This is the entry point for our Intro to Transfer Learning workshop. The whole goal during this workshop is to go through the first steps for image classification with Deep Learning and implement the building blocks from scratch.
If you want to do some reading beforehand, you'll find the following useful:
- About Tensorflow
- Different subsets of Tensorflow.
- Notions on image classification and re-training of classifiers in the original version of this workshop.
- Make sure you have wget and python3-venv installed:
sudo apt-get install wget python3-venv. - Make sure you create a new virtual environment (instructions below).
- Install the dependencies with the
requirements.txtin this repo. This includes the latest Tensorflow. We will be using features released on the latest version (1.12). - There are two sides for this workshop:
- Using the current code as-is and understanding its effects (which will serve as an intro).
- Getting our hands dirty, re-implementing the code. We will start from branch
ẁorkshop_init.
Python 3.5+ should work. Instructions to create a Python3 based environment are given below.
(If, instead of this, you are used to and prefer using docker, go right ahead! Please choose one based on v1.12 from here.))
Assuming you want to store your virtual environment under this same repo (FYI, other people prefer to do so under $HOME/.venvs):
~/TransferLearning$ sudo apt-get python3.6-venv
~/TransferLearning$ python3.6 -m venv venv-tf
~/TransferLearning$ source venv-tf/bin/activate
~/TransferLearning$ pip install -r requirements.txtOnly if you already have a GPU and working CUDA installation, do this in addition to the commands above:
~/TransferLearning$ pip install tensorflow-gpu(i.e. same as above and install tensorflow-gpu on top.)
Linux users: run this from the repo root folder:
~/TransferLearning$ bash sh/downloads.shThis script will:
- download a pretrained model and save it in the correct location.
- download a flower dataset, already split into training, validation and testing sets, and extract it to the correct location.
Whichever method you chose, your folder structure should be:
~/TransferLearning$ tree tf_files/
tf_files/
├── split_flowers
| ├── train
| │ ├── daisy
| │ │ ├── image12345.jpg
| │ │ └── ...
| │ ├── dandelion
| │ ├── roses
| │ ├── sunflowers
| │ └── tulips
| ├── val
| │ ├── daisy
| │ │ ├── image456779.jpg
| │ │ └── ...
| │ ├── dandelion
| │ ├── roses
| │ ├── sunflowers
| │ └── tulips
| └── test
| ├── daisy
| │ ├── image234673.jpg
| │ └── ...
| ├── dandelion
| ├── roses
| ├── sunflowers
| └── tulips
...
sh is your entry point; there's a bash script to showcase the usage of the retraining code.
scripts contains the retraining code; notice there have been major
modifications in the API (Tensorflow has changed a lot, thankfully)
since the original tutorial was released, so only the interface should feel
familiar.
tf_files contains data, models and tensorboard logs (training_summaries).
Notice after you have played with the current state of the code, we will be re-implementing data loading, model loading and training.
Notice tf_files/training_summaries contains the following structure:
experiment_label
└── training_timestamp1
└── events_file
This way you will be able to compare the loss function and other metrics from different trainings in the same Tensoboard plot. If you don't fully follow this currently, don't worry, it will be explained during the Workshop.
Once you've already launched a training, go to a second terminal and start your Tensorboard with:
~/TransferLearning$ tensorboard --logdir=tf_files/training_summaries &Go with your browser to http://localhost:6006 and enjoy!
Recap:
We'll be switching back and forth between the command line (to launch trainings) to Tensorboard (to see what's going on).
We will be re-implementing different modules to load images, load pre-trained models and train.
The ranch workshop contains incremental commits, showing building blocks one by one.
Modifications are under the same Apache license as the original Tensorflow code.