[Feature] Add segformer decode head and related train config (open-mm…

…lab#599)

* [Feature]Segformer re-implementation

* Using act_cfg and norm_cfg to control activation and normalization

* Split this PR into several little PRs

* Fix lint error

* Remove SegFormerHead

* [Feature] Add segformer decode head and related train config

* Add ade20K trainval support for segformer

1. Add related train and val configs;

2. Add AlignedResize;

* Set arg: find_unused_parameters = True

* parameters init refactor

* 1. Refactor segformer backbone parameters init;

2. Remove rebundant functions and unit tests;

* Remove rebundant codes

* Replace Linear Layer to 1X1 Conv

* Use nn.ModuleList to refactor segformer head.

* Remove local to_xtuple

* 1. Remove rebundant codes;

2. Modify module name;

* Refactor the backbone of segformer using mmcv.cnn.bricks.transformer.py

* Fix some code logic bugs.

* Add mit_convert.py to match pretrain keys of segformer.

* Resolve some comments.

* 1. Add some assert to ensure right params;

2. Support flexible peconv position;

* Add pe_index assert and fix unit test.

* 1. Add doc string for MixVisionTransformer;

2. Add some unit tests for MixVisionTransformer;

* Use hw_shape to pass shape of feature map.

* 1. Fix doc string of MixVisionTransformer;

2. Simplify MixFFN;

3. Modify H, W to hw_shape;

* Add more unit tests.

* Add doc string for shape convertion functions.

* Add some unit tests to improve code coverage.

* Fix Segformer backbone pretrain weights match bug.

* Modify configs of segformer.

* resolve the shape convertion functions doc string.

* Add pad_to_patch_size arg.

* Support progressive test with fewer memory cost.

* Modify default value of pad_to_patch_size arg.

* Temp code

* Using processor to refactor evaluation workflow.

* refactor eval hook.

* Fix process bar.

* Fix middle save argument.

* Modify some variable name of dataset evaluate api.

* Modify some viriable name of eval hook.

* Fix some priority bugs of eval hook.

* Fix some bugs about model loading and eval hook.

* Add ade20k 640x640 dataset.

* Fix related segformer configs.

* Depreciated efficient_test.

* Fix training progress blocked by eval hook.

* Depreciated old test api.

* Modify error patch size.

* Fix pretrain of mit_b0

* Fix the test api error.

* Modify dataset base config.

* Fix test api error.

* Modify outer api.

* Build a sampler test api.

* TODO: Refactor format_results.

* Modify variable names.

* Fix num_classes bug.

* Fix sampler index bug.

* Fix grammaly bug.

* Add part of benchmark results.

* Support batch sampler.

* More readable test api.

* Remove some command arg and fix eval hook bug.

* Support format-only arg.

* Modify format_results of datasets.

* Modify tool which use test apis.

* Update readme.

* Update readme of segformer.

* Updata readme of segformer.

* Update segformer readme and fix segformer mit_b4.

* Update readme of segformer.

* Clean AlignedResize related config.

* Clean code from pr open-mmlab#709

* Clean code from pr open-mmlab#709

* Add 512x512 segformer_mit-b5.

* Fix lint.

* Fix some segformer head bugs.

* Add segformer unit tests.

* Replace AlignedResize to ResizeToMultiple.

* Modify readme of segformer.

* Fix bug of ResizeToMultiple.

* Add ResizeToMultiple unit tests.

* Resolve conflict.

* Simplify the implementation of ResizeToMultiple.

* Update test results.

* Fix multi-scale test error when resize_ratio=1.75 and input size=640x640.

* Update segformer results.

* Update Segformer results.

* Fix some url bugs and pipelines bug.

* Move ckpt convertion to tools.

* Add segformer official pretrain weights usage.

* Clean redundant codes.

* Remove redundant codes.

* Unfied format.

* Add description for segformer converter.

* Update workers.

configs/_base_/models/segformer_mit-b0.py

@@ -0,0 +1,34 @@
+# model settings
+norm_cfg = dict(type='SyncBN', requires_grad=True)
+model = dict(
+    type='EncoderDecoder',
+    pretrained=None,
+    backbone=dict(
+        type='MixVisionTransformer',
+        in_channels=3,
+        embed_dims=32,
+        num_stages=4,
+        num_layers=[2, 2, 2, 2],
+        num_heads=[1, 2, 5, 8],
+        patch_sizes=[7, 3, 3, 3],
+        sr_ratios=[8, 4, 2, 1],
+        out_indices=(0, 1, 2, 3),
+        mlp_ratio=4,
+        qkv_bias=True,
+        drop_rate=0.0,
+        attn_drop_rate=0.0,
+        drop_path_rate=0.1),
+    decode_head=dict(
+        type='SegformerHead',
+        in_channels=[32, 64, 160, 256],
+        in_index=[0, 1, 2, 3],
+        channels=256,
+        dropout_ratio=0.1,
+        num_classes=19,
+        norm_cfg=norm_cfg,
+        align_corners=False,
+        loss_decode=dict(
+            type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0)),
+    # model training and testing settings
+    train_cfg=dict(),
+    test_cfg=dict(mode='whole'))

