focal_modulation_network.py

"""
Title: Focal Modulation: A replacement for Self-Attention
Author: [Aritra Roy Gosthipaty](https://twitter.com/ariG23498), [Ritwik Raha](https://twitter.com/ritwik_raha)
Date created: 2023/01/25
Last modified: 2026/01/27
Description: Image classification with Focal Modulation Networks.
Accelerator: GPU
Converted to Keras 3 by: [LakshmiKalaKadali](https://github.com/LakshmiKalaKadali)
"""

"""
## Introduction

This tutorial aims to provide a comprehensive guide to the implementation of
Focal Modulation Networks, as presented in
[Yang et al.](https://arxiv.org/abs/2203.11926).

This tutorial will provide a formal, minimalistic approach to implementing Focal
Modulation Networks and explore its potential applications in the field of Deep Learning.

**Problem statement**

The Transformer architecture ([Vaswani et al.](https://arxiv.org/abs/1706.03762)),
which has become the de facto standard in most Natural Language Processing tasks, has
also been applied to the field of computer vision, e.g. Vision
Transformers ([Dosovitskiy et al.](https://arxiv.org/abs/2010.11929v2)).

> In Transformers, the self-attention (SA) is arguably the key to its success which
enables input-dependent global interactions, in contrast to convolution operation which
constraints interactions in a local region with a shared kernel.

The **Attention** module is mathematically written as shown in **Equation 1**.

| ![Attention Equation](https://i.imgur.com/thdHvQx.png) |
| :--: |
| Equation 1: The mathematical equation of attention (Source: Aritra and Ritwik) |

Where:

- `Q` is the query
- `K` is the key
- `V` is the value
- `d_k` is the dimension of the key

With **self-attention**, the query, key, and value are all sourced from the input
sequence. Let us rewrite the attention equation for self-attention as shown in **Equation
2**.

| ![Self-Attention Equation](https://i.imgur.com/OFsmVdP.png) |
| :--: |
| Equation 2: The mathematical equation of self-attention (Source: Aritra and Ritwik) |

Upon looking at the equation of self-attention, we see that it is a quadratic equation.
Therefore, as the number of tokens increase, so does the computation time (cost too). To
mitigate this problem and make Transformers more interpretable, Yang et al.
have tried to replace the Self-Attention module with better components.

**The Solution**

Yang et al. introduce the Focal Modulation layer to serve as a
seamless replacement for the Self-Attention Layer. The layer boasts high
interpretability, making it a valuable tool for Deep Learning practitioners.

In this tutorial, we will delve into the practical application of this layer by training
the entire model on the CIFAR-10 dataset and visually interpreting the layer's
performance.

Note: We try to align our implementation with the
[official implementation](https://github.com/microsoft/FocalNet).
"""

"""
## Setup and Imports

Keras 3 allows this model to run on JAX, PyTorch, or TensorFlow. We use keras.ops for all mathematical operations to ensure the code remains backend-agnostic.
"""

import os

# Set backend before importing keras
os.environ["KERAS_BACKEND"] = "tensorflow"  # Or "torch" or "tensorflow"

import numpy as np
import keras
from keras import layers
from keras import ops
from matplotlib import pyplot as plt
from random import randint

# Set seed for reproducibility using Keras 3 utility.
keras.utils.set_random_seed(42)

"""
## Global Configuration

We do not have any strong rationale behind choosing these hyperparameters. Please feel
free to change the configuration and train the model.
"""

# --- GLOBAL CONFIGURATION ---
TRAIN_SLICE = 40000
BATCH_SIZE = 128  # 1024
INPUT_SHAPE = (32, 32, 3)
IMAGE_SIZE = 48
NUM_CLASSES = 10

LEARNING_RATE = 1e-4
WEIGHT_DECAY = 1e-4
EPOCHS = 20

"""
## Data Loading with PyDataset

Keras 3 introduces PyDataset as a standardized way to handle data.
It works identically across all backends and avoids the "Symbolic Tensor" issues often found
when using tf.data with JAX or PyTorch.
"""

