Merge pull request #247 from Nota-NetsPresso/dev

Update from dev branch: `v0.0.10` release

CHANGELOG.md

@@ -2,19 +2,48 @@
 
 ## New Features:
 
-No changes to highlight.
+- 
 
 ## Bug Fixes:
 
-No changes to highlight.
+- 
 
 ## Breaking Changes:
 
-No changes to highlight.
+-
 
 ## Other Changes:
 
-No changes to highlight.
+- 
+
+# v0.0.10
+
+## New Features:
+
+- Add a gpu option in `train_with_config` (only single-GPU supported) by `@deepkyu` in [PR 219](https://github.com/Nota-NetsPresso/netspresso-trainer/pull/219)
+- Support augmentation for classification task: cutmix, mixup by `@illian01` in [PR 221](https://github.com/Nota-NetsPresso/netspresso-trainer/pull/221)
+- Add model: MixNet by `@illian01` in [PR 229](https://github.com/Nota-NetsPresso/netspresso-trainer/pull/229)
+- Add `model.name` to get the exact nickname of the model by `@deepkyu` in [PR 243](https://github.com/Nota-NetsPresso/netspresso-trainer/pull/243/)
+- Add transforms: RandomErasing and TrivialAugmentationWide by `@illian01` in [PR 246](https://github.com/Nota-NetsPresso/netspresso-trainer/pull/246)
+
+## Bug Fixes:
+
+- Fix PIDNet model dataclass task field by `@illian01` in [PR 220](https://github.com/Nota-NetsPresso/netspresso-trainer/pull/220)
+- Fix default criterion value of classification `@illian01` in [PR 238](https://github.com/Nota-NetsPresso/netspresso-trainer/pull/238)
+- Fix model access of 2-stage detection pipeline to compat with distributed environment by `@illian` in [PR 239](https://github.com/Nota-NetsPresso/netspresso-trainer/pull/239)
+
+## Breaking Changes:
+
+- Enable dataset augmentation customizing by `@illian01` in [PR 201](https://github.com/Nota-NetsPresso/netspresso-trainer/pull/201)
+- Add postprocessor module by `@illian01` in [PR 223](https://github.com/Nota-NetsPresso/netspresso-trainer/pull/223)
+- Equalize the model backbone configuration format by `@illian01` in [PR 228](https://github.com/Nota-NetsPresso/netspresso-trainer/pull/228)
+- Separate FPN and PAFPN as neck module by `@illian01` in [PR 234](https://github.com/Nota-NetsPresso/netspresso-trainer/pull/234)
+- Auto-download pretrained checkpoint from AWS S3 by `@deepkyu` in [PR 244](https://github.com/Nota-NetsPresso/netspresso-trainer/pull/244)
+
+## Other Changes:
+
+- Update ruff rule (`W`) by `@deepkyu` in [PR 218](https://github.com/Nota-NetsPresso/netspresso-trainer/pull/218)
+- Integrate classification loss modules by `@illian01` in [PR 226](https://github.com/Nota-NetsPresso/netspresso-trainer/pull/226)
 
 # v0.0.9
 
@@ -121,6 +150,7 @@ This change is applied at [PR 151](https://github.com/Nota-NetsPresso/netspresso
 - Initialize loss and metric at same time with optimizer and lr schedulers by `@deepkyu` in [PR 138](https://github.com/Nota-NetsPresso/netspresso-trainer/pull/138)
 - Hotfix the error which shows 0 for validation loss and metrics by fixing the variable name by `@deepkyu` in [PR 140](https://github.com/Nota-NetsPresso/netspresso-trainer/pull/140)
 - Add missing field, `save_optimizer_state`, in `logging.yaml` by `@illian01` in [PR 149](https://github.com/Nota-NetsPresso/netspresso-trainer/pull/149)
+- Hotfix for pythonic config name (classification loss) by `@deepkyu` in [PR 242](https://github.com/Nota-NetsPresso/netspresso-trainer/pull/242)
 
 ## Breaking Changes:
 

config/augmentation/classification.yaml

@@ -1,31 +1,14 @@
 augmentation:
   img_size: &img_size 256
-  hsv_h: ~
-  hsv_s: ~
-  hsv_v: ~
-  degrees: ~
-  translate: ~
-  scale: ~ 
-  max_scale: ~
-  min_scale: ~  
-  crop_size_h: ~
-  crop_size_w: ~
-  resize_ratio0: ~
-  resize_ratiof: ~
-  resize_add: ~
-  shear: ~
-  perspective: ~
-  flipud: ~
-  fliplr: 0.5
-  mosaic: ~ 
-  mixup: 1.0 
-  copy_paste: ~ 
-  mixup_alpha: 0.0 
-  cutmix_alpha: 0.0 
-  mixup_switch_prob: 0.5 
-  color_jitter:
-    brightness: ~ 
-    contrast: ~ 
-    saturation: ~
-    hue: ~ 
-    colorjitter_p: ~ 
+  transforms:
+    - 
+      name: randomresizedcrop
+      size: *img_size
+      interpolation: bilinear
+    - 
+      name: randomhorizontalflip
+      p: 0.5
+  mix_transforms:
+    -
+      name: cutmix
+      alpha: 1.0

config/augmentation/detection.yaml

@@ -1,31 +1,7 @@
 augmentation:
   img_size: &img_size 512
-  hsv_h: ~
-  hsv_s: ~
-  hsv_v: ~
-  degrees: ~
-  translate: ~
-  scale: ~ 
-  max_scale: 2048
-  min_scale: 768
-  crop_size_h: 512
-  crop_size_w: 512
-  resize_ratio0: 0.5
-  resize_ratiof: 2.0
-  resize_add: 1
-  shear: ~
-  perspective: ~
-  flipud: ~
-  fliplr: 0.5
-  mosaic: ~
-  mixup: ~
-  copy_paste: ~
-  mixup_alpha: ~
-  cutmix_alpha: ~
-  mixup_switch_prob: ~
-  color_jitter:
-    brightness: 0.25
-    contrast: 0.25
-    saturation: 0.25
-    hue: 0.1
-    colorjitter_p: 0.5
+  transforms:
+    - 
+      name: resize
+      size: *img_size
+      interpolation: bilinear

config/augmentation/segmentation.yaml

@@ -1,31 +1,17 @@
 augmentation:
   img_size: &img_size 512
-  hsv_h: ~
-  hsv_s: ~
-  hsv_v: ~
-  degrees: ~
-  translate: ~
-  scale: ~ 
-  max_scale: 1024
-  min_scale: *img_size
-  crop_size_h: *img_size
-  crop_size_w: *img_size
-  resize_ratio0: 1.0
-  resize_ratiof: 1.5
-  resize_add: 1
-  shear: ~
-  perspective: ~
-  flipud: ~
-  fliplr: 0.5
-  mosaic: ~
-  mixup: ~
-  copy_paste: ~
-  mixup_alpha: ~
-  cutmix_alpha: ~
-  mixup_switch_prob: ~
-  color_jitter:
-    brightness: 0.25
-    contrast: 0.25
-    saturation: 0.25
-    hue: 0.1
-    colorjitter_p: 0.5
+  transforms:
+    - 
+      name: randomresizedcrop
+      size: *img_size
+      interpolation: bilinear
+    -
+      name: randomhorizontalflip
+      p: 0.5
+    -
+      name: colorjitter
+      brightness: 0.25
+      contrast: 0.25
+      saturation: 0.25
+      hue: 0.1
+      p: 0.5

config/augmentation/template/common.yaml

@@ -1,31 +1,27 @@
 augmentation:
-  img_size: &img_size 512
-  hsv_h: ~
-  hsv_s: ~
-  hsv_v: ~
-  degrees: ~
-  translate: ~
-  scale: ~ 
-  max_scale: 1024
-  min_scale: *img_size
-  crop_size_h: *img_size
-  crop_size_w: *img_size
-  resize_ratio0: 1.0
-  resize_ratiof: 1.5
-  resize_add: 1
-  shear: ~
-  perspective: ~
-  flipud: ~
-  fliplr: 0.5
-  mosaic: ~
-  mixup: ~
-  copy_paste: ~
-  mixup_alpha: ~
-  cutmix_alpha: ~
-  mixup_switch_prob: ~
-  color_jitter:
-    brightness: 0.25
-    contrast: 0.25
-    saturation: 0.25
-    hue: 0.1
-    colorjitter_p: 0.5
+  img_size: &img_size ~
+  transforms:
+    - 
+      name: randomresizedcrop
+      size: ~
+      interpolation: bilinear
+    -
+      name: randomhorizontalflip
+      p: ~
+    -
+      name: randomverticalflip
+      p: ~
+    -
+      name: colorjitter
+      brightness: ~
+      contrast: ~
+      saturation: ~
+      hue: ~
+      p: ~
+    -
+      name: resize
+      size: ~
+    -
+      name: pad
+      padding: ~
+  mix_transforms: ~

config/model/efficientformer/efficientformer-l1-classification.yaml

@@ -1,5 +1,6 @@
 model:
   task: classification
+  name: efficientformer_l1
   checkpoint: ./weights/efficientformer/efficientformer_l1_1000d.pth
   fx_model_checkpoint: ~
   resume_optimizer_checkpoint: ~
@@ -8,29 +9,44 @@ model:
     full: ~ # auto
     backbone:
       name: efficientformer
-      num_blocks: [3, 2, 6, 4]
-      hidden_sizes: [48, 96, 224, 448]
-      num_attention_heads: 8
-      attention_hidden_size: 256  # attention_hidden_size_splitted * num_attention_heads
-      attention_dropout_prob: 0.
-      attention_ratio: 4
-      attention_bias_resolution: 16
-      pool_size: 3
-      intermediate_ratio: 4
-      hidden_dropout_prob: 0.
-      hidden_activation_type: 'gelu'
-      layer_norm_eps: 1e-5
-      drop_path_rate: 0.
-      use_layer_scale: True
-      layer_scale_init_value: 1e-5
-      downsamples: [True, True, True, True]
-      down_patch_size: 3
-      down_stride: 2
-      down_pad: 1
-      vit_num: 1
+      params:
+        num_attention_heads: 8
+        attention_hidden_size: 256  # attention_hidden_size_splitted * num_attention_heads
+        attention_dropout_prob: 0.
+        attention_ratio: 4
+        attention_bias_resolution: 16
+        pool_size: 3
+        intermediate_ratio: 4
+        hidden_dropout_prob: 0.
+        hidden_activation_type: 'gelu'
+        layer_norm_eps: 1e-5
+        drop_path_rate: 0.
+        use_layer_scale: True
+        layer_scale_init_value: 1e-5
+        down_patch_size: 3
+        down_stride: 2
+        down_pad: 1
+        vit_num: 1
+      stage_params:
+        - 
+          num_blocks: 3
+          hidden_sizes: 48
+          downsamples: True
+        - 
+          num_blocks: 2
+          hidden_sizes: 96
+          downsamples: True
+        - 
+          num_blocks: 6
+          hidden_sizes: 224
+          downsamples: True
+        - 
+          num_blocks: 4
+          hidden_sizes: 448
+          downsamples: True
     head:
       name: fc
   losses:
-    - criterion: label_smoothing_cross_entropy
-      smoothing: 0.1
+    - criterion: cross_entropy
+      label_smoothing: 0.1
       weight: ~

config/model/efficientformer/efficientformer-l1-detection.yaml

@@ -1,5 +1,6 @@
 model:
   task: detection
+  name: efficientformer_l1
   checkpoint: ./weights/efficientformer/efficientformer_l1_1000d.pth
   fx_model_checkpoint: ~
   resume_optimizer_checkpoint: ~
@@ -8,26 +9,43 @@ model:
     full: ~ # auto
     backbone:
       name: efficientformer
-      num_blocks: [3, 2, 6, 4]
-      hidden_sizes: [48, 96, 224, 448]
-      num_attention_heads: 8
-      attention_hidden_size: 256  # attention_hidden_size_splitted * num_attention_heads
-      attention_dropout_prob: 0.
-      attention_ratio: 4
-      attention_bias_resolution: 16
-      pool_size: 3
-      intermediate_ratio: 4
-      hidden_dropout_prob: 0.
-      hidden_activation_type: 'gelu'
-      layer_norm_eps: 1e-5
-      drop_path_rate: 0.
-      use_layer_scale: True
-      layer_scale_init_value: 1e-5
-      downsamples: [True, True, True, True]
-      down_patch_size: 3
-      down_stride: 2
-      down_pad: 1
-      vit_num: 1
+      params:
+        num_attention_heads: 8
+        attention_hidden_size: 256  # attention_hidden_size_splitted * num_attention_heads
+        attention_dropout_prob: 0.
+        attention_ratio: 4
+        attention_bias_resolution: 16
+        pool_size: 3
+        intermediate_ratio: 4
+        hidden_dropout_prob: 0.
+        hidden_activation_type: 'gelu'
+        layer_norm_eps: 1e-5
+        drop_path_rate: 0.
+        use_layer_scale: True
+        layer_scale_init_value: 1e-5
+        down_patch_size: 3
+        down_stride: 2
+        down_pad: 1
+        vit_num: 1
+      stage_params:
+        - 
+          num_blocks: 3
+          hidden_sizes: 48
+          downsamples: True
+        - 
+          num_blocks: 2
+          hidden_sizes: 96
+          downsamples: True
+        - 
+          num_blocks: 6
+          hidden_sizes: 224
+          downsamples: True
+        - 
+          num_blocks: 4
+          hidden_sizes: 448
+          downsamples: True
+    neck:
+      name: fpn
     head:
       name: faster_rcnn
   losses: