How to understan the parameters in the application_resnet50() model?

[danli349]

Hello:

model <- application_resnet50(weights = "imagenet")
summary(model)

In the input layer the Output Shape is (None, 224, 224, 3)
In the conv1_conv layer, why the Output Shape is (None, 112, 112, 64)?
There is no max pooling layer before it.
Does it omit a (2, 2) max pooling layer, or is the kernel step size 2?
And the kernel size should be (7, 7), right?
How does the conv2_block1_1_bn(BatchNormalization) work layer work? why the parameters are 256?

Thanks a lot

[t-kalinowski]

The first conv1_conv layer is configured with stride=2.
https://github.com/keras-team/keras/blob/c8c2c6fc7585a56ca8c8dfd084f5dadad6ad1036/keras/src/applications/resnet.py#L152

One way towards answering the second question is to directly inspect the weights of the layer:

get_layer(model, "conv2_block1_1_bn")$weights
get_layer(model, "conv2_block1_1_bn")$weights |> sapply(\(w) setNames(list(w$shape), w$name)) 

$gamma
TensorShape([64])

$beta
TensorShape([64])

$moving_mean
TensorShape([64])

$moving_variance
TensorShape([64])

Please take a look at the reference page for layer_batch_normalization() for a description of how the different weight tensors are applied.

[danli349]

@t-kalinowski Thanks a lot