Skip to content

S
seminar-breakout
Commit 8132387f authored by Yuxin Wu's avatar Yuxin Wu
Browse files
Options

[MaskRCNN] from-scratch results

parent 0fa90c41
Hide whitespace changes
Inline Side-by-side
Showing with 33 additions and 21 deletions
examples/FasterRCNN/README.md
View file @ 8132387f
......@@ -10,6 +10,7 @@ with the support of:
+ Multi-GPU / distributed training
+ Cross-GPU BatchNorm (from [MegDet: A Large Mini-Batch Object Detector](https://arxiv.org/abs/1711.07240))
+ [Group Normalization](https://arxiv.org/abs/1803.08494)
+ Training from scratch (from [Rethinking ImageNet Pre-training](https://arxiv.org/abs/1811.08883))
## Dependencies
+ Python 3.3+; OpenCV.
......@@ -76,32 +77,37 @@ prediction will need to be run with the corresponding training configs.
## Results
These models are trained on trainval35k and evaluated on minival2014 using mAP@IoU=0.50:0.95.
All models are fine-tuned from ImageNet pre-trained R50/R101 models in the
[model zoo](http://models.tensorpack.com/FasterRCNN/), unless otherwise noted.
All models are fine-tuned from ImageNet pre-trained R50/R101 models in
[tensorpack model zoo](http://models.tensorpack.com/FasterRCNN/), unless otherwise noted.
All models are trained with 8 NVIDIA V100s, unless otherwise noted.
Performance in [Detectron](https://github.com/facebookresearch/Detectron/) can be roughly reproduced.
Mask R-CNN results contain both box and mask mAP.
| Backbone | mAP<br/>(box;mask) | Detectron mAP <sup>[1](#ft1)</sup><br/> (box;mask) | Time (on 8 V100s) | Configurations <br/> (click to expand) |
| - | - | - | - | - |
| R50-C4 | 33.1 | | 18h | <details><summary>super quick</summary>`MODE_MASK=False FRCNN.BATCH_PER_IM=64`<br/>`PREPROC.SHORT_EDGE_SIZE=600 PREPROC.MAX_SIZE=1024`<br/>`TRAIN.LR_SCHEDULE=[150000,230000,280000]` </details> |
| R50-C4 | 36.6 | 36.5 | 44h | <details><summary>standard</summary>`MODE_MASK=False` </details> |
| R50-FPN | 37.4 | 37.9 | 29h | <details><summary>standard</summary>`MODE_MASK=False MODE_FPN=True` </details> |
| R50-C4 | 38.2;33.3 [:arrow_down:](http://models.tensorpack.com/FasterRCNN/COCO-R50C4-MaskRCNN-Standard.npz) | 37.8;32.8 | 49h | <details><summary>standard</summary>this is the default </details> |
| R50-FPN | 38.5;35.2 [:arrow_down:](http://models.tensorpack.com/FasterRCNN/COCO-R50FPN-MaskRCNN-Standard.npz) | 38.6;34.5 | 30h | <details><summary>standard</summary>`MODE_FPN=True` </details> |
| R50-FPN | 42.0;36.3 | | 41h | <details><summary>+Cascade</summary>`MODE_FPN=True FPN.CASCADE=True` </details> |
| R50-FPN | 39.5;35.2 | 39.5;34.4<sup>[2](#ft2)</sup> | 33h | <details><summary>+ConvGNHead</summary>`MODE_FPN=True`<br/>`FPN.FRCNN_HEAD_FUNC=fastrcnn_4conv1fc_gn_head` </details> |
| R50-FPN | 40.0;36.2 [:arrow_down:](http://models.tensorpack.com/FasterRCNN/COCO-R50FPN-MaskRCNN-StandardGN.npz) | 40.3;35.7 | 40h | <details><summary>+GN</summary>`MODE_FPN=True`<br/>`FPN.NORM=GN BACKBONE.NORM=GN`<br/>`FPN.FRCNN_HEAD_FUNC=fastrcnn_4conv1fc_gn_head`<br/>`FPN.MRCNN_HEAD_FUNC=maskrcnn_up4conv_gn_head` |
| R101-C4 | 41.4;35.2 [:arrow_down:](http://models.tensorpack.com/FasterRCNN/COCO-R101C4-MaskRCNN-Standard.npz) | | 60h | <details><summary>standard</summary>`BACKBONE.RESNET_NUM_BLOCKS=[3,4,23,3]` </details> |
| R101-FPN | 40.4;36.6 [:arrow_down:](http://models.tensorpack.com/FasterRCNN/COCO-R101FPN-MaskRCNN-Standard.npz) | 40.9;36.4 | 38h | <details><summary>standard</summary>`MODE_FPN=True`<br/>`BACKBONE.RESNET_NUM_BLOCKS=[3,4,23,3]` </details> |
| R101-FPN | 46.5;40.1 [:arrow_down:](http://models.tensorpack.com/FasterRCNN/COCO-R101FPN-MaskRCNN-BetterParams.npz) <sup>[3](#ft3)</sup> | | 73h | <details><summary>+++</summary>`MODE_FPN=True FPN.CASCADE=True`<br/>`BACKBONE.RESNET_NUM_BLOCKS=[3,4,23,3]`<br/>`TEST.RESULT_SCORE_THRESH=1e-4`<br/>`PREPROC.TRAIN_SHORT_EDGE_SIZE=[640,800]`<br/>`TRAIN.LR_SCHEDULE=[420000,500000,540000]` </details> |
<a id="ft1">1</a>: Here we comapre models that have identical training & inference cost between the two implementation. However their numbers are different due to many small implementation details.
