jigsolving/pooling.py at master · prolleytroblems/jigsolving · GitHub

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
import numpy as np
from numba import cuda
from image_obj import Solution
from utils import param_check
from datetime import datetime

DEFAULTS={"debug_mode":False, "pool_method":"avg"}

def pool(images_or_solution, pooling, stride, **params):
    params=param_check(params, DEFAULTS)

    if params["debug_mode"]==True:
        start=datetime.now()

    if isinstance(images_or_solution, np.ndarray) or isinstance(images_or_solution, list):
        assert len(np.array(images_or_solution).shape)==4
        pooled=[]
        for image in images_or_solution:
            pooled.append(pool_image(image, pooling, stride, **params))
        if params["debug_mode"]==True:
            print("Pooling images, "+params["pool_method"]+": "+str((datetime.now()-start).seconds*1000+float((datetime.now()-start).microseconds)/1000)+" ms")
        return np.array(pooled, dtype=np.uint8)

    elif isinstance(images_or_solution, Solution):
        dims=images_or_solution.shape
        images_or_solution = images_or_solution.arrays
        pooled=[]
        for image in images_or_solution:
            pooled.append(pool_image(image, pooling, stride, **params))
        if params["debug_mode"]==True:
            print("Pooling solution, "+params["pool_method"]+": "+str((datetime.now()-start).seconds*1000+float((datetime.now()-start).microseconds)/1000)+" ms")
        return Solution(np.array(pooled, dtype=np.uint8), dims)

    else: raise TypeError("images_or_solution must be an ndarray or Solution instance")

def pool_image(image, pooling, stride, pool_method="avg", **params):
    """Apply pooling to a single image. \n
        image    ndarray \n
        pooling  tuple of int of len=2 \n
        stride   tuple of int of len=2 \n """

    if not(pooling[0]>=stride[0] and pooling[1]>=stride[1]): raise TypeError("Pooling and stride inputs should be ndarray-like")
    if not(len(image.shape)==3): raise TypeError("Image must be a 3D ndarray.")

    pooling=np.array(pooling, dtype=np.uint8)
    stride=np.array(stride, dtype=np.uint8)

    for axis in range(2):
        image=add_padding(image, axis, pooling, stride, "end")

    final_dims=((image.shape[0]-pooling[0]+stride[0])//stride[0], (image.shape[1]-pooling[1]+stride[1])//stride[1], 3)
    tpb=16
    bpgy=(final_dims[0]-1)//tpb+1
    bpgx=(final_dims[1]-1)//tpb+1

    dimage=cuda.to_device(np.ascontiguousarray(image))
    pooled=np.array(np.zeros(final_dims), dtype=np.uint8)
    dpooled=cuda.to_device(pooled)

    func = {"max": max_pool_unit, "avg": avg_pool_unit}

    func[pool_method][(bpgy, bpgx), (tpb, tpb)](dimage, pooling, stride, dpooled)

    dpooled.to_host()

    return pooled

@cuda.jit
def max_pool_unit(image, pooling, stride, pooled):
    y, x=cuda.grid(2)

    if y>pooled.shape[0] or x>pooled.shape[1]:
        return
    window=image[y*stride[1]:y*stride[1]+pooling[1], x*stride[0]:x*stride[0]+pooling[0], :]

    pooled[y,x,0]=0
    pooled[y,x,1]=0
    pooled[y,x,2]=0

    for i in range(pooling[0]):
        for j in range(pooling[1]):
            if window[i,j,0]>pooled[y,x,0]:
                pooled[y,x,0]=window[i,j,0]
            if window[i,j,1]>pooled[y,x,1]:
                pooled[y,x,1]=window[i,j,1]
            if window[i,j,2]>pooled[y,x,2]:
                pooled[y,x,2]=window[i,j,2]

@cuda.jit
def avg_pool_unit(image, pooling, stride, pooled):
    y, x=cuda.grid(2)

    if y>pooled.shape[0] or x>pooled.shape[1]:
        return
    window=image[y*stride[1]:y*stride[1]+pooling[1], x*stride[0]:x*stride[0]+pooling[0], :]

    pooled[y,x,0]=0
    pooled[y,x,1]=0
    pooled[y,x,2]=0

    for i in range(pooling[0]):
        for j in range(pooling[1]):
            for c in range(3):
                pooled[y,x,c]+=window[i,j,c]
    for c in range(3):
        window[i,j,c]=window[i,j,c]/(pooling[0]/pooling[1])

def add_padding(image, axis, pooling, stride, side="end"):
    assert axis==0 or axis==1
    if (image.shape[axis]-pooling[axis]+stride[axis])%stride[axis]==0:
        return image
    else:
        #add a blank row/column
        coeff=[0,0]
        coeff[-axis+1]=1
        if side=="end":
            image=np.concatenate((image, np.zeros(((image.shape[0]-1)*coeff[0]+1,
                    (image.shape[1]-1)*coeff[1]+1, 3), dtype=np.uint8)), axis=axis)
            image=add_padding(image, axis, pooling, stride, side="start")
        elif side=="start":
            image=np.concatenate((np.zeros(((image.shape[0]-1)*coeff[0]+1,
                    (image.shape[1]-1)*coeff[1]+1, 3), dtype=np.uint8), image), axis=axis)
            image=add_padding(image, axis, pooling, stride, side="end")
        return image