[TIPC] Add TIPC Benchmarks of SFNet and MobileSeg-MV3 (#2505)

* Add TIPC chains of SFNet and MobileSeg-MV3 * Fix SegFormer initialization bug * Update benchmark statistics of OCRNet_HRNetW18 and PPMatting * Update SegFormer configs
PaddlePaddle · Sep 2, 2022 · 7970483 · 7970483
1 parent dbc7a84
commit 7970483
Show file tree

Hide file tree

Showing 24 changed files with 299 additions and 86 deletions.
diff --git a/configs/segformer/segformer_b0_cityscapes_1024x1024_160k.yml b/configs/segformer/segformer_b0_cityscapes_1024x1024_160k.yml
@@ -4,9 +4,12 @@ batch_size: 2
 iters: 160000
 
 model:
-  type: SegFormer_B0
+  type: SegFormer
+  backbone:
+    type: MixVisionTransformer_B0
+    pretrained: https://bj.bcebos.com/paddleseg/dygraph/backbone/mix_vision_transformer_b0.tar.gz
+  embedding_dim: 256
   num_classes: 19
-  pretrained: https://bj.bcebos.com/paddleseg/dygraph/mix_vision_transformer_b0.tar.gz
 
 optimizer:
   _inherited_: False

diff --git a/configs/segformer/segformer_b0_cityscapes_1024x512_160k.yml b/configs/segformer/segformer_b0_cityscapes_1024x512_160k.yml
@@ -4,9 +4,12 @@ batch_size: 1
 iters: 160000
 
 model:
-  type: SegFormer_B0
+  type: SegFormer
+  backbone:
+    type: MixVisionTransformer_B0
+    pretrained: https://bj.bcebos.com/paddleseg/dygraph/backbone/mix_vision_transformer_b0.tar.gz
+  embedding_dim: 256
   num_classes: 19
-  pretrained: https://bj.bcebos.com/paddleseg/dygraph/mix_vision_transformer_b0.tar.gz
 
 optimizer:
   _inherited_: False

diff --git a/configs/segformer/segformer_b1_cityscapes_1024x1024_160k.yml b/configs/segformer/segformer_b1_cityscapes_1024x1024_160k.yml
@@ -4,9 +4,12 @@ batch_size: 2
 iters: 160000
 
 model:
-  type: SegFormer_B1
+  type: SegFormer
+  backbone:
+    type: MixVisionTransformer_B1
+    pretrained: https://bj.bcebos.com/paddleseg/dygraph/backbone/mix_vision_transformer_b1.tar.gz
+  embedding_dim: 256
   num_classes: 19
-  pretrained: https://bj.bcebos.com/paddleseg/dygraph/mix_vision_transformer_b1.tar.gz
 
 optimizer:
   _inherited_: False

diff --git a/configs/segformer/segformer_b1_cityscapes_1024x512_160k.yml b/configs/segformer/segformer_b1_cityscapes_1024x512_160k.yml
@@ -4,9 +4,12 @@ batch_size: 1
 iters: 160000
 
 model:
-  type: SegFormer_B1
+  type: SegFormer
+  backbone:
+    type: MixVisionTransformer_B1
+    pretrained: https://bj.bcebos.com/paddleseg/dygraph/backbone/mix_vision_transformer_b1.tar.gz
+  embedding_dim: 256
   num_classes: 19
-  pretrained: https://bj.bcebos.com/paddleseg/dygraph/mix_vision_transformer_b1.tar.gz
 
 optimizer:
   _inherited_: False

diff --git a/configs/segformer/segformer_b2_cityscapes_1024x1024_160k.yml b/configs/segformer/segformer_b2_cityscapes_1024x1024_160k.yml
@@ -4,9 +4,12 @@ batch_size: 2
 iters: 160000
 
 model:
-  type: SegFormer_B2
+  type: SegFormer
+  backbone:
+    type: MixVisionTransformer_B2
+    pretrained: https://bj.bcebos.com/paddleseg/dygraph/backbone/mix_vision_transformer_b2.tar.gz
+  embedding_dim: 768
   num_classes: 19
-  pretrained: https://bj.bcebos.com/paddleseg/dygraph/mix_vision_transformer_b2.tar.gz
 
 optimizer:
   _inherited_: False

diff --git a/configs/segformer/segformer_b2_cityscapes_1024x512_160k.yml b/configs/segformer/segformer_b2_cityscapes_1024x512_160k.yml
@@ -4,9 +4,12 @@ batch_size: 1
 iters: 160000
 
 model:
-  type: SegFormer_B2
+  type: SegFormer
+  backbone:
+    type: MixVisionTransformer_B2
+    pretrained: https://bj.bcebos.com/paddleseg/dygraph/backbone/mix_vision_transformer_b2.tar.gz
+  embedding_dim: 768
   num_classes: 19
-  pretrained: https://bj.bcebos.com/paddleseg/dygraph/mix_vision_transformer_b2.tar.gz
 
