io

Author: yxdragon

cnnumpy/__init__.py

@@ -1,2 +1,3 @@
 from .layer import *
-from .net import Net, read_net
+from .net import Net
+from .io import read_net, read_onnx

cnnumpy/io.py

@@ -0,0 +1,89 @@
+import json, re, numpy as np
+from .net import Net
+
+def read_net(path):
+	net = Net()
+	with open(path+'.lay') as f: lay = json.load(f)
+	with open(path+'.flw') as f: flw = json.load(f)
+	net.load_json(lay, flw)
+	net.load_weights(np.load(path+'.npy'))
+	return net
+
+def parse(matched):
+	gps = list(matched.groups())
+	if len(matched.groups())==0: return ''
+	for i in range(len(gps)):
+		if not '%' in gps[i]: continue
+		gps[i] = gps[i].replace('%',"'")
+		gps[i] = gps[i].replace(')',"')")
+		gps[i] = gps[i].replace(',',"',")
+	return str(gps)+'\n'
+
+pt = re.compile(r'.*%(.+?) .+(\(.+\))')
+conv = re.compile(r'.*%(.+?) .+?(Conv).+?strides=(\[\d+?, \d+?\]).+?(\(%.+?, %.+?, %.+?\)).+?\n')
+relu = re.compile(r'.*%(.+?) .+?(Relu)\(%(.+?)\).+?\n')
+sigmoid = re.compile(r'.*%(.+?) .+?(Sigmoid)\(%(.+?)\).+?\n')
+maxpool = re.compile(r'.*%(.+?) .+?(MaxPool).+?strides=(\[\d+?, \d+?\]).+?\(%(.+?)\).+?\n')
+upsample = re.compile(r'.*%.+? .+?Constant\[value=.+?(\d+\.?\d*) \[.+?\n.+?%(.+?) .+?(Upsample).+?\(%(.+?),.+?\n')
+flatten = re.compile(r'.*%.+?Constant.+?\n.+?Shape.+?\n.+?Gather.+?\n.+?Constant.+?\n.+?Unsqueeze.+?\n.+?Unsqueeze.+?\n.+?Concat.+?\n.+?%(.+?) .+?(Reshape)\(%(.+?),.+?\n')
+dense = re.compile(r'.*%(.+?) .+?(Gemm).+(\(%.+?, %.+?, %.+?\)).+?\n')
+concat = re.compile(r'.*%(.+?) .+?(Concat).+(\(%.+?\)).+?\n')
+add = re.compile(r'.*%(.+?) .+?(Add)(\(%.+?\)).+?\n')
+weight = re.compile(r'.*%(.+?) .+?(\(.+?\)).*\n')
+
+res = (flatten, upsample, conv, relu, sigmoid, maxpool, dense, concat, add, weight)
+
+def read_onnx(path):
+	with open(path+'.onnx') as f:
+		cont = f.read()
+	for i in res: cont = i.sub(parse, cont)
+	#for i in cont.split('\n'): print(i)
+	cont = [eval(i) for i in cont.split('\n') if len(i)>0 and i[0]=='[']
+	cont = [[eval(j) if ',' in j else j for j in i] for i in cont]
+
+	body = []
+	flow = []
+	key = {}
+	for i in cont:
+		if len(i)==2: key[i[0]] = i[1]
+		elif i[1]=='Conv':
+			num = len(body)
+			shp = [key[i[3][1]][j] for j in (1,0,2)] + [i[2][0]]
+			body.append(('conv_%s'%num, 'conv', shp))
+			flow.append((i[3][0], ['conv_%s'%num], i[0]))
+		elif i[1]=='Gemm':
+			num = len(body)
+			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)
+			body.append(('sigmoid_%s'%num, 'sigmoid', None))
+			flow.append((i[2], ['sigmoid_%s'%num], i[0]))
+		elif i[1]=='Relu':
+			num = len(body)
+			body.append(('relu_%s'%num, 'relu', None))
+			flow.append((i[2], ['relu_%s'%num], i[0]))
+		elif i[1]=='Add':
+			num = len(body)
+			body.append(('add_%s'%num, 'add', None))
+			flow.append((i[2], ['add_%s'%num], i[0]))
+		elif i[1]=='Concat':
+			num = len(body)
+			body.append(('concat_%s'%num, 'concat', None))
+			flow.append((i[2], ['concat_%s'%num], i[0]))
+		elif i[1]=='MaxPool':
+			num = len(body)
+			body.append(('maxpool_%s'%num, 'maxpool', [i[2][0]]))
+			flow.append((i[3], ['maxpool_%s'%num], i[0]))
+		elif i[2]=='Upsample':
+			num = len(body)
+			body.append(('upsample_%s'%num, 'upsample', [int(float(i[0]))]))
+			flow.append((i[3], ['upsample_%s'%num], i[1]))
+		elif i[1]=='Reshape':
+			num = len(body)
+			body.append(('flatten_%s'%num, 'flatten', None))
+			flow.append((i[2], ['flatten_%s'%num], i[0]))
+	net = Net()
+	net.load_json(body, flow)
+	net.load_weights(np.load(path+'.npy'))
+	return net

cnnumpy/net.py

@@ -1,6 +1,4 @@
 from .layer import layerkey as key
-import numpy as np
-import json
 
 class Net:
 	def __init__(self):
@@ -19,7 +17,7 @@ def forward(self, x):
 		for x, ls, y in self.cmds:
 			for l in ls:
 				out = x if l == ls[0] else y
-				if isinstance(out, list):
+				if not isinstance(out, str):
 					p = [rst[i] for i in out]
 				else: p = rst[out]
 				rst[y] = dic[l](p)
@@ -75,13 +73,5 @@ def load_weights(self, data):
 	def __call__(self, x):
 		return self.forward(x)
 
-def read_net(path):
-	net = Net()
-	with open(path+'.lay') as f: lay = json.load(f)
-	with open(path+'.flw') as f: flw = json.load(f)
-	net.load_json(lay, flw)
-	net.load_weights(np.load(path+'.npy'))
-	return net
-
 if __name__ == '__main__':
-	pass
+	pass