(x_train, y_train), (x_test, y_test) = keras.datasets.cifar10.load_data()
(x_train, y_train), (x_val, y_val) = (
    (x_train[:TRAIN_SLICE], y_train[:TRAIN_SLICE]),
    (x_train[TRAIN_SLICE:], y_train[TRAIN_SLICE:]),
)


class FocalDataset(keras.utils.PyDataset):
    def __init__(self, x_data, y_data, batch_size, shuffle=False, **kwargs):
        super().__init__(**kwargs)
        self.x_data = x_data
        self.y_data = y_data
        self.batch_size = batch_size
        self.shuffle = shuffle
        self.indices = np.arange(len(x_data))
        if self.shuffle:
            np.random.shuffle(self.indices)

    def __len__(self):
        return int(np.ceil(len(self.x_data) / self.batch_size))

    def __getitem__(self, idx):
        start = idx * self.batch_size
        end = min((idx + 1) * self.batch_size, len(self.x_data))
        batch_indices = self.indices[start:end]

        x_batch = self.x_data[batch_indices]
        y_batch = self.y_data[batch_indices]

        # Convert to backend-native tensors
        x_batch = ops.convert_to_tensor(x_batch, dtype="float32")
        y_batch = ops.convert_to_tensor(y_batch, dtype="int32")

        return x_batch, y_batch

    def on_epoch_end(self):
        if self.shuffle:
            np.random.shuffle(self.indices)


train_ds = FocalDataset(x_train, y_train, batch_size=BATCH_SIZE, shuffle=True)
val_ds = FocalDataset(x_val, y_val, batch_size=BATCH_SIZE, shuffle=False)
test_ds = FocalDataset(x_test, y_test, batch_size=BATCH_SIZE, shuffle=False)

"""
## Architecture

We pause here to take a quick look at the Architecture of the Focal Modulation Network.
**Figure 1** shows how every individual layer is compiled into a single model. This gives
us a bird's eye view of the entire architecture.

| ![Diagram of the model](https://i.imgur.com/v5HYV5R.png) |
| :--: |
| Figure 1: A diagram of the Focal Modulation model (Source: Aritra and Ritwik) |

We dive deep into each of these layers in the following sections. This is the order we
will follow:


- Patch Embedding Layer
- Focal Modulation Block
  - Multi-Layer Perceptron
  - Focal Modulation Layer
    - Hierarchical Contextualization
    - Gated Aggregation
  - Building Focal Modulation Block
- Building the Basic Layer

To better understand the architecture in a format we are well versed in, let us see how
the Focal Modulation Network would look when drawn like a Transformer architecture.

**Figure 2** shows the encoder layer of a traditional Transformer architecture where Self
Attention is replaced with the Focal Modulation layer.

The <font color="blue">blue</font> blocks represent the Focal Modulation block. A stack
of these blocks builds a single Basic Layer. The <font color="green">green</font> blocks
represent the Focal Modulation layer.

| ![The Entire Architecture](https://i.imgur.com/PduYD6m.png) |
| :--: |
| Figure 2: The Entire Architecture (Source: Aritra and Ritwik) |
"""

"""
## Patch Embedding Layer

The patch embedding layer is used to patchify the input images and project them into a
latent space. This layer is also used as the down-sampling layer in the architecture.
"""


class PatchEmbed(layers.Layer):
    """Image patch embedding layer, also acts as the down-sampling layer.

    Args:
        image_size (Tuple[int]): Input image resolution.
        patch_size (Tuple[int]): Patch spatial resolution.
        embed_dim (int): Embedding dimension.
    """

    def __init__(
        self, image_size=(224, 224), patch_size=(4, 4), embed_dim=96, **kwargs
    ):
        super().__init__(**kwargs)
        self.patch_resolution = [
            image_size[0] // patch_size[0],
            image_size[1] // patch_size[1],
        ]
        self.proj = layers.Conv2D(
            filters=embed_dim, kernel_size=patch_size, strides=patch_size
        )
        self.flatten = layers.Reshape(target_shape=(-1, embed_dim))
        self.norm = layers.LayerNormalization(epsilon=1e-7)

    def call(self, x):
        """Patchifies the image and converts into tokens.

        Args:
            x: Tensor of shape (B, H, W, C)

        Returns:
            A tuple of the processed tensor, height of the projected
            feature map, width of the projected feature map, number
            of channels of the projected feature map.
        """
        x = self.proj(x)
        shape = ops.shape(x)
        height, width, channels = shape[1], shape[2], shape[3]
        x = self.norm(self.flatten(x))
        return x, height, width, channels


"""
## Focal Modulation block

A Focal Modulation block can be considered as a single Transformer Block with the Self
Attention (SA) module being replaced with Focal Modulation module, as we saw in **Figure
2**.

