Segformer model and recipe

src/super_gradients/common/object_names.py

@@ -288,6 +288,12 @@ class Models:
     PP_YOLOE_M = "ppyoloe_m"
     PP_YOLOE_L = "ppyoloe_l"
     PP_YOLOE_X = "ppyoloe_x"
+    SEGFORMER_B0 = "segformer_b0"
+    SEGFORMER_B1 = "segformer_b1"
+    SEGFORMER_B2 = "segformer_b2"
+    SEGFORMER_B3 = "segformer_b3"
+    SEGFORMER_B4 = "segformer_b4"
+    SEGFORMER_B5 = "segformer_b5"
 
     DEKR_CUSTOM = "dekr_custom"
 

src/super_gradients/recipes/cityscapes_segformer_b0.yaml

@@ -0,0 +1,107 @@
+#  SegFormer-B0 segmentation training example with Cityscapes dataset.
+#  Reproduction of paper:
+#  Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar,  Jose M. Alvarez, Ping Luo
+#  "SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers"
+#  ( https://arxiv.org/pdf/2105.15203.pdf )
+#
+#  Official git repo:
+#      https://github.com/NVlabs/SegFormer
+#
+#  Imagenet-1k pre-trained backbone weights taken and adapted from:
+#      https://github.com/sithu31296/semantic-segmentation
+#
+# Instructions:
+#   1. We recommend preparing the data according to SG's CityScapes readme file:
+#      https://github.com/Deci-AI/super-gradients/blob/master/src/super_gradients/training/datasets/Dataset_Setup_Instructions.md
+#   2. Note: if you change the dataset's internal directory structure, make changes to the fields "list_file" and
+#      "labels_csv_path" of both "train_dataset_params" and "val_dataset_params" accordingly
+#   3. Edit the "data_root_dir" field below to point to the absolute path of the data root directory
+#   4. Edit the "ckpt_root_dir" field to the path where you want to save checkpoints and logs
+#   5. Move to the project root (where you will find the ReadMe and src folder)
+#   6. Run the command:
+#       python src/super_gradients/examples/train_from_recipe_example/train_from_recipe.py --config-name=cityscapes_segformer_b0
+#
+#
+# Imagenet-1K pre-trained backbone:
+#   MiT (Mix Transformer) B0:   https://deci-pretrained-models.s3.amazonaws.com/mit_backbones/mit_b0.pth
+#
+#   1. Download the weights from the above link and put them in a directory of your choice
+#   2. Below, insert the weights file's full path to checkpoint_params.checkpoint_path
+#   3. Ensure checkpoint_params.load_backbone: True
+
+
+defaults:
+  - training_hyperparams: default_train_params
+  - dataset_params: cityscapes_segformer_dataset_params
+  - checkpoint_params: default_checkpoint_params
+  - _self_
+
+architecture: segformer_b0
+
+data_root_dir: /data/cityscapes
+dataset_params:
+  train_dataset_params:
+    root_dir: ${data_root_dir}
+  val_dataset_params:
+    root_dir: ${data_root_dir}
+
+experiment_name: ${architecture}_cityscapes
+ckpt_root_dir:
+
+train_dataloader: cityscapes_train
+val_dataloader: cityscapes_val
+
+cityscapes_ignored_label: 19    # convenience parameter since it is used in many places in the YAML
+
+arch_params:
+  num_classes: 19
+
+checkpoint_params:
+  checkpoint_path:
+  load_backbone: True
+  load_weights_only: True
+  strict_load: no_key_matching
+
+load_checkpoint: False
+
+resume: False
+training_hyperparams:
+
+  resume: ${resume}
+
+  max_epochs: 400
+
+  lr_mode: poly
+  initial_lr: 0.0002   # for effective batch_size=8
+
+  optimizer: AdamW
+  zero_weight_decay_on_bias_and_bn: True
+
+  sync_bn: True
+
+  loss: cross_entropy
+  criterion_params:
+    ignore_index: ${cityscapes_ignored_label}
+
+  train_metrics_list:
+    - IoU:
+        num_classes: 20
+        ignore_index: ${cityscapes_ignored_label}
+
+  valid_metrics_list:
+    - IoU:
+        num_classes: 20
+        ignore_index: ${cityscapes_ignored_label}
+
+  metric_to_watch: IoU
+  greater_metric_to_watch_is_better: True
+
+multi_gpu: DDP
+num_gpus: 4
+
+
+# THE FOLLOWING PARAMS ARE DIRECTLY USED BY HYDRA
+hydra:
+  run:
+    # Set the output directory (i.e. where .hydra folder that logs all the input params will be generated)
+    dir: ${hydra_output_dir:${ckpt_root_dir}, ${experiment_name}}

