Fast-SCNN implemented (open-mmlab#58)

* init commit: fast_scnn

* 247917iters

* 4x8_80k

* configs placed in configs_unify.  4x8_80k exp.running.

* mmseg/utils/collect_env.py modified to support Windows

* study on lr

* bug in configs_unify/***/cityscapes.py fixed.

* lr0.08_100k

* lr_power changed to 1.2

* log_config by_epoch set to False.

* lr1.2

* doc strings added

* add fast_scnn backbone  test

* 80k 0.08,0.12

* add 450k

* fast_scnn test: fix BN bug.

* Add different config files into configs/

* .gitignore recovered.

* configs_unify del

* .gitignore recovered.

* delete sub-optimal config files of fast-scnn

* Code style improved.

* add docstrings to component modules of fast-scnn

* relevant files modified according to Jerry's instructions

* relevant files modified according to Jerry's instructions

* lint problems fixed.

* fast_scnn config extremely simplified.

* InvertedResidual

* fixed padding problems

* add unit test for inverted_residual

* add unit test for inverted_residual: debug 0

* add unit test for inverted_residual: debug 1

* add unit test for inverted_residual: debug 2

* add unit test for inverted_residual: debug 3

* add unit test for sep_fcn_head: debug 0

* add unit test for sep_fcn_head: debug 1

* add unit test for sep_fcn_head: debug 2

* add unit test for sep_fcn_head: debug 3

* add unit test for sep_fcn_head: debug 4

* add unit test for sep_fcn_head: debug 5

* FastSCNN type(dwchannels) changed to tuple.

* t changed to expand_ratio.

* Spaces fixed.

* Update mmseg/models/backbones/fast_scnn.py

Co-authored-by: Jerry Jiarui XU <[email protected]>

* Update mmseg/models/decode_heads/sep_fcn_head.py

Co-authored-by: Jerry Jiarui XU <[email protected]>

* Update mmseg/models/decode_heads/sep_fcn_head.py

Co-authored-by: Jerry Jiarui XU <[email protected]>

* Docstrings fixed.

* Docstrings fixed.

* Inverted Residual kept coherent with mmcl.

* Inverted Residual kept coherent with mmcl. Debug 0

* _make_layer parameters renamed.

* final commit

* Arg scale_factor deleted.

* Expand_ratio docstrings updated.

* final commit

* Readme for Fast-SCNN added.

* model-zoo.md modified.

* fast_scnn README updated.

* Move InvertedResidual module into mmseg/utils.

* test_inverted_residual module corrected.

* test_inverted_residual.py moved.

* encoder_decoder modified to avoid bugs when running PSPNet.
getting_started.md bug fixed.

* Revert "encoder_decoder modified to avoid bugs when running PSPNet. "

This reverts commit dd0aadf

Co-authored-by: Jerry Jiarui XU <[email protected]>

Author: johnzja

configs/_base_/models/fast_scnn.py

@@ -0,0 +1,58 @@
+# model settings
+norm_cfg = dict(type='SyncBN', requires_grad=True, momentum=0.01)
+model = dict(
+    type='EncoderDecoder',
+    backbone=dict(
+        type='FastSCNN',
+        downsample_dw_channels=(32, 48),
+        global_in_channels=64,
+        global_block_channels=(64, 96, 128),
+        global_block_strides=(2, 2, 1),
+        global_out_channels=128,
+        higher_in_channels=64,
+        lower_in_channels=128,
+        fusion_out_channels=128,
+        out_indices=(0, 1, 2),
+        norm_cfg=norm_cfg,
+        align_corners=False),
+    decode_head=dict(
+        type='DepthwiseSeparableFCNHead',
+        in_channels=128,
+        channels=128,
+        concat_input=False,
+        num_classes=19,
+        in_index=-1,
+        norm_cfg=norm_cfg,
+        align_corners=False,
+        loss_decode=dict(
+            type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.)),
+    auxiliary_head=[
+        dict(
+            type='FCNHead',
+            in_channels=128,
+            channels=32,
+            num_convs=1,
+            num_classes=19,
+            in_index=-2,
+            norm_cfg=norm_cfg,
+            concat_input=False,
+            align_corners=False,
+            loss_decode=dict(
+                type='CrossEntropyLoss', use_sigmoid=False, loss_weight=0.4)),
+        dict(
+            type='FCNHead',
+            in_channels=64,
+            channels=32,
+            num_convs=1,
+            num_classes=19,
+            in_index=-3,
+            norm_cfg=norm_cfg,
+            concat_input=False,
+            align_corners=False,
+            loss_decode=dict(
+                type='CrossEntropyLoss', use_sigmoid=False, loss_weight=0.4)),
+    ])
+
+# model training and testing settings
+train_cfg = dict()
+test_cfg = dict(mode='whole')

