Change some mechanism

README.md

@@ -31,7 +31,7 @@ python train.py -c pipeline.yaml
 ```
 python train.py \
     -c pipeline.yaml \
-    -o trainer.args.num_epochs=50 \
+    -o trainer.args.num_iterations=5000 \
     global.resume=checkpoint.pth
 ```
 **Notice: There are no spaces between keys and values in -o flag**

configs/classification/pipeline.yaml

@@ -1,94 +1,87 @@
 global:
-  debug: True
+  debug: true
   cfg_transform: configs/classification/transform.yaml
-  save_dir: runs
+  save_dir: /content/main/runs
   device: cuda:0
-  use_fp16: True
+  use_fp16: true
   pretrained: null
   resume: null
-
-
 trainer:
   name: ClassificationTrainer
   args:
-    num_epochs: 100
-    total_accumulate_steps: null
+    num_iterations: 10000
     clip_grad:  10.0
-    print_per_iter: 20
-    save_per_iter: 1000
-    evaluate_per_epoch :  1
+    print_interval: 20
+    save_interval: 1000
+    evaluate_interval :  1
     visualize_when_val :  True
-
-
 model:
   name: BaseTimmModel
   args:
     name: convnext_small
-    from_pretrained: True
+    from_pretrained: true
     num_classes: 2
-
-
 loss:
-  name: SmoothCELoss
-  args:
-    smoothing: 0.1
-
+  name: CELoss
 metrics:
-  - name: Accuracy
-    args:
-  - name: BalancedAccuracyMetric
-    args: 
-  - name: F1ScoreMetric
-    args:
-      average: weighted
-  - name: ConfusionMatrix
-    args:
-  - name: ErrorCases
-    args:
-
+- name: Accuracy
+  args: null
+- name: BalancedAccuracyMetric
+  args: null
+- name: F1ScoreMetric
+  args:
+    average: weighted
+- name: ConfusionMatrix
+  args: null
+- name: ErrorCases
+  args:
 optimizer:
   name: AdamW
   args:
-    lr: 0.001               # [adam: 1e-3 | sgd: 1e-2]
+    lr: 0.001
     weight_decay: 0.0005
-    betas: [0.937, 0.999]
-
+    betas:
+    - 0.937
+    - 0.999
 scheduler:
   name: SchedulerWrapper
   args:
     scheduler_name: cosine2
     t_initial: 7
     t_mul: 0.9
     eta_mul: 0.9
-    eta_min: 0.000001
-
-
-data: 
+    eta_min: 1.0e-06
+data:
   dataset:
     train:
       name: ImageFolderDataset
       args:
-        image_dir: data/dog-vs-cats/train
+        image_dir: /content/main/data/dog-vs-cats/train
         txt_classnames: configs/classification/classes.txt
-        test: False
-
     val:
       name: ImageFolderDataset
       args:
-        image_dir: data/dog-vs-cats/val
+        image_dir: /content/main/data/dog-vs-cats/val
         txt_classnames: configs/classification/classes.txt
-        test: True
-
   dataloader:
     train:
-      name: BalanceSampler
+      name: DataLoaderWithCollator
       args:
-        batch_size: 64
-        drop_last: True
-        shuffle: False
+        batch_size: 32
+        drop_last: true
+        shuffle: false
+        collate_fn:
+            name: MixupCutmixCollator
+            args:
+              mixup_alpha: 0.4
+              cutmix_alpha: 1.0
+              weight: [0.2, 0.2]
+        sampler:
+            name: BalanceSampler
+            args:
     val:
       name: DataLoaderWithCollator
       args:
-        batch_size: 64
-        drop_last: False
-        shuffle: True
+        batch_size: 32
+        drop_last: false
+        shuffle: true

configs/segmentation/eval.py → configs/semantic/eval.py

@@ -2,7 +2,7 @@
 
 mpl.use("Agg")
 from theseus.opt import Opts
-from theseus.segmentation.pipeline import Pipeline
+from theseus.semantic.pipeline import Pipeline
 
 if __name__ == "__main__":
     opts = Opts().parse_args()

configs/segmentation/infer.py → configs/semantic/infer.py

@@ -9,17 +9,17 @@
 import torch
 from datetime import datetime
 from theseus.opt import Config
-from theseus.segmentation.models import MODEL_REGISTRY
-from theseus.segmentation.augmentations import TRANSFORM_REGISTRY
-from theseus.segmentation.datasets import DATASET_REGISTRY, DATALOADER_REGISTRY
+from theseus.semantic.models import MODEL_REGISTRY
+from theseus.semantic.augmentations import TRANSFORM_REGISTRY
+from theseus.semantic.datasets import DATASET_REGISTRY, DATALOADER_REGISTRY
 
 from theseus.utilities.loading import load_state_dict
 from theseus.utilities.loggers import LoggerObserver, StdoutLogger
 from theseus.utilities.cuda import get_devices_info
 from theseus.utilities.getter import (get_instance, get_instance_recursively)
 
 from theseus.utilities.visualization.visualizer import Visualizer
-from theseus.segmentation.datasets.csv_dataset import CSVDataset
+from theseus.semantic.datasets.csv_dataset import CSVDataset
 
 @DATASET_REGISTRY.register()
 class TestCSVDataset(CSVDataset):

configs/segmentation/pipeline.yaml → configs/semantic/pipeline.yaml

@@ -1,20 +1,19 @@
 global:
   debug: true
-  cfg_transform: configs/segmentation/transform.yaml
+  cfg_transform: configs/semantic/transform.yaml
   save_dir: runs
   device: cuda:0
   use_fp16: true
   pretrained: null
   resume: null
 trainer:
-  name: SegmentationTrainer
+  name: SemanticTrainer
   args:
-    num_epochs: 50
-    total_accumulate_steps: null
+    num_iterations: 50000
     clip_grad: 10.0
-    print_per_iter: 20
-    save_per_iter: 1000
-    evaluate_per_epoch: 1
+    print_interval: 20
+    save_interval: 1000
+    evaluate_interval: 1
     visualize_when_val: true
 model:
   name: BaseSegModel
@@ -61,7 +60,7 @@ data:
         image_dir: data/images/train
         mask_dir: data/masks/train
         csv_path: data/train.csv
-        txt_classnames: configs/segmentation/classes.txt
+        txt_classnames: configs/semantic/classes.txt
         mosaic_size: 960 
         mosaic_prob: 0.4
     val:
@@ -70,7 +69,7 @@ data:
         image_dir: data/images/val
         mask_dir: data/masks/val
         csv_path: data/val.csv
-        txt_classnames: configs/segmentation/classes.txt
+        txt_classnames: configs/semantic/classes.txt
   dataloader:
     train:
       name: DataLoaderWithCollator

configs/segmentation/train.py → configs/semantic/train.py

@@ -2,7 +2,7 @@
 
 mpl.use("Agg")
 from theseus.opt import Opts
-from theseus.segmentation.pipeline import Pipeline
+from theseus.semantic.pipeline import Pipeline
 
 if __name__ == "__main__":
     opts = Opts().parse_args()

configs/segmentation/transform.yaml → configs/semantic/transform.yaml

@@ -27,8 +27,8 @@ train:
       #     value: 0
       - name: AlbNormalize
         args:
-          mean: [0.229, 0.224, 0.225]
-          std: [0.485, 0.456, 0.406]
+          mean: [0.485, 0.456, 0.406]
+          std: [0.229, 0.224, 0.225]
       - name: AlbToTensorV2
 val:
   name: AlbCompose
@@ -40,6 +40,6 @@ val:
           height: 640
       - name: AlbNormalize
         args:
-          mean: [0.229, 0.224, 0.225]
-          std: [0.485, 0.456, 0.406]
+          mean: [0.485, 0.456, 0.406]
+          std: [0.229, 0.224, 0.225]
       - name: AlbToTensorV2

requirements.txt

@@ -1,13 +1,13 @@
-albumentations
+albumentations>=1.1.0
 pyyaml>=5.1
 webcolors
 tensorboard
 tqdm
-ensemble-boxes
-timm
+timm>=0.5.4
 omegaconf
 pycocotools
-gdown==4.3.0
+gdown>=4.4.0
 grad-cam
 tabulate
 segmentation-models-pytorch
+opencv-python-headless==4.1.2.30 

setup.py

@@ -18,5 +18,7 @@
         "omegaconf",
         "gdown==3.13.0",
         "grad-cam",
+        "tabulate",
+        "segmentation-models-pytorch"
     ],
 )

theseus/base/datasets/__init__.py

@@ -2,7 +2,8 @@
 from torch.utils.data import DataLoader, Dataset
 from .dataloader import DataLoaderWithCollator
 from .dataset import ChainDataset, ConcatDataset, ImageDataset
-from .balance_sampler import BalanceSampler
+from .sampler import BalanceSampler
+from .collator import ChainCollatorWrapper
 
 DATASET_REGISTRY = Registry('DATASET')
 DATASET_REGISTRY.register(Dataset)
@@ -13,4 +14,5 @@
 DATALOADER_REGISTRY = Registry('DATALOADER')
 DATALOADER_REGISTRY.register(DataLoader)
 DATALOADER_REGISTRY.register(BalanceSampler)
+DATALOADER_REGISTRY.register(ChainCollatorWrapper)
 DATALOADER_REGISTRY.register(DataLoaderWithCollator)

theseus/base/datasets/collator.py

@@ -0,0 +1,21 @@
+from typing import List
+
+class BaseCollator(object):
+    """Base collator function
+    """
+    def __init__(self, **kwargs) -> None:
+        pass
+
+    def __call__(self, batch):
+        return batch
+
+class ChainCollatorWrapper(BaseCollator):
+    """Wrapper for list of collate functions
+    """
+    def __init__(self, pre_collate_fns: List, **kwargs):
+        self.pre_collate_fns = pre_collate_fns
+
+    def __call__(self, batch):
+        for fn in self.pre_collate_fns:
+            batch = fn(batch)
+        return batch