Keras、Pytorch实现卷积层替换
替换方法
- 修改输入数据维度
- 替换对应网络层(下文所示)
Keras
input = tf.random.normal(10, 512, 1)
conv1d = tf.keras.layers.Conv1D(24, 3, strides=1, use_bias=True, data_format="channels_last")(input)
input_2d = conv1d [..., np.newaxis] # 添加一个维度
conv2d = tf.keras.layers.Conv2D(24, (3,1), strides=1, use_bias=True, data_format="channels_last")(input_2d)
结果对比
Pytorch
import torch
input = torch.randn(20, 16, 50)
conv1d = torch.nn.Conv1d(16, 33, 3, stride=1) # Conv1d实现
output1d = torch.conv1d(input)
input_2d = input [..., np.newaxis] # 添加一个维度
conv2d = torch.nn.Conv2d(16, 33, (3,1), stride=1) # Conv2d实现Conv1d
output2d = conv2d(input_2d)