[Fix] fix

mmseg/core/evaluation/metrics.py

@@ -36,7 +36,8 @@ def calc_adaptive_fm(pred_label, label, beta=0.3):
     else:
         precision = area_intersection / torch.count_nonzero(binary_pred_label)
         recall = area_intersection / torch.count_nonzero(label)
-        adaptive_fm = (1 + beta) * precision * recall / (beta * precision + recall)
+        adaptive_fm = (1 + beta) * precision * recall / (
+                beta * precision + recall)
     return adaptive_fm
 
 
@@ -310,8 +311,7 @@ def eval_metrics(results,
     return ret_metrics
 
 
-def calc_sod_metrics(pred_label,
-                     label):
+def calc_sod_metrics(pred_label, label):
     if isinstance(pred_label, str):
         pred_label = torch.from_numpy(np.load(pred_label))
     else:
@@ -323,27 +323,25 @@ def calc_sod_metrics(pred_label,
     else:
         label = torch.from_numpy(label)
 
+    pred_label = pred_label.float()
     if pred_label.max() != pred_label.min():
-        pred_label = (pred_label - pred_label.min()) / (pred_label.max() - pred_label.min())
+        pred_label = (pred_label - pred_label.min()) / (
+                pred_label.max() - pred_label.min())
 
     mae = calc_mae(pred_label, label)
     adaptive_fm = calc_adaptive_fm(pred_label, label)
 
     return mae, adaptive_fm
 
 
-def pre_eval_to_sod_metrics(pre_eval_results,
-                            nan_to_num=None):
+def pre_eval_to_sod_metrics(pre_eval_results, nan_to_num=None):
     pre_eval_results = tuple(zip(*pre_eval_results))
     assert len(pre_eval_results) == 2
 
     mae = sum(pre_eval_results[0]) / len(pre_eval_results[0])
     adp_fm = sum(pre_eval_results[1]) / len(pre_eval_results[1])
 
-    ret_metrics = OrderedDict({
-        'MAE': mae.numpy(),
-        'adpFm': adp_fm.numpy()
-    })
+    ret_metrics = OrderedDict({'MAE': mae.numpy(), 'adpFm': adp_fm.numpy()})
     if nan_to_num is not None:
         ret_metrics = OrderedDict({
             metric: np.nan_to_num(metric_value, nan=nan_to_num)
@@ -352,9 +350,7 @@ def pre_eval_to_sod_metrics(pre_eval_results,
     return ret_metrics
 
 
-def eval_sod_metrics(results,
-                     gt_seg_maps,
-                     nan_to_num=None):
+def eval_sod_metrics(results, gt_seg_maps, nan_to_num=None):
     maes = []
     adp_fms = []
 

mmseg/datasets/custom.py

@@ -5,6 +5,7 @@
 
 import mmcv
 import numpy as np
+import torch.nn.functional as F
 from mmcv.utils import print_log
 from prettytable import PrettyTable
 from torch.utils.data import Dataset
@@ -262,7 +263,7 @@ def get_gt_seg_maps(self, efficient_test=None):
             self.gt_seg_map_loader(results)
             yield results['gt_semantic_seg']
 
-    def pre_eval(self, preds, indices):
+    def pre_eval(self, preds, indices, return_logit=False):
         """Collect eval result from each iteration.
 
         Args:
@@ -284,6 +285,14 @@ def pre_eval(self, preds, indices):
         pre_eval_results = []
 
         for pred, index in zip(preds, indices):
+            if return_logit:
+                if pred.shape[0] >= 2:
+                    pred = F.softmax(pred, dim=0)
+                    pred = pred.argmax(dim=0)
+                else:
+                    pred = F.sigmoid(pred)
+                    pred = pred.squeeze(0)
+                    pred = (pred > 0.5).int()
             seg_map = self.get_gt_seg_map_by_idx(index)
             pre_eval_results.append(
                 intersect_and_union(pred, seg_map, len(self.CLASSES),
@@ -358,6 +367,7 @@ def get_palette_for_custom_classes(self, class_names, palette=None):
     def evaluate(self,
                  results,
                  metric='mIoU',
+                 return_logit=False,
                  logger=None,
                  gt_seg_maps=None,
                  **kwargs):

mmseg/datasets/sod_custom.py

@@ -3,11 +3,12 @@
 
 import mmcv
 import numpy as np
+import torch.nn.functional as F
 from mmcv.utils import print_log
 from prettytable import PrettyTable
 
