resize image in python from one location to another. Not a clean one

Resize_image.py

from pathlib import Path
from tqdm import tqdm  
import pydicom
import numpy as np
from PIL import Image

sz=224
PATH = Path('big')
img_dir = PATH/'image'
TEST = False

sample = list(img_dir.iterdir())
if TEST:
    sample = sample[:1000]
    IMG_DIR = Path('sample')/str(sz)
    tloc = PATH/'sample'/str(sz)
else:
    IMG_DIR = Path('all')/str(sz)
    tloc = PATH/'all'/str(sz)
train, test = train_test_split(sample)

def scale_to(x, ratio, targ): 
    '''Calculate dimension of an image during scaling with aspect ratio'''
    return max(math.floor(x*ratio), targ)

def read_image_resize(tsize=sz, tloc=tloc):
    tloc.mkdir(parents=True, exist_ok=True)
    new_created = list(tloc.iterdir())
    to_be_created = set([x.stem for x in sample if x.is_file()]).difference([x.stem for x in new_created if x.is_file()])
    to_be_created = [PATH/'stage_1_train_images'/(f+'.dcm') for f in to_be_created]
    for loc in tqdm(to_be_created):
        img_arr = pydicom.read_file(loc.as_posix()).pixel_array
        img_arr = np.stack([img_arr] * 3, axis=2)
        img_arr = img_arr/img_arr.max()
        img_arr = (255*img_arr).clip(0, 255).astype(np.uint8)
        img_arr = Image.fromarray(img_arr)#.convert('RGB') # model expects 3 channel image
        r,c = img_arr.size
        ratio = tsize/min(r,c)
        sz = (scale_to(r, ratio, tsize), scale_to(c, ratio, tsize))
        img_arr = img_arr.resize(sz, Image.LINEAR)
        filepath = tloc / loc.name
#         np.save(filepath, img_arr, )
        img_arr.save((tloc/loc.stem).as_posix()+'.png')


read_image_resize()

It needs multiprocessing support. Copying 25k files (4Gb) took over 1hr.

With multi processing

from multiprocessing import Pool
from pathlib import Path
from tqdm import tqdm  
import pydicom
import numpy as np
from PIL import Image

img_dir = PATH/'stage_1_train_images'
sample = list(img_dir.iterdir())
SAMPLE_SIZE = 10
if TEST:
    sample = sample[:SAMPLE_SIZE]
train, test = train_test_split(sample)
batch_size=32
sz=224
tloc = Path('/tmp/pytorch')/str(sz)
tloc.mkdir(parents=True, exist_ok=True)

def scale_to(x, ratio, targ): 
    '''Calculate dimension of an image during scaling with aspect ratio'''
    return max(math.floor(x*ratio), targ)


def resize_image(loc, tsize=256, tloc=tloc):
    img_arr = pydicom.read_file(loc.as_posix()).pixel_array
    img_arr = np.stack([img_arr] * 3, axis=2)
    img_arr = img_arr/img_arr.max()
    img_arr = (255*img_arr).clip(0, 255).astype(np.uint8)
    img_arr = Image.fromarray(img_arr)#.convert('RGB') # model expects 3 channel image
    r,c = img_arr.size
    ratio = tsize/min(r,c)
    sz = (scale_to(r, ratio, tsize), scale_to(c, ratio, tsize))
    img_arr = img_arr.resize(sz, Image.LINEAR)
    filepath = tloc / loc.name
#         np.save(filepath, img_arr, )
    img_arr.save((tloc/loc.stem).as_posix()+'.png')
    
pool = Pool()                         # Create a multiprocessing Pool
pool.map(resize_image, tqdm(sample))