Let us recall how a focal modulation block is supposed to look like with the aid of the
**Figure 3**.


| ![Focal Modulation Block](https://i.imgur.com/bPYTSiB.png) |
| :--: |
| Figure 3: The isolated view of the Focal Modulation Block (Source: Aritra and Ritwik) |

The Focal Modulation Block consists of:
- Multilayer Perceptron
- Focal Modulation layer
"""

"""
### Multilayer Perceptron
"""


def MLP(in_features, hidden_features=None, out_features=None, mlp_drop_rate=0.0):
    hidden_features = hidden_features or in_features
    out_features = out_features or in_features
    return keras.Sequential(
        [
            layers.Dense(units=hidden_features, activation="gelu"),
            layers.Dense(units=out_features),
            layers.Dropout(rate=mlp_drop_rate),
        ]
    )


"""
### Focal Modulation layer

In a typical Transformer architecture, for each visual token (**query**) `x_i in R^C` in
an input feature map `X in R^{HxWxC}` a **generic encoding process** produces a feature
representation `y_i in R^C`.

The encoding process consists of **interaction** (with its surroundings for e.g. a dot
product), and **aggregation** (over the contexts for e.g weighted mean).

We will talk about two types of encoding here:
- Interaction and then Aggregation in **Self-Attention**
- Aggregation and then Interaction in **Focal Modulation**

**Self-Attention**

| ![Self-Attention Expression](https://i.imgur.com/heBYp0F.png) |
| :--: |
| **Figure 4**: Self-Attention module. (Source: Aritra and Ritwik) |

| ![Aggregation and Interaction for Self-Attention](https://i.imgur.com/j1k8Xmy.png) |
| :--: |
| **Equation 3:** Aggregation and Interaction in Self-Attention(Surce: Aritra and Ritwik)|

As shown in **Figure 4** the query and the key interact (in the interaction step) with
each other to output the attention scores. The weighted aggregation of the value comes
next, known as the aggregation step.

**Focal Modulation**

| ![Focal Modulation module](https://i.imgur.com/tmbLgQl.png) |
| :--: |
| **Figure 5**: Focal Modulation module. (Source: Aritra and Ritwik) |

| ![Aggregation and Interaction in Focal Modulation](https://i.imgur.com/gsvJfWp.png) |
| :--: |
| **Equation 4:** Aggregation and Interaction in Focal Modulation (Source: Aritra and Ritwik) |

**Figure 5** depicts the Focal Modulation layer. `q()` is the query projection
function. It is a **linear layer** that projects the query into a latent space. `m ()` is
the context aggregation function. Unlike self-attention, the
aggregation step takes place in focal modulation before the interaction step.
"""

"""
While `q()` is pretty straightforward to understand, the context aggregation function
`m()` is more complex. Therefore, this section will focus on `m()`.

| ![Context Aggregation](https://i.imgur.com/uqIRXI7.png)|
| :--: |
| **Figure 6**: Context Aggregation function `m()`. (Source: Aritra and Ritwik) |

The context aggregation function `m()` consists of two parts as shown in **Figure 6**:
- Hierarchical Contextualization
- Gated Aggregation
"""

"""
#### Hierarchical Contextualization

| ![Hierarchical Contextualization](https://i.imgur.com/q875c83.png)|
| :--: |
| **Figure 7**: Hierarchical Contextualization (Source: Aritra and Ritwik) |

In **Figure 7**, we see that the input is first projected linearly. This linear projection
produces `Z^0`. Where `Z^0` can be expressed as follows:

| ![Linear projection of z_not](https://i.imgur.com/pd0Z2Of.png) |
| :--: |
| Equation 5: Linear projection of `Z^0` (Source: Aritra and Ritwik) |

`Z^0` is then passed on to a series of Depth-Wise (DWConv) Conv and
[GeLU](https://keras.io/api/layers/activations/#gelu-function) layers. The
authors term each block of DWConv and GeLU as levels denoted by `l`. In **Figure 6** we
have two levels. Mathematically this is represented as:

| ![Levels of modulation](https://i.imgur.com/ijGD1Df.png) |
| :--: |
| Equation 6: Levels of the modulation layer (Source: Aritra and Ritwik) |

where `l in {1, ... , L}`

The final feature map goes through a Global Average Pooling Layer. This can be expressed
as follows:

| ![Avg Pool](https://i.imgur.com/MQzQhbo.png) |
| :--: |
| Equation 7: Average Pooling of the final feature (Source: Aritra and Ritwik)|
"""

"""
#### Gated Aggregation

| ![Gated Aggregation](https://i.imgur.com/LwrdDKo.png[/img)|
| :--: |
| **Figure 8**: Gated Aggregation (Source: Aritra and Ritwik) |

Now that we have `L+1` intermediate feature maps by virtue of the Hierarchical
Contextualization step, we need a gating mechanism that lets some features pass and
prohibits others. This can be implemented with the attention module.
Later in the tutorial, we will visualize these gates to better understand their
usefulness.

