我对深度学习、API和卷积网络都是新手,所以如果这些错误是幼稚的,请提前致歉。我试图建立一个简单的卷积神经网络分类。输入数据X有286
个样本,每个样本的500
时间点为4
维。维数是范畴变量的一个热门编码。我不知道该对Y
做些什么,所以我只是对样本做了一些聚类,然后对它们进行了热编码,以便为建模做实验。Y
目标数据有286
个样本,其中一个是6
类别的热编码。我的最终目标就是让它运行起来,这样我就可以找出如何将其更改为实际有用的学习问题,并使用隐藏层进行特征提取。在
我的问题是在最后一层中我不能让形状匹配。在
The model I made does the following:
(1) Inputs the data
(2) Convolutional layer
(3) Maxpooling layer
(4) Dropout regularization
(5) Large fully connected layer
(6) Output layer
import tensorflow as tf
import numpy as np
# Data Description
print(X[0,:])
# [[0 0 1 0]
# [0 0 1 0]
# [0 1 0 0]
# ...,
# [0 0 1 0]
# [0 0 1 0]
# [0 0 1 0]]
print(Y[0,:])
# [0 0 0 0 0 1]
X.shape, Y.shape
# ((286, 500, 4), (286, 6))
# Tensorboard callback
tensorboard= tf.keras.callbacks.TensorBoard()
# Build the model
# Input Layer taking in 500 time points with 4 dimensions
input_layer = tf.keras.layers.Input(shape=(500,4), name="sequence")
# 1 Dimensional Convolutional layer with 320 filters and a kernel size of 26
conv_layer = tf.keras.layers.Conv1D(320, 26, strides=1, activation="relu", )(input_layer)
# Maxpooling layer
maxpool_layer = tf.keras.layers.MaxPooling1D(pool_size=13, strides=13)(conv_layer)
# Dropout regularization
drop_layer = tf.keras.layers.Dropout(0.3)(maxpool_layer)
# Fully connected layer
dense_layer = tf.keras.layers.Dense(512, activation='relu')(drop_layer)
# Softmax activation to get probabilities for output layer
activation_layer = tf.keras.layers.Activation("softmax")(dense_layer)
# Output layer with probabilities
output = tf.keras.layers.Dense(num_classes)(activation_layer)
# Build model
model = tf.keras.models.Model(inputs=input_layer, outputs=output, name="conv_model")
model.compile(loss="categorical_crossentropy", optimizer="adam", callbacks=[tensorboard])
model.summary()
# _________________________________________________________________
# Layer (type) Output Shape Param #
# =================================================================
# sequence (InputLayer) (None, 500, 4) 0
# _________________________________________________________________
# conv1d_9 (Conv1D) (None, 475, 320) 33600
# _________________________________________________________________
# max_pooling1d_9 (MaxPooling1 (None, 36, 320) 0
# _________________________________________________________________
# dropout_9 (Dropout) (None, 36, 320) 0
# _________________________________________________________________
# dense_16 (Dense) (None, 36, 512) 164352
# _________________________________________________________________
# activation_7 (Activation) (None, 36, 512) 0
# _________________________________________________________________
# dense_17 (Dense) (None, 36, 6) 3078
# =================================================================
# Total params: 201,030
# Trainable params: 201,030
# Non-trainable params: 0
model.fit(X,Y, batch_size=128, epochs=100)
# ValueError: Error when checking target: expected dense_17 to have shape (None, 36, 6) but got array with shape (286, 6, 1)
Conv1D
的输出形状是一个3阶张量(batch, observations, kernels)
:在
然而,}将卷积与密集区分开就足以解决这个问题:
Dense
层需要一个2阶张量(batch, features)
。一个Flatten
、GlobalAveragePooling1D
或{Flatten
将把(batch, observations, kernels)
张量重塑为(batch, observations * kernels)
张量:GlobalAveragePooling1D
将平均(batch, observations, kernels)
张量中的所有观测值,得到一个(batch, kernels)
张量:您的tensorboard回调初始化似乎也有问题。这个很容易修好。在
对于时态数据处理,请看一下TimeDistributed wrapper。在
相关问题 更多 >
编程相关推荐