conv_block in resnets implementation

[Pagey]

Hey- kudos for the elegant implementation!- I'm having trouble understanding the purpose of these two functions:

def conv_block(inputs, num_filters):

Construct Block of Convolutions without Pooling

# x        : input into the block
# n_filters: number of filters
conv = Conv_1D_Block(inputs, num_filters, 3, 2)
conv = Conv_1D_Block(conv, num_filters, 3, 2)
return conv

def conv_block_bottleneck(inputs, num_filters):

Construct Block of Convolutions without Pooling

# x        : input into the block
# n_filters: number of filters
conv = Conv_1D_Block(inputs, num_filters, 3, 2)
conv = Conv_1D_Block(conv, num_filters, 3, 2)
conv = Conv_1D_Block(conv, num_filters, 3, 2)
return conv

which seem to add 2 or 3 consecutive convolutions with strides of 2 (i.e. they enforce a reduction of 4 or 8 in map sizes) after each block- i don't see evidence to this in the original resnet/resnext papers- which i believe only mentioned that the first convolutions of each blocks be with stride of 2- can you please elaborate on their purpose?

Thanks!

[Sakib1263]

Hi Pagey,

Yes, your concern is correct, stride length except the first conv block should be equal to 1. All codes have been updated and some new codes have been added.

Regards, Sakib