Batchnorm2d and Batchnormalization

[zhang-f]

I try to transfer the batchnorm2d layer, but it fails......I really want to know what's the difference between the two. Thanks! code:

# Create dummy data
data = torch.rand(1,3,2,2)
print(data)
data_keras = data.numpy()
data_pytorch = Variable(data)

da= nn.BatchNorm2d(3,momentum=0.99,eps=1e-5)

def lenet_keras():
    model = Sequential()
    model.add(layers.BatchNormalization(axis=1,momentum=0.99,epsilon= 1e-5,
                                center=True, scale=True, beta_initializer=initializers.get('zeros'), 
                                gamma_initializer=initializers.get('ones'),
                                input_shape = (3,2,2)))
    return model
le = lenet_keras()
le.summary()
wei = da.weight.detach().numpy()
bi = da.bias.detach().numpy()
rm = da.running_mean.detach().numpy()
rv = da.running_var.detach().numpy()
#print(bi)
#wei = np.transpose(wei,(2,3,1,0))
trans = [wei,bi,rv,rm]
print(trans)
weight = le.layers[0].set_weights(trans)
weight = le.layers[0].get_weights()
print(weight)
print(da(data_pytorch))
print(da(data_pytorch).shape)
dat = le.predict(data_keras)
print(dat)
print(dat.shape)

output:

tensor([[[[0.6875, 0.5928],
          [0.2540, 0.3808]],

         [[0.9787, 0.3589],
          [0.5710, 0.3397]],

         [[0.2124, 0.8893],
          [0.0698, 0.6545]]]])
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
batch_normalization_17 (Batc (None, 3, 2, 2)           12        
=================================================================
Total params: 12
Trainable params: 6
Non-trainable params: 6
_________________________________________________________________
[array([1., 1., 1.], dtype=float32), array([0., 0., 0.], dtype=float32), array([1., 1., 1.], dtype=float32), array([0., 0., 0.], dtype=float32)]
[array([1., 1., 1.], dtype=float32), array([0., 0., 0.], dtype=float32), array([1., 1., 1.], dtype=float32), array([0., 0., 0.], dtype=float32)]
tensor([[[[ 1.2217,  0.6677],
          [-1.3158, -0.5735]],

         [[ 1.6204, -0.7903],
          [ 0.0348, -0.8649]],

         [[-0.7396,  1.3113],
          [-1.1717,  0.6000]]]], grad_fn=<NativeBatchNormBackward>)
torch.Size([1, 3, 2, 2])
[[[[ -98.82477   -128.75443  ]
   [-235.9049    -195.80484  ]]

  [[  -6.7411804 -202.74635  ]
   [-135.65837   -208.8067   ]]

  [[-249.05997    -35.020264 ]
   [-294.15558   -109.255936 ]]]]
(1, 3, 2, 2)

[zhang-f]

Thanks, solved~