Currently supporting python 3.6, 3.7 and tensorflow ^2.1.
As years go by, Few Shot Learning (FSL) and especially Metric Learning is becoming a hot topic not only in academic papers but also in production applications.
While a lot of researcher nowadays tend to publish their code on github, there is still no easy framework to get started with FSL. Especially when it comes to benchmarking existing models on personal datasets it is not always easy to find its path into each single repo. Not mentioning the Tensorflow/PyTorch issue.
This repo aims at filling this gap by providing a single entry-point for Few Shot Learning. It is deeply inspired by Keras because it shares the same philosophy:
It was developed with a focus on enabling fast experimentation. Being able to go from idea to result with the least possible delay is key to doing good research.
Thus this repo mainly relies on two of the main high-level python packages for data science: Keras and Pandas. While Pandas may not seem very useful for researchers working with static dataset, it becomes a strong backbone in production applications when you always need to tinker with your data.
Few-shot learning is a task consisting in classifying unseen samples into n classes (so called n way task) where each classes is only described with few (from 1 to 5 in usual benchmarks) examples.
Most of the state-of-the-art algorithms try to sort of learn a metric into a well suited (optimized) feature space. Thus deep networks usually first encode the base images into a feature space onto which a distance or similarity is learnt.
This similarity is meant to be used to later classify samples according to their relative distance, either in a pair-wise manner where the nearest support set samples is used to classify the query sample (Voronoi diagram) or in a more advanced classifier. Indeed, this philosophy is most commonly known as the kernel trick where the kernel is actually the similarity learnt during training. Hence any kind of usual kernel based Machine Learning could potentially be plugged onto this learnt similarity (see the min_eigenvalue metric to track eigenvalues of the learnt similarity to see if it as actually a kernel).
There is no easy answer to the optimal choice of such a classifier in the feature space. This may depend on performance as well as on complexity and real application parameters. For instance if the support set is strongly imbalanced, you may not want to fit an advanced classifier onto it but rather use a raw nearest neighbor approach.
All these considerations lead to the need of a code architecture that will let you play with these parameters with your own data in order to take the best from them.
Amongst other, the original Siamese Nets is usually known as the network from Koch et al. This algorithm learns a pair-wise similarity between images. More precisely it uses a densely connected layers on top of the absolute difference between the two embeddings to predict 0 (different) or 1 (same).
Actually, and as it is now expressed in recent papers, the representation learning framework is as follows:
- a data augmentation module A
- an encoder network E
- a projection network P
- a loss L
This repo mimics this framework by proving model builders and notebooks to implement current SOTA algorithms and your own tweaks seamlessly:
- use
tf.data.Dataset.mapto apply data augmentation - define a
tf.Keras.Sequentialmodel for your encoder - define a
kernel, ie atf.keras.Layerwith two inputs and a real-valued output (see head models) - use any
support_layersto wrap the kernel and compute similarities in atf.keras.Sequentialmanner (see notebooks for instance). - use any loss chosen accordingly to the output of the
tf.keras.Sequentialmodel (GramMatrix or CentroidsMatrix for instance)
As an example, the TripletLoss algorithm uses indeed:
- data augmentation: whatever you want
- encoder: any backbone like ResNet50 or MobileNet
- kernel: the l2 norm:
k(x, x') = ||x - x'||^2 = tf.keras.layers.Lambda(lambda inputs: tf.reduce_sum(tf.square(inputs[0] - inputs[1]), axis=1)) - support_layer: triplet loss uses all the pair-wises distances, hence it is the GramMatrix
- loss:
k(a, p) + margin - k(a, n)with semi-hard mining (see triplet_loss)
This repos provides several tools for few-shot learning:
- Keras layers and models
- Keras sequences and Tensorflow datasets for training the models
- Notebooks with proven learning sequences
All these tools can be used all together or separately. One may want to stick with the keras model trained on regular
numpy arrays, with or without callbacks. When designing more advanced keras.Sequence or tf.data.Dataset for
training, it is advised (and some examples are provided) to use Pandas though it is not necessary at all.
Feel free to experiment and share your thought on this repo by contributing to it!
The notebooks section provides some examples. For instance, just run:
import tensorflow as tf
import tensorflow_datasets as tfds
from tensorflow.keras.models import Sequential
from keras_fsl.models.encoders import BasicCNN
from keras_fsl.layers import GramMatrix
from keras_fsl.losses.gram_matrix_losses import BinaryCrossentropy
from keras_fsl.metrics.gram_matrix_metrics import classification_accuracy, min_eigenvalue
from keras_fsl.utils.tensors import get_dummies
#%% Get data
train_dataset, val_dataset, test_dataset = [
dataset.shuffle(1024).batch(64).map(lambda x, y: (tf.image.convert_image_dtype(x, tf.float32), get_dummies(y)[0]))
for dataset in tfds.load(name="omniglot", split=["train[:90%]", "train[90%:]", "test"], as_supervised=True)
]
input_shape = next(tfds.as_numpy(train_dataset.take(1)))[0].shape[1:] # first shape is batch_size
#%% Training
encoder = BasicCNN(input_shape=input_shape)
support_layer = GramMatrix(kernel="DenseSigmoid")
model = Sequential([encoder, support_layer])
model.compile(optimizer="Adam", loss=BinaryCrossentropy(), metrics=[classification_accuracy(), min_eigenvalue])
model.fit(train_dataset, validation_data=val_dataset, epochs=5)