natagoh / sketchytest

some sketchy deep learning

Sketch Recognition with PyTorch

Authors: Natasha Goenawan and Mandy Lu adapted from Hnd Signs Recognition with PyTorch by Surag Nair, Olivier Moindrot and Guillaume Genthial*

Take the time to read the tutorials.

Note: all scripts must be run in folder pytorch/vision.

Requirements

We recommend using python3 and a virtual env. See instructions here.

virtualenv -p python3 .env
source .env/bin/activate
pip install -r requirements.txt

When you're done working on the project, deactivate the virtual environment with deactivate.

Task

Given an image of a hand doing a sign representing 0, 1, 2, 3, 4 or 5, predict the correct label.

Download the SKETCHES dataset

For the vision example, we will used the SKETCHES dataset created by [Eitz et al 2012]. The dataset is hosted on their website, download it here.

This will download the SKETCHES dataset (~524 MB) containing black and white sketches across 250 categories, which include airplane, alarm clock, angel, etc.

SKETCHES/
    airplane/
        1.png
        ...
    alarm clock/
        81.png
        ...

The images are named following {label}_{id}.png where the label is in [0, 250] and each label corresponds to [airplane, alarm clock, ..., zebra]. The training set contains ~16,000 images and the test set contains ~2,000 images.

Once the download is complete, move the dataset into data/SKETCHES. Run the script build_dataset.py which will resize the images to size (64, 64). The new resized dataset will be located by default in data/64x64_SKETCHES:

python build_dataset.py --data_dir data/SIGNS --output_dir data/64x64_SKETCHES

Quickstart (~10 min)

1. Build the dataset of size 64x64: make sure you complete this step before training

python build_dataset.py --data_dir data/SKETCHES --output_dir data/64x64_SKETCHES

2. Your first experiment We created a base_model directory for you under the experiments directory. It contains a file params.json which sets the hyperparameters for the experiment. It looks like

{
    "learning_rate": 1e-3,
    "batch_size": 32,
    "num_epochs": 10,
    ...
}

For every new experiment, you will need to create a new directory under experiments with a similar params.json file.

3. Train your experiment. Simply run

python train.py --data_dir data/64x64_SKETCHES --model_dir experiments/base_model

It will instantiate a model and train it on the training set following the hyperparameters specified in params.json. It will also evaluate some metrics on the validation set.

4. Your first hyperparameters search We created a new directory learning_rate in experiments for you. Now, run

python search_hyperparams.py --data_dir data/64x64_SKETCHES --parent_dir experiments/learning_rate

It will train and evaluate a model with different values of learning rate defined in search_hyperparams.py and create a new directory for each experiment under experiments/learning_rate/.

5. Display the results of the hyperparameters search in a nice format

python synthesize_results.py --parent_dir experiments/learning_rate

6. Evaluation on the test set Once you've run many experiments and selected your best model and hyperparameters based on the performance on the validation set, you can finally evaluate the performance of your model on the test set. Run

python evaluate.py --data_dir data/64x64_SKETCHES --model_dir experiments/base_model

Guidelines for more advanced use

We recommend reading through train.py to get a high-level overview of the training loop steps:

• loading the hyperparameters for the experiment (the params.json)
• loading the training and validation data
• creating the model, loss_fn and metrics
• training the model for a given number of epochs by calling train_and_evaluate(...)

You can then have a look at data_loader.py to understand:

• how jpg images are loaded and transformed to torch Tensors
• how the data_iterator creates a batch of data and labels and pads sentences

Once you get the high-level idea, depending on your dataset, you might want to modify

• model/net.py to change the neural network, loss function and metrics
• model/data_loader.py to suit the data loader to your specific needs
• train.py for changing the optimizer
• train.py and evaluate.py for some changes in the model or input require changes here

Once you get something working for your dataset, feel free to edit any part of the code to suit your own needs.

Resources

• PyTorch documentation
• Tutorials
• PyTorch warm-up