 optimizer:
   _inherited_: False

diff --git a/configs/segformer/segformer_b3_cityscapes_1024x1024_160k.yml b/configs/segformer/segformer_b3_cityscapes_1024x1024_160k.yml
@@ -4,9 +4,12 @@ batch_size: 2
 iters: 160000
 
 model:
-  type: SegFormer_B3
+  type: SegFormer
+  backbone:
+    type: MixVisionTransformer_B3
+    pretrained: https://bj.bcebos.com/paddleseg/dygraph/backbone/mix_vision_transformer_b3.tar.gz
+  embedding_dim: 768
   num_classes: 19
-  pretrained: https://bj.bcebos.com/paddleseg/dygraph/mix_vision_transformer_b3.tar.gz
 
 optimizer:
   _inherited_: False

diff --git a/configs/segformer/segformer_b3_cityscapes_1024x512_160k.yml b/configs/segformer/segformer_b3_cityscapes_1024x512_160k.yml
@@ -4,9 +4,12 @@ batch_size: 1
 iters: 160000
 
 model:
-  type: SegFormer_B3
+  type: SegFormer
+  backbone:
+    type: MixVisionTransformer_B3
+    pretrained: https://bj.bcebos.com/paddleseg/dygraph/backbone/mix_vision_transformer_b3.tar.gz
+  embedding_dim: 768
   num_classes: 19
-  pretrained: https://bj.bcebos.com/paddleseg/dygraph/mix_vision_transformer_b3.tar.gz
 
 optimizer:
   _inherited_: False

diff --git a/configs/segformer/segformer_b4_cityscapes_1024x1024_160k.yml b/configs/segformer/segformer_b4_cityscapes_1024x1024_160k.yml
@@ -4,9 +4,12 @@ batch_size: 2
 iters: 160000
 
 model:
-  type: SegFormer_B4
+  type: SegFormer
+  backbone:
+    type: MixVisionTransformer_B4
+    pretrained: https://bj.bcebos.com/paddleseg/dygraph/backbone/mix_vision_transformer_b4.tar.gz
+  embedding_dim: 768
   num_classes: 19
-  pretrained: https://bj.bcebos.com/paddleseg/dygraph/mix_vision_transformer_b4.tar.gz
 
 optimizer:
   _inherited_: False

diff --git a/configs/segformer/segformer_b4_cityscapes_1024x512_160k.yml b/configs/segformer/segformer_b4_cityscapes_1024x512_160k.yml
@@ -4,9 +4,12 @@ batch_size: 1
 iters: 160000
 
 model:
-  type: SegFormer_B4
+  type: SegFormer
+  backbone:
+    type: MixVisionTransformer_B4
+    pretrained: https://bj.bcebos.com/paddleseg/dygraph/backbone/mix_vision_transformer_b4.tar.gz
+  embedding_dim: 768
   num_classes: 19
-  pretrained: https://bj.bcebos.com/paddleseg/dygraph/mix_vision_transformer_b4.tar.gz
 
 optimizer:
   _inherited_: False

diff --git a/configs/segformer/segformer_b5_cityscapes_1024x1024_160k.yml b/configs/segformer/segformer_b5_cityscapes_1024x1024_160k.yml
@@ -4,9 +4,12 @@ batch_size: 1
 iters: 160000
 
 model:
-  type: SegFormer_B5
+  type: SegFormer
+  backbone:
+    type: MixVisionTransformer_B5
+    pretrained: https://bj.bcebos.com/paddleseg/dygraph/backbone/mix_vision_transformer_b5.tar.gz
+  embedding_dim: 768
   num_classes: 19
-  pretrained: https://bj.bcebos.com/paddleseg/dygraph/mix_vision_transformer_b5.tar.gz
 
 optimizer:
   _inherited_: False

diff --git a/configs/segformer/segformer_b5_cityscapes_1024x512_160k.yml b/configs/segformer/segformer_b5_cityscapes_1024x512_160k.yml
@@ -4,7 +4,11 @@ batch_size: 1
 iters: 160000
 
 model:
-  type: SegFormer_B5
+  type: SegFormer
+  backbone:
+    type: MixVisionTransformer_B5
+    pretrained: https://bj.bcebos.com/paddleseg/dygraph/backbone/mix_vision_transformer_b5.tar.gz
+  embedding_dim: 768
   num_classes: 19
 
 optimizer:

