model-r34-amf inferrence failed

[ccl-private]

Non-zero status code returned while running PRelu node. Name:'relu0' Status Message: relu0: right operand cannot broadcast on dim 0 LeftShape: {1,64,112,112}, RightShape: {64}

[ccl-private]

问题解决了～
`import cv2
import numpy as np
import os
import time

import mxnet as mx
from mxnet.contrib import onnx as onnx_mxnet
import onnx
from onnx import numpy_helper
import onnxruntime as ort
from onnx import helper
from onnx import TensorProto

#mxnet2onnx
def mxnet2onnx_test():
sym = './model-r34-amf/model-symbol.json'
params = './model-r34-amf/model-0000.params'

#NCHW
input_shape = [(1,3,112,112)]

onnx_file = './model-r34-amf/onnx/modelnew_onnx.onnx'

#返回转换后的onnx模型的路径
converted_model_path = onnx_mxnet.export_model(sym, params, input_shape, np.float, onnx_file)

#Check the model
onnx.checker.check_model(onnx.load(onnx_file))
print('The model is checked!')

#onnx model inferrence
def onnx_inferred_demo():
onnx_file = './model-r34-amf/onnx/modelnew2_onnx.onnx'

image_path='0.jpg'
img = cv2.imread(image_path)

img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)

#img=(img-127.5)
#img=img*0.0078125

img = np.transpose( img, (2,0,1) ) #HWC->CHW

ort_session = ort.InferenceSession(onnx_file)
input_name = ort_session.get_inputs()[0].name #'data'
outputs = ort_session.get_outputs()[0].name #'fc1'

input_blob = np.expand_dims(img, axis=0).astype(np.float32) #NCHW

out = ort_session.run([outputs], input_feed={input_name: input_blob})

print(out[0])

def createGraphMemberMap(graph_member_list):
member_map=dict();
for n in graph_member_list:
member_map[n.name]=n;
return member_map

#修改onnx的结构
def onnx_modify_demo():

onnx_file = './model-r34-amf/onnx/modelnew_onnx.onnx'

model = onnx.load(onnx_file)

print("load done~")

graph = model.graph
initializer_map = createGraphMemberMap(graph.initializer)
input_map = createGraphMemberMap(graph.input) #'data' 'scalar_op1' 'scalar_op2'
node_map = createGraphMemberMap(graph.node) #'_minusscalar0' '_mulscalar0'
output_map = createGraphMemberMap(graph.output) #'fc1'
#print(input_map)
#print(input_map.keys())

#====data, double to float====
print("data, double to float")
name='data'
#print(input_map[name])
graph.input.remove(input_map[name])
new_nv = helper.make_tensor_value_info(name, TensorProto.FLOAT, [1,3,112,112])
graph.input.extend([new_nv])
#input_map = createGraphMemberMap(graph.input)
#print(input_map[name])

#====Sub, double to float====
print("Sub, double to float")
#修改input的dtype
input_map = createGraphMemberMap(graph.input)
name='scalar_op2'
graph.input.remove(input_map[name])
new_nv = helper.make_tensor_value_info(name, TensorProto.FLOAT, [1])
graph.input.extend([new_nv])
#同时要修改initializer中初始值的type
initializer_map = createGraphMemberMap(graph.initializer)
graph.initializer.remove(initializer_map[name])
new_nv = helper.make_tensor(name, TensorProto.FLOAT, [1],[127.5])
graph.initializer.extend([new_nv])

#====Mul, double to float====
print("Mul, double to float")
input_map = createGraphMemberMap(graph.input)
name='scalar_op3'
graph.input.remove(input_map[name])
new_nv = helper.make_tensor_value_info(name, TensorProto.FLOAT, [1])
graph.input.extend([new_nv])

initializer_map = createGraphMemberMap(graph.initializer)
graph.initializer.remove(initializer_map[name])
new_nv = helper.make_tensor(name, TensorProto.FLOAT, [1],[0.0078125])
graph.initializer.extend([new_nv])

#====FC1, double to float====
print("FC1, double to float")
output_map = createGraphMemberMap(graph.output)
name='fc1'
graph.output.remove(output_map[name])
new_nv = helper.make_tensor_value_info(name, TensorProto.FLOAT, [1,512])
graph.output.extend([new_nv])


#===PReLU slop===
#C--->C*1*1
print("PReLU slop, C--->C*1*1")
for input_name in input_map.keys():
    if input_name.endswith('relu0_gamma') or input_name.endswith('relu1_gamma'):
        print(input_name)
        input_shape = input_map[input_name].type.tensor_type.shape.dim
        input_dim_val=input_shape[0].dim_value
        print(input_dim_val)
        
        graph.input.remove(input_map[input_name])
        new_nv = helper.make_tensor_value_info(input_name, TensorProto.FLOAT, [input_dim_val,1,1])
        graph.input.extend([new_nv])
        
        initializer_map = createGraphMemberMap(graph.initializer)
        graph.initializer.remove(initializer_map[input_name])
        weight_array = numpy_helper.to_array(initializer_map[input_name])
        print(weight_array.shape)
        weight_array=weight_array.astype(np.float) #np.float32与initializer中不匹配
        #===可以查看权重输出！！！
        
        b=[]
        for w in weight_array:
            b.append(w)
        new_nv = helper.make_tensor(input_name, TensorProto.FLOAT, [input_dim_val,1,1], b)
        graph.initializer.extend([new_nv])

for node in model.graph.node:
    if (node.op_type == "BatchNormalization"):
        for attr in node.attribute:
            if (attr.name == "spatial"):
                attr.i = 1

onnx.save(model, './model-r34-amf/onnx/modelnew2_onnx.onnx')

if name == 'main':

# mxnet2onnx_test() #==mxnet2onnx

# onnx_modify_demo() #===onnx修改===

onnx_inferred_demo() #===onnx前向推导===

`