add MGDA, DG, and normal training code

mseg_semantic/config/train/1080_release/mseg-mgda.yaml

@@ -0,0 +1,67 @@
+# difference with normal mseg.yaml is "use_apex: False", since apex model does not support model.no_sync()
+DATA:
+  dataset: [
+        ade20k-150-relabeled,
+        bdd-relabeled,
+        cityscapes-19-relabeled,
+        coco-panoptic-133-relabeled,
+        idd-39-relabeled,
+        mapillary-public65-relabeled,
+        sunrgbd-37-relabeled]
+  universal: True
+  use_mgda: False # to be determined at argument
+
+TRAIN:
+  use_naive_taxonomy: False
+  arch: hrnet
+  network_name: 
+  layers: 
+  sync_bn: True  # adopt sync_bn or not
+  train_h: 713
+  train_w: 713
+  scale_min: 0.5  # minimum random scale
+  scale_max: 2.0  # maximum random scale
+  short_size: 1080
+  rotate_min: -10  # minimum random rotate
+  rotate_max: 10  # maximum random rotate
+  zoom_factor: 8  # zoom factor for final prediction during training, be in [1, 2, 4, 8]
+  ignore_label: 255
+  aux_weight: 0.4
+  num_examples: 1000000
+  train_gpu: [0, 1, 2, 3, 4, 5, 6]
+  dataset_gpu_mapping: {
+        'ade20k-150-relabeled': [0],
+        'bdd-relabeled': [1],
+        'cityscapes-19-relabeled': [2],
+        'coco-panoptic-133-relabeled': [3],
+        'idd-39-relabeled': [4],
+        'mapillary-public65-relabeled': [5],
+        'sunrgbd-37-relabeled': [6],
+  }
+  workers: 64  # data loader workers
+  batch_size: 35 # batch size for training
+  batch_size_val: 1  # batch size for validation during training, memory and speed tradeoff
+  base_lr: 0.01
+  epochs: 10
+  start_epoch: 0
+  power: 0.9
+  momentum: 0.9
+  weight_decay: 0.0001
+  manual_seed:
+  print_freq: 10
+  save_freq: 1
+  save_path: default
+  weight:  # path to initial weight (default: none)
+  resume:  # path to latest checkpoint (default: none)
+  auto_resume: None # xx
+  evaluate: False  # evaluate on validation set, extra gpu memory needed and small batch_size_val is recommend
+Distributed:
+  dist_url: tcp://127.0.0.1:6795
+  dist_backend: 'nccl'
+  multiprocessing_distributed: True
+  world_size: 1
+  rank: 0
+  use_apex: False
+  opt_level: 'O0'
+  keep_batchnorm_fp32:
+  loss_scale:

mseg_semantic/config/train/1080_release/mseg-naive-baseline.yaml

@@ -0,0 +1,66 @@
+DATA:
+  dataset: [
+        ade20k-150,
+        bdd,
+        cityscapes-19,
+        coco-panoptic-133,
+        idd-39,
+        mapillary-public65,
+        sunrgbd-37]
+  universal: True
+  use_mgda: False # to be determined at argument
+
+TRAIN:
+  use_naive_taxonomy: True
+  arch: hrnet
+  network_name: 
+  layers: 
+  sync_bn: True  # adopt sync_bn or not
+  train_h: 713
+  train_w: 713
+  scale_min: 0.5  # minimum random scale
+  scale_max: 2.0  # maximum random scale
+  short_size: 1080
+  rotate_min: -10  # minimum random rotate
+  rotate_max: 10  # maximum random rotate
+  zoom_factor: 8  # zoom factor for final prediction during training, be in [1, 2, 4, 8]
+  ignore_label: 255
+  aux_weight: 0.4
+  num_examples: 1000000
+  train_gpu: [0, 1, 2, 3, 4, 5, 6]
+  dataset_gpu_mapping: {
+        'ade20k-150': [0],
+        'bdd': [1],
+        'cityscapes-19': [2],
+        'coco-panoptic-133': [3],
+        'idd-39': [4],
+        'mapillary-public65': [5],
+        'sunrgbd-37': [6],
+  }
+  workers: 64  # data loader workers
+  batch_size: 28 # batch size for training
+  batch_size_val: 1  # batch size for validation during training, memory and speed tradeoff
+  base_lr: 0.01
+  epochs: 10
+  start_epoch: 0
+  power: 0.9
+  momentum: 0.9
+  weight_decay: 0.0001
+  manual_seed:
+  print_freq: 10
+  save_freq: 1
+  save_path: default
+  weight:  # path to initial weight (default: none)
+  resume:  # path to latest checkpoint (default: none)
+  auto_resume: None # xx
+  evaluate: False  # evaluate on validation set, extra gpu memory needed and small batch_size_val is recommend
+Distributed:
+  dist_url: tcp://127.0.0.1:6795
+  dist_backend: 'nccl'
+  multiprocessing_distributed: True
+  world_size: 1
+  rank: 0
+  use_apex: True
+  opt_level: 'O0'
+  keep_batchnorm_fp32:
+  loss_scale:

