File size: 3,987 Bytes
c85d0ce
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
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
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
# The file source is from the [ESRGAN](https://github.com/xinntao/ESRGAN) project 
# forked by authors [joeyballentine](https://github.com/joeyballentine/ESRGAN) and [BlueAmulet](https://github.com/BlueAmulet/ESRGAN).

import gc

import numpy as np
import torch


def bgr_to_rgb(image: torch.Tensor) -> torch.Tensor:
    # flip image channels
    # https://github.com/pytorch/pytorch/issues/229
    out: torch.Tensor = image.flip(-3)
    # out: torch.Tensor = image[[2, 1, 0], :, :] #RGB to BGR #may be faster
    return out


def rgb_to_bgr(image: torch.Tensor) -> torch.Tensor:
    # same operation as bgr_to_rgb(), flip image channels
    return bgr_to_rgb(image)


def bgra_to_rgba(image: torch.Tensor) -> torch.Tensor:
    out: torch.Tensor = image[[2, 1, 0, 3], :, :]
    return out


def rgba_to_bgra(image: torch.Tensor) -> torch.Tensor:
    # same operation as bgra_to_rgba(), flip image channels
    return bgra_to_rgba(image)


def auto_split_upscale(
    lr_img: np.ndarray,
    upscale_function,
    scale: int = 4,
    overlap: int = 32,
    max_depth: int = None,
    current_depth: int = 1,
):
    # Attempt to upscale if unknown depth or if reached known max depth
    if max_depth is None or max_depth == current_depth:
        try:
            print(f"auto_split_upscale, current depth: {current_depth}")
            result, _ = upscale_function(lr_img, scale)
            return result, current_depth
        except RuntimeError as e:
            # Check to see if its actually the CUDA out of memory error
            if "CUDA" in str(e):
                # Collect garbage (clear VRAM)
                torch.cuda.empty_cache()
                gc.collect()
            # Re-raise the exception if not an OOM error
            else:
                raise RuntimeError(e)
        finally:
            # Free GPU memory and clean up resources
            torch.cuda.empty_cache()
            gc.collect()

    h, w, c = lr_img.shape

    # Split image into 4ths
    top_left = lr_img[: h // 2 + overlap, : w // 2 + overlap, :]
    top_right = lr_img[: h // 2 + overlap, w // 2 - overlap :, :]
    bottom_left = lr_img[h // 2 - overlap :, : w // 2 + overlap, :]
    bottom_right = lr_img[h // 2 - overlap :, w // 2 - overlap :, :]

    # Recursively upscale the quadrants
    # After we go through the top left quadrant, we know the maximum depth and no longer need to test for out-of-memory
    top_left_rlt, depth = auto_split_upscale(
        top_left,
        upscale_function,
        scale=scale,
        overlap=overlap,
        max_depth=max_depth,
        current_depth=current_depth + 1,
    )
    top_right_rlt, _ = auto_split_upscale(
        top_right,
        upscale_function,
        scale=scale,
        overlap=overlap,
        max_depth=depth,
        current_depth=current_depth + 1,
    )
    bottom_left_rlt, _ = auto_split_upscale(
        bottom_left,
        upscale_function,
        scale=scale,
        overlap=overlap,
        max_depth=depth,
        current_depth=current_depth + 1,
    )
    bottom_right_rlt, _ = auto_split_upscale(
        bottom_right,
        upscale_function,
        scale=scale,
        overlap=overlap,
        max_depth=depth,
        current_depth=current_depth + 1,
    )

    # Define output shape
    out_h = h * scale
    out_w = w * scale

    # Create blank output image
    output_img = np.zeros((out_h, out_w, c), np.uint8)

    # Fill output image with tiles, cropping out the overlaps
    output_img[: out_h // 2, : out_w // 2, :] = top_left_rlt[
        : out_h // 2, : out_w // 2, :
    ]
    output_img[: out_h // 2, -out_w // 2 :, :] = top_right_rlt[
        : out_h // 2, -out_w // 2 :, :
    ]
    output_img[-out_h // 2 :, : out_w // 2, :] = bottom_left_rlt[
        -out_h // 2 :, : out_w // 2, :
    ]
    output_img[-out_h // 2 :, -out_w // 2 :, :] = bottom_right_rlt[
        -out_h // 2 :, -out_w // 2 :, :
    ]

    return output_img, depth