TUM-Lens

TUM-Lens is a Android application that runs image classification and object detection models locally on the device for evaluation and showcasing purposes.

App is available in the Google Playstore here, where we have roughly 75 Downloads.

Screenshots

Classification	Details and Settings	Detection

---

Features

• Run classification and object detection on live camera stream
• Choose from multiple models
• Double tap to pause classification

---

Installing and Running TUM-Lens on your Android smartphone

• Install it regularly from Google Playstore

• Activate Developer Mode on your Android Device

  • Open the Settings app.
  • If you run Android 9 or higher, go to About Phone and scroll down to Build Number. Otherwise go to System, then *About Phone+ and scroll down to Build Number.
  • Now, tap 7 times on Build Number.
  • If you run into problems you can find additional information here: https://developer.android.com/studio/debug/dev-options

• Clone Repository and open it in Android Studio

  • Clone the project with git clone ... inside a local repository or use GitLab's Download Source Code option.
  • In Android Studio, choose File, New and Import Project ... and select the TUM_lens directory. Make sure that you do not select the root of this repository!

• Connect your phone via cable. It should appear under running devices within the drop down in the upper menu bar of Android Studio.

• Run the app Build and run the app by clicking the green play button in the upper menu bar.

---

Adding models

As a tool for rapid prototyping, TUM-Lens is set up in such a way that additional classification models can be added without changing the source code. However, the project must be re-build.

1. Place your .tflite files in the /assets directory.
2. Open nets.json in /assets.
3. Find the "nets": [ ... ] array in that file.
4. Add your model to that array. You may want to use an existing object for reference. Please note the section below on mandatory parameters and precautions.
5. Now, save the file and re-build the project.

Please note that model file names must not contain any . except for separating filename and extension. Otherwise, models will be compressed resulting in the app crashing!

Mandatory JSON Parameters

Parameter	Use
name	display name in model selector
top5accuracy	displayed as info in model selector
image_mean	float models require additional normalization in image pre-processing. Use 127.5f for float, but 0.0 for quantized networks to bypass the normalization.
image_std	use 127.5 for float and 1.0 for quantized networks.
probability_mean	float models don't need dequantization in post-processing, set to 0.0 to bypass normalization. Use 0.0 for quantized models as well
probability_std	use 1.0 for float and 255.0 for quantized networks
filename	name of .tflite model file in /assets
labels	name of .txt labels file in /assets

---

Model Zoo

You can find model that were already converted into the TF-Lite Model Format in the model_zoo directory. E-Mail me to add your model to that list :)

Model	Description
Inception V1	default Inception V1 net; converted for proof of concept
MobileNet V3 IIC	Transfer Learning use case for Intel Image Classification Kaggle Challenge; based on this repository; see section below for more details;
efficientnet-lite0-fp32	Float EfficientNet from TensorFlow Hub
efficientnet-lite0-int8	Quantized EfficientNet from TensorFlow Hub
mobilenet_v1_224	Float MobileNet V1 from TensorFlow Hub; Please note that this model is already integrated into the app by default!
mobilenet_v1_224_quant	Quantized MobileNet V1 from TensorFlow Hub

Find more image classification models for TF-Lite on TensorFlow Hub

---

Converting (Transfer Learning) Models to TF-Lite

We look at the procedure using the following example: https://gitlab.com/mircotroue/slim-transfer-learning/.

This Transfer Learning project is based around TF-Slim's MobileNet V3 Large, that has previously been trained on ImageNet. The new domain consists out of six classes:

1. buildings
2. forest
3. glacier
4. mountain
5. sea
6. street

For that, two adjustments were made to the code:

1. nets/mobilenet/mobilenet_v3.py: num_classes=6 (see here)
2. datasets/natural.py: _NUM_CLASSES = 6 (see here)

1 Exporting the Inference Graph

The default TF-Slim repository includes export_inference_graph.py. Place this file in the root directory of the cloned repository and run:

python export_inference_graph.py \
	--alsologtostderr \
	--model_name=mobilenet_v3_large \
	--dataset_name=natural \
	--output_file=/tmp/transferlearning_inferencegraph.pb

Please note

For the adjustment of the number of classes to take effect, we have to specifiy both, the model structure and the dataset used.

This section is based on the TF-Slim documentation.

2 Checkpoint Files

The model/ directory contains checkpoint files at different time steps, that we can use for our purpose.

Please note

While there are three types of .ckpt files, just use the common prefix for --input_checkpoint. TensorFlow will find the correct *.data* file by itself. For example: given model.ckpt-10000.meta, model.ckpt-10000.index and model.ckpt-10000.data-00000-of-00001, use model.ckpt-10000.

3 Freezing the Graph

Freezing the inference graph is the process of merging checkpoints [aka training] with GraphDef [aka model structure] into a single file.

Preparation

In case you don't know the name of your models output nodes, use TensorFlow's summarize_graph tool. It is one of the targets of the initial bazel build. You will find it here: tensorflow-master/bazel-bin/tensorflow/tools/graph_transform. Usage:

./summarize_graph --in_graph="tmp/transferlearning_inferencegraph.pb"
# [...] MobileNet/Predictions/Reshape_1

In case of MobileNet V3 Large, that we use here, we can just rely on the model description, found in the upper part of nets/mobilenet/mobilenet_v3.py

We now use the freeze_graph tool. If you pip installed TensorFlow, it should be automatically be available in your CLI.

freeze_graph \
  --input_graph=/tmp/transferlearning_inferencegraph.pb \
  --input_checkpoint=/tmp/model.ckpt-10000 \
  --input_binary=true \
  --output_graph=/tmp/frozen.pb \
  --output_node_names=MobileNet/Predictions/Reshape_1

4 Converting to TF-Lite

Please note

This part currently relies on TensorFlow v1.

Create a new python file, e.g. tflite_converter.py. Paste the following few lines:

import tensorflow as tf

# convert
converter = tf.compat.v1.lite.TFLiteConverter.from_frozen_graph(
    graph_def_file='../../../../../tmp/frozen.pb',
    input_arrays=['input'],
    input_shapes={'input' : [1, 224, 224, 3]},
    output_arrays=['MobileNet/Predictions/Reshape_1']
)
tflite_model = converter.convert()

# save
with open('transferlearning_model.tflite', 'wb') as f:
  f.write(tflite_model)

Now, run python tflite_converter.py. You will find the transferlearning_model.tflite file, that can be added to TUM Lens as described above.

This section is based on the TF-Lite documentation.