[Fix] fix fast scnn #606

xiexinch · 2021-06-18T11:37:29Z

Motivation

Train with current fast scnn config will get low performance, we may fix the config and code.

Modification

Set the loss_weight of decode head from 0.4 to 1, then the training result is correct.
For improving the performance, we match our network code with paddleseg. The key modification is that we change conv 1x1 to conv 3x3 with padding 1.
Add a param dw_act_cfg at the backbone __init__ function for customizing depthwise ConvModule.

Experiments:

iters	mIoU
160k	70.96

codecov · 2021-06-18T11:48:44Z

Codecov Report

Merging #606 (2b4e1bd) into master (0997224) will increase coverage by 0.00%.
The diff coverage is 100.00%.

❗ Current head 2b4e1bd differs from pull request most recent head 17886fb. Consider uploading reports for the commit 17886fb to get more accurate results

@@           Coverage Diff           @@
##           master     #606   +/-   ##
=======================================
  Coverage   85.17%   85.18%           
=======================================
  Files         105      105           
  Lines        5668     5670    +2     
  Branches      923      923           
=======================================
+ Hits         4828     4830    +2     
  Misses        662      662           
  Partials      178      178

Flag	Coverage Δ
unittests	`85.16% <100.00%> (+<0.01%)`	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
mmseg/models/decode_heads/psp_head.py	`100.00% <ø> (ø)`
mmseg/models/utils/inverted_residual.py	`100.00% <ø> (ø)`
mmseg/models/backbones/fast_scnn.py	`97.08% <100.00%> (+0.05%)`	⬆️
mmseg/models/decode_heads/sep_fcn_head.py	`100.00% <100.00%> (ø)`

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 0997224...17886fb. Read the comment docs.

Junjun2016 · 2021-06-18T16:08:06Z

mmseg/models/backbones/fast_scnn.py

        return nn.Sequential(*layers)

    def forward(self, x):
        x = self.bottleneck1(x)
        x = self.bottleneck2(x)
        x = self.bottleneck3(x)
-        x = torch.cat([x, *self.ppm(x)], dim=1)
+        x = torch.cat([x, *self.ppm(x)[::-1]], dim=1)


[::-1] is not necessary if we train from scratch.

Junjun2016 · 2021-06-18T16:09:31Z

configs/_base_/datasets/cityscapes.py

@@ -23,6 +23,7 @@
        img_scale=(2048, 1024),
        # img_ratios=[0.5, 0.75, 1.0, 1.25, 1.5, 1.75],
        flip=False,
+        # flip=True,


Delete # flip=True

Junjun2016 · 2021-06-18T16:19:25Z

configs/fastscnn/fast_scnn_4x8_160k_lr0.12_cityscapes.py

+]
+
+# Re-config the data sampler.
+data = dict(samples_per_gpu=4, workers_per_gpu=4)


If we use 8 GPUs and 4 batch size per GPU, please rename config name 4x8 to 8x4

Junjun2016 · 2021-06-18T16:19:49Z

configs/fastscnn/fast_scnn_4x8_320k_lr0.12_cityscapes.py

+]
+
+# Re-config the data sampler.
+data = dict(samples_per_gpu=4, workers_per_gpu=4)


If we use 8 GPUs and 4 batch size per GPU, please rename config name 4x8 to 8x4

Junjun2016 · 2021-06-18T16:21:46Z

mmseg/models/backbones/fast_scnn.py

        self.dsconv1 = DepthwiseSeparableConvModule(
            dw_channels1,
            dw_channels2,
            kernel_size=3,
            stride=2,
            padding=1,
-            norm_cfg=self.norm_cfg)
+            norm_cfg=self.norm_cfg,
+            dw_act_cfg=None)


We can have an ablation study for dw_act_cfg.

Junjun2016 · 2021-06-18T16:22:02Z

mmseg/models/backbones/fast_scnn.py

        self.dsconv2 = DepthwiseSeparableConvModule(
            dw_channels2,
            out_channels,
            kernel_size=3,
            stride=2,
            padding=1,
-            norm_cfg=self.norm_cfg)
+            norm_cfg=self.norm_cfg,
+            dw_act_cfg=None)


We can have an ablation study for dw_act_cfg.

Junjun2016 · 2021-06-18T16:24:14Z

mmseg/models/backbones/fast_scnn.py

@@ -136,11 +141,13 @@ def __init__(self,
            conv_cfg=self.conv_cfg,
            norm_cfg=self.norm_cfg,
            act_cfg=self.act_cfg,
-            align_corners=align_corners)
+            align_corners=True)


align_corners=True --> align_corners=align_corners
We can also have an ablation study for align_corners.

Junjun2016

Have an ablation study for DepthwiseSeparableConvModule, one with activation after depthwise convolution and one without.

Junjun2016

What are the results of our original model with more iteration (160k and 320k).

xiexinch · 2021-06-22T11:52:33Z

What are the results of our original model with more iteration (160k and 320k).

Training results of master branch model:

iters	mIoU
160k	69.69
320k	71.17

xvjiarui · 2021-06-27T17:58:20Z

We may describe what we changed in the PR message.
And update the link of the models.

Junjun2016 · 2021-07-02T05:55:12Z

configs/fastscnn/fast_scnn_8x4_160k_lr0.12_cityscapes.py

@@ -1,10 +1,10 @@
 _base_ = [


fast_scnn_8x4_160k_lr0.12_cityscapes.py --> fast_scnn_lr0.12_8x4_160k_cityscapes.py

Junjun2016 · 2021-07-02T05:57:08Z

mmseg/models/utils/inverted_residual.py

@@ -37,7 +37,8 @@ def __init__(self,
                 conv_cfg=None,
                 norm_cfg=dict(type='BN'),
                 act_cfg=dict(type='ReLU6'),
-                 with_cp=False):
+                 with_cp=False,
+                 **kwards):


Suggested change

**kwards):

**kwargs):

Junjun2016 · 2021-07-02T05:57:19Z

mmseg/models/utils/inverted_residual.py

@@ -55,7 +56,8 @@ def __init__(self,
                    kernel_size=1,
                    conv_cfg=conv_cfg,
                    norm_cfg=norm_cfg,
-                    act_cfg=act_cfg))
+                    act_cfg=act_cfg,
+                    **kwards))


Suggested change

**kwards))

**kwargs))

Junjun2016 · 2021-07-02T05:57:31Z

mmseg/models/utils/inverted_residual.py

@@ -67,14 +69,16 @@ def __init__(self,
                groups=hidden_dim,
                conv_cfg=conv_cfg,
                norm_cfg=norm_cfg,
-                act_cfg=act_cfg),
+                act_cfg=act_cfg,
+                **kwards),


Suggested change

**kwards),

**kwargs),

Junjun2016 · 2021-07-02T05:57:43Z

mmseg/models/utils/inverted_residual.py

            ConvModule(
                in_channels=hidden_dim,
                out_channels=out_channels,
                kernel_size=1,
                conv_cfg=conv_cfg,
                norm_cfg=norm_cfg,
-                act_cfg=None)
+                act_cfg=None,
+                **kwards)


Suggested change

**kwards)

**kwargs)

Junjun2016 · 2021-07-02T05:58:17Z

mmseg/models/decode_heads/psp_head.py

@@ -22,7 +22,7 @@ class PPM(nn.ModuleList):
    """

    def __init__(self, pool_scales, in_channels, channels, conv_cfg, norm_cfg,
-                 act_cfg, align_corners):
+                 act_cfg, align_corners, **kwards):


Suggested change

act_cfg, align_corners, **kwards):

act_cfg, align_corners, **kwargs):

Junjun2016 · 2021-07-02T05:58:27Z

mmseg/models/decode_heads/psp_head.py

@@ -41,7 +41,8 @@ def __init__(self, pool_scales, in_channels, channels, conv_cfg, norm_cfg,
                        1,
                        conv_cfg=self.conv_cfg,
                        norm_cfg=self.norm_cfg,
-                        act_cfg=self.act_cfg)))
+                        act_cfg=self.act_cfg,
+                        **kwards)))


Suggested change

**kwards)))

**kwargs)))

Junjun2016 · 2021-07-02T05:59:51Z

mmseg/models/decode_heads/sep_fcn_head.py

        for i in range(1, self.num_convs):
            self.convs[i] = DepthwiseSeparableConvModule(
                self.channels,
                self.channels,
                kernel_size=self.kernel_size,
                padding=self.kernel_size // 2,
-                norm_cfg=self.norm_cfg)
+                norm_cfg=self.norm_cfg,
+                dw_act_cfg=None)


Suggested change

dw_act_cfg=None)

dw_act_cfg=dw_act_cfg)

Junjun2016 · 2021-07-02T06:00:02Z

mmseg/models/decode_heads/sep_fcn_head.py


        if self.concat_input:
            self.conv_cat = DepthwiseSeparableConvModule(
                self.in_channels + self.channels,
                self.channels,
                kernel_size=self.kernel_size,
                padding=self.kernel_size // 2,
-                norm_cfg=self.norm_cfg)
+                norm_cfg=self.norm_cfg,
+                dw_act_cfg=None)