-from mmseg.core import calc_sod_metrics, eval_sod_metrics, \
-    pre_eval_to_sod_metrics
+from mmseg.core import (calc_sod_metrics, eval_sod_metrics,
+                        pre_eval_to_sod_metrics)
 from . import CustomDataset
 from .builder import DATASETS
 
@@ -21,7 +22,7 @@ class SODCustomDataset(CustomDataset):
     def __init__(self, **kwargs):
         super(SODCustomDataset, self).__init__(**kwargs)
 
-    def pre_eval(self, preds, indices):
+    def pre_eval(self, preds, indices, return_logit=False):
         """Collect eval result from each iteration.
 
         Args:
@@ -43,15 +44,22 @@ def pre_eval(self, preds, indices):
         pre_eval_results = []
 
         for pred, index in zip(preds, indices):
+            if return_logit:
+                if pred.shape[0] >= 2:
+                    pred = F.softmax(pred, dim=0)
+                    pred = pred[1]
+                else:
+                    pred = F.sigmoid(pred)
+                    pred = pred.squeeze(0)
             seg_map = self.get_gt_seg_map_by_idx(index)
-            pre_eval_results.append(
-                calc_sod_metrics(pred, seg_map))
+            pre_eval_results.append(calc_sod_metrics(pred, seg_map))
 
         return pre_eval_results
 
     def evaluate(self,
                  results,
                  logger=None,
+                 return_logit=False,
                  gt_seg_maps=None,
                  **kwargs):
         """Evaluate the dataset.
@@ -77,9 +85,7 @@ def evaluate(self,
                 results, str):
             if gt_seg_maps is None:
                 gt_seg_maps = self.get_gt_seg_maps()
-            ret_metrics = eval_sod_metrics(
-                results,
-                gt_seg_maps)
+            ret_metrics = eval_sod_metrics(results, gt_seg_maps)
         # test a list of pre_eval_results
         else:
             ret_metrics = pre_eval_to_sod_metrics(results)

mmseg/models/losses/cross_entropy_loss.py

@@ -36,14 +36,14 @@ def cross_entropy(pred,
 def _expand_onehot_labels(labels, label_weights, target_shape, ignore_index):
     """Expand onehot labels to match the size of prediction."""
     bin_labels = labels.new_zeros(target_shape)
-    valid_mask = (labels > 0) & (labels != ignore_index)
+    valid_mask = (labels >= 0) & (labels != ignore_index)
     inds = torch.nonzero(valid_mask, as_tuple=True)
 
     if inds[0].numel() > 0:
         if labels.dim() == 3:
-            bin_labels[inds[0], labels[valid_mask] - 1, inds[1], inds[2]] = 1
+            bin_labels[inds[0], labels[valid_mask], inds[1], inds[2]] = 1
         else:
-            bin_labels[inds[0], labels[valid_mask] - 1] = 1
+            bin_labels[inds[0], labels[valid_mask]] = 1
 
     valid_mask = valid_mask.unsqueeze(1).expand(target_shape).float()
     if label_weights is None:
@@ -83,8 +83,11 @@ def binary_cross_entropy(pred,
                 pred.dim() == 4 and label.dim() == 3), \
             'Only pred shape [N, C], label shape [N] or pred shape [N, C, ' \
             'H, W], label shape [N, H, W] are supported'
-        label, weight = _expand_onehot_labels(label, weight, pred.shape,
-                                              ignore_index)
+        if pred.shape[1] == 1:
+            pred = pred.squeeze(1)
+        else:
+            label, weight = _expand_onehot_labels(label, weight, pred.shape,
+                                                  ignore_index)
 
     # weighted element-wise losses
     if weight is not None: