GitHub Repository: keras-team/keras-io
Path: blob/master/examples/vision/ipynb/deit.ipynb
³²³⁶ views

Kernel: Python 3

Distilling Vision Transformers

Author: Sayak Paul
Date created: 2022/04/05
Last modified: 2022/04/08
Description: Distillation of Vision Transformers through attention.

Introduction

In the original Vision Transformers (ViT) paper (Dosovitskiy et al.), the authors concluded that to perform on par with Convolutional Neural Networks (CNNs), ViTs need to be pre-trained on larger datasets. The larger the better. This is mainly due to the lack of inductive biases in the ViT architecture -- unlike CNNs, they don't have layers that exploit locality. In a follow-up paper (Steiner et al.), the authors show that it is possible to substantially improve the performance of ViTs with stronger regularization and longer training.

Many groups have proposed different ways to deal with the problem of data-intensiveness of ViT training. One such way was shown in the Data-efficient image Transformers, (DeiT) paper (Touvron et al.). The authors introduced a distillation technique that is specific to transformer-based vision models. DeiT is among the first works to show that it's possible to train ViTs well without using larger datasets.

In this example, we implement the distillation recipe proposed in DeiT. This requires us to slightly tweak the original ViT architecture and write a custom training loop to implement the distillation recipe.

To run the example, you'll need TensorFlow Addons, which you can install with the following command:

pip install tensorflow-addons

To comfortably navigate through this example, you'll be expected to know how a ViT and knowledge distillation work. The following are good resources in case you needed a refresher:

Imports

In [0]:

from typing import List

import tensorflow as tf
import tensorflow_addons as tfa
import tensorflow_datasets as tfds
import tensorflow_hub as hub
from tensorflow import keras
from tensorflow.keras import layers

tfds.disable_progress_bar()
tf.keras.utils.set_random_seed(42)

Constants

In [0]:

# Model
MODEL_TYPE = "deit_distilled_tiny_patch16_224"
RESOLUTION = 224
PATCH_SIZE = 16
NUM_PATCHES = (RESOLUTION // PATCH_SIZE) ** 2
LAYER_NORM_EPS = 1e-6
PROJECTION_DIM = 192
NUM_HEADS = 3
NUM_LAYERS = 12
MLP_UNITS = [
    PROJECTION_DIM * 4,
    PROJECTION_DIM,
]
DROPOUT_RATE = 0.0
DROP_PATH_RATE = 0.1

# Training
NUM_EPOCHS = 20
BASE_LR = 0.0005
WEIGHT_DECAY = 0.0001

# Data
BATCH_SIZE = 256
AUTO = tf.data.AUTOTUNE
NUM_CLASSES = 5

You probably noticed that DROPOUT_RATE has been set 0.0. Dropout has been used in the implementation to keep it complete. For smaller models (like the one used in this example), you don't need it, but for bigger models, using dropout helps.

Load the `tf_flowers` dataset and prepare preprocessing utilities

The authors use an array of different augmentation techniques, including MixUp (Zhang et al.), RandAugment (Cubuk et al.), and so on. However, to keep the example simple to work through, we'll discard them.

In [0]:


def preprocess_dataset(is_training=True):
    def fn(image, label):
        if is_training:
            # Resize to a bigger spatial resolution and take the random
            # crops.
            image = tf.image.resize(image, (RESOLUTION + 20, RESOLUTION + 20))
            image = tf.image.random_crop(image, (RESOLUTION, RESOLUTION, 3))
            image = tf.image.random_flip_left_right(image)
        else:
            image = tf.image.resize(image, (RESOLUTION, RESOLUTION))
        label = tf.one_hot(label, depth=NUM_CLASSES)
        return image, label

    return fn


def prepare_dataset(dataset, is_training=True):
    if is_training:
        dataset = dataset.shuffle(BATCH_SIZE * 10)
    dataset = dataset.map(preprocess_dataset(is_training), num_parallel_calls=AUTO)
    return dataset.batch(BATCH_SIZE).prefetch(AUTO)


train_dataset, val_dataset = tfds.load(
    "tf_flowers", split=["train[:90%]", "train[90%:]"], as_supervised=True
)
num_train = train_dataset.cardinality()
num_val = val_dataset.cardinality()
print(f"Number of training examples: {num_train}")
print(f"Number of validation examples: {num_val}")

train_dataset = prepare_dataset(train_dataset, is_training=True)
val_dataset = prepare_dataset(val_dataset, is_training=False)

