From 90a441e217d8462ae24b02b4a5a2e72e7ddaaee3 Mon Sep 17 00:00:00 2001
From: FlyingQianMM <245467267@qq.com>
Date: Tue, 16 Jun 2020 21:16:01 +0800
Subject: [PATCH] add epsilon to denominator when computing loss (#292)

* add 0.00001 to denominator when computing loss

* use DEFAULT_EPSILON
---
 pdseg/loss.py | 22 ++++++++++++++++------
 1 file changed, 16 insertions(+), 6 deletions(-)

diff --git a/pdseg/loss.py b/pdseg/loss.py
index bd840610..c5ea306d 100644
--- a/pdseg/loss.py
+++ b/pdseg/loss.py
@@ -77,7 +77,7 @@ def softmax_with_loss(logit,
         weighted_label_one_hot.stop_gradient = True
 
     loss = loss * ignore_mask
-    avg_loss = fluid.layers.mean(loss) / fluid.layers.mean(ignore_mask)
+    avg_loss = fluid.layers.mean(loss) / (fluid.layers.mean(ignore_mask) + cfg.MODEL.DEFAULT_EPSILON)
 
     label.stop_gradient = True
     ignore_mask.stop_gradient = True
@@ -133,10 +133,12 @@ def multi_softmax_with_loss(logits,
         for i, logit in enumerate(logits):
             if label.shape[2] != logit.shape[2] or label.shape[
                     3] != logit.shape[3]:
-                label = fluid.layers.resize_nearest(label, logit.shape[2:])
-            logit_mask = (label.astype('int32') !=
+                logit_label = fluid.layers.resize_nearest(label, logit.shape[2:])
+            else:
+                logit_label = label
+            logit_mask = (logit_label.astype('int32') !=
                           cfg.DATASET.IGNORE_INDEX).astype('int32')
-            loss = softmax_with_loss(logit, label, logit_mask, num_classes)
+            loss = softmax_with_loss(logit, logit_label, logit_mask, num_classes, weight=weight)
             avg_loss += cfg.MODEL.MULTI_LOSS_WEIGHT[i] * loss
     else:
         avg_loss = softmax_with_loss(
@@ -148,7 +150,11 @@ def multi_dice_loss(logits, label, ignore_mask=None):
     if isinstance(logits, tuple):
         avg_loss = 0
         for i, logit in enumerate(logits):
-            logit_label = fluid.layers.resize_nearest(label, logit.shape[2:])
+            if label.shape[2] != logit.shape[2] or label.shape[
+                    3] != logit.shape[3]:
+                logit_label = fluid.layers.resize_nearest(label, logit.shape[2:])
+            else:
+                logit_label = label
             logit_mask = (logit_label.astype('int32') !=
                           cfg.DATASET.IGNORE_INDEX).astype('int32')
             loss = dice_loss(logit, logit_label, logit_mask)
@@ -162,7 +168,11 @@ def multi_bce_loss(logits, label, ignore_mask=None):
     if isinstance(logits, tuple):
         avg_loss = 0
         for i, logit in enumerate(logits):
-            logit_label = fluid.layers.resize_nearest(label, logit.shape[2:])
+            if label.shape[2] != logit.shape[2] or label.shape[
+                    3] != logit.shape[3]:
+                logit_label = fluid.layers.resize_nearest(label, logit.shape[2:])
+            else:
+                logit_label = label
             logit_mask = (logit_label.astype('int32') !=
                           cfg.DATASET.IGNORE_INDEX).astype('int32')
             loss = bce_loss(logit, logit_label, logit_mask)
-- 
GitLab