configs/segformer/readme.md

@@ -0,0 +1,73 @@
+# SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers
+
+## Introduction
+
+<!-- [ALGORITHM] -->
+
+```latex
+@article{xie2021segformer,
+  title={SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers},
+  author={Xie, Enze and Wang, Wenhai and Yu, Zhiding and Anandkumar, Anima and Alvarez, Jose M and Luo, Ping},
+  journal={arXiv preprint arXiv:2105.15203},
+  year={2021}
+}
+```
+
+## Results and models
+
+### ADE20k
+
+| Method | Backbone | Crop Size | Lr schd | Mem (GB) | Inf time (fps) | mIoU | mIoU(ms+flip) | config | download |
+| ------ | -------- | --------- | ------: | -------: | -------------- | ---: | ------------- | ------ | -------- |
+|Segformer | MIT-B0 | 512x512 | 160000 | 2.1 | 51.32 | 37.41 | 38.34 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/segformer/segformer_mit-b0_512x512_160k_ade20k.py) | [model](https://download.openmmlab.com/mmsegmentation/v0.5/segformer/segformer_mit-b0_512x512_160k_ade20k/segformer_mit-b0_512x512_160k_ade20k_20210726_101530-8ffa8fda.pth) &#124; [log](https://download.openmmlab.com/mmsegmentation/v0.5/segformer/segformer_mit-b0_512x512_160k_ade20k/segformer_mit-b0_512x512_160k_ade20k_20210726_101530.log.json) |
+|Segformer | MIT-B1 | 512x512 | 160000 | 2.6 | 47.66 | 40.97 | 42.54 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/segformer/segformer_mit-b1_512x512_160k_ade20k.py) | [model](https://download.openmmlab.com/mmsegmentation/v0.5/segformer/segformer_mit-b1_512x512_160k_ade20k/segformer_mit-b1_512x512_160k_ade20k_20210726_112106-d70e859d.pth) &#124; [log](https://download.openmmlab.com/mmsegmentation/v0.5/segformer/segformer_mit-b1_512x512_160k_ade20k/segformer_mit-b1_512x512_160k_ade20k_20210726_112106.log.json) |
+|Segformer | MIT-B2 | 512x512 | 160000 | 3.6 | 30.88 | 45.58 | 47.03 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/segformer/segformer_mit-b2_512x512_160k_ade20k.py) | [model](https://download.openmmlab.com/mmsegmentation/v0.5/segformer/segformer_mit-b2_512x512_160k_ade20k/segformer_mit-b2_512x512_160k_ade20k_20210726_112103-cbd414ac.pth) &#124; [log](https://download.openmmlab.com/mmsegmentation/v0.5/segformer/segformer_mit-b2_512x512_160k_ade20k/segformer_mit-b2_512x512_160k_ade20k_20210726_112103.log.json) |
+|Segformer | MIT-B3 | 512x512 | 160000 | 4.8 | 22.11 | 47.82 | 48.81 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/segformer/segformer_mit-b3_512x512_160k_ade20k.py) | [model](https://download.openmmlab.com/mmsegmentation/v0.5/segformer/segformer_mit-b3_512x512_160k_ade20k/segformer_mit-b3_512x512_160k_ade20k_20210726_081410-962b98d2.pth) &#124; [log](https://download.openmmlab.com/mmsegmentation/v0.5/segformer/segformer_mit-b3_512x512_160k_ade20k/segformer_mit-b3_512x512_160k_ade20k_20210726_081410.log.json) |
+|Segformer | MIT-B4 | 512x512 | 160000 | 6.1 | 15.45 | 48.46 | 49.76 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/segformer/segformer_mit-b4_512x512_160k_ade20k.py) | [model](https://download.openmmlab.com/mmsegmentation/v0.5/segformer/segformer_mit-b4_512x512_160k_ade20k/segformer_mit-b4_512x512_160k_ade20k_20210728_183055-7f509d7d.pth) &#124; [log](https://download.openmmlab.com/mmsegmentation/v0.5/segformer/segformer_mit-b4_512x512_160k_ade20k/segformer_mit-b4_512x512_160k_ade20k_20210728_183055.log.json) |
+|Segformer | MIT-B5 | 512x512 | 160000 | 7.2 | 11.89 | 49.13 | 50.22 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/segformer/segformer_mit-b5_512x512_160k_ade20k.py) | [model](https://download.openmmlab.com/mmsegmentation/v0.5/segformer/segformer_mit-b5_512x512_160k_ade20k/segformer_mit-b5_512x512_160k_ade20k_20210726_145235-94cedf59.pth) &#124; [log](https://download.openmmlab.com/mmsegmentation/v0.5/segformer/segformer_mit-b5_512x512_160k_ade20k/segformer_mit-b5_512x512_160k_ade20k_20210726_145235.log.json) |
+|Segformer | MIT-B5 | 640x640 | 160000 | 11.5 | 11.30 | 49.62 | 50.36 | [config](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/segformer/segformer_mit-b5_640x640_160k_ade20k.py) | [model](https://download.openmmlab.com/mmsegmentation/v0.5/segformer/segformer_mit-b5_640x640_160k_ade20k/segformer_mit-b5_640x640_160k_ade20k_20210801_121243-41d2845b.pth) &#124; [log](https://download.openmmlab.com/mmsegmentation/v0.5/segformer/segformer_mit-b5_640x640_160k_ade20k/segformer_mit-b5_640x640_160k_ade20k_20210801_121243.log.json) |
+
+Evaluation with AlignedResize:
+
+| Method | Backbone | Crop Size | Lr schd | mIoU | mIoU(ms+flip) |
+| ------ | -------- | --------- | ------: | ---: | ------------- |
+|Segformer | MIT-B0 | 512x512 | 160000 | 38.1  | 38.57 |
+|Segformer | MIT-B1 | 512x512 | 160000 | 41.64 | 42.76 |
+|Segformer | MIT-B2 | 512x512 | 160000 | 46.53 | 47.49 |
+|Segformer | MIT-B3 | 512x512 | 160000 | 48.46 | 49.14 |
+|Segformer | MIT-B4 | 512x512 | 160000 | 49.34 | 50.29 |
+|Segformer | MIT-B5 | 512x512 | 160000 | 50.08 | 50.72 |
+|Segformer | MIT-B5 | 640x640 | 160000 | 50.58 | 50.8  |
+
+We replace `AlignedResize` in original implementatiuon to `Resize + ResizeToMultiple`. If you want to test by
+using `AlignedResize`, you can change the dataset pipeline like this:
+
+```python
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='MultiScaleFlipAug',
+        img_scale=(2048, 512),
+        # img_ratios=[0.5, 0.75, 1.0, 1.25, 1.5, 1.75],
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            # resize image to multiple of 32, improve SegFormer by 0.5-1.0 mIoU.
+            dict(type='ResizeToMultiple', size_divisor=32),
+            dict(type='RandomFlip'),
+            dict(type='Normalize', **img_norm_cfg),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img']),
+        ])
+]
+```
+
+## How to use segformer official pretrain weights
+
+We convert the backbone weights from the official repo (https://github.com/NVlabs/SegFormer) with `tools/model_converters/mit_convert.py`.
+
+You may follow below steps to start segformer training preparation:
+
+1. Download segformer pretrain weights (Suggest put in `pretrain/`);
+2. Run convert script to convert official pretrain weights: `python tools/model_converters/mit_convert.py pretrain/mit_b0.pth pretrain/mit_b0.pth`;
+3. Modify `pretrained` of segformer model config, for example, `pretrained` of `segformer_mit-b0_512x512_160k_ade20k.py` is set to `pretrain/mit_b0.pth`;