First, we build the weights for aggregation. Here we apply a **linear layer** on the input
feature map that projects it into `L+1` dimensions.

| ![Gates](https://i.imgur.com/1CgEo1G.png) |
| :--: |
| Eqation 8: Gates (Source: Aritra and Ritwik) |

Next we perform the weighted aggregation over the contexts.

| ![z out](https://i.imgur.com/mpJ712R.png) |
| :--: |
| Eqation 9: Final feature map (Source: Aritra and Ritwik) |

To enable communication across different channels, we use another linear layer `h()`
to obtain the modulator

| ![Modulator](https://i.imgur.com/0EpT3Ti.png) |
| :--: |
| Eqation 10: Modulator (Source: Aritra and Ritwik) |

To sum up the Focal Modulation layer we have:

| ![Focal Modulation Layer](https://i.imgur.com/1QIhvYA.png) |
| :--: |
| Eqation 11: Focal Modulation Layer (Source: Aritra and Ritwik) |
"""


class FocalModulationLayer(layers.Layer):
    """The Focal Modulation layer includes query projection & context aggregation.

    Args:
        dim (int): Projection dimension.
        focal_window (int): Window size for focal modulation.
        focal_level (int): The current focal level.
        focal_factor (int): Factor of focal modulation.
        proj_drop_rate (float): Rate of dropout.
    """

    def __init__(
        self,
        dim,
        focal_window,
        focal_level,
        focal_factor=2,
        proj_drop_rate=0.0,
        **kwargs,
    ):
        super().__init__(**kwargs)
        self.dim, self.focal_level = dim, focal_level
        self.initial_proj = layers.Dense(units=(2 * dim) + (focal_level + 1))
        self.focal_layers = [
            keras.Sequential(
                [
                    layers.ZeroPadding2D(
                        padding=((focal_factor * i + focal_window) // 2)
                    ),
                    layers.Conv2D(
                        filters=dim,
                        kernel_size=(focal_factor * i + focal_window),
                        activation="gelu",
                        groups=dim,
                        use_bias=False,
                    ),
                ]
            )
            for i in range(focal_level)
        ]
        self.gap = layers.GlobalAveragePooling2D(keepdims=True)
        self.mod_proj = layers.Conv2D(filters=dim, kernel_size=1)
        self.proj = layers.Dense(units=dim)
        self.proj_drop = layers.Dropout(proj_drop_rate)

    def call(self, x, training=None):
        """Forward pass of the layer.

        Args:
            x: Tensor of shape (B, H, W, C)
        """
        x_proj = self.initial_proj(x)
        query, context, gates = ops.split(x_proj, [self.dim, 2 * self.dim], axis=-1)

        # Apply Softmax for numerical stability
        gates = ops.softmax(gates, axis=-1)
        self.gates = gates

        context = self.focal_layers[0](context)
        context_all = context * gates[..., 0:1]
        for i in range(1, self.focal_level):
            context = self.focal_layers[i](context)
            context_all = context_all + (context * gates[..., i : i + 1])

        context_global = ops.gelu(self.gap(context))
        context_all = context_all + (context_global * gates[..., self.focal_level :])

        self.modulator = self.mod_proj(context_all)
        x_out = query * self.modulator
        return self.proj_drop(self.proj(x_out), training=training)


"""
### The Focal Modulation block

Finally, we have all the components we need to build the Focal Modulation block. Here we
take the MLP and Focal Modulation layer together and build the Focal Modulation block.
"""


class FocalModulationBlock(layers.Layer):
    """Combine FFN and Focal Modulation Layer.

    Args:
        dim (int): Number of input channels.
        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
        drop (float): Dropout rate.
        focal_level (int): Number of focal levels.
        focal_window (int): Focal window size at first focal level
    """

    def __init__(
        self, dim, mlp_ratio=4.0, drop=0.0, focal_level=1, focal_window=3, **kwargs
    ):
        super().__init__(**kwargs)
        self.norm1 = layers.LayerNormalization(epsilon=1e-5)
        self.modulation = FocalModulationLayer(
            dim, focal_window, focal_level, proj_drop_rate=drop
        )
        self.norm2 = layers.LayerNormalization(epsilon=1e-5)
        self.mlp = MLP(dim, int(dim * mlp_ratio), mlp_drop_rate=drop)

    def call(self, x, height=None, width=None, channels=None, training=None):
        """Processes the input tensor through the focal modulation block.

        Args:
            x : Inputs of the shape (B, L, C)
            height (int): The height of the feature map
            width (int): The width of the feature map
            channels (int): The number of channels of the feature map

        Returns:
            The processed tensor.
        """
        res = x
        x = ops.reshape(x, (-1, height, width, channels))
        x = self.modulation(x, training=training)
        x = ops.reshape(x, (-1, height * width, channels))
        x = res + x
        return x + self.mlp(self.norm2(x), training=training)


"""
## The Basic Layer

The basic layer consists of a collection of Focal Modulation blocks. This is
illustrated in **Figure 9**.

| ![Basic Layer](https://i.imgur.com/UcZV0K6.png) |
| :--: |
| **Figure 9**: Basic Layer, a collection of focal modulation blocks. (Source: Aritra and Ritwik) |

Notice how in **Fig. 9** there are more than one focal modulation blocks denoted by `Nx`.
This shows how the Basic Layer is a collection of Focal Modulation blocks.
"""


class BasicLayer(layers.Layer):
    """Collection of Focal Modulation Blocks.

    Args:
        dim (int): Dimensions of the model.
        out_dim (int): Dimension used by the Patch Embedding Layer.
        input_res (Tuple[int]): Input image resolution.
        depth (int): The number of Focal Modulation Blocks.
        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
        drop (float): Dropout rate.
        downsample (keras.layers.Layer): Downsampling layer at the end of the layer.
        focal_level (int): The current focal level.
        focal_window (int): Focal window used.
    """

    def __init__(
        self,
        dim,
        out_dim,
        input_res,
        depth,
        mlp_ratio=4.0,
        drop=0.0,
        downsample=None,
        focal_level=1,
        focal_window=1,
        **kwargs,
    ):
        super().__init__(**kwargs)
        self.blocks = [
            FocalModulationBlock(dim, mlp_ratio, drop, focal_level, focal_window)
            for _ in range(depth)
        ]
        self.downsample = (
            downsample(image_size=input_res, patch_size=(2, 2), embed_dim=out_dim)
            if downsample
            else None
        )

    def call(self, x, height=None, width=None, channels=None, training=None):
        """Forward pass of the layer.

        Args:
            x : Tensor of shape (B, L, C)
            height (int): Height of feature map
            width (int): Width of feature map
            channels (int): Embed Dim of feature map

        Returns:
            A tuple of the processed tensor, changed height, width, and
            dim of the tensor.
        """
        for block in self.blocks:
            x = block(
                x, height=height, width=width, channels=channels, training=training
            )
        if self.downsample:
            x = ops.reshape(x, (-1, height, width, channels))
            x, height, width, channels = self.downsample(x)
        return x, height, width, channels


"""
## The Focal Modulation Network model

This is the model that ties everything together.
It consists of a collection of Basic Layers with a classification head.
For a recap of how this is structured refer to **Figure 1**.
"""


class FocalModulationNetwork(keras.Model):
    """The Focal Modulation Network.

    Parameters:
        image_size (Tuple[int]): Spatial size of images used.
        patch_size (Tuple[int]): Patch size of each patch.
        num_classes (int): Number of classes used for classification.
        embed_dim (int): Patch embedding dimension.
        depths (List[int]): Depth of each Focal Transformer block.
    """

    def __init__(
        self,
        image_size=(48, 48),
        patch_size=(4, 4),
        num_classes=10,
        embed_dim=64,
        depths=[2, 3, 2],
        **kwargs,
    ):
        super().__init__(**kwargs)
        # Preprocessing integrated in model for backend-agnostic behavior
        self.rescaling = layers.Rescaling(1.0 / 255.0)
        self.resizing_larger = layers.Resizing(image_size[0] + 10, image_size[1] + 10)
        self.random_crop = layers.RandomCrop(image_size[0], image_size[1])
        self.resizing_target = layers.Resizing(image_size[0], image_size[1])
        self.random_flip = layers.RandomFlip("horizontal")

        self.patch_embed = PatchEmbed(image_size, patch_size, embed_dim)
        self.basic_layers = []
        for i in range(len(depths)):
            d = embed_dim * (2**i)
            self.basic_layers.append(
                BasicLayer(
                    dim=d,
                    out_dim=d * 2 if i < len(depths) - 1 else None,
                    input_res=(image_size[0] // (2**i), image_size[1] // (2**i)),
                    depth=depths[i],
                    downsample=PatchEmbed if i < len(depths) - 1 else None,
                )
            )
        self.norm = layers.LayerNormalization(epsilon=1e-7)
        self.avgpool = layers.GlobalAveragePooling1D()
        self.head = layers.Dense(num_classes, activation="softmax")

    def call(self, x, training=None):
        """Forward pass of the layer.

        Args:
            x: Tensor of shape (B, H, W, C)

        Returns:
            The logits.
        """
        x = self.rescaling(x)
        if training:
            x = self.resizing_larger(x)
            x = self.random_crop(x)
            x = self.random_flip(x)
        else:
            x = self.resizing_target(x)

        x, h, w, c = self.patch_embed(x)
        for layer in self.basic_layers:
            x, h, w, c = layer(x, height=h, width=w, channels=c, training=training)
        return self.head(self.avgpool(self.norm(x)))


"""
## Train the model

Now with all the components in place and the architecture actually built, we are ready to
put it to good use.

In this section, we train our Focal Modulation model on the CIFAR-10 dataset.
"""

"""
### Visualization Callback

A key feature of the Focal Modulation Network is explicit input-dependency. This means
the modulator is calculated by looking at the local features around the target location,
so it depends on the input. In very simple terms, this makes interpretation easy. We can
simply lay down the gating values and the original image, next to each other to see how
the gating mechanism works.

The authors of the paper visualize the gates and the modulator in order to focus on the
interpretability of the Focal Modulation layer. Below is a visualization
callback that shows the gates and modulator of a specific layer in the model while the
model trains.

We will notice later that as the model trains, the visualizations get better.

The gates appear to selectively permit certain aspects of the input image to pass
through, while gently disregarding others, ultimately leading to improved classification
accuracy.
"""


def display_grid(test_images, gates, modulator):
    """Displays the image with the gates and modulator overlayed.

    Args:
        test_images: A batch of test images.
        gates: The gates of the Focal Modualtion Layer.
        modulator: The modulator of the Focal Modulation Layer.
    """
    test_images_np = ops.convert_to_numpy(test_images) / 255.0
    gates_np = ops.convert_to_numpy(gates)
    mod_np = ops.convert_to_numpy(ops.norm(modulator, axis=-1))

    num_gates = gates_np.shape[-1]
    idx = randint(0, test_images_np.shape[0] - 1)
    fig, ax = plt.subplots(1, num_gates + 2, figsize=((num_gates + 2) * 4, 4))

    ax[0].imshow(test_images_np[idx])
    ax[0].set_title("Original")
    ax[0].axis("off")
    for i in range(num_gates):
        ax[i + 1].imshow(test_images_np[idx])
        ax[i + 1].imshow(gates_np[idx, ..., i], cmap="inferno", alpha=0.6)
        ax[i + 1].set_title(f"Gate {i+1}")
        ax[i + 1].axis("off")

    ax[-1].imshow(test_images_np[idx])
    ax[-1].imshow(mod_np[idx], cmap="inferno", alpha=0.6)
    ax[-1].set_title("Modulator")
    ax[-1].axis("off")
    plt.show()
    plt.close()


"""
### TrainMonitor
"""

# Fetch test batch for callback
test_batch_images, _ = test_ds[0]


class TrainMonitor(keras.callbacks.Callback):
    def __init__(self, epoch_interval=10):
        super().__init__()
        self.epoch_interval = epoch_interval
        self.upsampler = layers.UpSampling2D(size=(4, 4), interpolation="bilinear")

    def on_epoch_end(self, epoch, logs=None):
        if (epoch + 1) % self.epoch_interval == 0:
            _ = self.model(test_batch_images, training=False)
            layer = self.model.basic_layers[1].blocks[-1].modulation
            display_grid(
                test_batch_images,
                self.upsampler(layer.gates),
                self.upsampler(layer.modulator),
            )


"""
### Learning Rate scheduler
"""


class WarmUpCosine(keras.optimizers.schedules.LearningRateSchedule):
    def __init__(self, lr_base, total_steps, warmup_steps):
        super().__init__()
        self.lr_base, self.total_steps, self.warmup_steps = (
            lr_base,
            total_steps,
            warmup_steps,
        )

    def __call__(self, step):
        step = ops.cast(step, "float32")
        cos_lr = (
            0.5
            * self.lr_base
            * (
                1
                + ops.cos(
                    np.pi
                    * (step - self.warmup_steps)
                    / (self.total_steps - self.warmup_steps)
                )
            )
        )
        warmup_lr = (self.lr_base / self.warmup_steps) * step
        return ops.where(
            step < self.warmup_steps,
            warmup_lr,
            ops.where(step > self.total_steps, 0.0, cos_lr),
        )


total_steps = (len(x_train) // BATCH_SIZE) * EPOCHS
scheduled_lrs = WarmUpCosine(LEARNING_RATE, total_steps, int(total_steps * 0.15))

"""
### Initialize, compile and train the model
"""

model = FocalModulationNetwork(image_size=(IMAGE_SIZE, IMAGE_SIZE))
model.compile(
    optimizer=keras.optimizers.AdamW(
        learning_rate=scheduled_lrs, weight_decay=WEIGHT_DECAY, clipnorm=1.0
    ),
    loss="sparse_categorical_crossentropy",
    metrics=["accuracy"],
)

history = model.fit(
    train_ds,
    validation_data=val_ds,
    epochs=EPOCHS,
    callbacks=[TrainMonitor(epoch_interval=5)],
)

"""
## Plot loss and accuracy
"""
plt.figure(figsize=(12, 4))
plt.subplot(1, 2, 1)
plt.plot(history.history["loss"], label="Train Loss")
plt.plot(history.history["val_loss"], label="Val Loss")
plt.legend()
plt.subplot(1, 2, 2)
plt.plot(history.history["accuracy"], label="Train Acc")
plt.plot(history.history["val_accuracy"], label="Val Acc")
plt.legend()
plt.show()

"""
## Test visualizations

Let's test our model on some test images and see how the gates look like.
"""
test_images, test_labels = next(iter(test_ds))

_ = model(test_images, training=False)

target_layer = model.basic_layers[1].blocks[-1].modulation
gates = target_layer.gates
modulator = target_layer.modulator

upsampler = layers.UpSampling2D(size=(4, 4), interpolation="bilinear")
gates_upsampled = upsampler(gates)
modulator_upsampled = upsampler(modulator)

for row in range(5):
    display_grid(
        test_images=test_images,
        gates=gates_upsampled,
        modulator=modulator_upsampled,
    )

"""
## Conclusion

The proposed architecture, the Focal Modulation Network
architecture is a mechanism that allows different
parts of an image to interact with each other in a way that depends on the image itself.
It works by first gathering different levels of context information around each part of
the image (the "query token"), then using a gate to decide which context information is
most relevant, and finally combining the chosen information in a simple but effective
way.

This is meant as a replacement of Self-Attention mechanism from the Transformer
architecture. The key feature that makes this research notable is not the conception of
attention-less networks, but rather the introduction of a equally powerful architecture
that is interpretable.

The authors also mention that they created a series of Focal Modulation Networks
(FocalNets) that significantly outperform Self-Attention counterparts and with a fraction
of parameters and pretraining data.

The FocalNets architecture has the potential to deliver impressive results and offers a
simple implementation. Its promising performance and ease of use make it an attractive
alternative to Self-Attention for researchers to explore in their own projects. It could
potentially become widely adopted by the Deep Learning community in the near future.

## Acknowledgement

We would like to thank [PyImageSearch](https://pyimagesearch.com/) for providing with a
Colab Pro account, [JarvisLabs.ai](https://cloud.jarvislabs.ai/) for GPU credits,
and also Microsoft Research for providing an
[official implementation](https://github.com/microsoft/FocalNet) of their paper.
We would also like to extend our gratitude to the first author of the
paper [Jianwei Yang](https://twitter.com/jw2yang4ai) who reviewed this tutorial
extensively.
"""

"""
## Relevant Chapters from Deep Learning with Python
- [Chapter 8: Image classification](https://deeplearningwithpython.io/chapters/chapter08_image-classification)
"""