def get_output_for(self, input, deterministic=False, **kwargs):
def _phase_shift(input,r):
bsize,c,a,b = input.shape[0],1,self.output_shape[2]//r,self.output_shape[3]//r
X = T.reshape(input, (bsize,r,r,a,b))
X = T.transpose(X, (0, 3,4,1,2)) # bsize, a, b, r2,r1
X = T.split(x=X,splits_size=[1]*a,n_splits=a,axis=1) # a, [bsize, b, r, r]
X = [T.reshape(x,(bsize,b,r,r))for x in X]
X = T.concatenate(X,axis=2) # bsize, b, a*r, r
X = T.split(x=X,splits_size =[1]*b,n_splits=b,axis=1) # b, [bsize, a*r, r]
X = [T.reshape(x,(bsize,a*r,r))for x in X]
X = T.concatenate(X,axis=2) # bsize, a*r, b*r
return X.dimshuffle(0,'x',1,2)
Xc = T.split(x=input,splits_size =[input.shape[1]//self.c]*self.c,n_splits=self.c,axis=1)
return T.concatenate([_phase_shift(xc,self.r) for xc in Xc],axis=1)
# Multiscale Dilated Convolution Block
# This function (not a layer in and of itself, though you could make it one) returns a set of concatenated conv2d and dilatedconv2d layers.
# Each layer uses the same basic filter W, operating at a different dilation factor (or taken as the mean of W for the 1x1 conv).
# The channel-wise output of each layer is weighted by a set of coefficients, which are initialized to 1 / the total number of dilation scales,
# meaning that were starting by taking an elementwise mean. These should be learnable parameters.
# NOTES: - I'm considering changing the variable names to be more descriptive, and look less like ridiculous academic code. It's on the to-do list.
# - I keep the bias and nonlinearity out of the default definition for this layer, as I expect it to be batchnormed and nonlinearized in the model config.
评论列表
文章目录
