This repository has been archived by the owner on Aug 15, 2019. It is now read-only.
-
Notifications
You must be signed in to change notification settings - Fork 949
/
image_ops.ts
306 lines (280 loc) · 12.2 KB
/
image_ops.ts
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
/**
* @license
* Copyright 2018 Google Inc. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
import {ForwardFunc} from '../engine';
import {ENV} from '../environment';
import {nonMaxSuppressionImpl} from '../kernels/non_max_suppression_impl';
import {Tensor, Tensor1D, Tensor2D, Tensor3D, Tensor4D} from '../tensor';
import {convertToTensor} from '../tensor_util_env';
import {TensorLike} from '../types';
import * as util from '../util';
import {op} from './operation';
/**
* Bilinear resize a batch of 3D images to a new shape.
*
* @param images The images, of rank 4 or rank 3, of shape
* `[batch, height, width, inChannels]`. If rank 3, batch of 1 is assumed.
* @param size The new shape `[newHeight, newWidth]` to resize the
* images to. Each channel is resized individually.
* @param alignCorners Defaults to False. If true, rescale
* input by `(new_height - 1) / (height - 1)`, which exactly aligns the 4
* corners of images and resized images. If false, rescale by
* `new_height / height`. Treat similarly the width dimension.
*/
/** @doc {heading: 'Operations', subheading: 'Images', namespace: 'image'} */
function resizeBilinear_<T extends Tensor3D|Tensor4D>(
images: T|TensorLike, size: [number, number], alignCorners = false): T {
const $images = convertToTensor(images, 'images', 'resizeBilinear');
util.assert(
$images.rank === 3 || $images.rank === 4,
() => `Error in resizeBilinear: x must be rank 3 or 4, but got ` +
`rank ${$images.rank}.`);
util.assert(
size.length === 2,
() => `Error in resizeBilinear: new shape must 2D, but got shape ` +
`${size}.`);
let batchImages = $images as Tensor4D;
let reshapedTo4D = false;
if ($images.rank === 3) {
reshapedTo4D = true;
batchImages =
$images.as4D(1, $images.shape[0], $images.shape[1], $images.shape[2]);
}
const [newHeight, newWidth] = size;
const forward: ForwardFunc<Tensor4D> = (backend, save) =>
backend.resizeBilinear(batchImages, newHeight, newWidth, alignCorners);
const backward = (dy: Tensor4D, saved: Tensor[]) => {
return {
batchImages: () => ENV.engine.runKernel(
backend =>
backend.resizeBilinearBackprop(dy, batchImages, alignCorners),
{})
};
};
const res = ENV.engine.runKernel(forward, {batchImages}, backward);
if (reshapedTo4D) {
return res.as3D(res.shape[1], res.shape[2], res.shape[3]) as T;
}
return res as T;
}
/**
* NearestNeighbor resize a batch of 3D images to a new shape.
*
* @param images The images, of rank 4 or rank 3, of shape
* `[batch, height, width, inChannels]`. If rank 3, batch of 1 is assumed.
* @param size The new shape `[newHeight, newWidth]` to resize the
* images to. Each channel is resized individually.
* @param alignCorners Defaults to False. If true, rescale
* input by `(new_height - 1) / (height - 1)`, which exactly aligns the 4
* corners of images and resized images. If false, rescale by
* `new_height / height`. Treat similarly the width dimension.
