Question

我在Pytorch和Tensorflow上建立了简单的召集网络。

我将重量推向从Pytorch受过预训的层的Tensorflow层,根据stride,大不相同。参数

If the stride is set to 1, the l2 norm between the outputs is approximately same (0.8435), but when the stride is set to 2 or higher, the l2 norm is significantly high (156.6889).

是否有任何人解释为什么在Pytorch和Tensorflow对stride的计算有所不同?

# implementation on Pytorch
nn.Conv2d(3, 32, kernel_size=3, padding=1, stride=1, bias=False)

# implementation on Tensorflow
tf.keras.layers.Conv2D(filters=32, kernel_size=(3, 3), strides=1, padding="same", use_bias=False)

-> this results almost same output.

# implementation on Pytorch
nn.Conv2d(3, 32, kernel_size=3, padding=1, stride=2, bias=False)

# implementation on Tensorflow
tf.keras.layers.Conv2D(filters=32, kernel_size=(3, 3), strides=2, padding="same", use_bias=False)

-> this results different output.

Answer 1

众所周知,Py Torch和Tensorflow的粉碎 con混凝土结果之间存在一些差异。简言之,这些结果可能略有不同:边界效应、权重初始化——精确性、衰减和组群聚集、数字准确性和计算相关问题。

例如,str2是指可能产生大量产出差异的3px过滤器移动。以前讨论过的其他细节:here:

友情链接