本文是对TVM官方关于如何使用TVM编译TensorFlow模型文档的翻译整理,并记录了实现时遇到的小坑。
TVM部署TensorFlow模型
本文介绍如何使用TVM部署TensorFlow模型。
在开始之前,首先需要安装TensorFlow的Python包。
Python及TensorFlow环境
以下例程需要Python3.5以上的环境才能运行,请使用3.5以上Python的版本。
我在使用Python3.6.5与TensorFlow1.14.0运行例程时遇到如下报错:
Traceback (most recent call last):
File "from_tensorflow.py", line 129, in <module>
shape=shape_dict)
File "/home/$USER/local/tvm/python/tvm/relay/frontend/tensorflow.py", line 2413, in from_tensorflow
mod, params = g.from_tensorflow(graph, layout, shape, outputs)
File "/home/$USER/local/tvm/python/tvm/relay/frontend/tensorflow.py", line 2058, in from_tensorflow
op = self._convert_operator(node.op, inputs, attr, graph)
File "/home/$USER/local/tvm/python/tvm/relay/frontend/tensorflow.py", line 2376, in _convert_operator
sym = convert_map[op_name](inputs, attrs, self._params)
File "/home/$USER/local/tvm/python/tvm/relay/frontend/tensorflow.py", line 562, in _impl
extras={'method': "BILINEAR"})(inputs, attr)
File "/home/$USER/local/tvm/python/tvm/relay/frontend/tensorflow.py", line 155, in __call__
return _get_relay_op(op_name)(*inputs, **new_attrs)
TypeError: resize() got an unexpected keyword argument 'half_pixel_centers'
回退TensorFlow版本至1.12.0后报错消失。TVM社区关于本问题的探讨:https://discuss.tvm.ai/t/typeerror-when-running-the-from-tensorflow-example/3046
关于TensorFlow的安装请参考https://www.tensorflow.org/install
此处通过pip进行安装即可:
pip3 install TensorFlow==1.12.0
# 导入 tvm, relay
import tvm
from tvm import relay
# 导入 os and numpy
import numpy as np
import os.path
# 导入 Tensorflow imports
import tensorflow as tf
# Tensorflow 效用函数
import tvm.relay.testing.tf as tf_testing
# 相关文件的在线地址(此处使用了dmlc在GitHub上的数据)
repo_base = 'https://github.com/dmlc/web-data/raw/master/tensorflow/models/InceptionV1/'
# 测试用图
img_name = 'elephant-299.jpg'
image_url = os.path.join(repo_base, img_name)
教程
有关TensorFlow的各种模型的更多详细信息,请参阅 docs/frontend/tensorflow.md 。
model_name = 'classify_image_graph_def-with_shapes.pb'
model_url = os.path.join(repo_base, model_name)
# 图像标签
map_proto = 'imagenet_2012_challenge_label_map_proto.pbtxt'
map_proto_url = os.path.join(repo_base, map_proto)
# 可读的图像标签
label_map = 'imagenet_synset_to_human_label_map.txt'
label_map_url = os.path.join(repo_base, label_map)
# 目标设置
# 如果使用cuda,可以使用以下推荐配置。
#target = 'cuda'
#target_host = 'llvm'
#layout = "NCHW"
#ctx = tvm.gpu(0)
target = 'llvm'
target_host = 'llvm'
layout = None
ctx = tvm.cpu(0)
下载所需文件
下列程序将下载上面所列的所需文件
from tvm.contrib.download import download_testdata
img_path = download_testdata(image_url, img_name, module='data')
model_path = download_testdata(model_url, model_name, module=['tf', 'InceptionV1'])
map_proto_path = download_testdata(map_proto_url, map_proto, module='data')
label_path = download_testdata(label_map_url, label_map, module='data')
导入模型
从protobuf文件创建TensorFlow图定义
with tf.gfile.FastGFile(model_path, 'rb') as f:
graph_def = tf.GraphDef()
graph_def.ParseFromString(f.read())
graph = tf.import_graph_def(graph_def, name='')
# 调用效用函数,将图定义导入默认图。
graph_def = tf_testing.ProcessGraphDefParam(graph_def)
# 向图增加shape。
with tf.Session() as sess:
graph_def = tf_testing.AddShapesToGraphDef(sess, 'softmax')
图像解码
官方注解
TensorFlow前端导入不支持JpegDecode等处理操作,所以我们绕过JpegDecode(只返回源节点)。因此,我们需要向TVM提供已解码的帧。
from PIL import Image
image = Image.open(img_path).resize((299, 299))
x =