卷积神经网络

layers.Conv2D

In [1]: import tensorflow as tf
   ...: from tensorflow.keras import layers

########## x必须ndim=4, shape:[b, h, w, c]
In [2]: x = tf.random.normal([1, 32, 32, 3])

In [3]: layer = layers.Conv2D(4, kernel_size=5, strides=1, padding='valid')

In [4]: out = layer(x)

In [5]: out.shape
Out[5]: TensorShape([1, 28, 28, 4])

In [6]: layer = layers.Conv2D(4, kernel_size=5, strides=1, padding='same')

In [7]: out = layer(x)

In [8]: out.shape
Out[8]: TensorShape([1, 32, 32, 4])

In [9]: layer = layers.Conv2D(4, kernel_size=5, strides=2, padding='same')

In [10]: out = layer(x)

In [11]: out.shape
Out[11]: TensorShape([1, 16, 16, 4])
# weight & bias
In [12]: layer.kernel
Out[12]: 
<tf.Variable 'conv2d_2/kernel:0' shape=(5, 5, 3, 4) dtype=float32, numpy=
array([[[[ 0.07923736,  0.14217006, -0.0109257 ,  0.08888994],
         [-0.01091294, -0.16224794,  0.18029307, -0.01917681],
         [ 0.1477335 ,  0.13419123,  0.13923149, -0.18107778]],...

nn.conv2d

In [14]: w = tf.random.normal([5, 5, 3, 4])

In [15]: b = tf.zeros([4])

In [16]: x.shape
Out[16]: TensorShape([1, 32, 32, 3])

In [17]: out = tf.nn.conv2d(x, w, strides=1, padding='VALID')

In [18]: out.shape
Out[18]: TensorShape([1, 28, 28, 4])

In [19]: out = out + b

In [20]: out.shape
Out[20]: TensorShape([1, 28, 28, 4])

In [21]: out = tf.nn.conv2d(x, w, strides=2, padding='VALID')

In [22]: out.shape
Out[22]: TensorShape([1, 14, 14, 4])

Gradient

标签：layer,卷积,shape,神经网络,TensorShape,tf,out,Out
来源： https://blog.csdn.net/qq_46456049/article/details/112995733