mseg_semantic/config/train/1080_release/mseg-relabeled-1m.yaml

@@ -0,0 +1,66 @@
+DATA:
+  dataset: [
+        ade20k-150-relabeled,
+        bdd-relabeled,
+        cityscapes-19-relabeled,
+        coco-panoptic-133-relabeled,
+        idd-39-relabeled,
+        mapillary-public65-relabeled,
+        sunrgbd-37-relabeled]
+  universal: True
+  use_mgda: False # to be determined at argument
+
+TRAIN:
+  use_naive_taxonomy: False
+  arch: hrnet
+  network_name: 
+  layers: 
+  sync_bn: True  # adopt sync_bn or not
+  train_h: 713
+  train_w: 713
+  scale_min: 0.5  # minimum random scale
+  scale_max: 2.0  # maximum random scale
+  short_size: 1080 # image resolution is 1080p at training
+  rotate_min: -10  # minimum random rotate
+  rotate_max: 10  # maximum random rotate
+  zoom_factor: 8  # zoom factor for final prediction during training, be in [1, 2, 4, 8]
+  ignore_label: 255
+  aux_weight: 0.4
+  num_examples: 1000000 # 1 Million crops per dataset is default training duration
+  train_gpu: [0, 1, 2, 3, 4, 5, 6]
+  dataset_gpu_mapping: {
+        'ade20k-150-relabeled': [0],
+        'bdd-relabeled': [1],
+        'cityscapes-19-relabeled': [2],
+        'coco-panoptic-133-relabeled': [3],
+        'idd-39-relabeled': [4],
+        'mapillary-public65-relabeled': [5],
+        'sunrgbd-37-relabeled': [6]
+  }
+  workers: 64  # data loader workers
+  batch_size: 14 # batch size for training
+  batch_size_val: 1  # batch size for validation during training, memory and speed tradeoff
+  base_lr: 0.01
+  epochs: 10
+  start_epoch: 0
+  power: 0.9
+  momentum: 0.9
+  weight_decay: 0.0001
+  manual_seed:
+  print_freq: 10
+  save_freq: 1
+  save_path: default
+  weight:  # path to initial weight (default: none)
+  resume:  # path to latest checkpoint (default: none)
+  auto_resume: None # xx
+  evaluate: False  # evaluate on validation set, extra gpu memory needed and small batch_size_val is recommend
+Distributed:
+  dist_url: tcp://127.0.0.1:6795
+  dist_backend: 'nccl'
+  multiprocessing_distributed: True
+  world_size: 1
+  rank: 0
+  use_apex: True
+  opt_level: 'O0'
+  keep_batchnorm_fp32:
+  loss_scale:

mseg_semantic/config/train/1080_release/mseg-relabeled-3m.yaml

@@ -0,0 +1,66 @@
+DATA:
+  dataset: [
+        ade20k-150-relabeled,
+        bdd-relabeled,
+        cityscapes-19-relabeled,
+        coco-panoptic-133-relabeled,
+        idd-39-relabeled,
+        mapillary-public65-relabeled,
+        sunrgbd-37-relabeled]
+  universal: True
+  use_mgda: False # to be determined at argument
+
+TRAIN:
+  use_naive_taxonomy: False
+  arch: hrnet
+  network_name: 
+  layers: 
+  sync_bn: True  # adopt sync_bn or not
+  train_h: 713
+  train_w: 713
+  scale_min: 0.5  # minimum random scale
+  scale_max: 2.0  # maximum random scale
+  short_size: 1080 # image resolution is 1080p for training
+  rotate_min: -10  # minimum random rotate
+  rotate_max: 10  # maximum random rotate
+  zoom_factor: 8  # zoom factor for final prediction during training, be in [1, 2, 4, 8]
+  ignore_label: 255
+  aux_weight: 0.4
+  num_examples: 3000000
+  train_gpu: [0, 1, 2, 3, 4, 5, 6]
+  dataset_gpu_mapping: {
+        'ade20k-150-relabeled': [0],
+        'bdd-relabeled': [1],
+        'cityscapes-19-relabeled': [2],
+        'coco-panoptic-133-relabeled': [3],
+        'idd-39-relabeled': [4],
+        'mapillary-public65-relabeled': [5],
+        'sunrgbd-37-relabeled': [6],
+  }
+  workers: 64  # data loader workers
+  batch_size: 35 # batch size for training
+  batch_size_val: 1  # batch size for validation during training, memory and speed tradeoff
+  base_lr: 0.01
+  epochs: 10
+  start_epoch: 0
+  power: 0.9
+  momentum: 0.9
+  weight_decay: 0.0001
+  manual_seed:
+  print_freq: 10
+  save_freq: 1
+  save_path: default
+  weight:  # path to initial weight (default: none)
+  resume:  # path to latest checkpoint (default: none)
+  auto_resume: None # xx
+  evaluate: False  # evaluate on validation set, extra gpu memory needed and small batch_size_val is recommend
+Distributed:
+  dist_url: tcp://127.0.0.1:6795
+  dist_backend: 'nccl'
+  multiprocessing_distributed: True
+  world_size: 1
+  rank: 0
+  use_apex: True
+  opt_level: 'O0'
+  keep_batchnorm_fp32:
+  loss_scale:

mseg_semantic/config/train/1080_release/mseg-unrelabeled.yaml

@@ -0,0 +1,67 @@
+DATA:
+  dataset: [
+        ade20k-150,
+        bdd,
+        cityscapes-19,
+        coco-panoptic-133,
+        idd-39,
+        mapillary-public65,
+        sunrgbd-37]
+  universal: True
+  use_multiple_datasets: True
+  use_mgda: False # to be determined at argument
+
+TRAIN:
+  use_naive_taxonomy: False
+  arch: hrnet
+  network_name: 
+  layers: 
+  sync_bn: True  # adopt sync_bn or not
+  train_h: 713
+  train_w: 713
+  scale_min: 0.5  # minimum random scale
+  scale_max: 2.0  # maximum random scale
+  short_size: 1080
+  rotate_min: -10  # minimum random rotate
+  rotate_max: 10  # maximum random rotate
+  zoom_factor: 8  # zoom factor for final prediction during training, be in [1, 2, 4, 8]
+  ignore_label: 255
+  aux_weight: 0.4
+  num_examples: 1000000
+  train_gpu: [0, 1, 2, 3, 4, 5, 6]
+  dataset_gpu_mapping: {
+        'ade20k-150': [0],
+        'bdd': [1],
+        'cityscapes-19': [2],
+        'coco-panoptic-133': [3],
+        'idd-39': [4],
+        'mapillary-public65': [5],
+        'sunrgbd-37': [6],
+  }
+  workers: 64  # data loader workers
+  batch_size: 35 # batch size for training
+  batch_size_val: 1  # batch size for validation during training, memory and speed tradeoff
+  base_lr: 0.01
+  epochs: 10
+  start_epoch: 0
+  power: 0.9
+  momentum: 0.9
+  weight_decay: 0.0001
+  manual_seed:
+  print_freq: 10
+  save_freq: 1
+  save_path: default
+  weight:  # path to initial weight (default: none)
+  resume:  # path to latest checkpoint (default: none)
+  auto_resume: None # xx
+  evaluate: False  # evaluate on validation set, extra gpu memory needed and small batch_size_val is recommend
+Distributed:
+  dist_url: tcp://127.0.0.1:6795
+  dist_backend: 'nccl'
+  multiprocessing_distributed: True
+  world_size: 1
+  rank: 0
+  use_apex: True
+  opt_level: 'O0'
+  keep_batchnorm_fp32:
+  loss_scale:

mseg_semantic/config/train/1080_release/single_oracle.yaml

@@ -0,0 +1,53 @@
+DATA:
+  dataset: single
+  universal: False
+  use_mgda: False # to be determined at argument
+
+TRAIN:
+  use_naive_taxonomy: False
+  arch: hrnet
+  network_name: 
+  layers: 
+  sync_bn: True  # adopt sync_bn or not
+  train_h: 713
+  train_w: 713
+  scale_min: 0.5  # minimum random scale
+  scale_max: 2.0  # maximum random scale
+  short_size: 1080
+  rotate_min: -10  # minimum random rotate
+  rotate_max: 10  # maximum random rotate
+  zoom_factor: 8  # zoom factor for final prediction during training, be in [1, 2, 4, 8]
+  ignore_label: 255
+  aux_weight: 0.4
+  num_examples: 1000000
+  train_gpu: [0, 1, 2, 3, 4, 5, 6, 7]
+  dataset_gpu_mapping: {
+    'single': [0, 1, 2, 3, 4, 5, 6, 7],
+  }
+  workers: 32  # data loader workers
+  batch_size: 32 # batch size for training
+  batch_size_val: 1  # batch size for validation during training, memory and speed tradeoff
+  base_lr: 0.01
+  epochs: 10
+  start_epoch: 0
+  power: 0.9
+  momentum: 0.9
+  weight_decay: 0.0001
+  manual_seed:
+  print_freq: 10
+  save_freq: 1
+  save_path: default
+  weight:  # path to initial weight (default: none)
+  resume:  # path to latest checkpoint (default: none)
+  auto_resume: None
+  evaluate: False  # evaluate on validation set, extra gpu memory needed and small batch_size_val is recommend
+Distributed:
+  dist_url: tcp://127.0.0.1:6795
+  dist_backend: 'nccl'
+  multiprocessing_distributed: True
+  world_size: 1
+  rank: 0
+  use_apex: True
+  opt_level: 'O0'
+  keep_batchnorm_fp32:
+  loss_scale: