Image-Resizing-Tests.ipynb

resizing_correctness_tests.py

import os
import shutil
import sys
from pathlib import Path

import cv2
import numpy as np
import nvidia.dali.fn as fn
import nvidia.dali.types as types
import torch, torchvision, nvidia.dali, numpy, cv2

ver = lambda x: f"{x.__name__.ljust(14)}: {x.__version__.rjust(3)}"
print("\n".join(list(map(ver, [torch, torchvision, nvidia.dali, numpy, cv2]))))

from nvidia.dali import pipeline_def
from PIL import Image
from PIL.ImageDraw import ImageDraw
from torchvision.transforms.functional import InterpolationMode
from torchvision.transforms.functional import resize as torch_resize
from torchvision.transforms.functional import to_pil_image, to_tensor

mkdir = lambda x: Path(x).mkdir()


def create_test_img(size=(128, 128), outline_width: int = 3):
    assert size[0] == size[1], f"Square image dimensions expected"

    im = Image.new("RGB", size, color=(255, 255, 255))
    d = ImageDraw(im)
    qtr = int(size[0] / 4)
    pad = int(0.015625 * size[0])
    d.ellipse((qtr, qtr, qtr * 3, qtr * 3), fill=None, outline=0, width=outline_width)
    d.rectangle(
        ((pad, pad), (size[0] - pad, size[1] - pad)), fill=None, outline=0, width=outline_width
    )

    return im


def resize_opencv(img: Image.Image, size=(128, 128), interpolation=cv2.INTER_LINEAR):
    # im = cv2.cvtColor(np.array(img), cv2.COLOR_BGR2RGB)
    im = np.array(img)
    im = cv2.resize(im, size, interpolation)
    return Image.fromarray(im)


def resize_torch(
    img: Image.Image,
    size=(128, 128),
    interpolation=InterpolationMode.BICUBIC,
    antialias: bool = None,
):
    im = torch_resize(to_tensor(img), size, interpolation, antialias=antialias)
    #     return im
    return to_pil_image(im)


def resize_dali(img: str, size=(128, 128), interpolation=types.INTERP_LINEAR, device="gpu"):
    # Setup directory structure for DALI
    if Path("tmp-test").exists():
        shutil.rmtree("tmp-test")
    mkdir("tmp-test")
    mkdir("tmp-test/0")
    img.save("tmp-test/0/test-img.png", quality=100)

    # Define pipeline
    @pipeline_def
    def test():
        imgs, _ = fn.readers.file(file_root="tmp-test")
        imgs = fn.decoders.image(imgs, device="mixed" if device == "gpu" else "cpu")
        imgs = fn.resize(
            imgs,
            device="gpu",
            resize_x=size[0],
            resize_y=size[0],
            mode="stretch",
            interp_type=interpolation,
        )
        return imgs, _

    # Build Pipeline
    pipe = test(batch_size=1, num_threads=1, device_id=0)
    pipe.build()

    # Run Pipeline and get outputs
    img = pipe.run()[0]
    img = img.as_cpu().as_array()[0]  # 0th element of the batch
    assert img.ndim == 3

    return Image.fromarray(img)


img = create_test_img((2048, 2048), outline_width=6)
size = (128, 128)  # Output size


# -- DALI
# --- Correct outputs
resize_dali(img, size, interpolation=types.INTERP_GAUSSIAN)
resize_dali(img, size, interpolation=types.INTERP_LANCZOS3)
resize_dali(img, size, interpolation=types.INTERP_TRIANGULAR)

# --- Wrong outputs
resize_dali(img, size, interpolation=types.INTERP_CUBIC)
resize_dali(img, size, interpolation=types.INTERP_LINEAR)
resize_dali(img, size, interpolation=types.INTERP_NN)


# -- PIL
# --- Correct outputs
img.resize(size, Image.BILINEAR)
img.resize(size, Image.BICUBIC)
img.resize(size, Image.LANCZOS)
img.resize(size, Image.BOX)
img.resize(size, Image.HAMMING)

# --- Wrong outputs
# None!

# -- Torch
# --- Correct outputs
resize_torch(img, size, interpolation=InterpolationMode.BILINEAR, antialias=True)

# --- Wrong outputs
resize_torch(img, size, interpolation=InterpolationMode.BILINEAR)
resize_torch(img, size, interpolation=InterpolationMode.BICUBIC)
resize_torch(img, size, interpolation=InterpolationMode.BICUBIC, antialias=True)

# -- OpenCV
# --- Correct outputs
# None!
# --- Wrong outputs
resize_opencv(img, size, cv2.INTER_NEAREST)
resize_opencv(img, size, cv2.INTER_LINEAR)
resize_opencv(img, size, cv2.INTER_AREA)
resize_opencv(img, size, cv2.INTER_CUBIC)
resize_opencv(img, size, cv2.INTER_LANCZOS4)