configs/segformer/segformer_mit-b0_512x512_160k_ade20k.py

@@ -0,0 +1,33 @@
+_base_ = [
+    '../_base_/models/segformer_mit-b0.py', '../_base_/datasets/ade20k.py',
+    '../_base_/default_runtime.py', '../_base_/schedules/schedule_160k.py'
+]
+
+model = dict(
+    pretrained='pretrain/mit_b0.pth', decode_head=dict(num_classes=150))
+
+# optimizer
+optimizer = dict(
+    _delete_=True,
+    type='AdamW',
+    lr=0.00006,
+    betas=(0.9, 0.999),
+    weight_decay=0.01,
+    paramwise_cfg=dict(
+        custom_keys={
+            'pos_block': dict(decay_mult=0.),
+            'norm': dict(decay_mult=0.),
+            'head': dict(lr_mult=10.)
+        }))
+
+lr_config = dict(
+    _delete_=True,
+    policy='poly',
+    warmup='linear',
+    warmup_iters=1500,
+    warmup_ratio=1e-6,
+    power=1.0,
+    min_lr=0.0,
+    by_epoch=False)
+
+data = dict(samples_per_gpu=2, workers_per_gpu=2)

configs/segformer/segformer_mit-b1_512x512_160k_ade20k.py

@@ -0,0 +1,8 @@
+_base_ = ['./segformer_mit-b0_512x512_160k_ade20k.py']
+
+# model settings
+model = dict(
+    pretrained='pretrain/mit_b1.pth',
+    backbone=dict(
+        embed_dims=64, num_heads=[1, 2, 5, 8], num_layers=[2, 2, 2, 2]),
+    decode_head=dict(in_channels=[64, 128, 320, 512]))

configs/segformer/segformer_mit-b2_512x512_160k_ade20k.py

@@ -0,0 +1,8 @@
+_base_ = ['./segformer_mit-b0_512x512_160k_ade20k.py']
+
+# model settings
+model = dict(
+    pretrained='pretrain/mit_b2.pth',
+    backbone=dict(
+        embed_dims=64, num_heads=[1, 2, 5, 8], num_layers=[3, 4, 6, 3]),
+    decode_head=dict(in_channels=[64, 128, 320, 512]))

configs/segformer/segformer_mit-b3_512x512_160k_ade20k.py

@@ -0,0 +1,8 @@
+_base_ = ['./segformer_mit-b0_512x512_160k_ade20k.py']
+
+# model settings
+model = dict(
+    pretrained='pretrain/mit_b3.pth',
+    backbone=dict(
+        embed_dims=64, num_heads=[1, 2, 5, 8], num_layers=[3, 4, 18, 3]),
+    decode_head=dict(in_channels=[64, 128, 320, 512]))

