只考虑一维，毕竟大部分图像cnn网络的两个维度也是同样参数。
# -*- coding: utf-8 -*-
"""
Created on Tue Mar 26 20:33:32 2019
@author: BigFly
"""
upper = lambda x: int(x) + int( (x-int(x))>0 )
class Layer(object):
def __init__(self, name, kernel, stride, pad=[0,0], padmode="VALID"):
self.name = name
self.kernel = kernel
self.stride = stride
self.pad = pad
self.padmode = padmode
self.used = False
padding = "%s(%-2d,%-2d)"%(self.padmode,self.pad[0],self.pad[1])
self.details="['%s':\t k_%-2d s_%-2d p_%s unused]"\
%(self.name, self.kernel, self.stride, padding)
def handle(self,shape):
reshape = shape+self.pad[0]+self.pad[1]
outshape = 0
if self.padmode.lower() == 'valid':
outshape = upper((reshape-self.kernel+1)/self.stride)
realshape = (outshape-1)*self.stride+self.kernel
addpad =[0,realshape-reshape]
elif self.padmode.lower() == 'same':
outshape = upper(reshape/self.stride)
realshape = (outshape-1)*self.stride+self.kernel
dshape=realshape-reshape
addpad = [dshape//2, dshape-dshape//2]
self.padding=[self.pad[i]+addpad[i] for i in [0,1]]
self.inshape=shape
self.reshape=realshape
self.outshape=outshape
self.used = True
self.receptive_field()
self.process="['%s':\t in:%-4d (%-4d) -> out:%-4d]"\
%(self.name, self.inshape,self.reshape,self.outshape)
padding = "%s(%-2d,%-2d)"%(self.padmode,self.padding[0],self.padding[1])
self.details="['%s':\t k_%-2d s_%-2d p_%s used]"\
%(self.name, self.kernel, self.stride, padding)
return outshape
def prints(self):
print(self.details)
if self.used:
print(self.process)
print(self.receptive_field)
def receptive_field(self):
self.receptive_field=[(-self.padding[0]+i*self.stride,
-self.padding[0]+i*self.stride+self.kernel-1)
for i in range(self.outshape) ]
class Networks(object):
def __init__(self,inshape,layers):
self.inshape= inshape
self.layersnum= len(layers)
layershape= inshape
for layer in layers:
layershape = layer.handle(layershape)
self.details =['input: %d'%(inshape)]+\
[layer.details for layer in layers]
self.process =['input: %d'%(inshape)]+\
[layer.process for layer in layers]+\
['output: %d'%(layershape)]
self.receptive_field = self.clac_receptive_field()
def prints(self):
print("\n".join(self.details))
print("----------------")
print("\n".join(self.process))
print("----------------")
for i in range(self.layersnum):
print("%s \t "%(layers[i].name),self.receptive_field[i])
print("original range : ",self.receptive_field[-1])
print("----------------")
def clac_receptive_field(self):
receptive_field=[layers[0].receptive_field]
realkernel,realstride = layers[0].kernel,layers[0].stride
start = layers[0].receptive_field[0][0]
end = layers[0].receptive_field[-1][1]
if self.layersnum>1:
for layer in layers[1:]:
cur_rf=[]
for rf in layer.receptive_field:
cur_rf.append((start+rf[0]*realstride,
start+rf[1]*realstride+realkernel-1))
receptive_field.append(cur_rf)
start = cur_rf[0][0]
end = min(end,cur_rf[-1][1])
realstride, realkernel = realstride*layer.stride, \
realstride*(layer.kernel-1)+realkernel
receptive_field.append([0,end])
return receptive_field
layers=[ Layer("conv1", 11, 4, [0,0]), \
Layer("pool1", 3, 2, [0,1]), \
Layer("conv2", 5, 1, [2,2]), \
Layer("pool2", 3, 2, [0,0]), \
Layer("conv3", 3, 1, [1,1]), \
Layer("conv4", 3, 1, [1,1]), \
Layer("conv5", 3, 1, [1,1]), \
Layer("pool5", 3, 2, [0,0])]
cnnf=Networks(224,layers)
cnnf.prints()