History

Swarup Ghosh 847fd0f096 Moved train tpu to training and added example script		2020-02-12 12:58:37 +05:30
..
README.md	Moved train tpu to training and added example script	2020-02-12 12:58:37 +05:30
generate_train_labels.py	Moved train tpu to training and added example script	2020-02-12 12:58:37 +05:30
requirements.txt	Moved train tpu to training and added example script	2020-02-12 12:58:37 +05:30
train_on_gpu.py	Moved train tpu to training and added example script	2020-02-12 12:58:37 +05:30

README.md

Steps to train on a GPU using VGGFace2 dataset

Gathering Dataset

Create your account on http://zeus.robots.ox.ac.uk/vgg_face2/login/ once you login goto http://zeus.robots.ox.ac.uk/vgg_face2/ and download all the files mentioned there.

Train Data_v1. 36G. MD5: 88813c6b15de58afc8fa75ea83361d7f.
Train_Images_v1. The training image list, e.g., 'n000002/0001_01.jpg'.

Prepare Dataset

Create Class Labels File

For the train and test data pipeline, we have to generate a text file containing line-seperated class labels.

Please make sure to set the correct path of the train folder in generate_train_labels.py script

python3 generate_train_labels.py

This script will generate VGGFace2-class_labels_train.txt file which is needed further.

Create TF-Records

Use the prepare_tfrecords.py script from tfrecords-faster repository to generate TFRecords.

Please make sure that the paths are correct for train dataset.

Start Training

Also make sure that the paths are correct in train_on_gpu.py script:

IMAGE_SIZE = (152, 152)
CHANNELS = 3
NUM_CLASSES = 8631 # VGGFace_train_v1 dataset

BATCH_SIZE = 8 # tune it as per your GPU memory
LEARN_RATE = 0.01 * (BATCH_SIZE / 128)
MOMENTUM = 0.9
EPOCHS = 15

CL_PATH = './VGGFace2-class_labels_train.txt'
DATASET_PATH = './tfrecords'
TB_PATH = './tensorboard_logs'

Then, run:

python3 train_on_gpu.py