configs/fastscnn/README.md

@@ -0,0 +1,18 @@
+# Fast-SCNN for Semantic Segmentation
+
+## Introduction
+```
+@article{poudel2019fast,
+  title={Fast-scnn: Fast semantic segmentation network},
+  author={Poudel, Rudra PK and Liwicki, Stephan and Cipolla, Roberto},
+  journal={arXiv preprint arXiv:1902.04502},
+  year={2019}
+}
+```
+
+## Results and models
+
+### Cityscapes
+|   Method   | Backbone  | Crop Size | Lr schd | Mem (GB) | Inf time (fps) | mIoU  | mIoU(ms+flip) |                                                                                                                                                                                              download                                                                                              |
+|------------|-----------|-----------|--------:|----------|----------------|------:|---------------|----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
+| Fast-SCNN  | Fast-SCNN | 512x1024  |   80000 |      8.4 |          63.61 | 69.06 | -             | [model](https://openmmlab.oss-cn-hangzhou.aliyuncs.com/mmsegmentation/v0.5/fast_scnn/fast_scnn_4x8_80k_lr0.12_cityscapes-cae6c46a.pth) | [log](https://openmmlab.oss-cn-hangzhou.aliyuncs.com/mmsegmentation/v0.5/fast_scnn/fast_scnn_4x8_80k_lr0.12_cityscapes-20200807_165744.log.json)     |

configs/fastscnn/fast_scnn_4x8_80k_lr0.12_cityscapes.py

@@ -0,0 +1,10 @@
+_base_ = [
+    '../_base_/models/fast_scnn.py', '../_base_/datasets/cityscapes.py',
+    '../_base_/default_runtime.py', '../_base_/schedules/schedule_80k.py'
+]
+
+# Re-config the data sampler.
+data = dict(samples_per_gpu=8, workers_per_gpu=4)
+
+# Re-config the optimizer.
+optimizer = dict(type='SGD', lr=0.12, momentum=0.9, weight_decay=4e-5)

docs/getting_started.md

@@ -338,7 +338,7 @@ The final output filename will be `psp_r50_512x1024_40ki_cityscapes-{hash id}.pt
 We provide a script to convert model to [ONNX](https://github.com/onnx/onnx) format. The converted model could be visualized by tools like [Netron](https://github.com/lutzroeder/netron). Besides, we also support comparing the output results between Pytorch and ONNX model.
 
 ```shell
-python tools/pytorch2onnx.py ${CONFIG_FILE} --checkpoint ${CHECKPOINT_FILE} --output_file ${ONNX_FILE} [--shape ${INPUT_SHAPE} --verify]
+python tools/pytorch2onnx.py ${CONFIG_FILE} --checkpoint ${CHECKPOINT_FILE} --output-file ${ONNX_FILE} [--shape ${INPUT_SHAPE} --verify]
 ```
 
 **Note**: This tool is still experimental. Some customized operators are not supported for now.

docs/model_zoo.md

@@ -81,11 +81,14 @@ Please refer to [ANN](https://github.com/open-mmlab/mmsegmentation/blob/master/c
 
 Please refer to [OCRNet](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/ocrnet) for details.
 
+### Fast-SCNN
+
+Please refer to [Fast-SCNN](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/fastscnn) for details.
+
 ### ResNeSt
 
 Please refer to [ResNeSt](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/resnest) for details.
 
-
 ### Mixed Precision (FP16) Training
 
 Please refer [Mixed Precision (FP16) Training](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/fp16/README.md) for details.

mmseg/models/backbones/__init__.py

@@ -1,6 +1,10 @@
+from .fast_scnn import FastSCNN
 from .hrnet import HRNet
 from .resnest import ResNeSt
 from .resnet import ResNet, ResNetV1c, ResNetV1d
 from .resnext import ResNeXt
 
-__all__ = ['ResNet', 'ResNetV1c', 'ResNetV1d', 'ResNeXt', 'HRNet', 'ResNeSt']
+__all__ = [
+    'ResNet', 'ResNetV1c', 'ResNetV1d', 'ResNeXt', 'HRNet', 'FastSCNN',
+    'ResNeSt'
+]