*/
/** @doc {heading: 'Operations', subheading: 'Images', namespace: 'image'} */
function resizeNearestNeighbor_<T extends Tensor3D|Tensor4D>(
images: T|TensorLike, size: [number, number], alignCorners = false): T {
const $images = convertToTensor(images, 'images', 'resizeNearestNeighbor');
util.assert(
$images.rank === 3 || $images.rank === 4,
() => `Error in resizeNearestNeighbor: x must be rank 3 or 4, but got ` +
`rank ${$images.rank}.`);
util.assert(
size.length === 2,
() =>
`Error in resizeNearestNeighbor: new shape must 2D, but got shape ` +
`${size}.`);
util.assert(
$images.dtype === 'float32' || $images.dtype === 'int32',
() => '`images` must have `int32` or `float32` as dtype');
let batchImages = $images as Tensor4D;
let reshapedTo4D = false;
if ($images.rank === 3) {
reshapedTo4D = true;
batchImages =
$images.as4D(1, $images.shape[0], $images.shape[1], $images.shape[2]);
}
const [newHeight, newWidth] = size;
const forward: ForwardFunc<Tensor4D> = (backend, save) =>
backend.resizeNearestNeighbor(
batchImages, newHeight, newWidth, alignCorners);
const backward = (dy: Tensor4D, saved: Tensor[]) => {
return {
batchImages: () => ENV.engine.runKernel(
backend => backend.resizeNearestNeighborBackprop(
dy, batchImages, alignCorners),
{})
};
};
const res = ENV.engine.runKernel(forward, {batchImages}, backward);
if (reshapedTo4D) {
return res.as3D(res.shape[1], res.shape[2], res.shape[3]) as T;
}
return res as T;
}
/**
* Performs non maximum suppression of bounding boxes based on
* iou (intersection over union)
*
* @param boxes a 2d tensor of shape `[numBoxes, 4]`. Each entry is
* `[y1, x1, y2, x2]`, where `(y1, x1)` and `(y2, x2)` are the corners of
* the bounding box.
* @param scores a 1d tensor providing the box scores of shape `[numBoxes]`.
* @param maxOutputSize The maximum number of boxes to be selected.
* @param iouThreshold A float representing the threshold for deciding whether
* boxes overlap too much with respect to IOU. Must be between [0, 1].
* Defaults to 0.5 (50% box overlap).
* @param scoreThreshold A threshold for deciding when to remove boxes based
* on score. Defaults to -inf, which means any score is accepted.
* @return A 1D tensor with the selected box indices.
*/
/** @doc {heading: 'Operations', subheading: 'Images', namespace: 'image'} */
function nonMaxSuppression_(
boxes: Tensor2D|TensorLike, scores: Tensor1D|TensorLike,
maxOutputSize: number, iouThreshold = 0.5,
scoreThreshold = Number.NEGATIVE_INFINITY): Tensor1D {
const $boxes = convertToTensor(boxes, 'boxes', 'nonMaxSuppression');
const $scores = convertToTensor(scores, 'scores', 'nonMaxSuppression');
const inputs = nonMaxSuppSanityCheck(
$boxes, $scores, maxOutputSize, iouThreshold, scoreThreshold);
maxOutputSize = inputs.maxOutputSize;
iouThreshold = inputs.iouThreshold;
scoreThreshold = inputs.scoreThreshold;
return ENV.engine.runKernel(
b => b.nonMaxSuppression(
$boxes, $scores, maxOutputSize, iouThreshold, scoreThreshold),
{$boxes});
}
/** This is the async version of `nonMaxSuppression` */
async function nonMaxSuppressionAsync_(
boxes: Tensor2D|TensorLike, scores: Tensor1D|TensorLike,
maxOutputSize: number, iouThreshold = 0.5,
scoreThreshold = Number.NEGATIVE_INFINITY): Promise<Tensor1D> {
const $boxes = convertToTensor(boxes, 'boxes', 'nonMaxSuppressionAsync');
const $scores = convertToTensor(scores, 'scores', 'nonMaxSuppressionAsync');
const inputs = nonMaxSuppSanityCheck(
$boxes, $scores, maxOutputSize, iouThreshold, scoreThreshold);
maxOutputSize = inputs.maxOutputSize;
iouThreshold = inputs.iouThreshold;
scoreThreshold = inputs.scoreThreshold;
const boxesVals = await $boxes.data();
const scoresVals = await $scores.data();
const res = nonMaxSuppressionImpl(
boxesVals, scoresVals, maxOutputSize, iouThreshold, scoreThreshold);
if ($boxes !== boxes) {
$boxes.dispose();
}
if ($scores !== scores) {
$scores.dispose();
}
return res;
}
function nonMaxSuppSanityCheck(
boxes: Tensor2D, scores: Tensor1D, maxOutputSize: number,
iouThreshold: number, scoreThreshold: number):
{maxOutputSize: number, iouThreshold: number, scoreThreshold: number} {
if (iouThreshold == null) {
iouThreshold = 0.5;
}
if (scoreThreshold == null) {
scoreThreshold = Number.NEGATIVE_INFINITY;
}
const numBoxes = boxes.shape[0];
maxOutputSize = Math.min(maxOutputSize, numBoxes);
util.assert(
0 <= iouThreshold && iouThreshold <= 1,
() => `iouThreshold must be in [0, 1], but was '${iouThreshold}'`);
util.assert(
boxes.rank === 2,
() => `boxes must be a 2D tensor, but was of rank '${boxes.rank}'`);
util.assert(
boxes.shape[1] === 4,
() =>
`boxes must have 4 columns, but 2nd dimension was ${boxes.shape[1]}`);
util.assert(scores.rank === 1, () => 'scores must be a 1D tensor');
util.assert(
scores.shape[0] === numBoxes,
() => `scores has incompatible shape with boxes. Expected ${numBoxes}, ` +
`but was ${scores.shape[0]}`);
return {maxOutputSize, iouThreshold, scoreThreshold};
}
/**
* Extracts crops from the input image tensor and resizes them using bilinear
* sampling or nearest neighbor sampling (possibly with aspect ratio change)
* to a common output size specified by crop_size.
*
* @param image 4d tensor of shape `[batch,imageHeight,imageWidth, depth]`,
* where imageHeight and imageWidth must be positive, specifying the
* batch of images from which to take crops
* @param boxes 2d float32 tensor of shape `[numBoxes, 4]`. Each entry is
* `[y1, x1, y2, x2]`, where `(y1, x1)` and `(y2, x2)` are the normalized
* coordinates of the box in the boxInd[i]'th image in the batch
* @param boxInd 1d int32 tensor of shape `[numBoxes]` with values in range
* `[0, batch)` that specifies the image that the `i`-th box refers to.
* @param cropSize 1d int32 tensor of 2 elements `[cropHeigh, cropWidth]`
* specifying the size to which all crops are resized to.
* @param method Optional string from `'bilinear' | 'nearest'`,
* defaults to bilinear, which specifies the sampling method for resizing
* @param extrapolationValue A threshold for deciding when to remove boxes based
* on score. Defaults to 0.
* @return A 4D tensor of the shape `[numBoxes,cropHeight,cropWidth,depth]`
*/
/** @doc {heading: 'Operations', subheading: 'Images', namespace: 'image'} */
function cropAndResize_(
image: Tensor4D|TensorLike,
boxes: Tensor2D|TensorLike,
boxInd: Tensor1D|TensorLike,
cropSize: [number, number],
method?: 'bilinear'|'nearest',
extrapolationValue?: number,
): Tensor4D {
const $image = convertToTensor(image, 'image', 'cropAndResize', 'float32');
const $boxes = convertToTensor(boxes, 'boxes', 'cropAndResize', 'float32');
const $boxInd = convertToTensor(boxInd, 'boxInd', 'cropAndResize', 'int32');
method = method || 'bilinear';
extrapolationValue = extrapolationValue || 0;
const numBoxes = $boxes.shape[0];
util.assert(
$image.rank === 4,
() => 'Error in cropAndResize: image must be rank 4,' +
`but got rank ${$image.rank}.`);
util.assert(
$boxes.rank === 2 && $boxes.shape[1] === 4,
() => `Error in cropAndResize: boxes must be have size [${numBoxes},4] ` +
`but had shape ${$boxes.shape}.`);
util.assert(
$boxInd.rank === 1 && $boxInd.shape[0] === numBoxes,
() => `Error in cropAndResize: boxInd must be have size [${numBoxes}] ` +
`but had shape ${$boxes.shape}.`);
util.assert(
cropSize.length === 2,
() => `Error in cropAndResize: cropSize must be of length 2, but got ` +
`length ${cropSize.length}.`);
util.assert(
cropSize[0] >= 1 && cropSize[1] >= 1,
() => `cropSize must be atleast [1,1], but was ${cropSize}`);
util.assert(
method === 'bilinear' || method === 'nearest',
() => `method must be bilinear or nearest, but was ${method}`);
const forward: ForwardFunc<Tensor4D> = (backend, save) =>
backend.cropAndResize(
$image, $boxes, $boxInd, cropSize, method, extrapolationValue);
const res = ENV.engine.runKernel(forward, {$image, $boxes});
return res as Tensor4D;
}
export const resizeBilinear = op({resizeBilinear_});
export const resizeNearestNeighbor = op({resizeNearestNeighbor_});
export const nonMaxSuppression = op({nonMaxSuppression_});
export const nonMaxSuppressionAsync = nonMaxSuppressionAsync_;
export const cropAndResize = op({cropAndResize_});