<a id="ft2">2</a>: Numbers taken from [Group Normalization](https://arxiv.org/abs/1803.08494)
| Backbone | mAP<br/>(box;mask) | Detectron mAP <sup>[1](#ft1)</sup><br/> (box;mask) | Time (on 8 V100s) | Configurations <br/> (click to expand) |
| - | - | - | - | - |
| R50-C4 | 33.1 | | 18h | <details><summary>super quick</summary>`MODE_MASK=False FRCNN.BATCH_PER_IM=64`<br/>`PREPROC.SHORT_EDGE_SIZE=600 PREPROC.MAX_SIZE=1024`<br/>`TRAIN.LR_SCHEDULE=[150000,230000,280000]` </details> |
| R50-C4 | 36.6 | 36.5 | 44h | <details><summary>standard</summary>`MODE_MASK=False` </details> |
| R50-FPN | 37.4 | 37.9 | 29h | <details><summary>standard</summary>`MODE_MASK=False MODE_FPN=True` </details> |
| R50-C4 | 38.2;33.3 [:arrow_down:](http://models.tensorpack.com/FasterRCNN/COCO-R50C4-MaskRCNN-Standard.npz) | 37.8;32.8 | 49h | <details><summary>standard</summary>this is the default </details> |
| R50-FPN | 38.5;35.2 [:arrow_down:](http://models.tensorpack.com/FasterRCNN/COCO-R50FPN-MaskRCNN-Standard.npz) | 38.6;34.5 | 30h | <details><summary>standard</summary>`MODE_FPN=True` </details> |
| R50-FPN | 42.0;36.3 | | 41h | <details><summary>+Cascade</summary>`MODE_FPN=True FPN.CASCADE=True` </details> |
| R50-FPN | 39.5;35.2 | 39.5;34.4<sup>[2](#ft2)</sup> | 33h | <details><summary>+ConvGNHead</summary>`MODE_FPN=True`<br/>`FPN.FRCNN_HEAD_FUNC=fastrcnn_4conv1fc_gn_head` </details> |
| R50-FPN | 40.0;36.2 [:arrow_down:](http://models.tensorpack.com/FasterRCNN/COCO-R50FPN-MaskRCNN-StandardGN.npz) | 40.3;35.7 | 40h | <details><summary>+GN</summary>`MODE_FPN=True`<br/>`FPN.NORM=GN BACKBONE.NORM=GN`<br/>`FPN.FRCNN_HEAD_FUNC=fastrcnn_4conv1fc_gn_head`<br/>`FPN.MRCNN_HEAD_FUNC=maskrcnn_up4conv_gn_head` |
| R101-C4 | 41.4;35.2 [:arrow_down:](http://models.tensorpack.com/FasterRCNN/COCO-R101C4-MaskRCNN-Standard.npz) | | 60h | <details><summary>standard</summary>`BACKBONE.RESNET_NUM_BLOCKS=[3,4,23,3]` </details> |
| R101-FPN | 40.4;36.6 [:arrow_down:](http://models.tensorpack.com/FasterRCNN/COCO-R101FPN-MaskRCNN-Standard.npz) | 40.9;36.4 | 38h | <details><summary>standard</summary>`MODE_FPN=True`<br/>`BACKBONE.RESNET_NUM_BLOCKS=[3,4,23,3]` </details> |
| R101-FPN | 46.5;40.1 [:arrow_down:](http://models.tensorpack.com/FasterRCNN/COCO-R101FPN-MaskRCNN-BetterParams.npz) <sup>[3](#ft3)</sup> | | 73h | <details><summary>3x+Cascade+TrainAug</summary>`MODE_FPN=True FPN.CASCADE=True`<br/>`BACKBONE.RESNET_NUM_BLOCKS=[3,4,23,3]`<br/>`TEST.RESULT_SCORE_THRESH=1e-4`<br/>`PREPROC.TRAIN_SHORT_EDGE_SIZE=[640,800]`<br/>`TRAIN.LR_SCHEDULE=[420000,500000,540000]` </details> |
