Tensorflow Object Detection - Convert .pb file to tflite

Question

I try to convert a frozen SSD mobilenet v2 model to TFLITE format for android usage. Here are all my steps:

1. I retrain with TF Object Detection API 's train.py file using the ssd_mobilenet_v2_coco_2018_03_29 model frok the model zoo. (OK)

2. Export the trained model.ckpt to frozen model file using export_inference_graph.p y also provided by TF Object Detection API. (OK)

3. Test the frozen graph in python with GPU and also with only CPU allowed. It works. (OK)

Here comes the downside, i try to use the following code:

import tensorflow as tf
tf.enable_eager_execution()
saved_model_dir = 'inference_graph/saved_model/'
converter = tf.contrib.lite.TFLiteConverter.from_saved_model(saved_model_dir,input_arrays=input_arrays,output_arrays=output_arrays,input_shapes={"image_tensor": [1, 832, 832, 3]})
converter.post_training_quantize = True

First I tried without adding input shapes parameter to the function, but it didn't work. Since that time I read that you can write there anything it doesn't matter.

The output till this line:

INFO:tensorflow:Saver not created because there are no variables in the graph to restore
INFO:tensorflow:The specified SavedModel has no variables; no checkpoints were restored.
INFO:tensorflow:The given SavedModel MetaGraphDef contains SignatureDefs with the following keys: {'serving_default'}
INFO:tensorflow:input tensors info: 
INFO:tensorflow:Tensor's key in saved_model's tensor_map: inputs
INFO:tensorflow: tensor name: image_tensor:0, shape: (-1, -1, -1, 3), type: DT_UINT8
INFO:tensorflow:output tensors info: 
INFO:tensorflow:Tensor's key in saved_model's tensor_map: num_detections
INFO:tensorflow: tensor name: num_detections:0, shape: (-1), type: DT_FLOAT
INFO:tensorflow:Tensor's key in saved_model's tensor_map: detection_boxes
INFO:tensorflow: tensor name: detection_boxes:0, shape: (-1, 100, 4), type: DT_FLOAT
INFO:tensorflow:Tensor's key in saved_model's tensor_map: detection_scores
INFO:tensorflow: tensor name: detection_scores:0, shape: (-1, 100), type: DT_FLOAT
INFO:tensorflow:Tensor's key in saved_model's tensor_map: detection_classes
INFO:tensorflow: tensor name: detection_classes:0, shape: (-1, 100), type: DT_FLOAT
INFO:tensorflow:Saver not created because there are no variables in the graph to restore
INFO:tensorflow:The specified SavedModel has no variables; no checkpoints were restored.
INFO:tensorflow:Froze 0 variables.
INFO:tensorflow:Converted 0 variables to const ops.

Then i wanna convert:

tflite_quantized_model = converter.convert()

This is the output:

---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
<ipython-input-6-61a136476642> in <module>
----> 1 tflite_quantized_model = converter.convert()

~/.local/lib/python3.5/site-packages/tensorflow/contrib/lite/python/lite.py in convert(self)
    451           input_tensors=self._input_tensors,
    452           output_tensors=self._output_tensors,
--> 453           **converter_kwargs)
    454     else:
    455       # Graphs without valid tensors cannot be loaded into tf.Session since they

~/.local/lib/python3.5/site-packages/tensorflow/contrib/lite/python/convert.py in toco_convert_impl(input_data, input_tensors, output_tensors, *args, **kwargs)
    340   data = toco_convert_protos(model_flags.SerializeToString(),
    341                              toco_flags.SerializeToString(),
--> 342                              input_data.SerializeToString())
    343   return data
    344 

~/.local/lib/python3.5/site-packages/tensorflow/contrib/lite/python/convert.py in toco_convert_protos(model_flags_str, toco_flags_str, input_data_str)
    133     else:
    134       raise RuntimeError("TOCO failed see console for info.\n%s\n%s\n" %
--> 135                          (stdout, stderr))
    136 
    137 

RuntimeError: TOCO failed see console for info.

I can't copy here the console output hence it's more than the 30000 character limit but here you can see it: https://pastebin.com/UyT2x2Vk

Please please help at this point, what should I do to make it work :(

My config: Ubuntu 16.04, Tensorflow-GPU 1.12

Thanks In andvance!

Answer 1

Had the same issue last week, resolved it by following the steps described here .

Basically the issue is that their main script does not support SSD models. I did not use bazel to do this, but the tflite_convert utility.

Careful with the export_tflite_ssd_graph.py script, read all its options before using it (mainly the --max_detections that saved my life).

Hope this helps.

Edit: Your step 2 is invalid. A saved_model cannot be converted to a tflite model if it contains a SSD. You need to export the trained model.ckpt using the export_tflite_ssd_graph.py script and use the .pb file created to convert it to tflite with the tflite_convert tool.

Answer 2

Your .pb file is not in right format. Here is the solution: https://github.com/peace195/tensorflow-lite-yolo-v3

We need do 2 steps:

1. Convert .weights to SavedModel.

2. Using tflite_convert to convert from saved model to tflite format.

Please use docker to setup environment and follow the instruction carefully.