This package provides only a few metric learning losses below;
- ArcFace
- AdaCos
- CircleLoss
I have been greatly inspired by PyTorch Metric Learning.
$ pip install tf-simple-metric-learningProvided layers are implemented via tf.keras.layers.Layer API, enables;
from tf_simple_metric_learning.layers import ArcFace
arcface = ArcFace(num_classes=NUM_CLASSES, margin=MARGIN, scale=SCALE)Example notebook is in examples directory. Implement CircleLossCL (Class-level label version) layer top of EfficientNet and train it for Cars196 dataset;
import tensorflow as tf
from tf_simple_metric_learning.layers import ArcFace, AdaCos, CircleLossCL
inputs = tf.keras.layers.Input([*IMAGE_SIZE, 3], dtype=tf.uint8)
x = tf.cast(inputs, dtype=tf.float32)
x = tf.keras.applications.efficientnet.preprocess_input(x)
net = tf.keras.applications.EfficientNetB0(include_top=False, weights='imagenet', pooling='avg')
embeds = net(x)
labels = tf.keras.layers.Input([], dtype=tf.int32)
labels_onehot = tf.one_hot(labels, depth=num_classes)
# Create metric learning layer
# metric_layer = ArcFace(num_classes=num_classes, margin=0.5, scale=64)
# metric_layer = AdaCos(num_classes=num_classes)
metric_layer = CircleLossCL(num_classes=num_classes, margin=0.25, scale=256)
logits = metric_layer([embeds, labels_onehot])
model = tf.keras.Model(inputs=[inputs, labels], outputs=logits)
model.summary()Note that you should feed labels as input into model in training because these layers require labels to forward.
In evaluation or prediction, above model requires both images and labels but labels is ignored in those metric learning layers. We only need to use dummy labels (ignored) with the target images because we can't access labels in evaluation or prediction.