DPT Model Card

Table of Contents:

• Load trained model
• Model init parameters
• Dataset

Load trained model

1. Install requirements.

pip install -U segmentation_models_pytorch albumentations

2. Run inference.

import torch
import requests
import numpy as np
import albumentations as A
import segmentation_models_pytorch as smp

from PIL import Image

device = "cuda" if torch.cuda.is_available() else "cpu"

# Load pretrained model and preprocessing function
checkpoint = "smp-hub/dpt-large-ade20k"
model = smp.from_pretrained(checkpoint).eval().to(device)
preprocessing = A.Compose.from_pretrained(checkpoint)

# Load image
url = "https://huggingface.co/datasets/hf-internal-testing/fixtures_ade20k/resolve/main/ADE_val_00000001.jpg"
image = Image.open(requests.get(url, stream=True).raw)

# Preprocess image
np_image = np.array(image)
normalized_image = preprocessing(image=np_image)["image"]
input_tensor = torch.as_tensor(normalized_image)
input_tensor = input_tensor.permute(2, 0, 1).unsqueeze(0)  # HWC -> BCHW
input_tensor = input_tensor.to(device)

# Perform inference
with torch.no_grad():
    output_mask = model(input_tensor)

# Postprocess mask
mask = torch.nn.functional.interpolate(
    output_mask, size=(image.height, image.width), mode="bilinear", align_corners=False
)
mask = mask.argmax(1).cpu().numpy()  # argmax over predicted classes (channels dim)

Model init parameters

model_init_params = {
    "encoder_name": "tu-vit_large_patch16_384",
    "encoder_depth": 4,
    "encoder_weights": None,
    "encoder_output_indices": None,
    "decoder_intermediate_channels": (256, 512, 1024, 1024),
    "decoder_fusion_channels": 256,
    "dynamic_img_size": True,
    "in_channels": 3,
    "classes": 150,
    "activation": None,
    "aux_params": None
}

Dataset

Dataset name: ADE20K

More Information

• Library: https://github.com/qubvel/segmentation_models.pytorch
• Docs: https://smp.readthedocs.io/en/latest/

This model has been pushed to the Hub using the PytorchModelHubMixin