Model description
This repo contains the model for the notebook Image Classification using BigTransfer (BiT).
Full credits go to Sayan Nath
Reproduced by Rushi Chaudhari
BigTransfer (also known as BiT) is a state-of-the-art transfer learning method for image classification.
Dataset
The Flower Dataset is A large set of images of flowers
Training procedure
Training hyperparameters
The following hyperparameters were used during training:
RESIZE_TO = 384
CROP_TO = 224
BATCH_SIZE = 64
STEPS_PER_EPOCH = 10
AUTO = tf.data.AUTOTUNE # optimise the pipeline performance
NUM_CLASSES = 5 # number of classes
SCHEDULE_LENGTH = (
500 # we will train on lower resolution images and will still attain good results
)
SCHEDULE_BOUNDARIES = [
200,
300,
400,
]
The hyperparamteres like SCHEDULE_LENGTH
and SCHEDULE_BOUNDARIES
are determined based on empirical results. The method has been explained in the original paper and in their Google AI Blog Post.
The SCHEDULE_LENGTH
is aslo determined whether to use MixUp Augmentation or not. You can also find an easy MixUp Implementation in Keras Coding Examples.
Training results
- Downloads last month
- 15
Inference API (serverless) does not yet support tf-keras models for this pipeline type.