Fix(epoch_metric): fix epoch metric hand out

Makefile

@@ -49,4 +49,4 @@ pipeline:
 	bash ./tools/train.sh \
 		--config-path=./configs/default.py \
 		--cuda-visible-devices=0,1,2,3 \
-		--is-distributed=false
+		--is-distributed=true

configs/default.py

@@ -37,9 +37,9 @@
 )
 
 DATA = dict(
-    train=dict(type=__data_type, path=__psga_dirpath, phase="train", fold=__fold, tiles_intersection=0.5,
-               micron_tile_size=__microns_tile_size, crop_emptiness_degree=0.9, label_binning=True,
-               subsample_tiles_count=10, balance_subsample=True),
+    # train=dict(type=__data_type, path=__psga_dirpath, phase="train", fold=__fold, tiles_intersection=0.5,
+    #            micron_tile_size=__microns_tile_size, crop_emptiness_degree=0.9, label_binning=True,
+    #            subsample_tiles_count=10, balance_subsample=True),
 
     val=dict(type=__data_type, path=__psga_dirpath, phase="val", fold=__fold, tiles_intersection=0.0,
              micron_tile_size=__microns_tile_size, crop_emptiness_degree=0.95, label_binning=True),
@@ -96,7 +96,7 @@
          train=False, unfreeze_epoch=3),
     dict(type="NormalizationLockHook", train=False, requires_grad=None),
 
-    dict(type="ModifiedPytorchDPHook"),
+    dict(type="ModifiedPytorchDDPHook"),
     # dict(type="OptimizerHook", name="base"),
 
     dict(type="EpochMetricHook", handle=dict(qwk="qwk_metric"))

scripts/dev2.py

@@ -11,24 +11,6 @@ def show(image):
     plt.imshow(image)
     plt.show()
 
-from collections import Counter, defaultdict
-from numpy.random import choice
-import random
-from typing import Iterable, List
-
-labels = [5, 5, 5, 5, 5, 4, 4, 4, 4, 4, 4, 4, 3, 3, 3, 2, 2, 2, 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1] + [10] * 30 + [332] * 2  + [111] * 11111
-# count = 20
-
-
-
-
-
-indices = balanced_subsample(labels, 20)
-print(Counter(np.array(labels)[indices]))
-
-
-
-a = 4
 
 
 

src/psga/data_source/provider/kaggle/dataset.py

@@ -136,7 +136,7 @@ def __getitem__(self, item: int) -> Dict[str, Union[torch.Tensor, np.ndarray, in
         selected_indices = np.asarray(list(range(image_tiles.shape[0])))
         if self._subsample_tiles_count:
             if self._balance_subsample:
-                selected_indices = balanced_subsample(labels, count=self._subsample_tiles_count)
+                selected_indices = np.asarray(balanced_subsample(labels, count=self._subsample_tiles_count))
             else:
                 selected_indices = np.random.choice(list(range(image_tiles.shape[0])),
                                                     size=self._subsample_tiles_count, replace=False)

src/psga/train/contrib/hooks/epoch_metric.py

@@ -13,7 +13,6 @@
 from ppln.hooks.registry import HOOKS
 from ppln.hooks.base import BaseHook
 from ppln.hooks.priority import Priority
-from ppln.utils.misc import get_dist_info
 
 from ..utils.dist import all_gather_cpu
 
@@ -61,11 +60,9 @@ def after_epoch(self, runner: Runner) -> NoReturn:
                 inputs = inputs.index_select(self._inputs_dim, indices)
                 targets = targets.index_select(self._targets_dim, indices)
 
-                rank, _ = get_dist_info()
-                if rank == 0:
-                    func_name = self._handle[metric_name]
-                    metric = runner.batch_processor.estimate(func_name, inputs, targets)
-                    runner.log_buffer.output[metric_name] = metric.item()
+                func_name = self._handle[metric_name]
+                metric = runner.batch_processor.estimate(func_name, inputs, targets)
+                runner.log_buffer.output[metric_name] = metric.item()
 
     def _flat(self, data: List[Dict[str, torch.Tensor]], key: str, dim: int) -> torch.Tensor:
         tensor = torch.cat([batch[key] for batch in data], dim=dim)

src/psga/train/tile_classifier/batch.py

@@ -13,13 +13,13 @@
     decode_ordinal_targets
 )
 from ..utils import to_chunks
-import numpy as np
-import matplotlib.pyplot as plt
-def show(image):
-    plt.figure()
-    image = image * np.array([32.75003156, 59.0587483, 40.4596739]) + np.array([236.71301302, 217.72548731, 229.42839042])
-    plt.imshow(image)
-    plt.show()
+# import numpy as np
+# import matplotlib.pyplot as plt
+# def show(image):
+#     plt.figure()
+#     image = image * np.array([32.75003156, 59.0587483, 40.4596739]) + np.array([236.71301302, 217.72548731, 229.42839042])
+#     plt.imshow(image)
+#     plt.show()
 # viz = np.moveaxis(batch_images.cpu().numpy(), 1, -1)