生成.tflite文件过程中遇到的问题及解决方案
问题一：
官方给出的例子1：
https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/lite/toco/g3doc/python_api.md#example-2-export-with-variables
import tensorflow as tf
img = tf.placeholder(name="img", dtype=tf.float32, shape=(1, 64, 64, 3))
val = img + tf.constant([1., 2., 3.]) + tf.constant([1., 4., 4.])
out = tf.identity(val, name="out")
with tf.Session() as sess:
tflite_model = tf.contrib.lite.toco_convert(sess.graph_def, [img], [out])
open("converteds_model.tflite", "wb").write(tflite_model)运行之后报错：
---------------------------------------------------------------------------
NameError Traceback (most recent call last)
1-003f6775a35f> in ()
4 out = tf.identity(val, name="out")
5 with tf.Session() as sess:
----> 6 tflite_model = tf.contrib.lite.toco_convert(sess.graph_def, [img], [out])
7 open("converteds_model.tflite", "wb").write(tflite_model)
~/.conda/envs/tensorflow/lib/python3.6/site-packages/tensorflow/contrib/lite/python/lite.py in toco_convert(input_data, input_tensors, output_tensors, inference_type, input_format, output_format, quantized_input_stats, drop_control_dependency)
210 data = toco_convert_protos(model.SerializeToString(),
211 toco.SerializeToString(),
--> 212 input_data.SerializeToString())
213 return data
214
~/.conda/envs/tensorflow/lib/python3.6/site-packages/tensorflow/contrib/lite/python/lite.py in toco_convert_protos(model_flags_str, toco_flags_str, input_data_str)
103 model_flags_str, toco_flags_str, input_data_str)
104
--> 105 with tempfile.NamedTemporaryFile() as fp_toco, \
106 tempfile.NamedTemporaryFile() as fp_model, \
107 tempfile.NamedTemporaryFile() as fp_input, \
NameError: name 'tempfile' is not defined 原因：It seems it has to do with the interface sealing in tensorflow(where we try to prevent exposing symbols unrelated to the interfaces)。
解决方案：manually put back imported modules：
import tensorflow as tf
#----------------Solution Begin----------------
# manually put back imported modules
import tempfile
import subprocess
tf.contrib.lite.tempfile = tempfile
tf.contrib.lite.subprocess = subprocess
#-----------------Solution End-----------------
img = tf.placeholder(name="img", dtype=tf.float32, shape=(1, 64, 64, 3))
val = img + tf.constant([1., 2., 3.]) + tf.constant([1., 4., 4.])
out = tf.identity(val, name="out")
with tf.Session() as sess:
tflite_model = tf.contrib.lite.toco_convert(sess.graph_def, [img], [out])
open("converteds_model.tflite", "wb").write(tflite_model)运行之后就能在工作目录下看到生成的converteds_model.tflite了。
问题二：
官方给出的例子2：
https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/lite/toco/g3doc/python_api.md#example-2-export-with-variables
import tensorflow as tf
img = tf.placeholder(name="img", dtype=tf.float32, shape=(1, 64, 64, 3))
var = tf.get_variable("weights", dtype=tf.float32, shape=(1,64,64,3))
val = img + var
def canonical_name(x):
return x.name.split(":")[0]
out = tf.identity(val, name="out")
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
out_tensors = [out]
frozen_graphdef = tf.graph_util.convert_variables_to_constants(
sess, sess.graph_def, map(canonical_name, out_tensors))
tflite_model = tf.contrib.lite.toco_convert(
frozen_graphdef, [img], out_tensors)
open("converted_model.tflite", "wb").write(tflite_model)运行之后报错：
---------------------------------------------------------------------------
TypeError Traceback (most recent call last)
3-f81c3f0d99d5> in ()
13 out_tensors = [out]
14 frozen_graphdef = tf.graph_util.convert_variables_to_constants(
---> 15 sess, sess.graph_def, map(canonical_name, out_tensors))
16 tflite_model = tf.contrib.lite.toco_convert(
17 frozen_graphdef, [img], out_tensors)
~/.conda/envs/tensorflow/lib/python3.6/site-packages/tensorflow/python/framework/graph_util_impl.py in convert_variables_to_constants(sess, input_graph_def, output_node_names, variable_names_whitelist, variable_names_blacklist)
230 # This graph only includes the nodes needed to evaluate the output nodes, and
231 # removes unneeded nodes like those involved in saving and assignment.