diff --git a/paddleseg/models/segformer.py b/paddleseg/models/segformer.py
@@ -127,51 +127,3 @@ def forward(self, x):
                 mode='bilinear',
                 align_corners=self.align_corners)
         ]
-
-
-@manager.MODELS.add_component
-def SegFormer_B0(**kwargs):
-    return SegFormer(
-        backbone=manager.BACKBONES['MixVisionTransformer_B0'](),
-        embedding_dim=256,
-        **kwargs)
-
-
-@manager.MODELS.add_component
-def SegFormer_B1(**kwargs):
-    return SegFormer(
-        backbone=manager.BACKBONES['MixVisionTransformer_B1'](),
-        embedding_dim=256,
-        **kwargs)
-
-
-@manager.MODELS.add_component
-def SegFormer_B2(**kwargs):
-    return SegFormer(
-        backbone=manager.BACKBONES['MixVisionTransformer_B2'](),
-        embedding_dim=768,
-        **kwargs)
-
-
-@manager.MODELS.add_component
-def SegFormer_B3(**kwargs):
-    return SegFormer(
-        backbone=manager.BACKBONES['MixVisionTransformer_B3'](),
-        embedding_dim=768,
-        **kwargs)
-
-
-@manager.MODELS.add_component
-def SegFormer_B4(**kwargs):
-    return SegFormer(
-        backbone=manager.BACKBONES['MixVisionTransformer_B4'](),
-        embedding_dim=768,
-        **kwargs)
-
-
-@manager.MODELS.add_component
-def SegFormer_B5(**kwargs):
-    return SegFormer(
-        backbone=manager.BACKBONES['MixVisionTransformer_B5'](),
-        embedding_dim=768,
-        **kwargs)
diff --git a/test_tipc/README.md b/test_tipc/README.md
@@ -47,6 +47,8 @@
 | HRNetW48Contrast | HRNetW48Contrast | 支持 | - | - | - |
 | ENCNet | ENCNet | 支持 | - | - | - |
 | ESPNetV2 | ESPNetV2 | 支持 | - | - | - |
+| SFNet | SFNet | 支持 | - | - | - |
+| MobileSeg | MobileSeg-MV3 | 支持 | - | - | - |
 
