尝试使用toco转换TF模型时出错

Question

I have a TF model which I need to convert to int8 for implementation in an embedded device. I am attempting to use toco to convert it, yet I am having trouble determining how to instantiate toco with the input args properly:

<USER>:~/.local/bin$ python3 toco \
>   --input_file=<PATH>/frozen_graph.pb \
>   --output_file=<PATH>/tflite_graph.tflite \
>   --input_format=TENSORFLOW_GRAPHDEF \
>   --output_format=TFLITE \
>   --inference_type=QUANTIZED_UINT8 \
>   --output_arrays=Sidmoid \
>   --input_arrays=IteratorGetNext \
>   --input_shapes=1:16:16:1 \
>   --mean_values=128 \
>   --std_values=127
WARNING:tensorflow:From <USER>.local/lib/python3.5/site-packages/tensorflow/contrib/learn/python/learn/datasets/base.py:198: retry (from tensorflow.contrib.learn.python.learn.datasets.base) is deprecated and will be removed in a future version.
Instructions for updating:
Use the retry module or similar alternatives.
2018-04-27 11:40:33.699249: F tensorflow/contrib/lite/toco/model_cmdline_flags.cc:240] Check failed: mean_values.size() == model_flags->input_arrays_size() 
Aborted (core dumped)

I understand there is an issue with the mean_values input arg, however I don't understand the intended usage of:

--mean_values="" string mean_values parameter for image models, comma-separated list of doubles, used to compute input activations from input pixel data. Each entry in the list should match an entry in --input_arrays.

I have tried also ",128,128," , as my interpretation is that you wouldn't want a mean value for your batch size/channel dims, but the documentation says that it needs an input for each input in the list should match an element in input_arrays.

I am also unsure what to put for the input_arrays arg, as most places I've seen code examples just use input but to me it seems you'd need the name of the node in the graph.

I think I'm misunderstanding something here, but I just can't seem to figure it out. Any help would be appreciated!

Answer 1

For anyone wondering, it was the documentation that slightly confused me: It says that "Input_shapes" is a string of "Shapes corresponding to --input_arrays, colon separated..." however the dimensions should still be comma separated:

bazel-bin/third_party/tensorflow/contrib/lite/toco/toco \
  --input_file=<my_path>/frozen_eval_graph.pb \
  --output_file=<my_path>/tflite_graph.tflite \
  --input_format=TENSORFLOW_GRAPHDEF \
  --output_format=TFLITE \
  --inference_type=QUANTIZED_UINT8 \
  --output_arrays=Sigmoid \
  --input_arrays=Reshape \
  --input_shapes=1,16,16,1 \
  --mean_values=128 \
  --std_values=127 \
  --variable_batch=true