add keras model to tf pb model for opencv dnn

tools/keras_to_pb.py

@@ -0,0 +1,143 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+Created on Wed Sep  4 00:00:43 2019
+keras to pd for opencv dnn 
+@author: chineseocr
+"""
+
+
+import tensorflow as tf
+import os 
+from keras import backend as K
+from tensorflow.python.framework import graph_util,graph_io
+def keras_to_pb(kerasmodel,outputDir,modelName='model.pd',outName = "output_"):
+    if not os.path.exists(outputDir):
+        os.makedirs(outputDir)
+    out_nodes = []
+    for i in range(len(kerasmodel.outputs)):
+        out_nodes.append(outName + str(i + 1))
+        tf.identity(kerasmodel.outputs[i],outName + str(i + 1))
+    sess = K.get_session()
+    init_graph = sess.graph.as_graph_def()
+    main_graph = graph_util.convert_variables_to_constants(sess,init_graph,out_nodes)
+    graph_io.write_graph(main_graph,outputDir,name = modelName,as_text = False)
+
+
+def pd_to_pbtxt(pdPath):
+    with tf.gfile.FastGFile(pdPath,'rb') as f:
+       graph_def = tf.GraphDef()
+       graph_def.ParseFromString(f.read())
+
+    for i in reversed(range(len(graph_def.node))):
+        if graph_def.node[i].op == 'Const':
+            del graph_def.node[i]
+        for attr in ['T', 'data_format', 'Tshape', 'N', 'Tidx', 'Tdim',
+                     'use_cudnn_on_gpu', 'Index', 'Tperm', 'is_training',
+                     'Tpaddings']:
+            if attr in graph_def.node[i].attr:
+                del graph_def.node[i].attr[attr]
+
+    path,filename = os.path.split(pdPath)
+    filename = filename.replace('.pb','.pbtxt')
+    tf.train.write_graph(graph_def,path, filename, as_text=True) 
+
+
+def remove_node(txt,name):
+    index = txt.find('node {\n  '+ name )
+    ind = index
+    punc = []
+    flag=False
+    if index>=0:
+        for i in range(ind,len(txt)):
+                if txt[i]=='{':
+                    punc.append('{')
+                elif txt[i]=='}':
+                        if '{' in punc:
+                            punc.pop(-1)
+                        if len(punc)==0:
+                            flag=True
+                            break
+
+
+
+    if flag:
+        txt=txt.replace(txt[index:i+1],'')
+
+    return txt
+
+
+
+if __name__=='__main__':
+    ## demo vgg16 to pb 
+
+    """
+    Open model.pbtxt and remove nodes with names strided_slice,flatten/Shape, flatten/strided_slice, flatten/Prod, flatten/stack.
+    Replace the node
+    node {
+      name: "flatten/Reshape"
+      op: "Reshape"
+      input: "block5_pool/MaxPool"
+      input: "flatten/stack"
+    }
+    
+    on
+    
+    node {
+      name: "flatten/Reshape"
+      op: "Flatten"
+      input: "block5_pool/MaxPool"
+    }
+    
+    """
+    def pbtxt_adjust(pbtxt):
+        with open(pbtxt) as f:
+            txt = f.read()
+
+        nodename = 'node {\n  name: "flatten/Reshape"'
+        ## replace 
+        index = txt.find(nodename)
+        if index>0:
+            for ind in range(index,len(txt)):
+                if txt[ind]=='}':
+                   break
+            replacestr = txt[index:ind+1]
+            txt=txt.replace(replacestr,replacestr.replace('op: "Reshape"','op: "Flatten"'))
+
+
+        ## del node 
+        delnamelist = [ 'name: "flatten/Shape"',
+                       'name: "flatten/strided_slice"',
+                       'name: "flatten/Prod"',
+                       'name: "flatten/stack"'
+                       ]
+
+        for delname in delnamelist:
+            txt = remove_node(txt,delname)
+
+        with open(pbtxt,'w') as f:
+            f.write(txt)
+
+
+    from keras.applications.vgg16 import VGG16
+    import cv2
+    import numpy as np
+    vgg = VGG16(weights=None)
+    name = vgg.name
+    modelName = name+'.pb'
+    outputDir=os.path.join('/tmp/',name)
+    keras_to_pb(vgg,outputDir,modelName=modelName,outName = "output_")
+    pb = os.path.join(outputDir,modelName)
+    pd_to_pbtxt(pb)
+    pbtxt = os.path.join(outputDir,name+'.pbtxt')
+    pbtxt_adjust(pbtxt)
+
+    dnn = cv2.dnn.readNetFromTensorflow(pb,pbtxt)
+    inputBlob = np.zeros((1,3,224,224))
+    dnn.setInput(inputBlob)
+    pred = dnn.forward()
+    print('dnn:',pred[0][:10])
+    print('vgg:',vgg.predict(np.zeros((1,224,224,3)))[0][:10])
+
+
+