Introduction

Transfer learning is a technique where a pre-trained model, initially trained on a large dataset for a specific task, is fine-tuned for a different but related task.

Instead of starting from scratch, we leverage the pre-trained model and adapt it to our specific problem. This approach can significantly reduce training time and data requirements.

In this project, I performed transfer learning on the SSD MobileNet V2 FPNLite 320x320 model using Tensorflow's Object Detection API to allow it to detect glasses on a person's face in real-time.

Setup Guide

This guide assumes that you already have Python, CUDA, and CUDNN installed.

Installing Protocol Buffers (protoc)

Protocol Buffers are a data serialization format used by TensorFlow to define and serialize data structures, especially for exchanging data between different components of a TensorFlow-based system.

1. Go to the Protocol Buffers release page on Github and download the .zip file for your OS: https://github.com/protocolbuffers/protobuf/releases

2. Extract the directory to an installion location of your choice.

3. Add the address of the bin subdirectory to the environment variables/path of your system. A restart may be necessary for the changes to apply successfully.

Installing Tensorflow Object Detection API

1. Clone the Tensorflow Models repository:

git clone https://github.com/tensorflow/models.git

2. Navigate to the models subdirectory:

cd models/research

3. Run the following command:

protoc object_detection/protos/*.proto --python_out=.

4. Copy the .py setup file to the current directory:

cp object_detection/packages/tf2/setup.py .

5. Run the setup. I recommend doing this in a new Python virtual environment to avoid conflicts:

python -m pip install .

6. Verify the install by running the following imports in Python:

import tensorflow as tf
from object_detection.utils import config_util
from object_detection.protos import pipeline_pb2
from google.protobuf import text_format

Possible Installation Errors

1. After you run "python -m pip install .", you may get the following error:

Error:
Collecting PyYAML==5.4.1
  Using cached PyYAML-5.4.1.tar.gz (175 kB)
  Installing build dependencies ... done
  Getting requirements to build wheel ... error
  error: subprocess-exited-with-error

  × Getting requirements to build wheel did not run successfully.
  │ exit code: 1

This is caused by the setup not checking if wheel is installed and/or an incorrect version of Cython being installed. To solve this:

pip install wheel
pip install "Cython<3.0" "pyyaml<6" --no-build-isolation

2. Importing from object_detection may throw a protobuf error. This can be solved by installing an older version of protobuf:

pip install --upgrade protobuf<=3.20

3. You may also get a warning message along the lines of WARNING: Ignoring invalid distribution. To remove this, navigate to C:\users\MYNAME\appdata\local\programs\python\python310\lib\site-packages or your equivalent packages directory and run:

 rm ~rotobuf-3.19.6.dist-info

Installing labelImg Python Package

pip install labelImg