Implementing the DeiT variants of ViT

Since DeiT is an extension of ViT it'd make sense to first implement ViT and then extend it to support DeiT's components.

First, we'll implement a layer for Stochastic Depth (Huang et al.) which is used in DeiT for regularization.

In [0]:

# Referred from: github.com:rwightman/pytorch-image-models.
class StochasticDepth(layers.Layer):
    def __init__(self, drop_prop, **kwargs):
        super().__init__(**kwargs)
        self.drop_prob = drop_prop

    def call(self, x, training=True):
        if training:
            keep_prob = 1 - self.drop_prob
            shape = (tf.shape(x)[0],) + (1,) * (len(tf.shape(x)) - 1)
            random_tensor = keep_prob + tf.random.uniform(shape, 0, 1)
            random_tensor = tf.floor(random_tensor)
            return (x / keep_prob) * random_tensor
        return x

Now, we'll implement the MLP and Transformer blocks.

In [0]:


def mlp(x, dropout_rate: float, hidden_units: List):
    """FFN for a Transformer block."""
    # Iterate over the hidden units and
    # add Dense => Dropout.
    for (idx, units) in enumerate(hidden_units):
        x = layers.Dense(
            units,
            activation=tf.nn.gelu if idx == 0 else None,
        )(x)
        x = layers.Dropout(dropout_rate)(x)
    return x


def transformer(drop_prob: float, name: str) -> keras.Model:
    """Transformer block with pre-norm."""
    num_patches = NUM_PATCHES + 2 if "distilled" in MODEL_TYPE else NUM_PATCHES + 1
    encoded_patches = layers.Input((num_patches, PROJECTION_DIM))

    # Layer normalization 1.
    x1 = layers.LayerNormalization(epsilon=LAYER_NORM_EPS)(encoded_patches)

    # Multi Head Self Attention layer 1.
    attention_output = layers.MultiHeadAttention(
        num_heads=NUM_HEADS,
        key_dim=PROJECTION_DIM,
        dropout=DROPOUT_RATE,
    )(x1, x1)
    attention_output = (
        StochasticDepth(drop_prob)(attention_output) if drop_prob else attention_output
    )

    # Skip connection 1.
    x2 = layers.Add()([attention_output, encoded_patches])

    # Layer normalization 2.
    x3 = layers.LayerNormalization(epsilon=LAYER_NORM_EPS)(x2)

    # MLP layer 1.
    x4 = mlp(x3, hidden_units=MLP_UNITS, dropout_rate=DROPOUT_RATE)
    x4 = StochasticDepth(drop_prob)(x4) if drop_prob else x4

    # Skip connection 2.
    outputs = layers.Add()([x2, x4])

    return keras.Model(encoded_patches, outputs, name=name)

We'll now implement a ViTClassifier class building on top of the components we just developed. Here we'll be following the original pooling strategy used in the ViT paper -- use a class token and use the feature representations corresponding to it for classification.

