TensorFlow Addons Losses: TripletSemiHardLoss

View on TensorFlow.org View source on GitHub Download notebook


This notebook will demonstrate how to use the TripletSemiHardLoss function in TensorFlow Addons.



As first introduced in the FaceNet paper, TripletLoss is a loss function that trains a neural network to closely embed features of the same class while maximizing the distance between embeddings of different classes. To do this an anchor is chosen along with one negative and one positive sample. fig3

The loss function is described as a Euclidean distance function:


Where A is our anchor input, P is the positive sample input, N is the negative sample input, and alpha is some margin we use to specify when a triplet has become too "easy" and we no longer want to adjust the weights from it.

SemiHard Online Learning

As shown in the paper, the best results are from triplets known as "Semi-Hard". These are defined as triplets where the negative is farther from the anchor than the positive, but still produces a positive loss. To efficiently find these triplets we utilize online learning and only train from the Semi-Hard examples in each batch.


pip install -q -U tensorflow-addons
import io
import numpy as np
import tensorflow as tf
import tensorflow_addons as tfa
import tensorflow_datasets as tfds

Prepare the Data

def _normalize_img(img, label):
    img = tf.cast(img, tf.float32) / 255.
    return (img, label)

train_dataset, test_dataset = tfds.load(name="mnist", split=['train', 'test'], as_supervised=True)

# Build your input pipelines
train_dataset = train_dataset.shuffle(1024).batch(32)
train_dataset = train_dataset.map(_normalize_img)

test_dataset = test_dataset.batch(32)
test_dataset = test_dataset.map(_normalize_img)
Downloading and preparing dataset mnist/3.0.1 (download: 11.06 MiB, generated: 21.00 MiB, total: 32.06 MiB) to /home/kbuilder/tensorflow_datasets/mnist/3.0.1...

Warning:absl:Dataset mnist is hosted on GCS. It will automatically be downloaded to your
local data directory. If you'd instead prefer to read directly from our public
GCS bucket (recommended if you're running on GCP), you can instead pass
`try_gcs=True` to `tfds.load` or set `data_dir=gs://tfds-data/datasets`.

Dataset mnist downloaded and prepared to /home/kbuilder/tensorflow_datasets/mnist/3.0.1. Subsequent calls will reuse this data.

Build the Model


model = tf.keras.Sequential([
    tf.keras.layers.Conv2D(filters=64, kernel_size=2, padding='same', activation='relu', input_shape=(28,28,1)),
    tf.keras.layers.Conv2D(filters=32, kernel_size=2, padding='same', activation='relu'),
    tf.keras.layers.Dense(256, activation=None), # No activation on final dense layer
    tf.keras.layers.Lambda(lambda x: tf.math.l2_normalize(x, axis=1)) # L2 normalize embeddings


Train and Evaluate

# Compile the model

# Train the network
history = model.fit(
Epoch 1/5
1875/1875 [==============================] - 7s 4ms/step - loss: 0.5907
Epoch 2/5
1875/1875 [==============================] - 6s 3ms/step - loss: 0.4570
Epoch 3/5
1875/1875 [==============================] - 6s 3ms/step - loss: 0.4242
Epoch 4/5
1875/1875 [==============================] - 6s 3ms/step - loss: 0.4048
Epoch 5/5
1875/1875 [==============================] - 6s 3ms/step - loss: 0.3928

# Evaluate the network
results = model.predict(test_dataset)
# Save test embeddings for visualization in projector
np.savetxt("vecs.tsv", results, delimiter='\t')

out_m = io.open('meta.tsv', 'w', encoding='utf-8')
for img, labels in tfds.as_numpy(test_dataset):
    [out_m.write(str(x) + "\n") for x in labels]

  from google.colab import files

Embedding Projector

The vector and metadata files can be loaded and visualized here: https://projector.tensorflow.org/

You can see the results of our embedded test data when visualized with UMAP: embedding