 ## 3. 测试工具简介
 ### 目录介绍

diff --git a/test_tipc/configs/mobileseg_mv3/mobileseg_mobilenetv3_cityscapes_1024x512_80k.yml b/test_tipc/configs/mobileseg_mv3/mobileseg_mobilenetv3_cityscapes_1024x512_80k.yml
@@ -0,0 +1,48 @@
+_base_: '../_base_/cityscapes.yml'
+
+batch_size: 4
+iters: 80000
+
+optimizer:
+  weight_decay: 5.0e-4
+
+lr_scheduler:
+  warmup_iters: 1000
+  warmup_start_lr: 1.0e-5
+  learning_rate: 0.005
+
+loss:
+  types:
+    - type: OhemCrossEntropyLoss
+      min_kept: 130000
+    - type: OhemCrossEntropyLoss
+      min_kept: 130000
+    - type: OhemCrossEntropyLoss
+      min_kept: 130000
+  coef: [1, 1, 1]
+
+train_dataset:
+  transforms:
+    - type: ResizeStepScaling
+      min_scale_factor: 0.5
+      max_scale_factor: 2.0
+      scale_step_size: 0.25
+    - type: RandomPaddingCrop
+      crop_size: [1024, 512]
+    - type: RandomHorizontalFlip
+    - type: RandomDistort
+      brightness_range: 0.5
+      contrast_range: 0.5
+      saturation_range: 0.5
+    - type: Normalize
+  mode: train
+
+model:
+  type: MobileSeg
+  backbone:
+    type: MobileNetV3_large_x1_0  # out channels: [24, 40, 112, 160]
+    pretrained: https://paddleseg.bj.bcebos.com/dygraph/backbone/mobilenetv3_large_x1_0_ssld.tar.gz
+  cm_bin_sizes: [1, 2, 4]
+  cm_out_ch: 128
+  arm_out_chs: [32, 64, 128]
+  seg_head_inter_chs: [32, 32, 32]
diff --git a/test_tipc/configs/mobileseg_mv3/train_infer_python.txt b/test_tipc/configs/mobileseg_mv3/train_infer_python.txt
@@ -0,0 +1,61 @@
+===========================train_params===========================
+model_name:mobileseg_mv3
+python:python3.7
+gpu_list:0|0,1
+Global.use_gpu:null|null
+--precision:null
+--iters:lite_train_lite_infer=20|lite_train_whole_infer=20|whole_train_whole_infer=1000
+--save_dir:
+--batch_size:lite_train_lite_infer=2|lite_train_whole_infer=2|whole_train_whole_infer=4
+--model_path:null
+train_model_name:best_model/model.pdparams
+train_infer_img_dir:test_tipc/data/cityscapes/cityscapes_val_5.list
+null:null
+##
+trainer:norm
+norm_train:train.py --config test_tipc/configs/mobileseg_mv3/mobileseg_mobilenetv3_cityscapes_1024x512_80k.yml --save_interval 500 --seed 100 --num_workers 8
+pact_train:null
+fpgm_train:null
+distill_train:null
+null:null
+null:null
+##
+===========================eval_params===========================
+eval:val.py --config test_tipc/configs/mobileseg_mv3/mobileseg_mobilenetv3_cityscapes_1024x512_80k.yml --num_workers 8
+null:null
+##
+===========================export_params===========================
+--save_dir:
+--model_path:
+norm_export:export.py --config test_tipc/configs/mobileseg_mv3/mobileseg_mobilenetv3_cityscapes_1024x512_80k.yml
+quant_export:null
+fpgm_export:null
+distill_export:null
+export1:null
+export2:null
+===========================infer_params===========================
+infer_model:./test_tipc/output/mobileseg_mv3/model.pdparams
+infer_export:export.py --config test_tipc/configs/mobileseg_mv3/mobileseg_mobilenetv3_cityscapes_1024x512_80k.yml
+infer_quant:False
+inference:deploy/python/infer.py
+--device:cpu|gpu
+--enable_mkldnn:True
+--cpu_threads:6
+--batch_size:1
+--use_trt:False
+--precision:fp32
+--config:
+--image_path:./test_tipc/data/cityscapes/cityscapes_val_5.list
+--save_log_path:null
+--benchmark:True
+--save_dir:
+--model_name:mobileseg_mv3
+===========================infer_benchmark_params==========================
+random_infer_input:[{float32,[3,1024,512]}]
+===========================train_benchmark_params==========================
+batch_size:4
+fp_items:fp32|fp16
+epoch:400
+--profiler_options:'batch_range=[10,20];state=GPU;tracer_option=Default;profile_path=model.profile'
+flags:FLAGS_eager_delete_tensor_gb=0.0;FLAGS_fraction_of_gpu_memory_to_use=0.98;FLAGS_conv_workspace_size_limit=4096;FLAGS_cudnn_deterministic=False
+log_iters:15;set_cv_threads:true;skip_iters:2;repeats:500
diff --git a/test_tipc/configs/ocrnet_hrnetw18/train_infer_python.txt b/test_tipc/configs/ocrnet_hrnetw18/train_infer_python.txt
@@ -55,7 +55,7 @@ random_infer_input:[{float32,[3,512,1024]}]
 ===========================train_benchmark_params==========================
 batch_size:2
 fp_items:fp32|fp16
-epoch:500
+epoch:600
 --profiler_options:'batch_range=[10,20];state=GPU;tracer_option=Default;profile_path=model.profile'
 flags:FLAGS_eager_delete_tensor_gb=0.0;FLAGS_fraction_of_gpu_memory_to_use=0.98;FLAGS_conv_workspace_size_limit=4096;FLAGS_cudnn_deterministic=False
-log_iters:8;set_cv_threads:true;skip_iters:3;repeats:500
+log_iters:15;set_cv_threads:true;skip_iters:3;repeats:500
diff --git a/test_tipc/configs/ppmatting/train_infer_python.txt b/test_tipc/configs/ppmatting/train_infer_python.txt
@@ -55,7 +55,7 @@ random_infer_input:[{float32,[3,512,512]}]
 ===========================train_benchmark_params==========================
 batch_size:2
 fp_items:fp32|fp16
-epoch:250
+epoch:400
 --profiler_options:'batch_range=[10,20];state=GPU;tracer_option=Default;profile_path=model.profile'
 flags:FLAGS_eager_delete_tensor_gb=0.0;FLAGS_fraction_of_gpu_memory_to_use=0.98;FLAGS_conv_workspace_size_limit=4096;FLAGS_cudnn_deterministic=False
-log_iters:8;repeats:500;task:mat
+log_iters:12;repeats:500;task:mat
diff --git a/test_tipc/configs/segformer_b0/segformer_b0_cityscapes_1024x1024_160k.yml b/test_tipc/configs/segformer_b0/segformer_b0_cityscapes_1024x1024_160k.yml
@@ -4,9 +4,12 @@ batch_size: 2
 iters: 160000
 
 model:
-  type: SegFormer_B0
+  type: SegFormer
+  backbone:
+    type: MixVisionTransformer_B0
+    pretrained: https://bj.bcebos.com/paddleseg/dygraph/backbone/mix_vision_transformer_b0.tar.gz
+  embedding_dim: 256
   num_classes: 19
-  pretrained: https://bj.bcebos.com/paddleseg/dygraph/mix_vision_transformer_b0.tar.gz
 
 optimizer:
   _inherited_: False