Skip to content

EfficientNetV2-S

EfficientNetV2-S is the smallest and most efficient model in the EfficientNetV2 family. Introduced in the paper "EfficientNetV2: Smaller Models and Faster Training", EfficientNetV2-S achieves state-of-the-art performance on image classification tasks, and it can be trained much faster and has a smaller model size of up to 6.8 times when compared to previous state-of-the-art models. It uses a combination of advanced techniques such as Swish activation function, Squeeze-and-Excitation blocks, and efficient channel attention to optimize its performance and efficiency.

Overall

  • Framework: PyTorch
  • Model format: ONNX
  • Model task: Image classification
  • Source: torchvision

Usages

from furiosa.models.vision import EfficientNetV2s
from furiosa.runtime.sync import create_runner

image = "tests/assets/cat.jpg"

effnetv2s = EfficientNetV2s()
with create_runner(effnetv2s.model_source()) as runner:
    inputs, _ = effnetv2s.preprocess(image)
    outputs = runner.run(inputs)
    effnetv2s.postprocess(outputs)

Inputs

The input is a 3-channel image of 384x384 (height, width).

  • Data Type: numpy.float32
  • Tensor Shape: [1, 3, 384, 384]
  • Memory Format: NCHW, where:
    • N - batch size
    • C - number of channels
    • H - image height
    • W - image width
  • Color Order: BGR
  • Optimal Batch Size (minimum: 1): <= 8

Outputs

The output is a numpy.float32 tensor with the shape ([1,]), including a class id. postprocess() transforms the class id to a label string.

Pre/Postprocessing

furiosa.models.vision.EfficientNetV2s class provides preprocess and postprocess methods that convert input images to input tensors and the model outputs to labels respectively. You can find examples at EfficientNetV2-S Usage.

furiosa.models.vision.EfficientNetV2s.preprocess

Read and preprocess an image located at image_path.

Parameters:

Name Type Description Default
image Union[str, Path, ArrayLike]

A path of an image.

 required
with_scaling bool

Whether to apply model-specific techniques that involve scaling the model's input and converting its data type to float32. Refer to the code to gain a precise understanding of the techniques used. Defaults to False.

 False

Returns:

Type Description
Tuple[ndarray, None]

The first element of the tuple is a numpy array that meets the input requirements of the model. The second element of the tuple is unused in this model and has no value. To learn more information about the output numpy array, please refer to Inputs.

furiosa.models.vision.EfficientNetV2s.postprocess

Convert the outputs of a model to a label string, such as car and cat.

Parameters:

Name Type Description Default
model_outputs Sequence[ArrayLike]

the outputs of the model. Please learn more about the output of model, please refer to Outputs.

 required

Returns:

Name Type Description
str str

A classified label, e.g., "tabby, tabby cat".

Notes on the source field of this model

There is a significant update in sdk version 0.9.0, which involves that the Furiosa's quantization tool adopts DFG(Data Flow Graph) as its output format instead of onnx. DFG is an IR of FuriosaAI that supports more diverse quantization schemes than onnx and is more specialized for FuriosaAI’s Warboy.

The EfficientNetV2-S we offer has been quantized with furiosa-sdk 0.9.0 and thus formatted in DFG. The ONNX file in the source field is the original f32 model, not yet quantized.

In case you need to use a different batch size or start from scratch, you can either start from the DFG or use the original ONNX file (and repeat the quantization process).