Source code for ts.torch_handler.image_segmenter

"""
Module for image segmentation default handler
"""
import torch
from torchvision import transforms as T
from .vision_handler import VisionHandler

class ImageSegmenter(VisionHandler):
    """
    ImageSegmenter handler class. This handler takes a batch of images
    and returns output shape as [N K H W],
    where N - batch size, K - number of classes, H - height and W - width.
    """

    image_processing = T.Compose([
        T.Resize(256),
        T.CenterCrop(224),
        T.ToTensor(),
        T.Normalize(
            mean=[0.485, 0.456, 0.406],
            std=[0.229, 0.224, 0.225]
        )])

    def postprocess(self, data):
        # Returning the class for every pixel makes the response size too big
        # (> 24mb). Instead, we'll only return the top class for each image
        data = data['out']
        data = torch.nn.functional.softmax(data, dim=1)
        data = torch.max(data, dim=1)
        data = torch.stack([data.indices.type(data.values.dtype), data.values], dim=3)

        return data.tolist()