configs/segformer/segformer_mit-b4_512x512_160k_ade20k.py

@@ -0,0 +1,8 @@
+_base_ = ['./segformer_mit-b0_512x512_160k_ade20k.py']
+
+# model settings
+model = dict(
+    pretrained='pretrain/mit_b4.pth',
+    backbone=dict(
+        embed_dims=64, num_heads=[1, 2, 5, 8], num_layers=[3, 8, 27, 3]),
+    decode_head=dict(in_channels=[64, 128, 320, 512]))

configs/segformer/segformer_mit-b5_512x512_160k_ade20k.py

@@ -0,0 +1,8 @@
+_base_ = ['./segformer_mit-b0_512x512_160k_ade20k.py']
+
+# model settings
+model = dict(
+    pretrained='pretrain/mit_b5.pth',
+    backbone=dict(
+        embed_dims=64, num_heads=[1, 2, 5, 8], num_layers=[3, 6, 40, 3]),
+    decode_head=dict(in_channels=[64, 128, 320, 512]))

configs/segformer/segformer_mit-b5_640x640_160k_ade20k.py

@@ -0,0 +1,44 @@
+_base_ = ['./segformer_mit-b0_512x512_160k_ade20k.py']
+
+# dataset settings
+img_norm_cfg = dict(
+    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
+crop_size = (640, 640)
+train_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(type='LoadAnnotations', reduce_zero_label=True),
+    dict(type='Resize', img_scale=(2048, 640), ratio_range=(0.5, 2.0)),
+    dict(type='RandomCrop', crop_size=crop_size, cat_max_ratio=0.75),
+    dict(type='RandomFlip', prob=0.5),
+    dict(type='PhotoMetricDistortion'),
+    dict(type='Normalize', **img_norm_cfg),
+    dict(type='Pad', size=crop_size, pad_val=0, seg_pad_val=255),
+    dict(type='DefaultFormatBundle'),
+    dict(type='Collect', keys=['img', 'gt_semantic_seg']),
+]
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='MultiScaleFlipAug',
+        img_scale=(2048, 640),
+        # img_ratios=[0.5, 0.75, 1.0, 1.25, 1.5, 1.75],
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='RandomFlip'),
+            dict(type='Normalize', **img_norm_cfg),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img']),
+        ])
+]
+data = dict(
+    train=dict(pipeline=train_pipeline),
+    val=dict(pipeline=test_pipeline),
+    test=dict(pipeline=test_pipeline))
+
+# model settings
+model = dict(
+    pretrained='pretrain/mit_b5.pth',
+    backbone=dict(
+        embed_dims=64, num_heads=[1, 2, 5, 8], num_layers=[3, 6, 40, 3]),
+    decode_head=dict(in_channels=[64, 128, 320, 512]))

mmseg/datasets/pipelines/transforms.py

@@ -6,6 +6,63 @@
 from ..builder import PIPELINES
 
 
+@PIPELINES.register_module()
+class ResizeToMultiple(object):
+    """Resize images & seg to multiple of divisor.
+
+    Args:
+        size_divisor (int): images and gt seg maps need to resize to multiple
+            of size_divisor. Default: 32.
+        interpolation (str, optional): The interpolation mode of image resize.
+            Default: None
+    """
+
+    def __init__(self, size_divisor=32, interpolation=None):
+        self.size_divisor = size_divisor
+        self.interpolation = interpolation
+
+    def __call__(self, results):
+        """Call function to resize images, semantic segmentation map to
+        multiple of size divisor.
+
+        Args:
+            results (dict): Result dict from loading pipeline.
+
+        Returns:
+            dict: Resized results, 'img_shape', 'pad_shape' keys are updated.
+        """
+        # Align image to multiple of size divisor.
+        img = results['img']
+        img = mmcv.imresize_to_multiple(
+            img,
+            self.size_divisor,
+            scale_factor=1,
+            interpolation=self.interpolation
+            if self.interpolation else 'bilinear')
+
+        results['img'] = img
+        results['img_shape'] = img.shape
+        results['pad_shape'] = img.shape
+
+        # Align segmentation map to multiple of size divisor.
+        for key in results.get('seg_fields', []):
+            gt_seg = results[key]
+            gt_seg = mmcv.imresize_to_multiple(
+                gt_seg,
+                self.size_divisor,
+                scale_factor=1,
+                interpolation='nearest')
+            results[key] = gt_seg
+
+        return results
+
+    def __repr__(self):
+        repr_str = self.__class__.__name__
+        repr_str += (f'(size_divisor={self.size_divisor}, '
+                     f'interpolation={self.interpolation})')
+        return repr_str
+
+
 @PIPELINES.register_module()
 class Resize(object):
     """Resize images & seg.