--> 232 inference_graph = extract_sub_graph(input_graph_def, output_node_names)
233
234 found_variables = {}
~/.conda/envs/tensorflow/lib/python3.6/site-packages/tensorflow/python/framework/graph_util_impl.py in extract_sub_graph(graph_def, dest_nodes)
174 _assert_nodes_are_present(name_to_node, dest_nodes)
175
--> 176 nodes_to_keep = _bfs_for_reachable_nodes(dest_nodes, name_to_input_name)
177
178 nodes_to_keep_list = sorted(
~/.conda/envs/tensorflow/lib/python3.6/site-packages/tensorflow/python/framework/graph_util_impl.py in _bfs_for_reachable_nodes(target_nodes, name_to_input_name)
138 nodes_to_keep = set()
139 # Breadth first search to find all the nodes that we should keep.
--> 140 next_to_visit = target_nodes[:]
141 while next_to_visit:
142 n = next_to_visit[0]
TypeError: 'map' object is not subscriptable
原因：笔者用的是Python3.6，直接用map(canonical_name, out_tensors))会报错
解决方案：将map(canonical_name, out_tensors))改为list(map(canonical_name, out_tensors)))
然后运行会报和例子1同样的错误，解决方法相同：
import tensorflow as tf
#----------------Solution Begin----------------
# manually put back imported modules
import tempfile
import subprocess
tf.contrib.lite.tempfile = tempfile
tf.contrib.lite.subprocess = subprocess
#-----------------Solution End-----------------
img = tf.placeholder(name="img", dtype=tf.float32, shape=(1, 64, 64, 3))
var = tf.get_variable("weights", dtype=tf.float32, shape=(1,64,64,3))
val = img + var
def canonical_name(x):
return x.name.split(":")[0]
out = tf.identity(val, name="out")
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
out_tensors = [out]
frozen_graphdef = tf.graph_util.convert_variables_to_constants(
#----------------Solution Begin----------------
sess, sess.graph_def, list(map(canonical_name, out_tensors)))
#-----------------Solution End-----------------
tflite_model = tf.contrib.lite.toco_convert(
frozen_graphdef, [img], out_tensors)
open("converted_model.tflite", "wb").write(tflite_model)问题三：.tflite转换失败
在将训练好的模型转换为.tflite格式要经过两个步骤，首先是进行freeze_graph转换为.pb格式，再进行toco转换为.tflite格式。第一步进行freeze的时候不管是使用训练的graph.pbtxt还是使用测试用的eval.pbtxt都是可以转换成功的，因此一般进行freeze_graph的时候是不会报错的，错误会在第二步使用toco将.pb格式的模型转换为.tflite格式的模型的时候暴露出来。
使用toco将.pb格式的模型转换为.tflite格式的模型可能出现的错误形式如图1-3所示：
图1 toco转换失败-错误形式1
图2 toco转换失败-错误形式2
图3 toco转换失败-错误形式3
原因：
在训练好对模型（目前tensorflow训练生成模型格式是V2版本的，其形式如图4所示）进行转换的时候流程不对，直接使用了训练用的graph.pbtxt。正确的做法是先生成evaluate时用的eval.pbtxt，然后进行bazel-bin/tensorflow/python/tools/freeze_graph，最后再进行bazel-bin/tensorflow/contrib/lite/toco/toco。专门介绍从数据准备，到训练，到模型转换，再到部署到手机的整个流程的详见Tensorflow部署到移动端。
图4
更详细介绍错误产生过程
mobilenet_v1
freeze_graph，下面的graph.pbtxt是train的时候生成的，用它来进行freeze是不会报错的，如图5上半部分所示：
bazel-bin/tensorflow/python/tools/freeze_graph \
--input_graph=/home/lg/Desktop/mobilenet_v1/checkpoint/graph.pbtxt \
--input_checkpoint=/home/lg/Desktop/mobilenet_v1/checkpoint/model.ckpt-20000 \
--input_binary=false \
--output_graph=/home/lg/Desktop/mobilenet_v1/frozen_mobilenet_v1_244.pb \
