1],
      strides=expected_output.strides
  )
  
  print("Cos. Shape Input ",cos_view.shape)
  ## Trainning
  for i in range(epochs):
      print('Epoch', i, '/', epochs)
      UFCNN_1.fit(cos_view, expected_output_view, verbose=1,epochs=1,shuffle=False, batch_size=batch_size)    
  
  print('Predicting')
  predicted_output = UFCNN_1.predict(cos_view, batch_size=batch_size)
  rmse = np.sqrt(((predicted_output - cos_view) ** 2).mean(axis=None))
  print ("RMSE ", rmse)
  

  回复于 2024-04-20T21:55:49+08:00