mmseg/models/backbones/mit.py

@@ -11,7 +11,7 @@
 
 from ...utils import get_root_logger
 from ..builder import BACKBONES
-from ..utils import PatchEmbed, mit_convert, nchw_to_nlc, nlc_to_nchw
+from ..utils import PatchEmbed, nchw_to_nlc, nlc_to_nchw
 
 
 class MixFFN(BaseModule):
@@ -159,7 +159,13 @@ def forward(self, x, hw_shape, identity=None):
         if identity is None:
             identity = x_q
 
-        out = self.attn(query=x_q, key=x_kv, value=x_kv)[0]
+        # `need_weights=True` will let nn.MultiHeadAttention
+        # `return attn_output, attn_output_weights.sum(dim=1) / num_heads`
+        # The `attn_output_weights.sum(dim=1)` may cause cuda error. So, we set
+        # `need_weights=False` to ignore `attn_output_weights.sum(dim=1)`.
+        # This issue - `https://github.com/pytorch/pytorch/issues/37583` report
+        # the error that large scale tensor sum operation may cause cuda error.
+        out = self.attn(query=x_q, key=x_kv, value=x_kv, need_weights=False)[0]
 
         return identity + self.dropout_layer(self.proj_drop(out))
 
@@ -387,17 +393,9 @@ def init_weights(self):
                 self.pretrained, logger=logger, map_location='cpu')
             if 'state_dict' in checkpoint:
                 state_dict = checkpoint['state_dict']
-            elif 'model' in checkpoint:
-                state_dict = checkpoint['model']
             else:
                 state_dict = checkpoint
 
-            if self.pretrain_style == 'official':
-                # Because segformer backbone is not support by mmcls,
-                # so we need to convert pretrain weights to match this
-                # implementation.
-                state_dict = mit_convert(state_dict)
-
             self.load_state_dict(state_dict, False)
 
     def forward(self, x):

mmseg/models/decode_heads/__init__.py

@@ -16,6 +16,7 @@
 from .point_head import PointHead
 from .psa_head import PSAHead
 from .psp_head import PSPHead
+from .segformer_head import SegformerHead
 from .sep_aspp_head import DepthwiseSeparableASPPHead
 from .sep_fcn_head import DepthwiseSeparableFCNHead
 from .setr_mla_head import SETRMLAHead
@@ -26,5 +27,6 @@
     'FCNHead', 'PSPHead', 'ASPPHead', 'PSAHead', 'NLHead', 'GCHead', 'CCHead',
     'UPerHead', 'DepthwiseSeparableASPPHead', 'ANNHead', 'DAHead', 'OCRHead',
     'EncHead', 'DepthwiseSeparableFCNHead', 'FPNHead', 'EMAHead', 'DNLHead',
-    'PointHead', 'APCHead', 'DMHead', 'LRASPPHead', 'SETRUPHead', 'SETRMLAHead'
+    'PointHead', 'APCHead', 'DMHead', 'LRASPPHead', 'SETRUPHead',
+    'SETRMLAHead', 'SegformerHead'
 ]