--output_node_names=MobilenetV1/Predictions/Reshape_1 \
--checkpoint_version=2toco，企图利用graph.pbtxt freeze之后的.pb文件转换为.tflite文件：
企图转换为FLOAT的.tflite，报错如图5下半部分所示：
bazel-bin/tensorflow/contrib/lite/toco/toco \
--input_file=/home/lg/Desktop/mobilenet_v1/frozen_mobilenet_v1_244.pb \
--input_format=TENSORFLOW_GRAPHDEF \
--output_format=TFLITE \
--output_file=/home/lg/Desktop/mobilenet_v1/frozen_graph_mobilenet_v1.tflite \
--inference_type=FLOAT \
--input_type=FLOAT \
--input_arrays='shuffle_batch/(shuffle_batch)' \
--output_arrays=MobilenetV1/Predictions/Reshape_1 \
--input_shapes=1,244,244,3企图转换为QUANTIZED_UINT8的.tflite，报错如图6所示：
bazel-bin/tensorflow/contrib/lite/toco/toco \
--input_file=/home/lg/Desktop/mobilenet_v1/frozen_mobilenet_v1_244.pb --input_format=TENSORFLOW_GRAPHDEF \
--output_format=TFLITE \
--output_file=/home/lg/Desktop/mobilenet_v1/frozen_graph_mobilenet_v1.tflite --inference_type=QUANTIZED_UINT8 \
--input_type=QUANTIZED_UINT8 \
--input_arrays='shuffle_batch/(shuffle_batch)' \
--output_arrays=MobilenetV1/Predictions/Reshape_1 \
--input_shapes=1,244,244,3
图5 错误的利用train的graph.pbtxt进行freeze_graph、toco
图6 错误的利用train的graph.pbtxt进行freeze_graph、toco
正确的做法
首先产生eval.pbtxt才能保证后续的freeze_graph、toco的正确。如何产生eval.pbtxt详见Tensorflow部署到移动端。
利用eval.pbtxt进行正确的freeze graph：
freeze_graph
bazel-bin/tensorflow/python/tools/freeze_graph \
--input_graph=/home/lg/Desktop/mobilenet_v1/mobilenet_v1_eval.pbtxt \
--input_checkpoint=/home/lg/Desktop/mobilenet_v1/checkpoint/model.ckpt-20000 \
--input_binary=false \
--output_graph=/home/lg/Desktop/mobilenet_v1/frozen_mobilenet_v1_224.pb \
--output_node_names=MobilenetV1/Predictions/Reshape_1 \
--checkpoint_version=2toco
利用eval.pbtxt进行正确的freeze graph之后生成正确的.pb文件，再对.pb文件进行toco：
保持FLOAT类型的toco操作：
bazel-bin/tensorflow/contrib/lite/toco/toco \
--input_file=/home/lg/Desktop/mobilenet_v1/frozen_mobilenet_v1_224.pb \
--input_format=TENSORFLOW_GRAPHDEF \
--output_format=TFLITE \
--output_file=/home/lg/Desktop/mobilenet_v1/frozen_graph_mobilenet_v1-FLOAT.tflite \
--inference_type=FLOAT \
--input_type=FLOAT \
--input_arrays=Placeholder \
--output_arrays=MobilenetV1/Predictions/Reshape_1 \
--input_shapes=1,224,224,3进行量化为QUANTIZED_UINT8的toco操作：
bazel-bin/tensorflow/contrib/lite/toco/toco \
--input_file=/home/lg/Desktop/mobilenet_v1/frozen_mobilenet_v1_224.pb \
--input_format=TENSORFLOW_GRAPHDEF \
--output_format=TFLITE \
--output_file=/home/lg/Desktop/mobilenet_v1/frozen_graph_mobilenet_v1-QUANTIZED_UINT8.tflite \
--inference_type=QUANTIZED_UINT8 \
--input_type=QUANTIZED_UINT8 \
--input_arrays=Placeholder \
--output_arrays=MobilenetV1/Predictions/Reshape_1 \
--input_shapes=1,224,224,3 \
--default_ranges_min=0.0 \
–default_ranges_max=255.0Inception_v3
同样的，下面的graph.pbtxt也是train的时候生成的，用它来进行freeze也是不会报错的：
freeze_graph
bazel-bin/tensorflow/python/tools/freeze_graph \
--input_graph=/home/lg/Desktop/inception_v3/checkpoint/graph.pbtxt \
