habana_frameworks.mediapipe.fn.Resize

Class:

• habana_frameworks.mediapipe.fn.Resize(**kwargs)

Define graph call:

• __call__(input)

Parameter:

• input - Input tensor to operator. Supported dimensions: minimum = 3, maximum = 4. Supported data types: INT8, UINT8, UINT16. 3D tensor or 4D tensor with last dimension size as 1, containing image data. Input image data in (N*C)xWxHx1 data layout with (N*C) being the fastest changing dimension.

Description:

Resize an image by the scale factor using either bilinear, cubic, nearest neighbor sampling.

Supported backend:

• HPU

Keyword Arguments

kwargs	Description
mode	Interpolation mode to be selected. Type: ResizeInterpolationMode_t Default: 1 Optional: yes Supported modes: RESIZE_INTER_NEAREST: 0 Does nearest neighbor interpolation for 1D tensor and N-neighbor interpolation for N-D tensor RESIZE_INTER_LINEAR: 1 Does linear interpolation for 1D tensor and N-linear interpolation for N-D tensor RESIZE_INTER_CUBIC: 2 Does cubic interpolation for 1D tensor and N-cubic interpolation for N-D tensor
cubicCoeffA	Is valid only if mode is cubic The coefficient ‘a’ used in cubic interpolation. Two common choice are -0.5 and -0.75. Check https://ieeexplore.ieee.org/document/1163711 for details. Type: float Default: -0.75 Optional: yes
scaleDim1	Scaling along width axis. Type: float Default: 0.0 Optional: yes
scaleDim2	Scaling along height axis. Type: float Default: 0.0 Optional: yes
scaleDim3	Scaling along batch axis. Type: float Default: 0.0 Optional: yes
coordTransMode	Describes how to transform the coordinate in the resized tensor to the coordinate in the original tensor. Type: ResizeCoordinateTransformationMode_t Default: HALF_PIXEL_MODE Optional: yes Supported modes: if HALF_PIXEL_MODE then x_original = (x_resized + 0.5) / scale - 0.5 if PYTORCH_HALF_PIXEL_MODE then x_original = length_resized > 1 ? (x_resized + 0.5) / scale - 0.5 : 0 if ALIGN_CORNERS_MODE then x_original = x_resized * (length_original - 1) / (length_resized - 1) if ASYMMETRIC_MODE then x_original = x_resized / scale if TF_HALF_PIXEL_FOR_NN_MODE then x_original = (x_resized + 0.5) / scale
nearestMode	Is valid only if mode is nearest interpolation. It indicates how to get nearest pixel in input tensor. Type: ResizeNearestMode_t Default: ROUND_PREFER_FLOOR Optional: yes Supported modes: ROUND_PREFER_FLOOR (also know as round half down) ROUND_PREFER_CEIL (also know as round half up) FLOOR CEIL
excludeOutside	If set to True then the weight of sampling locations outside the tensor will be set to 0 and the weight will be re-normalized so that their sum is 1.0. Type: bool Default: False Optional: yes
useScales	Determine whether scaling will be applied to each dimensions. If False then output tensor size is determined by size. Type: bool Default: False Optional: yes
size1	Width of the output tensor. This is set if useScales is set to false. Type: int Default: 0 Optional: yes
size2	Height of the output tensor. This is set if useScales is set to false. Type: int Default: 0 Optional: yes
size3	Batch of the output tensor. This is set if useScales is set to false. Type: int Default: 1 Optional: yes
dtype	Output data type. Type: habana_frameworks.mediapipe.media_types.dtype Default: UINT8 Optional: yes Supported data type: INT8 UINT8 UINT16

Note

1. All Input/Output tensor data type must have same data type.

2. Input/Output tensor should be in (N*C)xWxH data layout. (N*C) being the fastest changing dimension.

3. Resize linear mode supports only bilinear interpolation along dimensions 1 (Width) and 2 (Height) only.

4. Resize neighbor mode supports simultaneous scaling of three non fast changing dimensions.

5. Resize cubic mode supports only bicubic interpolation along dimensions 1 (Width) and 2 (Height) only.

Example: Resize Operator

The following code snippet shows usage of Resize operator. Reshape and transpose operations are performed before resize operation to arrange tensor layout to (N*C)WH1 as required by resize operation. Post reshape and transpose convert the tensor layout to CWHN.

from habana_frameworks.mediapipe import fn
from habana_frameworks.mediapipe.mediapipe import MediaPipe
from habana_frameworks.mediapipe.media_types import imgtype as it
from habana_frameworks.mediapipe.media_types import dtype as dt
import matplotlib.pyplot as plt

# Create media pipeline derived class
class myMediaPipe(MediaPipe):
    def __init__(self, device, dir, queue_depth, batch_size, channel, img_h, img_w, resize_h, resize_w):
        super(
            myMediaPipe,
            self).__init__(
            device,
            queue_depth,
            batch_size,
            self.__class__.__name__)

        self.input = fn.ReadImageDatasetFromDir(shuffle=False,
                                                dir=dir,
                                                format="jpg")

        # WHCN
        self.decode = fn.ImageDecoder(device="hpu",
                                      output_format=it.RGB_P,
                                      resize=[img_w, img_h])

        # WHCN -> NCWH
        self.pre_transpose = fn.Transpose(permutation=[3, 2, 0, 1],
                                          tensorDim=4,
                                          dtype=dt.UINT8)

        # NCWH -> (N*C)WH1
        self.pre_reshape = fn.Reshape(size=[batch_size*channel, img_w, img_h, 1],
                                      tensorDim=4,
                                      layout='',
                                      dtype=dt.UINT8)

        self.resize = fn.Resize(mode=1,
                                size1=resize_w,
                                size2=resize_h,
                                size3=1,
                                dtype=dt.UINT8)

        # (N*C)WH1 -> NCWH
        self.post_reshape = fn.Reshape(size=[batch_size, channel, resize_w, resize_h],
                                       tensorDim=4,
                                       layout='',
                                       dtype=dt.UINT8)

        # NCWH -> CWHN
        self.post_transpose = fn.Transpose(permutation=[1, 2, 3, 0],
                                           tensorDim=4,
                                           dtype=dt.UINT8)

    def definegraph(self):
        images, labels = self.input()
        images = self.decode(images)
        images = self.pre_transpose(images)
        images = self.pre_reshape(images)
        images = self.resize(images)
        images = self.post_reshape(images)
        images = self.post_transpose(images)
        return images, labels

def display_images(images, batch_size, cols):
    rows = (batch_size + 1) // cols
    plt.figure(figsize=(10, 10))
    for i in range(batch_size):
        ax = plt.subplot(rows, cols, i + 1)
        plt.imshow(images[i])
        plt.axis("off")
    plt.show()


def main():
    batch_size = 6
    channels = 3
    img_width = 200
    img_height = 200
    resize_width = 300
    resize_height = 300
    img_dir = "/path/to/images"
    queue_depth = 2
    columns = 3

    # Create media pipeline object
    pipe = myMediaPipe('hpu', img_dir, queue_depth, batch_size,
                        channels, img_height, img_width, resize_width, resize_height)

    # Build media pipeline
    pipe.build()

    # Initialize media pipeline iterator
    pipe.iter_init()

    # Run media pipeline
    images, labels = pipe.run()

    # Copy data to host from device as numpy array
    images = images.as_cpu().as_nparray()
    labels = labels.as_cpu().as_nparray()

    # Display images
    display_images(images, batch_size, columns)

if __name__ == "__main__":
    main()

Resized Images 1

1

Licensed under a CC BY SA 4.0 license. The images used here are taken from https://data.caltech.edu/records/mzrjq-6wc02.