Suggested change

dw_act_cfg=None)

dw_act_cfg=dw_act_cfg)

Junjun2016 · 2021-07-02T06:08:50Z

mmseg/models/backbones/fast_scnn.py

@@ -290,6 +309,8 @@ class FastSCNN(BaseModule):
            dict(type='ReLU')
        align_corners (bool): align_corners argument of F.interpolate.
            Default: False
+        dw_act_cfg (dict): Activation config of depthwise ConvModule. If it is


'default' is a string value at DepthwiseSeparableConvModule. In LearningToDownsample module, I want to set it to None.

mmseg/models/backbones/fast_scnn.py

* [Refactor] Match paddle seg weight * Match inference * fix exp setting * delete comment and rename config files * replace hard code with config parameters * fix ppm concat order * remove hardcode * update result * fix typo * complement docstring * complement FutureFusionModule docstring * modify log link

* polish README description * polish * Update README_cn.md

xiexinch added 3 commits June 7, 2021 21:20

[Refactor] Match paddle seg weight

20d2748

Match inference

e344c6b

fix exp setting

93d72da

Junjun2016 reviewed Jun 18, 2021

View reviewed changes

delete comment and rename config files

ad1a4b4

xiexinch added 2 commits June 24, 2021 19:55

replace hard code with config parameters

4b35e35

fix ppm concat order

9e3dfee

xiexinch added 4 commits July 2, 2021 11:01

remove hardcode

7cd7e15

merge upstream master

db54ce7

update result

ddebd49

merge master

ced58ba

xiexinch changed the title ~~[Fix] fix fast scnn training config~~ [Fix] fix fast scnn Jul 2, 2021

Junjun2016 reviewed Jul 2, 2021

View reviewed changes

mmseg/models/backbones/fast_scnn.py Show resolved Hide resolved

open-mmlab deleted a comment from Junjun2016 Jul 2, 2021

xiexinch added 4 commits July 2, 2021 14:36

fix typo

106e5d4

complement docstring

1b76893

complement FutureFusionModule docstring

018bcee

modify log link

17886fb

Junjun2016 merged commit 7e1d853 into open-mmlab:master Jul 2, 2021

sibozhang pushed a commit to sibozhang/mmsegmentation that referenced this pull request Mar 22, 2024

[Docs] Polish README description (open-mmlab#606)

989f822

* polish README description * polish * Update README_cn.md

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[Fix] fix fast scnn #606

[Fix] fix fast scnn #606

xiexinch commented Jun 18, 2021 •

edited

Loading

codecov bot commented Jun 18, 2021 •

edited

Loading

Junjun2016 Jun 18, 2021

Junjun2016 Jun 18, 2021

Junjun2016 Jun 18, 2021

Junjun2016 Jun 18, 2021

Junjun2016 Jun 18, 2021

Junjun2016 Jun 18, 2021

Junjun2016 Jun 18, 2021

Junjun2016 left a comment

Junjun2016 left a comment

xiexinch commented Jun 22, 2021 •

edited

Loading

xvjiarui commented Jun 27, 2021

Junjun2016 Jul 2, 2021

Junjun2016 Jul 2, 2021

Junjun2016 Jul 2, 2021

Junjun2016 Jul 2, 2021

Junjun2016 Jul 2, 2021

Junjun2016 Jul 2, 2021

Junjun2016 Jul 2, 2021

Junjun2016 Jul 2, 2021

Junjun2016 Jul 2, 2021

Junjun2016 Jul 2, 2021

xiexinch Jul 2, 2021

	act_cfg, align_corners, **kwards):
	act_cfg, align_corners, **kwargs):

[Fix] fix fast scnn #606

[Fix] fix fast scnn #606

Conversation

xiexinch commented Jun 18, 2021 • edited Loading

Motivation

Modification

codecov bot commented Jun 18, 2021 • edited Loading

Codecov Report

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Junjun2016 left a comment

Choose a reason for hiding this comment

Junjun2016 left a comment

Choose a reason for hiding this comment

xiexinch commented Jun 22, 2021 • edited Loading

xvjiarui commented Jun 27, 2021

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

xiexinch commented Jun 18, 2021 •

edited

Loading

codecov bot commented Jun 18, 2021 •

edited

Loading

xiexinch commented Jun 22, 2021 •

edited

Loading