In [0]:


class ViTClassifier(keras.Model):
    """Vision Transformer base class."""

    def __init__(self, **kwargs):
        super().__init__(**kwargs)

        # Patchify + linear projection + reshaping.
        self.projection = keras.Sequential(
            [
                layers.Conv2D(
                    filters=PROJECTION_DIM,
                    kernel_size=(PATCH_SIZE, PATCH_SIZE),
                    strides=(PATCH_SIZE, PATCH_SIZE),
                    padding="VALID",
                    name="conv_projection",
                ),
                layers.Reshape(
                    target_shape=(NUM_PATCHES, PROJECTION_DIM),
                    name="flatten_projection",
                ),
            ],
            name="projection",
        )

        # Positional embedding.
        init_shape = (
            1,
            NUM_PATCHES + 1,
            PROJECTION_DIM,
        )
        self.positional_embedding = tf.Variable(
            tf.zeros(init_shape), name="pos_embedding"
        )

        # Transformer blocks.
        dpr = [x for x in tf.linspace(0.0, DROP_PATH_RATE, NUM_LAYERS)]
        self.transformer_blocks = [
            transformer(drop_prob=dpr[i], name=f"transformer_block_{i}")
            for i in range(NUM_LAYERS)
        ]

        # CLS token.
        initial_value = tf.zeros((1, 1, PROJECTION_DIM))
        self.cls_token = tf.Variable(
            initial_value=initial_value, trainable=True, name="cls"
        )

        # Other layers.
        self.dropout = layers.Dropout(DROPOUT_RATE)
        self.layer_norm = layers.LayerNormalization(epsilon=LAYER_NORM_EPS)
        self.head = layers.Dense(
            NUM_CLASSES,
            name="classification_head",
        )

    def call(self, inputs, training=True):
        n = tf.shape(inputs)[0]

        # Create patches and project the patches.
        projected_patches = self.projection(inputs)

        # Append class token if needed.
        cls_token = tf.tile(self.cls_token, (n, 1, 1))
        cls_token = tf.cast(cls_token, projected_patches.dtype)
        projected_patches = tf.concat([cls_token, projected_patches], axis=1)

        # Add positional embeddings to the projected patches.
        encoded_patches = (
            self.positional_embedding + projected_patches
        )  # (B, number_patches, projection_dim)
        encoded_patches = self.dropout(encoded_patches)

        # Iterate over the number of layers and stack up blocks of
        # Transformer.
        for transformer_module in self.transformer_blocks:
            # Add a Transformer block.
            encoded_patches = transformer_module(encoded_patches)

        # Final layer normalization.
        representation = self.layer_norm(encoded_patches)

        # Pool representation.
        encoded_patches = representation[:, 0]

        # Classification head.
        output = self.head(encoded_patches)
        return output

This class can be used standalone as ViT and is end-to-end trainable. Just remove the distilled phrase in MODEL_TYPE and it should work with vit_tiny = ViTClassifier(). Let's now extend it to DeiT. The following figure presents the schematic of DeiT (taken from the DeiT paper):

Apart from the class token, DeiT has another token for distillation. During distillation, the logits corresponding to the class token are compared to the true labels, and the logits corresponding to the distillation token are compared to the teacher's predictions.

In [0]:


class ViTDistilled(ViTClassifier):
    def __init__(self, regular_training=False, **kwargs):
        super().__init__(**kwargs)
        self.num_tokens = 2
        self.regular_training = regular_training

        # CLS and distillation tokens, positional embedding.
        init_value = tf.zeros((1, 1, PROJECTION_DIM))
        self.dist_token = tf.Variable(init_value, name="dist_token")
        self.positional_embedding = tf.Variable(
            tf.zeros(
                (
                    1,
                    NUM_PATCHES + self.num_tokens,
                    PROJECTION_DIM,
                )
            ),
            name="pos_embedding",
        )

        # Head layers.
        self.head = layers.Dense(
            NUM_CLASSES,
            name="classification_head",
        )
        self.head_dist = layers.Dense(
            NUM_CLASSES,
            name="distillation_head",
        )

    def call(self, inputs, training=True):
        n = tf.shape(inputs)[0]

        # Create patches and project the patches.
        projected_patches = self.projection(inputs)

        # Append the tokens.
        cls_token = tf.tile(self.cls_token, (n, 1, 1))
        dist_token = tf.tile(self.dist_token, (n, 1, 1))
        cls_token = tf.cast(cls_token, projected_patches.dtype)
        dist_token = tf.cast(dist_token, projected_patches.dtype)
        projected_patches = tf.concat(
            [cls_token, dist_token, projected_patches], axis=1
        )

        # Add positional embeddings to the projected patches.
        encoded_patches = (
            self.positional_embedding + projected_patches
        )  # (B, number_patches, projection_dim)
        encoded_patches = self.dropout(encoded_patches)

        # Iterate over the number of layers and stack up blocks of
        # Transformer.
        for transformer_module in self.transformer_blocks:
            # Add a Transformer block.
            encoded_patches = transformer_module(encoded_patches)

        # Final layer normalization.
        representation = self.layer_norm(encoded_patches)

        # Classification heads.
        x, x_dist = (
            self.head(representation[:, 0]),
            self.head_dist(representation[:, 1]),
        )

        if not training or self.regular_training:
            # During standard train / finetune, inference average the classifier
            # predictions.
            return (x + x_dist) / 2

        elif training:
            # Only return separate classification predictions when training in distilled
            # mode.
            return x, x_dist

Let's verify if the ViTDistilled class can be initialized and called as expected.

In [0]:

deit_tiny_distilled = ViTDistilled()

dummy_inputs = tf.ones((2, 224, 224, 3))
outputs = deit_tiny_distilled(dummy_inputs, training=False)
print(outputs.shape)

Implementing the trainer

Unlike what happens in standard knowledge distillation (Hinton et al.), where a temperature-scaled softmax is used as well as KL divergence, DeiT authors use the following loss function:

Here,

CE is cross-entropy
psi is the softmax function
Z_s denotes student predictions
y denotes true labels
y_t denotes teacher predictions

In [0]:


class DeiT(keras.Model):
    # Reference:
    # https://keras.io/examples/vision/knowledge_distillation/
    def __init__(self, student, teacher, **kwargs):
        super().__init__(**kwargs)
        self.student = student
        self.teacher = teacher

        self.student_loss_tracker = keras.metrics.Mean(name="student_loss")
        self.dist_loss_tracker = keras.metrics.Mean(name="distillation_loss")

    @property
    def metrics(self):
        metrics = super().metrics
        metrics.append(self.student_loss_tracker)
        metrics.append(self.dist_loss_tracker)
        return metrics

    def compile(
        self,
        optimizer,
        metrics,
        student_loss_fn,
        distillation_loss_fn,
    ):
        super().compile(optimizer=optimizer, metrics=metrics)
        self.student_loss_fn = student_loss_fn
        self.distillation_loss_fn = distillation_loss_fn

    def train_step(self, data):
        # Unpack data.
        x, y = data

        # Forward pass of teacher
        teacher_predictions = tf.nn.softmax(self.teacher(x, training=False), -1)
        teacher_predictions = tf.argmax(teacher_predictions, -1)

        with tf.GradientTape() as tape:
            # Forward pass of student.
            cls_predictions, dist_predictions = self.student(x / 255.0, training=True)

            # Compute losses.
            student_loss = self.student_loss_fn(y, cls_predictions)
            distillation_loss = self.distillation_loss_fn(
                teacher_predictions, dist_predictions
            )
            loss = (student_loss + distillation_loss) / 2

        # Compute gradients.
        trainable_vars = self.student.trainable_variables
        gradients = tape.gradient(loss, trainable_vars)

        # Update weights.
        self.optimizer.apply_gradients(zip(gradients, trainable_vars))

        # Update the metrics configured in `compile()`.
        student_predictions = (cls_predictions + dist_predictions) / 2
        self.compiled_metrics.update_state(y, student_predictions)
        self.dist_loss_tracker.update_state(distillation_loss)
        self.student_loss_tracker.update_state(student_loss)

