mlx-image

Image models based on Apple MLX framework for Apple Silicon machines.

Why?

Apple MLX framework is a great tool to run machine learning models on Apple Silicon machines.

This repository is meant to convert image models from timm/torchvision to Apple MLX framework. The weights are just converted from .pth to .npz/.safetensors and the models are not trained again.

How to install

pip install mlx-image

Models

Model weights are available on the mlx-vision community on HuggingFace.

To load a model with pre-trained weights:

from mlxim.model import create_model

# loading weights from HuggingFace (https://huggingface.co/mlx-vision/resnet18-mlxim)
model = create_model("resnet18") # pretrained weights loaded from HF

# loading weights from local file
model = create_model("resnet18", weights="path/to/resnet18/model.safetensors")

To list all available models:

from mlxim.model import list_models
list_models()

Supported models

List of all models available in mlx-image:

• ResNet: resnet18, resnet34, resnet50, resnet101, resnet152, wide_resnet50_2, wide_resnet101_2
• ViT:

  • supervised: vit_base_patch16_224, vit_base_patch16_224.swag_lin, vit_base_patch16_384.swag_e2e, vit_base_patch32_224, vit_large_patch16_224, vit_large_patch16_224, vit_large_patch16_224.swag_lin, vit_large_patch16_512.swag_e2e, vit_huge_patch14_224.swag_lin, vit_huge_patch14_518.swag_e2e

  • DINO v1: vit_base_patch16_224.dino, vit_small_patch16_224.dino, vit_small_patch8_224.dino, vit_base_patch8_224.dino

  • DINO v2: vit_small_patch14_518.dinov2, vit_base_patch14_518.dinov2, vit_large_patch14_518.dinov2

• Swin: swin_tiny_patch4_window7_224, swin_small_patch4_window7_224, swin_base_patch4_window7_224, swin_v2_tiny_patch4_window8_256, swin_v2_small_patch4_window8_256, swin_v2_base_patch4_window8_256
• RegNet: regnet_x_400mf, regnet_x_800mf, regnet_x_1_6gf, regnet_x_3_2gf, regnet_x_8gf, regnet_x_16gf, regnet_x_32gf, regnet_y_400mf, regnet_y_800mf, regnet_y_1_6gf, regnet_y_3_2gf, regnet_y_8gf, regnet_y_16gf, regnet_y_32gf, regnet_y_128gf
• EfficientNet: efficientnet_b0, efficientnet_b1, efficientnet_b2, efficientnet_b3, efficientnet_b4, efficientnet_b5, efficientnet_b6, efficientnet_b7
• MobileNet: mobilenet_v2, mobilenet_v3_large, mobilenet_v3_small

Warning: The regnet_y_128gf model couldn't be tested due to computational limitations.

ImageNet-1K Results

Go to results-imagenet-1k.csv to check every model converted to mlx-image and its performance on ImageNet-1K with different settings.

TL;DR performance is comparable to the original models from PyTorch implementations.

Similarity to PyTorch and other familiar tools

mlx-image tries to be as close as possible to PyTorch:

• DataLoader -> you can define your own collate_fn and also use num_workers to speed up data loading
• Dataset -> mlx-image already supports LabelFolderDataset (the good and old PyTorch ImageFolder) and FolderDataset (a generic folder with images in it)
• ModelCheckpoint -> keeps track of the best model and saves it to disk (similar to PyTorchLightning). It also suggests early stopping

Training

Training is similar to PyTorch. Here's an example of how to train a model:

import mlx.nn as nn
import mlx.optimizers as optim
from mlxim.model import create_model
from mlxim.data import LabelFolderDataset, DataLoader

train_dataset = LabelFolderDataset(
    root_dir="path/to/train",
    class_map={0: "class_0", 1: "class_1", 2: ["class_2", "class_3"]}
)
train_loader = DataLoader(
    dataset=train_dataset,
    batch_size=32,
    shuffle=True,
    num_workers=4
)
model = create_model("resnet18") # pretrained weights loaded from HF
optimizer = optim.Adam(learning_rate=1e-3)

def train_step(model, inputs, targets):
    logits = model(inputs)
    loss = mx.mean(nn.losses.cross_entropy(logits, target))
    return loss

model.train()
for epoch in range(10):
    for batch in train_loader:
        x, target = batch
        train_step_fn = nn.value_and_grad(model, train_step)
        loss, grads = train_step_fn(x, target)
        optimizer.update(model, grads)
        mx.eval(model.state, optimizer.state)

Validation

The validation.py script is run every time a pth model is converted to mlx and it's used to check if the model performs similarly to the original one on ImageNet-1K.

I use the configuration file config/validation.yaml to set the parameters for the validation script.

You can download the ImageNet-1K validation set from mlx-vision space on HuggingFace at this link.

Contributing

This is a work in progress, so any help is appreciated.

I am working on it in my spare time, so I can't guarantee frequent updates.

If you love coding and want to contribute, follow the instructions in CONTRIBUTING.md.

Additional Resources

• mlx-vision community
• HuggingFace doc

To-Dos

[ ] inference script (similar to train/validation)

[ ] DenseNet

[x] MobileNet

Contact

If you have any questions, please email riccardomusmeci92@gmail.com.