search

sydsim / personlab-tf

implementation of PersonLab(https://arxiv.org/abs/1803.08225) using TF-slim
103 stars 19 forks source link

Conversion to TF lite is failing #8

Open sandeepjana opened 5 years ago

sandeepjana commented 5 years ago

First, thanks for detailed description of training steps. I trained the model for a few epochs.

My final goal is to obtain tflite model. So, I converted the ckpt model to pb by following this SO answer: https://stackoverflow.com/a/45868106/1595171

Then tried toco converter to convert the model to tflite, like this:

toco \ --graph_def_file=logs/samples/output_graph_122303.pb \ --output_file=output_graph_122303.tflite \ --output_format=TFLITE \ --input_shape=1,224,224,3 \ --input_array=input \ --output_array='split:0', 'split:1', 'mul_60:0', 'mul_63:0', 'mul_61:0', 'mul_64:0', 'mul_62:0', 'mul_65:0' \ --inference_type=FLOAT \ --inference_input_type=FLOAT

I got the following error: ValueError: Input 0 of node MobilenetV2/Conv/weights/Assign was passed float from MobilenetV2/Conv/weights:0 incompatible with expected float_ref.

I googled for quite sometime but nothing worked out for me.

sydsim commented 5 years ago

sorry but I can't help you because I don't know about the tf lite. but it seems that types float and float_ref doesn't match is the cause of the error.