keras/mit_b3_ade20k_512

Model Overview

A Keras model implementing the MixTransformer architecture to be used as a backbone for the SegFormer architecture. This model is supported in both KerasCV and KerasHub. KerasCV will no longer be actively developed, so please try to use KerasHub.

References:

• SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers # noqa: E501
• Based on the TensorFlow implementation from DeepVision # noqa: E501

Links

• MiT Quickstart Notebook: coming soon
• MiT API Documentation: coming soon

Installation

Keras and KerasHub can be installed with:

pip install -U -q keras-Hub
pip install -U -q keras>=3

Jax, TensorFlow, and Torch come preinstalled in Kaggle Notebooks. For instructions on installing them in another environment see the Keras Getting Started page.

Presets

The following model checkpoints are provided by the Keras team. Weights have been ported from https://dl.fbaipublicfiles.com/segment_anything/. Full code examples for each are available below. Here's the table formatted similarly to the given pattern:

Here's the updated table with the input resolutions included in the descriptions:

Preset name	Parameters	Description
mit_b0_ade20k_512	3.32M	MiT (MixTransformer) model with 8 transformer blocks, trained on the ADE20K dataset with an input resolution of 512x512 pixels.
mit_b1_ade20k_512	13.16M	MiT (MixTransformer) model with 8 transformer blocks, trained on the ADE20K dataset with an input resolution of 512x512 pixels.
mit_b2_ade20k_512	24.20M	MiT (MixTransformer) model with 16 transformer blocks, trained on the ADE20K dataset with an input resolution of 512x512 pixels.
mit_b3_ade20k_512	44.08M	MiT (MixTransformer) model with 28 transformer blocks, trained on the ADE20K dataset with an input resolution of 512x512 pixels.
mit_b4_ade20k_512	60.85M	MiT (MixTransformer) model with 41 transformer blocks, trained on the ADE20K dataset with an input resolution of 512x512 pixels.
mit_b5_ade20k_640	81.45M	MiT (MixTransformer) model with 52 transformer blocks, trained on the ADE20K dataset with an input resolution of 640x640 pixels.
mit_b0_cityscapes_1024	3.32M	MiT (MixTransformer) model with 8 transformer blocks, trained on the Cityscapes dataset with an input resolution of 1024x1024 pixels.
mit_b1_cityscapes_1024	13.16M	MiT (MixTransformer) model with 8 transformer blocks, trained on the Cityscapes dataset with an input resolution of 1024x1024 pixels.
mit_b2_cityscapes_1024	24.20M	MiT (MixTransformer) model with 16 transformer blocks, trained on the Cityscapes dataset with an input resolution of 1024x1024 pixels.
mit_b3_cityscapes_1024	44.08M	MiT (MixTransformer) model with 28 transformer blocks, trained on the Cityscapes dataset with an input resolution of 1024x1024 pixels.
mit_b4_cityscapes_1024	60.85M	MiT (MixTransformer) model with 41 transformer blocks, trained on the Cityscapes dataset with an input resolution of 1024x1024 pixels.
mit_b5_cityscapes_1024	81.45M	MiT (MixTransformer) model with 52 transformer blocks, trained on the Cityscapes dataset with an input resolution of 1024x1024 pixels.

Example Usage

Using the class with a backbone:

import tensorflow as tf
import keras_cv
import numpy as np

images = np.ones(shape=(1, 96, 96, 3))
labels = np.zeros(shape=(1, 96, 96, 1))
backbone = keras_cv.models.MiTBackbone.from_preset("mit_b3_ade20k_512")

# Evaluate model
model(images)

# Train model
model.compile(
     optimizer="adam",
     loss=keras.losses.BinaryCrossentropy(from_logits=False),
     metrics=["accuracy"],
)
model.fit(images, labels, epochs=3)

Example Usage with Hugging Face URI

Using the class with a backbone:

import tensorflow as tf
import keras_cv
import numpy as np

images = np.ones(shape=(1, 96, 96, 3))
labels = np.zeros(shape=(1, 96, 96, 1))
backbone = keras_cv.models.MiTBackbone.from_preset("hf://keras/mit_b3_ade20k_512")

# Evaluate model
model(images)

# Train model
model.compile(
     optimizer="adam",
     loss=keras.losses.BinaryCrossentropy(from_logits=False),
     metrics=["accuracy"],
)
model.fit(images, labels, epochs=3)