Feature Pyramid Networks #685
Conversation
| # offset | ||
| gt_loc[:, :2] = (gt_loc[:, :2] - rois_yx[mask]) / \ | ||
| rois_hw[mask] / std[0] | ||
| gt_loc[:, 2:] = xp.log(gt_loc[:, 2:] / rois_hw[mask]) / std[1] |
There was a problem hiding this comment.
I find it convenient to see this operation defined in a function because I use this operation myself when writing my own loss function.
Could you define one or use the one included already in ChainerCV?
I think your style seems to be simpler than the one below.
There was a problem hiding this comment.
Similar method is also used in SSD. So, I agree with you that we should add a common utility for this functionality. Note that I cannot use bbox2loc here since they are not exactly same.
|
train/eval/demo scripts are working now. |
|
Although we discussed |
|
OK |
| _canonical_level = 2 | ||
| _canonical_scale = 224 | ||
| _roi_size = 7 | ||
| _roi_sample_ratio = 2 |
chainercv/links/model/fpn/head.py
Outdated
| shape :math:`(R_l,)`. | ||
| std (tuple of floats): Two coefficients used for encoding | ||
| bounding boxes. | ||
| bboxes (list of arrays): A lisf of arrays whose shape is |
chainercv/links/model/fpn/head.py
Outdated
| bboxes (list of arrays): A lisf of arrays whose shape is | ||
| :math:`(R_n, 4)`, where :math:`R_n` is the number of | ||
| ground truth bounding boxes. | ||
| labels (list of arrays): A lisf of arrays whose shape is |
chainercv/links/model/fpn/rpn.py
Outdated
| sizes (list of tuples of two ints): A list of | ||
| :math:`(H_n, W_n)`, where :math:`H_n` and :math:`W_n` | ||
| are height and width of the :math:`n`-th image. | ||
| bboxes (list of arrays): A lisf of arrays whose shape is |
chainercv/links/model/fpn/head.py
Outdated
| :math:`(R_n,)`. | ||
|
|
||
| Returns: | ||
| tuple of four arrays: |
examples/fpn/README.md
Outdated
| | Faster-RCNN FPN ResNet50 | 36.7 % [2] | 37.1 % | | ||
| | Faster-RCNN FPN ResNet50 | 39.4 % [2] | 39.5 % | | ||
|
|
||
| Scores are mean of mean Average Precision (mmAP). |
There was a problem hiding this comment.
the mean of mean Average Precision
examples/fpn/README.md
Outdated
| You can train the model with the following code. | ||
| Note that this code requires `chainermn` module. | ||
| ``` | ||
| $ mpi4exec -n <#gpu> python train_multi.py [--model faster_rcnn_fpn_resnet50|faster_rcnn_fpn_101] [--batchsize <batchsize>] |
There was a problem hiding this comment.
README(s) of other models also use mpi4exec. So, I guess this is ok.
There was a problem hiding this comment.
is mpi4exec a valid command? I never used it.
There was a problem hiding this comment.
I changed my mind since the official document of ChainerMN uses mpiexec. Should I update all README(s) in this PR?
There was a problem hiding this comment.
It is ok to change all READMEs in this PR.
| sed -e 's/return_crowded=True)/return_crowded=True).slice[:20]/' -i eval_coco.py | ||
|
|
||
| $PYTHON eval_coco.py --model faster_rcnn_fpn_resnet50 --gpu 0 --batchsize 2 | ||
| $PYTHON eval_coco.py --model faster_rcnn_fpn_resnet10 --gpu 0 --batchsize 2 |
No description provided.