| R101-FPN<br/>(From Scratch) | 47.5;41.2 [:arrow_down:](http://models.tensorpack.com/FasterRCNN/COCO-R101FPN-MaskRCNN-ScratchGN.npz) | 47.4;40.5<sup>[4](#ft4)</sup> | 45h (on 48 V100s) | <details><summary>9x+GN+Cascade+TrainAug</summary>`MODE_FPN=True FPN.CASCADE=True`<br/>`BACKBONE.RESNET_NUM_BLOCKS=[3,4,23,3]`<br/>`FPN.NORM=GN BACKBONE.NORM=GN`<br/>`FPN.FRCNN_HEAD_FUNC=fastrcnn_4conv1fc_gn_head`<br/>`PREPROC.TRAIN_SHORT_EDGE_SIZE=[640,800]`<br/>`TRAIN.LR_SCHEDULE=[1500000,1580000,1620000]`<br/>`BACKBONE.FREEZE_AT=0`</details> |
<a id="ft1">1</a>: Numbers taken from [Detectron Model Zoo](https://github.com/facebookresearch/Detectron/blob/master/MODEL_ZOO.md).
We comapre models that have identical training & inference cost between the two implementation. However their numbers can be different due to many small implementation details.
<a id="ft2">2</a>: Numbers taken from Table 5 in [Group Normalization](https://arxiv.org/abs/1803.08494)
<a id="ft3">3</a>: Our mAP is __10+ point__ better than the official model in [matterport/Mask_RCNN](https://github.com/matterport/Mask_RCNN/releases/tag/v2.0) with the same R101-FPN backbone.
<a id="ft4">4</a>: This entry does not use ImageNet pre-training. Detectron numbers are taken from Fig. 5 in [Rethinking ImageNet Pre-training](https://arxiv.org/abs/1811.08883).
Note that our training strategy is slightly different: we enable cascade throughout the entire training.
## Notes
[NOTES.md](NOTES.md) has some notes about implementation details & speed.
tensorpack/dataflow/parallel_map.py
View file @ 8132387f
......@@ -116,6 +116,8 @@ class MultiThreadMapData(_ParallelMapData):
You can use **strict mode**, where `MultiThreadMapData.__iter__()`
is guaranteed to produce the exact set which `df.__iter__()`
produces. Although the order of data still isn't preserved.
The behavior of strict mode is undefined if the dataflow is infinite.
"""
class _Worker(StoppableThread):
def __init__(self, inq, outq, evt, map_func):
......@@ -219,6 +221,8 @@ class MultiProcessMapDataZMQ(_ParallelMapData, _MultiProcessZMQDataFlow):
You can use **strict mode**, where `MultiProcessMapData.__iter__()`
is guaranteed to produce the exact set which `df.__iter__()`
produces. Although the order of data still isn't preserved.
The behavior of strict mode is undefined if the dataflow is infinite.
"""
class _Worker(mp.Process):
def __init__(self, identity, map_func, pipename, hwm):
......@@ -405,8 +409,10 @@ if __name__ == '__main__':
return x
ds = Zero(100)
ds = MultiThreadMapData(ds, 50, f, buffer_size=50, strict=False)
ds = MultiThreadMapData(ds, 50, f, buffer_size=50, strict=True)
ds.reset_state()
for k in ds:
for idx, k in enumerate(ds):
print("Bang!", k)
if idx == 100:
break
print("END!")
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment