Add light models to guarantee the backward compatibility

openvinotoolkit · Dec 11, 2023 · 985fca2 · 985fca2
1 parent c021f18
commit 985fca2
Show file tree

Hide file tree

Showing 6 changed files with 271 additions and 4 deletions.
diff --git a/src/otx/recipe/classification/efficientnet_b0_light.yaml b/src/otx/recipe/classification/efficientnet_b0_light.yaml
@@ -0,0 +1,77 @@
+# @package _global_
+defaults:
+  - override /base: classification
+  - override /callbacks: classification
+  - override /data: mmpretrain
+  - override /model: mmpretrain
+data:
+  train_subset:
+    batch_size: 64
+    transforms:
+      - type: LoadImageFromFile
+      - backend: cv2
+        scale: 224
+        type: RandomResizedCrop
+      - type: PackInputs
+  val_subset:
+    batch_size: 64
+    transforms:
+      - type: LoadImageFromFile
+      - backend: cv2
+        edge: short
+        scale: 256
+        type: ResizeEdge
+      - crop_size: 224
+        type: CenterCrop
+      - type: PackInputs
+  test_subset:
+    batch_size: 64
+    transforms:
+      - type: LoadImageFromFile
+      - backend: cv2
+        edge: short
+        scale: 256
+        type: ResizeEdge
+      - crop_size: 224
+        type: CenterCrop
+      - type: PackInputs
+model:
+  otx_model:
+    config:
+      backbone:
+        version: b0
+        pretrained: true
+        type: OTXEfficientNet
+      head:
+        in_channels: 1280
+        loss:
+          loss_weight: 1.0
+          type: CrossEntropyLoss
+        num_classes: 1000
+        topk:
+          - 1
+          - 5
+        type: LinearClsHead
+      neck:
+        type: GlobalAveragePooling
+      data_preprocessor:
+        mean:
+          - 123.675
+          - 116.28
+          - 103.53
+        std:
+          - 58.395
+          - 57.12
+          - 57.375
+        to_rgb: True
+        type: ClsDataPreprocessor
+      type: ImageClassifier
+  optimizer:
+    _target_: torch.optim.SGD
+    _partial_: true
+    lr: 0.0049
+    momentum: 0.9
+    weight_decay: 0.0001
+  scheduler:
+    factor: 0.5
+    patience: 1
diff --git a/src/otx/recipe/classification/efficientnet_v2_light.yaml b/src/otx/recipe/classification/efficientnet_v2_light.yaml
@@ -0,0 +1,79 @@
+# @package _global_
+defaults:
+  - override /base: classification
+  - override /callbacks: classification
+  - override /data: mmpretrain
+  - override /model: mmpretrain
+data:
+  train_subset:
+    batch_size: 64
+    transforms:
+      - type: LoadImageFromFile
+      - backend: cv2
+        scale: 224
+        type: RandomResizedCrop
+      - direction: horizontal
+        prob: 0.5
+        type: RandomFlip
+      - type: PackInputs
+  val_subset:
+    batch_size: 64
+    transforms:
+      - type: LoadImageFromFile
+      - backend: cv2
+        edge: short
+        scale: 256
+        type: ResizeEdge
+      - crop_size: 224
+        type: CenterCrop
+      - type: PackInputs
+  test_subset:
+    batch_size: 64
+    transforms:
+      - type: LoadImageFromFile
+      - backend: cv2
+        edge: short
+        scale: 256
+        type: ResizeEdge
+      - crop_size: 224
+        type: CenterCrop
+      - type: PackInputs
+model:
+  otx_model:
+    config:
+      backbone:
+        pretrained: true
+        type: OTXEfficientNetV2
+      head:
+        in_channels: 1280
+        loss:
+          loss_weight: 1.0
+          type: CrossEntropyLoss
+        num_classes: 1000
+        topk:
+          - 1
+          - 5
+        type: LinearClsHead
+      neck:
+        type: GlobalAveragePooling
+      data_preprocessor:
+        mean:
+          - 123.675
+          - 116.28
+          - 103.53
+        std:
+          - 58.395
+          - 57.12
+          - 57.375
+        to_rgb: True
+        type: ClsDataPreprocessor
+      type: ImageClassifier
+  optimizer:
+    _target_: torch.optim.SGD
+    _partial_: true
+    lr: 0.0071
+    momentum: 0.9
+    weight_decay: 0.0001
+  scheduler:
+    factor: 0.5
+    patience: 1
diff --git a/src/otx/recipe/classification/mobilenet_v3_large_light.yaml b/src/otx/recipe/classification/mobilenet_v3_large_light.yaml
@@ -0,0 +1,78 @@
+# @package _global_
+defaults:
+  - override /base: classification
+  - override /callbacks: classification
+  - override /data: mmpretrain
+  - override /model: mmpretrain
+data:
+  train_subset:
+    batch_size: 64
+    transforms:
+      - type: LoadImageFromFile
+      - backend: cv2
+        scale: 224
+        type: RandomResizedCrop
+      - direction: horizontal
+        prob: 0.5
+        type: RandomFlip
+      - type: PackInputs
+  val_subset:
+    batch_size: 64
+    transforms:
+      - type: LoadImageFromFile
+      - backend: cv2
+        edge: short
+        scale: 256
+        type: ResizeEdge
+      - crop_size: 224
+        type: CenterCrop
+      - type: PackInputs
+  test_subset:
+    batch_size: 64
+    transforms:
+      - type: LoadImageFromFile
+      - backend: cv2
+        edge: short
+        scale: 256
+        type: ResizeEdge
+      - crop_size: 224
+        type: CenterCrop
+      - type: PackInputs
+model:
+  otx_model:
+    config:
+      backbone:
+        type: OTXMobileNetV3
+      head:
+        in_channels: 960
+        loss:
+          loss_weight: 1.0
+          type: CrossEntropyLoss
+        num_classes: 1000
+        topk:
+          - 1
+          - 5
+        type: LinearClsHead
+      neck:
+        type: GlobalAveragePooling
+      data_preprocessor:
+        mean:
+          - 123.675
+          - 116.28
+          - 103.53
+        std:
+          - 58.395
+          - 57.12
+          - 57.375
+        to_rgb: True
+        type: ClsDataPreprocessor
+      type: ImageClassifier
+  optimizer:
+    _target_: torch.optim.SGD
+    _partial_: true
+    lr: 0.0058
+    momentum: 0.9
+    weight_decay: 0.0001
+  scheduler:
+    factor: 0.5
+    patience: 1
diff --git a/src/otx/recipe/classification/otx_efficientnet_b0.yaml b/src/otx/recipe/classification/otx_efficientnet_b0.yaml
@@ -43,15 +43,26 @@ model:
         pretrained: true
         type: OTXEfficientNet
       head:
+        act_cfg:
+          type: HSwish
+        dropout_rate: 0.2
         in_channels: 1280
+        init_cfg:
+          bias: 0.0
+          layer: Linear
+          mean: 0.0
+          std: 0.01
+          type: Normal
         loss:
           loss_weight: 1.0
           type: CrossEntropyLoss
+        mid_channels:
+          - 1280
         num_classes: 1000
         topk:
           - 1
           - 5
-        type: LinearClsHead
+        type: StackedLinearClsHead
       neck:
         type: GlobalAveragePooling
       data_preprocessor:

diff --git a/src/otx/recipe/classification/otx_efficientnet_v2.yaml b/src/otx/recipe/classification/otx_efficientnet_v2.yaml
@@ -45,15 +45,26 @@ model:
         pretrained: true
         type: OTXEfficientNetV2
       head:
+        act_cfg:
+          type: HSwish
+        dropout_rate: 0.2
         in_channels: 1280
+        init_cfg:
+          bias: 0.0
+          layer: Linear
+          mean: 0.0
+          std: 0.01
+          type: Normal
         loss:
           loss_weight: 1.0
           type: CrossEntropyLoss
+        mid_channels:
+          - 1280
         num_classes: 1000
         topk:
           - 1
           - 5
-        type: LinearClsHead
+        type: StackedLinearClsHead
       neck:
         type: GlobalAveragePooling
       data_preprocessor:

diff --git a/src/otx/recipe/classification/otx_mobilenet_v3_large.yaml b/src/otx/recipe/classification/otx_mobilenet_v3_large.yaml
@@ -44,15 +44,26 @@ model:
       backbone:
         type: OTXMobileNetV3
       head:
-        in_channels: 960
+        act_cfg:
+          type: HSwish
+        dropout_rate: 0.2
+        in_channels: 1280
+        init_cfg:
+          bias: 0.0
+          layer: Linear
+          mean: 0.0
+          std: 0.01
+          type: Normal
         loss:
           loss_weight: 1.0
           type: CrossEntropyLoss
+        mid_channels:
+          - 1280
         num_classes: 1000
         topk:
           - 1
           - 5
-        type: LinearClsHead
+        type: StackedLinearClsHead
       neck:
         type: GlobalAveragePooling
       data_preprocessor: