Add RCNNLossLayer, RCNNDetectionLayer for Faster(er) R-CNN #3496
Conversation
qingqing01
left a comment
qingqing01
left a comment
There was a problem hiding this comment.
@pkuyym Please help to review.
| size_t topK, | ||
| real confThreshold, | ||
| real nmsThreshold, | ||
| vector<size_t>* indices) { |
There was a problem hiding this comment.
I see applyNMSFast is similar with SSD:
void applyNMSFast(const vector<NormalizedBBox>& bboxes,
const real* confScoreData,
size_t classIdx,
size_t topK,
real confThreshold,
real nmsThreshold,
size_t numPriorBBoxes,
size_t numClasses,
vector<size_t>* indices) {
vector<pair<real, size_t>> scores;
for (size_t i = 0; i < numPriorBBoxes; ++i) {
size_t confOffset = i * numClasses + classIdx;
if (confScoreData[confOffset] > confThreshold)
scores.push_back(std::make_pair(confScoreData[confOffset], i));
}
// ...
}觉得可以写成下面，依据confScoreData来判断：
void applyNMSFast(const vector<NormalizedBBox>& bboxes,
size_t topK,
real confThreshold,
real nmsThreshold,
size_t numClasses,
vector<size_t>* indices,
const real* confScoreData,
size_t classIdx,
size_t numPriorBBoxes) {
vector<pair<real, size_t>> scores;
if (confScoreData) {
for (size_t i = 0; i < numPriorBBoxes; ++i) {
size_t confOffset = i * numClasses + classIdx;
if (confScoreData[confOffset] > confThreshold)
scores.push_back(std::make_pair(confScoreData[confOffset], i));
}
} else {
for (size_t i = 0; i < bboxes.size(); ++i) {
scores.push_back(std::make_pair(bboxes[i].first, i));
}
}
// ...
}| } | ||
| } | ||
|
|
||
| size_t getDetectionIndices( |
There was a problem hiding this comment.
目测和applyNMSFast类似，和SSD里的代码可复用~
| decodedBBox.yMax = decodedBBoxCenterY + decodedBBoxHeight / 2; | ||
|
|
||
| return decodedBBox; | ||
| } |
There was a problem hiding this comment.
上面代码也类似，不过为了方便调试收敛效果，代码优化也行。
| std::vector<real> roiLocData(4); // RoI location | ||
| for (size_t j = 0; j < 4; ++j) { | ||
| roiLocData[j] = *(roisData + n * roiDim + 1 + j); | ||
| } |
There was a problem hiding this comment.
int batchIdx = *(roisData + n * roiDim);
std::vector<real> roiLocData(4); // RoI location
for (size_t j = 0; j < 4; ++j) {
roiLocData[j] = *(roisData + n * roiDim + 1 + j);
}
==>
roisData += roiDim;
int batchIdx = *roisData;
std::vector<real> roiLocData(roisData+ 1, roisData+ 5);
std::vector的初始化：http://www.cplusplus.com/reference/vector/vector/vector/
| for (size_t j = 0; j < 4; ++j) { | ||
| predLocData[j] = *(locPredData + n * numClasses_ * 4 + c * 4 + j); | ||
| } | ||
| real predConfData = *(confPredData + n * numClasses_ + c); |
There was a problem hiding this comment.
同样代码可以短一些：
locPredData += numClasses_ * 4;
for (size_t c = 0; c < numClasses_; ++c) {
if (c == backgroundId_) continue;
std::vector<real> predLocData(locPredData + c * 4, locPredData + c * 4 + 4);
real predConfData = *(confPredData + c);
// ...| * contains the prior-box data. The rest two input layers are | ||
| * layers for generating bounding-box location offset and the | ||
| * classification confidence. | ||
| * - Output: The predict bounding boxes. |
There was a problem hiding this comment.
The predicted bounding boxes.
|
|
||
| @config_layer('rcnn_loss') | ||
| class RCNNLossLayer(LayerBase): | ||
| def __init__(self, name, inputs, loss_ratio, num_classes, background_id=0): |
| loss_ratio, | ||
| num_classes, | ||
| background_id=0, | ||
| name=None): |
There was a problem hiding this comment.
it's better to add some comments for the difference between rnn_loss and multibox_loss, if you understand.
|
感谢您给PaddlePaddle贡献代码。由于Paddle V1/V2版本已不再维护，相关代码也已从develop分支上删除，因此关闭您的PR，欢迎您向Paddle最新版-Fluid贡献代码。 |
3 participants
Add RCNNLossLayer, RCNNDetectionLayer for Faster(er) R-CNN.