please help to understand the extract_features funtion in net.py

[BKZero]

i read the paper and i think when we extract feature for compressing channels of one conv layer, we randomly select some patches of the input of the layer, which will generate a series of points of the output of the layer, and the patches are corresponding to the points. but in the extract_features funtion, i think the points and the patches are selected randomly and independently, so i confused.

for name in names: #i think the names[0] is the input layer name and names[1] is the output layer name
                # pad = pads[name]
                shape = shapes[name]
                feat = self.blobs_data(name)
                if 0: print(name, self.blobs_shape(name))
                if inner or len(self.blobs_shape(name))==2 or (shape[0] == 1 and shape[1] == 1):
                    feats_dict[name][fc_idx:(fc_idx + nPicsPerBatch)] = feat.reshape((self.num, -1))
                    continue
                if batch >= dcfgs.nBatches and name in self.convs:
                    continue
                # TODO!!! different patch for different image per batch
                if save:
                    if not frozen_points or (batch, name, "randx") not in points_dict:
                        #embed()
                        randx = np.random.randint(0, shape[0]-0, nPointsPerLayer)
                        randy = np.random.randint(0, shape[1]-0, nPointsPerLayer)
                        #when the loop number comes to 1, the randx should be the same with loop number=1, right?
                        if dcfgs.dic.option == cfgs.pruning_options.resnet:
                            branchrandxy = None
                            branch1name = '_branch1'
                            branch2cname = '_branch2c'` ```

i think the names[0] is the input layer name and names[1] is the output layer name
when the loop number comes to 1, the randx should be the same with loop number=1, right?

[BKZero]

https://github.com/yihui-he/channel-pruning/blob/master/lib/net.py

[ethanhe42]

you can look at PocketFlow They implemented both channel pruning and AMC. Their code is well documented.