Skip to content

deepnets

This module provides tools to easily interact with deep learning models.

OTBTF is needed to use this module.

Super-resolution

The SR4RS model can be applied over any scenes.core.Source instance. We recall that this model is intended to be used over Sentinel-2 optical images. For example, here is how we perform the super-resolution of a Theia S2 product:

import scenes
archive = "SENTINEL2B_..._T31TEJ_D_V1-8.zip"
s2_scene = scenes.sentinel.Sentinel22AScene(archive)
s2_image = s2_scene.get_10m_bands()
sr = scenes.deepnets.sr4rs(s2_image)  # pyotb.core.OTBObject
sr.write("sr.tif")

inference(dic)

Generic function to perform deep nets inference.

When OTBTF is not found, a warning message is printed. Args: dic: otb parameters dict

Returns:

Type Description

pyotb App instance. When OTBTF is not found, None is returned.
Source code in scenes/deepnets.py 
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
def inference(dic: Dict[str, any]):
    """
    Generic function to perform deep nets inference.

    When OTBTF is not found, a warning message is printed.
    Args:
        dic: otb parameters dict

    Returns:
        pyotb App instance. When OTBTF is not found, None is returned.

    """
    output = None
    try:
        output = pyotb.TensorflowModelServe(dic)
    except AttributeError:
        print("OTBTF module has not been found in the system! It is mandatory "
              "to use deepnets. ")
    return output

sr4rs(input_image, model_url=SR4RS_MODEL_URL, tmp_dir='/tmp')

Applies the SR4RS model for super-resolution

See https://github.com/remicres/sr4rs for implementation details.

Parameters:

Name Type Description Default
input_image Union[str, OTBObject]

pyotb Input

 required
model_url str

SR4RS pre-trained model URL. Must point to a online .zip file.

 SR4RS_MODEL_URL
tmp_dir str

directory for temporary files.

 '/tmp'

Returns:

Type Description
App

pyotb Output
Source code in scenes/deepnets.py 
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
def sr4rs(
        input_image: Union[str, pyotb.core.OTBObject],
        model_url: str = SR4RS_MODEL_URL,
        tmp_dir: str = "/tmp"
) -> pyotb.core.App:
    """
    Applies the SR4RS model for super-resolution

    See https://github.com/remicres/sr4rs for implementation details.

    Args:
        input_image: pyotb Input
        model_url: SR4RS pre-trained model URL. Must point to a online .zip
            file.
        tmp_dir: directory for temporary files.

    Returns:
        pyotb Output

    """
    efield = 512
    gen_fcn = 64
    ratio = 0.25
    rfield = int((efield + 2 * gen_fcn) * ratio)

    # download the model if not already here
    tmp_zip = os.path.join(tmp_dir, os.path.basename(model_url))
    tmp_unzipped = os.path.splitext(tmp_zip)[0]
    if not os.path.exists(tmp_unzipped):
        download(model_url, out_filename=tmp_zip)
        with zipfile.ZipFile(tmp_zip, 'r') as zip_ref:
            print("Unzipping model...")
            zip_ref.extractall(tmp_dir)

    return inference({
        "source1.il": input_image,
        "source1.rfieldx": rfield,
        "source1.rfieldy": rfield,
        "source1.placeholder": "lr_input",
        "model.dir": tmp_unzipped,
        "model.fullyconv": "on",
        "output.names": f"output_{gen_fcn}",
        "output.efieldx": efield,
        "output.efieldy": efield,
        "output.spcscale": ratio,
    })