        # Return a dict of performance.
        results = {m.name: m.result() for m in self.metrics}
        return results

    def test_step(self, data):
        # Unpack the data.
        x, y = data

        # Compute predictions.
        y_prediction = self.student(x / 255.0, training=False)

        # Calculate the loss.
        student_loss = self.student_loss_fn(y, y_prediction)

        # Update the metrics.
        self.compiled_metrics.update_state(y, y_prediction)
        self.student_loss_tracker.update_state(student_loss)

        # Return a dict of performance.
        results = {m.name: m.result() for m in self.metrics}
        return results

    def call(self, inputs):
        return self.student(inputs / 255.0, training=False)

Load the teacher model

This model is based on the BiT family of ResNets (Kolesnikov et al.) fine-tuned on the tf_flowers dataset. You can refer to this notebook to know how the training was performed. The teacher model has about 212 Million parameters which is about 40x more than the student.

In [0]:

!wget -q https://github.com/sayakpaul/deit-tf/releases/download/v0.1.0/bit_teacher_flowers.zip
!unzip -q bit_teacher_flowers.zip

In [0]:

bit_teacher_flowers = keras.models.load_model("bit_teacher_flowers")

Training through distillation

In [0]:

deit_tiny = ViTDistilled()
deit_distiller = DeiT(student=deit_tiny, teacher=bit_teacher_flowers)

lr_scaled = (BASE_LR / 512) * BATCH_SIZE
deit_distiller.compile(
    optimizer=tfa.optimizers.AdamW(weight_decay=WEIGHT_DECAY, learning_rate=lr_scaled),
    metrics=["accuracy"],
    student_loss_fn=keras.losses.CategoricalCrossentropy(
        from_logits=True, label_smoothing=0.1
    ),
    distillation_loss_fn=keras.losses.SparseCategoricalCrossentropy(from_logits=True),
)
_ = deit_distiller.fit(train_dataset, validation_data=val_dataset, epochs=NUM_EPOCHS)

If we had trained the same model (the ViTClassifier) from scratch with the exact same hyperparameters, the model would have scored about 59% accuracy. You can adapt the following code to reproduce this result:

vit_tiny = ViTClassifier()

inputs = keras.Input((RESOLUTION, RESOLUTION, 3))
x = keras.layers.Rescaling(scale=1./255)(inputs)
outputs = deit_tiny(x)
model = keras.Model(inputs, outputs)

model.compile(...)
model.fit(...)

Notes

Through the use of distillation, we're effectively transferring the inductive biases of a CNN-based teacher model.
Interestingly enough, this distillation strategy works better with a CNN as the teacher model rather than a Transformer as shown in the paper.
The use of regularization to train DeiT models is very important.
ViT models are initialized with a combination of different initializers including truncated normal, random normal, Glorot uniform, etc. If you're looking for end-to-end reproduction of the original results, don't forget to initialize the ViTs well.
If you want to explore the pre-trained DeiT models in TensorFlow and Keras with code for fine-tuning, check out these models on TF-Hub.

Acknowledgements

Ross Wightman for keeping timm updated with readable implementations. I referred to the implementations of ViT and DeiT a lot during implementing them in TensorFlow.
Aritra Roy Gosthipaty who implemented some portions of the ViTClassifier in another project.
Google Developers Experts program for supporting me with GCP credits which were used to run experiments for this example.

Example available on HuggingFace:

Trained Model	Demo

Distilling Vision Transformers

Introduction

Imports

Constants

Load the `tf_flowers` dataset and prepare preprocessing utilities

Implementing the DeiT variants of ViT

Implementing the trainer

Load the teacher model

Training through distillation

Notes

Acknowledgements

Product

Resources

Company

Distilling Vision Transformers

Introduction

Imports

Constants

Load the tf_flowers dataset and prepare preprocessing utilities

Implementing the DeiT variants of ViT

Implementing the trainer

Load the teacher model

Training through distillation

Notes

Acknowledgements

Load the `tf_flowers` dataset and prepare preprocessing utilities