InsightFace： how to train 112*96

(1)修改insightface\src\image_iter.py： def net_sample(sef) line124 return之前加一句img=img[:,8:104,:]

修改前

header, img = recordio.unpack(s)

return header.label, img, None, None

修改后

header, img = recordio.unpack(s)

img = img[:,8:104,:]#added by zuoqing

return header.label, img, None, None

(2)修改insightface\src\image_iter.py： def next(self) line215 self.postprocess_data(datum)之前加一句datum=datum[:,8:104,:]

修改前

#print(datum.shape)

batch_data[i][:] = self.postprocess_data(datum)

修改后

#print(datum.shape)

datum = datum[:,8:104,:]# added by zuoqing

batch_data[i][:] = self.postprocess_data(datum)

(3)修改datasets\faces_emore\property：112,112,改成了112,96 （同理，你用其他数据训练也得改）

(4)修改insightface\src\eval\verification.py： def load_bin(path,image_size) line193 tranpose之前加一句img=img[:,8:104,:]

修改前

img = mx.image.imdecode(_bin)

img = nd.transpose(img, axes=(2, 0, 1))

修改后

img = mx.image.imdecode(_bin)

img = img[:,8:104,:]#added by zuoqing

img = nd.transpose(img, axes=(2, 0, 1))

(5)修改insightface\src\symbols\symbol_utils.py里面get_fc1把kernel改成(7,6)

改完之后显存会少使用一些，可以使用更大的batch_size