File size: 15,409 Bytes
e9f9fd3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
"Image transformations for data augmentation. All transforms are done on the tensor level"
from ..torch_core import *
from .image import *
from .image import _affine_mult

__all__ = ['brightness', 'contrast', 'crop', 'crop_pad', 'cutout', 'dihedral', 'dihedral_affine', 'flip_affine', 'flip_lr',
           'get_transforms', 'jitter', 'pad', 'perspective_warp', 'rand_pad', 'rand_crop', 'rand_zoom', 'rgb_randomize', 'rotate', 'skew', 'squish',
           'rand_resize_crop', 'symmetric_warp', 'tilt', 'zoom', 'zoom_crop']

_pad_mode_convert = {'reflection':'reflect', 'zeros':'constant', 'border':'replicate'}

#NB: Although TfmLighting etc can be used as decorators, that doesn't work in Windows,
#    so we do it manually for now.

def _brightness(x, change:uniform):
    "Apply `change` in brightness of image `x`."
    return x.add_(scipy.special.logit(change))
brightness = TfmLighting(_brightness)

def _contrast(x, scale:log_uniform):
    "Apply `scale` to contrast of image `x`."
    return x.mul_(scale)
contrast = TfmLighting(_contrast)

def _rotate(degrees:uniform):
    "Rotate image by `degrees`."
    angle = degrees * math.pi / 180
    return [[float(cos(angle)), float(-sin(angle)), 0.],
            [float(sin(angle)),  float(cos(angle)), 0.],
            [0.        ,  0.        , 1.]]
rotate = TfmAffine(_rotate)

def _get_zoom_mat(sw:float, sh:float, c:float, r:float)->AffineMatrix:
    "`sw`,`sh` scale width,height - `c`,`r` focus col,row."
    return [[sw, 0,  c],
            [0, sh,  r],
            [0,  0, 1.]]

def _zoom(scale:uniform=1.0, row_pct:uniform=0.5, col_pct:uniform=0.5):
    "Zoom image by `scale`. `row_pct`,`col_pct` select focal point of zoom."
    s = 1-1/scale
    col_c = s * (2*col_pct - 1)
    row_c = s * (2*row_pct - 1)
    return _get_zoom_mat(1/scale, 1/scale, col_c, row_c)
zoom = TfmAffine(_zoom)

def _squish(scale:uniform=1.0, row_pct:uniform=0.5, col_pct:uniform=0.5):
    "Squish image by `scale`. `row_pct`,`col_pct` select focal point of zoom."
    if scale <= 1:
        col_c = (1-scale) * (2*col_pct - 1)
        return _get_zoom_mat(scale, 1, col_c, 0.)
    else:
        row_c = (1-1/scale) * (2*row_pct - 1)
        return _get_zoom_mat(1, 1/scale, 0., row_c)
squish = TfmAffine(_squish)

def _jitter(c, magnitude:uniform):
    "Replace pixels by random neighbors at `magnitude`."
    c.flow.add_((torch.rand_like(c.flow)-0.5)*magnitude*2)
    return c
jitter = TfmCoord(_jitter)

def _flip_lr(x):
    "Flip `x` horizontally."
    #return x.flip(2)
    if isinstance(x, ImagePoints):
        x.flow.flow[...,0] *= -1
        return x
    return tensor(np.ascontiguousarray(np.array(x)[...,::-1]))
flip_lr = TfmPixel(_flip_lr)

def _flip_affine() -> TfmAffine:
    "Flip `x` horizontally."
    return [[-1, 0, 0.],
            [0,  1, 0],
            [0,  0, 1.]]
flip_affine = TfmAffine(_flip_affine)

def _dihedral(x, k:partial(uniform_int,0,7)):
    "Randomly flip `x` image based on `k`."
    flips=[]
    if k&1: flips.append(1)
    if k&2: flips.append(2)
    if flips: x = torch.flip(x,flips)
    if k&4: x = x.transpose(1,2)
    return x.contiguous()
dihedral = TfmPixel(_dihedral)

def _dihedral_affine(k:partial(uniform_int,0,7)):
    "Randomly flip `x` image based on `k`."
    x = -1 if k&1 else 1
    y = -1 if k&2 else 1
    if k&4: return [[0, x, 0.],
                    [y, 0, 0],
                    [0, 0, 1.]]
    return [[x, 0, 0.],
            [0, y, 0],
            [0, 0, 1.]]
dihedral_affine = TfmAffine(_dihedral_affine)

def _pad_coord(x, row_pad:int, col_pad:int, mode='zeros'):
    #TODO: implement other padding modes than zeros?
    h,w = x.size
    pad = torch.Tensor([w/(w + 2*col_pad), h/(h + 2*row_pad)])
    x.flow = FlowField((h+2*row_pad, w+2*col_pad) , x.flow.flow * pad[None])
    return x

def _pad_default(x, padding:int, mode='reflection'):
    "Pad `x` with `padding` pixels. `mode` fills in space ('zeros','reflection','border')."
    mode = _pad_mode_convert[mode]
    return F.pad(x[None], (padding,)*4, mode=mode)[0]

def _pad_image_points(x, padding:int, mode='reflection'):
    return _pad_coord(x, padding, padding, mode)

def _pad(x, padding:int, mode='reflection'):
    f_pad = _pad_image_points if isinstance(x, ImagePoints) else  _pad_default
    return f_pad(x, padding, mode)

pad = TfmPixel(_pad, order=-10)

def _cutout(x, n_holes:uniform_int=1, length:uniform_int=40):
    "Cut out `n_holes` number of square holes of size `length` in image at random locations."
    h,w = x.shape[1:]
    for n in range(n_holes):
        h_y = np.random.randint(0, h)
        h_x = np.random.randint(0, w)
        y1 = int(np.clip(h_y - length / 2, 0, h))
        y2 = int(np.clip(h_y + length / 2, 0, h))
        x1 = int(np.clip(h_x - length / 2, 0, w))
        x2 = int(np.clip(h_x + length / 2, 0, w))
        x[:, y1:y2, x1:x2] = 0
    return x

cutout = TfmPixel(_cutout, order=20)

def _rgb_randomize(x, channel:int=None, thresh:float=0.3):
    "Randomize one of the channels of the input image"
    if channel is None: channel = np.random.randint(0, x.shape[0] - 1)
    x[channel] = torch.rand(x.shape[1:]) * np.random.uniform(0, thresh)
    return x

rgb_randomize = TfmPixel(_rgb_randomize)

def _minus_epsilon(row_pct:float, col_pct:float, eps:float=1e-7):
    if row_pct==1.: row_pct -= 1e-7
    if col_pct==1.: col_pct -= 1e-7
    return row_pct,col_pct

def _crop_default(x, size, row_pct:uniform=0.5, col_pct:uniform=0.5):
    "Crop `x` to `size` pixels. `row_pct`,`col_pct` select focal point of crop."
    rows,cols = tis2hw(size)
    row_pct,col_pct = _minus_epsilon(row_pct,col_pct)
    row = int((x.size(1)-rows+1) * row_pct)
    col = int((x.size(2)-cols+1) * col_pct)
    return x[:, row:row+rows, col:col+cols].contiguous()

def _crop_image_points(x, size, row_pct=0.5, col_pct=0.5):
    h,w = x.size
    rows,cols = tis2hw(size)
    row_pct,col_pct = _minus_epsilon(row_pct,col_pct)
    x.flow.flow.mul_(torch.Tensor([w/cols, h/rows])[None])
    row = int((h-rows+1) * row_pct)
    col = int((w-cols+1) * col_pct)
    x.flow.flow.add_(-1 + torch.Tensor([w/cols-2*col/cols, h/rows-2*row/rows])[None])
    x.size = (rows, cols)
    return x

def _crop(x, size, row_pct:uniform=0.5, col_pct:uniform=0.5):
    f_crop = _crop_image_points if isinstance(x, ImagePoints) else _crop_default
    return f_crop(x, size, row_pct, col_pct)

crop = TfmPixel(_crop)

def _crop_pad_default(x, size, padding_mode='reflection', row_pct:uniform = 0.5, col_pct:uniform = 0.5):
    "Crop and pad tfm - `row_pct`,`col_pct` sets focal point."
    padding_mode = _pad_mode_convert[padding_mode]
    size = tis2hw(size)
    if x.shape[1:] == torch.Size(size): return x
    rows,cols = size
    row_pct,col_pct = _minus_epsilon(row_pct,col_pct)
    if x.size(1)<rows or x.size(2)<cols:
        row_pad = max((rows-x.size(1)+1)//2, 0)
        col_pad = max((cols-x.size(2)+1)//2, 0)
        x = F.pad(x[None], (col_pad,col_pad,row_pad,row_pad), mode=padding_mode)[0]
    row = int((x.size(1)-rows+1)*row_pct)
    col = int((x.size(2)-cols+1)*col_pct)
    x = x[:, row:row+rows, col:col+cols]
    return x.contiguous() # without this, get NaN later - don't know why

def _crop_pad_image_points(x, size, padding_mode='reflection', row_pct = 0.5, col_pct = 0.5):
    size = tis2hw(size)
    rows,cols = size
    if x.size[0]<rows or x.size[1]<cols:
        row_pad = max((rows-x.size[0]+1)//2, 0)
        col_pad = max((cols-x.size[1]+1)//2, 0)
        x = _pad_coord(x, row_pad, col_pad)
    return crop(x,(rows,cols), row_pct, col_pct)

def _crop_pad(x, size, padding_mode='reflection', row_pct:uniform = 0.5, col_pct:uniform = 0.5):
    f_crop_pad = _crop_pad_image_points if isinstance(x, ImagePoints) else _crop_pad_default
    return f_crop_pad(x, size, padding_mode, row_pct, col_pct)

crop_pad = TfmCrop(_crop_pad)

def _image_maybe_add_crop_pad(img, tfms):
    tfm_names = [tfm.__name__ for tfm in tfms]
    return [crop_pad()] + tfms if 'crop_pad' not in tfm_names else tfms
Image._maybe_add_crop_pad = _image_maybe_add_crop_pad

rand_pos = {'row_pct':(0,1), 'col_pct':(0,1)}

def rand_pad(padding:int, size:int, mode:str='reflection'):
    "Fixed `mode` `padding` and random crop of `size`"
    return [pad(padding=padding,mode=mode),
            crop(size=size, **rand_pos)]

def rand_zoom(scale:uniform=1.0, p:float=1.):
    "Randomized version of `zoom`."
    return zoom(scale=scale, **rand_pos, p=p)

def rand_crop(*args, padding_mode='reflection', p:float=1.):
    "Randomized version of `crop_pad`."
    return crop_pad(*args, **rand_pos, padding_mode=padding_mode, p=p)

def zoom_crop(scale:float, do_rand:bool=False, p:float=1.0):
    "Randomly zoom and/or crop."
    zoom_fn = rand_zoom if do_rand else zoom
    crop_fn = rand_crop if do_rand else crop_pad
    return [zoom_fn(scale=scale, p=p), crop_fn()]

def _find_coeffs(orig_pts:Points, targ_pts:Points)->Tensor:
    "Find 8 coeff mentioned [here](https://web.archive.org/web/20150222120106/xenia.media.mit.edu/~cwren/interpolator/)."
    matrix = []
    #The equations we'll need to solve.
    for p1, p2 in zip(targ_pts, orig_pts):
        matrix.append([p1[0], p1[1], 1, 0, 0, 0, -p2[0]*p1[0], -p2[0]*p1[1]])
        matrix.append([0, 0, 0, p1[0], p1[1], 1, -p2[1]*p1[0], -p2[1]*p1[1]])

    A = FloatTensor(matrix)
    B = FloatTensor(orig_pts).view(8, 1)
    #The 8 scalars we seek are solution of AX = B
    return torch.linalg.solve(A,B)[:,0]

def _apply_perspective(coords:FlowField, coeffs:Points)->FlowField:
    "Transform `coords` with `coeffs`."
    size = coords.flow.size()
    #compress all the dims expect the last one ang adds ones, coords become N * 3
    coords.flow = coords.flow.view(-1,2)
    #Transform the coeffs in a 3*3 matrix with a 1 at the bottom left
    coeffs = torch.cat([coeffs, FloatTensor([1])]).view(3,3)
    coords.flow = torch.addmm(coeffs[:,2], coords.flow, coeffs[:,:2].t())
    coords.flow.mul_(1/coords.flow[:,2].unsqueeze(1))
    coords.flow = coords.flow[:,:2].view(size)
    return coords

_orig_pts = [[-1,-1], [-1,1], [1,-1], [1,1]]

def _do_perspective_warp(c:FlowField, targ_pts:Points, invert=False):
    "Apply warp to `targ_pts` from `_orig_pts` to `c` `FlowField`."
    if invert: return _apply_perspective(c, _find_coeffs(targ_pts, _orig_pts))
    return _apply_perspective(c, _find_coeffs(_orig_pts, targ_pts))

def _perspective_warp(c, magnitude:partial(uniform,size=8)=0, invert=False):
    "Apply warp of `magnitude` to `c`."
    magnitude = magnitude.view(4,2)
    targ_pts = [[x+m for x,m in zip(xs, ms)] for xs, ms in zip(_orig_pts, magnitude)]
    return _do_perspective_warp(c, targ_pts, invert)
perspective_warp = TfmCoord(_perspective_warp)

def _symmetric_warp(c, magnitude:partial(uniform,size=4)=0, invert=False):
    "Apply symmetric warp of `magnitude` to `c`."
    m = listify(magnitude, 4)
    targ_pts = [[-1-m[3],-1-m[1]], [-1-m[2],1+m[1]], [1+m[3],-1-m[0]], [1+m[2],1+m[0]]]
    return _do_perspective_warp(c, targ_pts, invert)
symmetric_warp = TfmCoord(_symmetric_warp)

def _tilt(c, direction:uniform_int, magnitude:uniform=0, invert=False):
    "Tilt `c` field with random `direction` and `magnitude`."
    orig_pts = [[-1,-1], [-1,1], [1,-1], [1,1]]
    if direction == 0:   targ_pts = [[-1,-1], [-1,1], [1,-1-magnitude], [1,1+magnitude]]
    elif direction == 1: targ_pts = [[-1,-1-magnitude], [-1,1+magnitude], [1,-1], [1,1]]
    elif direction == 2: targ_pts = [[-1,-1], [-1-magnitude,1], [1,-1], [1+magnitude,1]]
    elif direction == 3: targ_pts = [[-1-magnitude,-1], [-1,1], [1+magnitude,-1], [1,1]]
    coeffs = _find_coeffs(targ_pts, _orig_pts) if invert else _find_coeffs(_orig_pts, targ_pts)
    return _apply_perspective(c, coeffs)
tilt = TfmCoord(_tilt)

def _skew(c, direction:uniform_int, magnitude:uniform=0, invert=False):
    "Skew `c` field with random `direction` and `magnitude`."
    orig_pts = [[-1,-1], [-1,1], [1,-1], [1,1]]
    if direction == 0:   targ_pts = [[-1-magnitude,-1], [-1,1], [1,-1], [1,1]]
    elif direction == 1: targ_pts = [[-1,-1-magnitude], [-1,1], [1,-1], [1,1]]
    elif direction == 2: targ_pts = [[-1,-1], [-1-magnitude,1], [1,-1], [1,1]]
    elif direction == 3: targ_pts = [[-1,-1], [-1,1+magnitude], [1,-1], [1,1]]
    elif direction == 4: targ_pts = [[-1,-1], [-1,1], [1+magnitude,-1], [1,1]]
    elif direction == 5: targ_pts = [[-1,-1], [-1,1], [1,-1-magnitude], [1,1]]
    elif direction == 6: targ_pts = [[-1,-1], [-1,1], [1,-1], [1+magnitude,1]]
    elif direction == 7: targ_pts = [[-1,-1], [-1,1], [1,-1], [1,1+magnitude]]
    coeffs = _find_coeffs(targ_pts, _orig_pts) if invert else _find_coeffs(_orig_pts, targ_pts)
    return _apply_perspective(c, coeffs)
skew = TfmCoord(_skew)

def get_transforms(do_flip:bool=True, flip_vert:bool=False, max_rotate:float=10., max_zoom:float=1.1,
                   max_lighting:float=0.2, max_warp:float=0.2, p_affine:float=0.75,
                   p_lighting:float=0.75, xtra_tfms:Optional[Collection[Transform]]=None)->Collection[Transform]:
    "Utility func to easily create a list of flip, rotate, `zoom`, warp, lighting transforms."
    res = [rand_crop()]
    if do_flip:    res.append(dihedral_affine() if flip_vert else flip_lr(p=0.5))
    if max_warp:   res.append(symmetric_warp(magnitude=(-max_warp,max_warp), p=p_affine))
    if max_rotate: res.append(rotate(degrees=(-max_rotate,max_rotate), p=p_affine))
    if max_zoom>1: res.append(rand_zoom(scale=(1.,max_zoom), p=p_affine))
    if max_lighting:
        res.append(brightness(change=(0.5*(1-max_lighting), 0.5*(1+max_lighting)), p=p_lighting))
        res.append(contrast(scale=(1-max_lighting, 1/(1-max_lighting)), p=p_lighting))
    #       train                   , valid
    return (res + listify(xtra_tfms), [crop_pad()])

def _compute_zs_mat(sz:TensorImageSize, scale:float, squish:float,
                   invert:bool, row_pct:float, col_pct:float)->AffineMatrix:
    "Utility routine to compute zoom/squish matrix."
    orig_ratio = math.sqrt(sz[1]/sz[0])
    for s,r,i in zip(scale,squish, invert):
        s,r = 1/math.sqrt(s),math.sqrt(r)
        if s * r <= 1 and s / r <= 1: #Test if we are completely inside the picture
            w,h = (s/r, s*r) if i else (s*r,s/r)
            col_c = (1-w) * (2*col_pct - 1)
            row_c = (1-h) * (2*row_pct - 1)
            return _get_zoom_mat(w, h, col_c, row_c)

    #Fallback, hack to emulate a center crop without cropping anything yet.
    if orig_ratio > 1: return _get_zoom_mat(1/orig_ratio**2, 1, 0, 0.)
    else:              return _get_zoom_mat(1, orig_ratio**2, 0, 0.)

def _zoom_squish(c, scale:uniform=1.0, squish:uniform=1.0, invert:rand_bool=False,
                row_pct:uniform=0.5, col_pct:uniform=0.5):
    #This is intended for scale, squish and invert to be of size 10 (or whatever) so that the transform
    #can try a few zoom/squishes before falling back to center crop (like torchvision.RandomResizedCrop)
    m = _compute_zs_mat(c.size, scale, squish, invert, row_pct, col_pct)
    return _affine_mult(c, FloatTensor(m))
zoom_squish = TfmCoord(_zoom_squish)

def rand_resize_crop(size:int, max_scale:float=2., ratios:Tuple[float,float]=(0.75,1.33)):
    "Randomly resize and crop the image to a ratio in `ratios` after a zoom of `max_scale`."
    return [zoom_squish(scale=(1.,max_scale,8), squish=(*ratios,8), invert=(0.5,8), row_pct=(0.,1.), col_pct=(0.,1.)),
            crop(size=size)]