一、简介
海思平台算子精度比对工具所依赖的标杆数据生成环境搭建指导文档，基于ubuntu操作系统，目前仅支持 Caffe 。
二、环境准备
1.caffe安装
参考：
- caffe 推理场景dump脚本使用方法
（1）功能说明
用于获取caffe的运行结果，可以获取每一层的结果，输入需要是bin文件。
示例1：dump模型每个节点的npy结果文件
python3.7 caffe_dump.py -m resnet50.prototxt -w resnet50.caffemodel -i test.bin -n 'data' -o ./output_d
其中 -n 指定的名称为模型的实际输入名称，-i 指定的输入文件需要与输入名称一一对应，且大小以及数据类型相匹配。
示例1：npy格式的dump文件会生成到 -o 指定的路径中，如下图：
caffe_dump.py
# coding=utf-8
import caffe
import sys
import argparse
import os
import caffe.proto.caffe_pb2 as caffe_pb2
import google.protobuf.text_format
import json
import numpy as np
import time
TIME_LENGTH = 1000
FILE_PERMISSION_FLAG = 0o600
class CaffeProcess:
def __init__(self):
parse = argparse.ArgumentParser()
parse.add_argument("-w", dest="weight_file_path",
help="<Required> the caffe weight file path",
required=True)
parse.add_argument("-m", dest="model_file_path",
help="<Required> the caffe model file path",
required=True)
parse.add_argument("-o", dest="output_path", help="<Required> the output path",
required=True)
parse.add_argument("-i", "--input_bins", dest="input_bins", help="input_bins bins. e.g. './a.bin;./c.bin'",
required=True)
parse.add_argument("-n", "--input_names", dest="input_names",
help="input nodes name. e.g. 'input_0;input_1'",
required=True)
args, _ = parse.parse_known_args(sys.argv[1:])
self.weight_file_path = os.path.realpath(args.weight_file_path)
self.model_file_path = os.path.realpath(args.model_file_path)
self.input_bins = args.input_bins.split(";")
self.input_names = args.input_names.split(";")
self.output_path = os.path.realpath(args.output_path)
self.net_param = None
self.cur_layer_idx = -1
@staticmethod
def _check_file_valid(path, is_file):
if not os.path.exists(path):
print('Error: The path "' + path + '" does not exist.')
exit(-1)
if is_file:
if not os.path.isfile(path):
print('Error: The path "' + path + '" is not a file.')
exit(-1)
else:
if not os.path.isdir(path):
print('Error: The path "' + path + '" is not a directory.')
exit(-1)
def _check_arguments_valid(self):
self._check_file_valid(self.model_file_path, True)
self._check_file_valid(self.weight_file_path, True)
self._check_file_valid(self.output_path, False)
for input_file in self.input_bins:
self._check_file_valid(input_file, True)
@staticmethod
def calDataSize(shape):
dataSize = 1
for dim in shape:
dataSize *= dim
return dataSize
def _load_inputs(self, net):
inputs_map = {}
for layer_name, blob in net.blobs.items():
if layer_name in self.input_names:
input_bin = np.fromfile(
self.input_bins[self.input_names.index(layer_name)], np.float32)
input_bin_shape = blob.data.shape
if self.calDataSize(input_bin_shape) == self.calDataSize(input_bin.shape):
input_bin = input_bin.reshape(input_bin_shape)
else:
print("Error: input node data size %d not match with input bin data size %d.", self.calDataSize(
input_bin_shape), self.calDataSize(input_bin.shape))
exit(-1)
inputs_map[layer_name] = input_bin
return inputs_map
def process(self):
"""
Function Description:
process the caffe net, save result as dump data
"""
# check path valid
self._check_arguments_valid()
# load model and weight file
net = caffe.Net(self.model_file_path, self.weight_file_path,
caffe.TEST)
inputs_map = self._load_inputs(net)
for key, value in inputs_map.items():
net.blobs[key].data[...] = value
# process
net.forward()
# read prototxt file
net_param = caffe_pb2.NetParameter()
with open(self.model_file_path, 'rb') as model_file:
google.protobuf.text_format.Parse(model_file.read(), net_param)
for layer in net_param.layer:
name = layer.name.replace("/", "_").replace(".", "_")
index = 0
for top in layer.top:
data = net.blobs[top].data[...]
file_name = name + "." + str(index) + "." + str(
round(time.time() * 1000000)) + ".npy"
output_dump_path = os.path.join(self.output_path, file_name)
np.save(output_dump_path, data)
os.chmod(output_dump_path, FILE_PERMISSION_FLAG)
print('The dump data of "' + layer.name
+ '" has been saved to "' + output_dump_path + '".')
index += 1
if __name__ == "__main__":
caffe_process = CaffeProcess()
caffe_process.process()