adding resnet18, gap

Author: yunlongdong

.gitignore

@@ -4,3 +4,4 @@ __pycache__
 *egg*
 build/*
 *.npy
+*.onnx

cnnumpy/io.py

@@ -26,6 +26,7 @@ def parse(matched):
 
 conv = re.compile(r'.*%(.+?) .+?(Conv).+?dilations=(\[\d+?, \d+?\]).+?strides=(\[\d+?, \d+?\]).+?(\(%.+?, %.+?, %.+?\)).+?\n')
 relu = re.compile(r'.*%(.+?) .+?(Relu)\(%(.+?)\).+?\n')
+gap = re.compile(r'.*%(.+?) .+?(GlobalAveragePool)\(%(.+?)\).+?\n')
 sigmoid = re.compile(r'.*%(.+?) .+?(Sigmoid)\(%(.+?)\).+?\n')
 maxpool = re.compile(r'.*%(.+?) .+?(MaxPool).+?kernel_shape=(\[\d+?, \d+?\]).+?strides=(\[\d+?, \d+?\]).+?\(%(.+?)\).+?\n')
 upsample = re.compile(r'.*%.+? .+?Constant\[value=.+?(\d+\.?\d*) \[.+?\n.+?%(.+?) .+?(Upsample).+?\(%(.+?),.+?\n')
@@ -36,7 +37,7 @@ def parse(matched):
 add = re.compile(r'.*%(.+?) .+?(Add)(\(%.+?\)).+?\n')
 weight = re.compile(r'.*%(.+?) .+?(\(.*?\)).*\n')
 
-res = (flatten, upsample, conv, relu, sigmoid, maxpool, dense, concat, add, batchnorm, weight)
+res = (flatten, upsample, conv, relu, gap, sigmoid, maxpool, dense, concat, add, batchnorm, weight)
 
 def read_onnx(path):
 	with open(path+'.txt') as f:
@@ -58,7 +59,7 @@ def read_onnx(path):
 			flow.append((i[4][0], ['conv_%s'%num], i[0]))
 		elif i[1]=='Gemm':
 			num = len(body)
-			body.append(('dense_%s'%num, 'dense', key[i[2][1][::-1]]))
+			body.append(('dense_%s'%num, 'dense', key[i[2][1]][::-1]))
 			flow.append((i[2][0], ['dense_%s'%num], i[0]))
 		elif i[1]=='Sigmoid':
 			num = len(body)
@@ -68,6 +69,10 @@ def read_onnx(path):
 			num = len(body)
 			body.append(('relu_%s'%num, 'relu', None))
 			flow.append((i[2], ['relu_%s'%num], i[0]))
+		elif i[1]=='GlobalAveragePool':
+			num = len(body)
+			body.append(('gap_%s'%num, 'gap', None))
+			flow.append((i[2], ['gap_%s'%num], i[0]))
 		elif i[1]=='Add':
 			num = len(body)
 			body.append(('add_%s'%num, 'add', None))
@@ -93,8 +98,9 @@ def read_onnx(path):
 			body.append(('flatten_%s'%num, 'flatten', None))
 			flow.append((i[2], ['flatten_%s'%num], i[0]))
 
-	#for i in body: print(i)
-	#for i in flow: print(i)
+	# for i in body: print(i)
+	# print("===============	")
+	# for i in flow: print(i)
 
 	net = Net()
 	net.load_json(body, flow)

cnnumpy/layer.py

@@ -94,11 +94,18 @@ def __init__(self, w=2, stride=2):
         self.w = w
         self.stride = stride
 
-    def para(self): return (self.stride,)
+    def para(self): return (self.w, self.stride)
 
     def forward(self, x):
         return maxpool(x, (self.w, self.w), (self.stride, self.stride))
 
+class GlobalAveragePool(Layer):
+    name = 'gap'
+    def __init__(self): pass    
+
+    def forward(self, x):
+        return x.mean(axis=(-2, -1))
+
 class UpSample(Layer):
     name = 'upsample'
     def __init__(self, k):
@@ -150,7 +157,7 @@ def load(self, buf):
 
 layerkey = {'dense':Dense, 'conv':Conv2d, 'relu':ReLU, 'batchnorm':BatchNorm,
     'flatten':Flatten, 'sigmoid':Sigmoid, 'softmax': Softmax,
-    'maxpool':Maxpool, 'upsample':UpSample, 'concat':Concatenate, 'add':Add}
+    'maxpool':Maxpool, 'upsample':UpSample, 'concat':Concatenate, 'add':Add, 'gap':GlobalAveragePool}
 
 if __name__ == "__main__":
     pass

demo/readme.md

@@ -12,3 +12,8 @@ Demo for edge detector with model HED. Run ```python main.py``` inside folder ``
 The detected result is shown as following: 
 ![](https://raw.githubusercontent.com/Image-Py/cnnumpy/master/demo/hed_edge_detector/rst.png)
 
+## Resnet18 trained on ImageNet
+Demo for image classification with resnet18. ```python main.py``` inside folder ```resnet18```. 2 files (same level folder with ```main.py```) are needed (res.txt, [res.npy]())
+The detected result is shown as following: 
+![](https://raw.githubusercontent.com/Image-Py/cnnumpy/master/demo/resnet18/rst.png)
+