--input_checkpoint=/home/lg/Desktop/inception_v3/checkpoint/model.ckpt-20000 \
--input_binary=false \
--output_graph=/home/lg/Desktop/inception_v3/frozen_inception_v3_299.pb \
--output_node_names=InceptionV3/Predictions/Reshape_1 \
--checkpoint_version=2toco，企图利用graph.pbtxt freeze之后的.pb文件转换为.tflite文件：
企图转换为FLOAT的.tflite，报错如图7所示：
bazel-bin/tensorflow/contrib/lite/toco/toco \
--input_file=/home/lg/Desktop/frozen_inception_v3_299.pb \
--input_format=TENSORFLOW_GRAPHDEF \
--output_format=TFLITE \
--output_file=/home/lg/Desktop/frozen_graph_inception_v3.tflite --inference_type=FLOAT \
--input_type=FLOAT \
--input_arrays=input \
--output_arrays=InceptionV3/Predictions/Reshape_1 \
--input_shapes=1,299,299,3
图7 错误的利用train的graph.pbtxt进行freeze_graph、toco
企图转换为QUANTIZED_UINT8 的.tflite，报错如图8所示：
bazel-bin/tensorflow/contrib/lite/toco/toco \
--input_file=/home/lg/Desktop/inception_v3/frozen_inception_v3_299.pb \
--input_format=TENSORFLOW_GRAPHDEF \
--output_format=TFLITE \
--output_file=/home/lg/Desktop/inception_v3/frozen_graph_inception_v3.tflite \
--inference_type=QUANTIZED_UINT8 \
--input_type=QUANTIZED_UINT8 \
--input_arrays=input \
--output_arrays=InceptionV3/Predictions/Reshape_1 \
--input_shapes=1,299,299,3
图8 错误的利用train的graph.pbtxt进行freeze_graph、toco
正确的做法
首先产生eval.pbtxt才能保证后续的freeze_graph、toco的正确。如何产生eval.pbtxt详见Tensorflow部署到移动端。
利用eval.pbtxt进行正确的freeze graph：
freeze_graph
bazel-bin/tensorflow/python/tools/freeze_graph \
--input_graph=/home/lg/Desktop/inception_v3/inception_v3_eval.pbtxt \
--input_checkpoint=/home/lg/Desktop/inception_v3/checkpoint/model.ckpt-20000 \
--input_binary=false \
--output_graph=/home/lg/Desktop/inception_v3/frozen_inception_v3_299.pb \
--output_node_names=InceptionV3/Predictions/Reshape_1 \
--checkpoint_version=2toco
利用eval.pbtxt进行正确的freeze graph之后生成正确的.pb文件，再对.pb文件进行toco：
保持FLOAT类型的toco操作：
bazel-bin/tensorflow/contrib/lite/toco/toco \
--input_file=/home/lg/Desktop/inception_v3/frozen_inception_v3_299.pb \
--input_format=TENSORFLOW_GRAPHDEF \
--output_format=TFLITE \
--output_file=/home/lg/Desktop/inception_v3/frozen_graph_inception_v3.tflite \
--inference_type=FLOAT \
--input_type=FLOAT \
--input_arrays=Placeholder \
--output_arrays=InceptionV3/Predictions/Reshape_1 \
--input_shapes=1,299,299,3进行量化为QUANTIZED_UINT8的toco操作：
bazel-bin/tensorflow/contrib/lite/toco/toco \
--input_file=/home/lg/Desktop/inception_v3/frozen_inception_v3_299.pb \
--input_format=TENSORFLOW_GRAPHDEF \
--output_format=TFLITE \
--output_file=/home/lg/Desktop/inception_v3/frozen_graph_inception_v3.tflite \
--inference_type=QUANTIZED_UINT8 \
--input_type=QUANTIZED_UINT8 \
--input_arrays=Placeholder \
--output_arrays=InceptionV3/Predictions/Reshape_1 \
--input_shapes=1,299,299,3 \
--default_ranges_min=0.0 \
--default_ranges_max=255.0参考文献
- https://github.com/tensorflow/tensorflow/issues/15410
- https://blog.csdn.net/liugan528/article/details/80044455
更多资料请移步github：
https://github.com/GarryLau