src/super_gradients/recipes/cityscapes_segformer_b1.yaml

@@ -0,0 +1,109 @@
+#  SegFormer-B1 segmentation training example with Cityscapes dataset.
+#  Reproduction of paper:
+#  Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar,  Jose M. Alvarez, Ping Luo
+#  "SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers"
+#  ( https://arxiv.org/pdf/2105.15203.pdf )
+#
+#  Official git repo:
+#      https://github.com/NVlabs/SegFormer
+#
+#
+#  Imagenet-1k pre-trained backbone weights taken and adapted from:
+#      https://github.com/sithu31296/semantic-segmentation
+#
+#
+# Instructions:
+#   1. We recommend preparing the data according to SG's CityScapes readme file:
+#      https://github.com/Deci-AI/super-gradients/blob/master/src/super_gradients/training/datasets/Dataset_Setup_Instructions.md
+#   2. Note: if you change the dataset's internal directory structure, make changes to the fields "list_file" and
+#      "labels_csv_path" of both "train_dataset_params" and "val_dataset_params" accordingly
+#   3. Edit the "data_root_dir" field below to point to the absolute path of the data root directory
+#   4. Edit the "ckpt_root_dir" field to the path where you want to save checkpoints and logs
+#   5. Move to the project root (where you will find the ReadMe and src folder)
+#   6. Run the command:
+#       python src/super_gradients/examples/train_from_recipe_example/train_from_recipe.py --config-name=cityscapes_segformer_b1
+#
+#
+# Imagenet-1K pre-trained backbone:
+#   MiT (Mix Transformer) B1:   https://deci-pretrained-models.s3.amazonaws.com/mit_backbones/mit_b1.pth
+#
+#   1. Download the weights from the above link and put them in a directory of your choice
+#   2. Below, insert the weights file's full path to checkpoint_params.checkpoint_path
+#   3. Ensure checkpoint_params.load_backbone: True
+
+
+defaults:
+  - training_hyperparams: default_train_params
+  - dataset_params: cityscapes_segformer_dataset_params
+  - checkpoint_params: default_checkpoint_params
+  - _self_
+
+architecture: segformer_b1
+
+data_root_dir: /data/cityscapes
+dataset_params:
+  train_dataset_params:
+    root_dir: ${data_root_dir}
+  val_dataset_params:
+    root_dir: ${data_root_dir}
+
+experiment_name: ${architecture}_cityscapes
+ckpt_root_dir:
+
+train_dataloader: cityscapes_train
+val_dataloader: cityscapes_val
+
+cityscapes_ignored_label: 19    # convenience parameter since it is used in many places in the YAML
+
+arch_params:
+  num_classes: 19
+
+checkpoint_params:
+  checkpoint_path:
+  load_backbone: True
+  load_weights_only: True
+  strict_load: no_key_matching
+
+load_checkpoint: False
+
+resume: False
+training_hyperparams:
+
+  resume: ${resume}
+
+  max_epochs: 400
+
+  lr_mode: poly
+  initial_lr: 0.0002   # for effective batch_size=8
+
+  optimizer: AdamW
+  zero_weight_decay_on_bias_and_bn: True
+
+  sync_bn: True
+
+  loss: cross_entropy
+  criterion_params:
+    ignore_index: ${cityscapes_ignored_label}
+
+  train_metrics_list:
+    - IoU:
+        num_classes: 20
+        ignore_index: ${cityscapes_ignored_label}
+
+  valid_metrics_list:
+    - IoU:
+        num_classes: 20
+        ignore_index: ${cityscapes_ignored_label}
+
+  metric_to_watch: IoU
+  greater_metric_to_watch_is_better: True
+
+multi_gpu: DDP
+num_gpus: 4
+
+
+# THE FOLLOWING PARAMS ARE DIRECTLY USED BY HYDRA
+hydra:
+  run:
+    # Set the output directory (i.e. where .hydra folder that logs all the input params will be generated)
+    dir: ${hydra_output_dir:${ckpt_root_dir}, ${experiment_name}}

