Tensorflow Lite iOS Camera example does not work with retrained MobileNet model

Question

I'm trying to run the Tensorflow Lite Camera example with a retrained Mobilenet model.

I successfully run the iOS camera app according to the instructions and this fix . The app works as expected with the model mobilenet_v1_1.0_224.tflite .

I install Tensorflow:

pip3 install -U virtualenv
virtualenv --system-site-packages -p python3 ./venv
source ./venv/bin/activate
pip install --upgrade pip
pip install --upgrade tensorflow==1.12.0
pip install --upgrade tensorflow-hub==40.6.2

I now want to retrain the model using the flowers set . I download the flowers folder and run:

python retrain.py \
    --bottleneck_dir=bottleneck \
    --how_many_training_steps=400 \
    --model_dir=model \
    --output_graph=pola_retrained.pb \
    --output_labels=pola_retrained_labels.txt \
    --tfhub_module https://tfhub.dev/google/imagenet/mobilenet_v1_100_224/quantops/feature_vector/1 \
    --image_dir=flower_photos

Note: I can successfully test the retrained model using label_image.py script.

I convert the retrained model to its tflite format:

toco \
  --graph_def_file=pola_retrained.pb \
  --input_format=TENSORFLOW_GRAPHDEF \
  --output_format=TFLITE \
  --output_file=mobilenet_v1_1.0_224.tflite \
  --inference_type=FLOAT \
  --input_type=FLOAT \
  --input_arrays=Placeholder \
  --output_arrays=final_result \
  --input_shapes=1,224,224,3

I copy both the new model and the labels file to the iOS app. I modify the app parameters in CameraExampleViewController.mm as follows:

// These dimensions need to match those the model was trained with.
const int wanted_input_width = 224;
const int wanted_input_height = 224;
const int wanted_input_channels = 3;
const float input_mean = 128.0f;
const float input_std = 128.0f;
const std::string input_layer_name = "input";
const std::string output_layer_name = "final_result";

The app crashes. The index of the recognized object is outside of the range of trained objects . The confidence level is above 1.

Answer 1

The Tensorflow Lite Camera example hardcodes the output tensor size to 1000 . If you test the example with a retrained model having a smaller number of outputs, the iOS app crashes. Replace the following code in CameraExampleViewController.mm :

const int output_size = 1000;

with:

// read output size from the output sensor
const int output_tensor_index = interpreter->outputs()[0];
TfLiteTensor* output_tensor = interpreter->tensor(output_tensor_index);
TfLiteIntArray* output_dims = output_tensor->dims;
assert(output_dims->size == 2);
const int output_size = output_dims->data[1]-output_dims->data[0];

The code above fixes the problem by reading the output size from the model output dimensions. The appropriate PR has been submitted.