def f_pool(self, x, spec, in_dim):
layer_type, dims = spec
num_filters = in_dim[0]
if "globalmeanpool" == layer_type:
y, output_size = global_meanpool_2d(x, num_filters)
# scale the variance to match normal conv layers with xavier init
y = y * np.float32(in_dim[-1]) * np.float32(np.sqrt(3))
else:
assert dims[0] != 1 or dims[1] != 1
y, output_size = maxpool_2d(x, in_dim,
poolsize=(dims[1], dims[1]),
poolstride=(dims[0], dims[0]))
return y, output_size
def f_pool(self, x, spec, in_dim):
layer_type, dims = spec
num_filters = in_dim[0]
if "globalmeanpool" == layer_type:
y, output_size = global_meanpool_2d(x, num_filters)
# scale the variance to match normal conv layers with xavier init
y = y * np.float32(in_dim[-1]) * np.float32(np.sqrt(3))
else:
assert dims[0] != 1 or dims[1] != 1
y, output_size = maxpool_2d(x, in_dim,
poolsize=(dims[1], dims[1]),
poolstride=(dims[0], dims[0]))
return y, output_size