src/super_gradients/recipes/cityscapes_segformer_b2.yaml

@@ -0,0 +1,109 @@
+#  SegFormer-B2 segmentation training example with Cityscapes dataset.
+#  Reproduction of paper:
+#  Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar,  Jose M. Alvarez, Ping Luo
+#  "SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers"
+#  ( https://arxiv.org/pdf/2105.15203.pdf )
+#
+#  Official git repo:
+#      https://github.com/NVlabs/SegFormer
+#
+#
+#  Imagenet-1k pre-trained backbone weights taken and adapted from:
+#      https://github.com/sithu31296/semantic-segmentation
+#
+#
+# Instructions:
+#   1. We recommend preparing the data according to SG's CityScapes readme file:
+#      https://github.com/Deci-AI/super-gradients/blob/master/src/super_gradients/training/datasets/Dataset_Setup_Instructions.md
+#   2. Note: if you change the dataset's internal directory structure, make changes to the fields "list_file" and
+#      "labels_csv_path" of both "train_dataset_params" and "val_dataset_params" accordingly
+#   3. Edit the "data_root_dir" field below to point to the absolute path of the data root directory
+#   4. Edit the "ckpt_root_dir" field to the path where you want to save checkpoints and logs
+#   5. Move to the project root (where you will find the ReadMe and src folder)
+#   6. Run the command:
+#       python src/super_gradients/examples/train_from_recipe_example/train_from_recipe.py --config-name=cityscapes_segformer_b2
+#
+#
+# Imagenet-1K pre-trained backbone:
+#   MiT (Mix Transformer) B2:   https://deci-pretrained-models.s3.amazonaws.com/mit_backbones/mit_b2.pth
+#
+#   1. Download the weights from the above link and put them in a directory of your choice
+#   2. Below, insert the weights file's full path to checkpoint_params.checkpoint_path
+#   3. Ensure checkpoint_params.load_backbone: True
+
+
+defaults:
+  - training_hyperparams: default_train_params
+  - dataset_params: cityscapes_segformer_dataset_params
+  - checkpoint_params: default_checkpoint_params
+  - _self_
+
+architecture: segformer_b2
+
+data_root_dir: /data/cityscapes
+dataset_params:
+  train_dataset_params:
+    root_dir: ${data_root_dir}
+  val_dataset_params:
+    root_dir: ${data_root_dir}
+
+experiment_name: ${architecture}_cityscapes
+ckpt_root_dir:
+
+train_dataloader: cityscapes_train
+val_dataloader: cityscapes_val
+
+cityscapes_ignored_label: 19    # convenience parameter since it is used in many places in the YAML
+
+arch_params:
+  num_classes: 19
+
+checkpoint_params:
+  checkpoint_path:
+  load_backbone: True
+  load_weights_only: True
+  strict_load: no_key_matching
+
+load_checkpoint: False
+
+resume: False
+training_hyperparams:
+
+  resume: ${resume}
+
+  max_epochs: 400
+
+  lr_mode: poly
+  initial_lr: 0.0002   # for effective batch_size=8
+
+  optimizer: AdamW
+  zero_weight_decay_on_bias_and_bn: True
+
+  sync_bn: True
+
+  loss: cross_entropy
+  criterion_params:
+    ignore_index: ${cityscapes_ignored_label}
+
+  train_metrics_list:
+    - IoU:
+        num_classes: 20
+        ignore_index: ${cityscapes_ignored_label}
+
+  valid_metrics_list:
+    - IoU:
+        num_classes: 20
+        ignore_index: ${cityscapes_ignored_label}
+
+  metric_to_watch: IoU
+  greater_metric_to_watch_is_better: True
+
+multi_gpu: DDP
+num_gpus: 4
+
+
+# THE FOLLOWING PARAMS ARE DIRECTLY USED BY HYDRA
+hydra:
+  run:
+    # Set the output directory (i.e. where .hydra folder that logs all the input params will be generated)
+    dir: ${hydra_output_dir:${ckpt_root_dir}, ${experiment_name}}