Issue #690: added doc regarding reduction strategy in loss

mathpluscode · mathpluscode · commit 531b68112ae0 · 2021-04-02T01:28:57.000+01:00
diff --git a/deepreg/loss/image.py b/deepreg/loss/image.py
@@ -29,6 +29,8 @@ def __init__(
         Init.
 
         :param reduction: using SUM reduction over batch axis,
+            this is for supporting multi-device training,
+            and the loss will be divided by global batch size,
             calling the loss like `loss(y_true, y_pred)` will return a scalar tensor.
         :param name: name of the loss
         """
@@ -70,6 +72,8 @@ def __init__(
         :param num_bins: number of bins for intensity, the default value is empirical.
         :param sigma_ratio: a hyper param for gaussian function
         :param reduction: using SUM reduction over batch axis,
+            this is for supporting multi-device training,
+            and the loss will be divided by global batch size,
             calling the loss like `loss(y_true, y_pred)` will return a scalar tensor.
         :param name: name of the loss
         """
@@ -201,6 +205,8 @@ def __init__(
         :param smooth_nr: small constant added to numerator in case of zero covariance.
         :param smooth_dr: small constant added to denominator in case of zero variance.
         :param reduction: using SUM reduction over batch axis,
+            this is for supporting multi-device training,
+            and the loss will be divided by global batch size,
             calling the loss like `loss(y_true, y_pred)` will return a scalar tensor.
         :param name: name of the loss
         """
diff --git a/deepreg/loss/label.py b/deepreg/loss/label.py
@@ -34,6 +34,8 @@ def __init__(
         :param scales: list of scalars or None, if None, do not apply any scaling.
         :param kernel: gaussian or cauchy.
         :param reduction: using SUM reduction over batch axis,
+            this is for supporting multi-device training,
+            and the loss will be divided by global batch size,
             calling the loss like `loss(y_true, y_pred)` will return a scalar tensor.
         :param name: str, name of the loss.
         """
@@ -132,6 +134,8 @@ def __init__(
         :param scales: list of scalars or None, if None, do not apply any scaling.
         :param kernel: gaussian or cauchy.
         :param reduction: using SUM reduction over batch axis,
+            this is for supporting multi-device training,
+            and the loss will be divided by global batch size,
             calling the loss like `loss(y_true, y_pred)` will return a scalar tensor.
         :param name: str, name of the loss.
         """
@@ -206,6 +210,8 @@ def __init__(
         :param scales: list of scalars or None, if None, do not apply any scaling.
         :param kernel: gaussian or cauchy.
         :param reduction: using SUM reduction over batch axis,
+            this is for supporting multi-device training,
+            and the loss will be divided by global batch size,
             calling the loss like `loss(y_true, y_pred)` will return a scalar tensor.
         :param name: str, name of the loss.
         """
@@ -272,6 +278,8 @@ def __init__(
         :param scales: list of scalars or None, if None, do not apply any scaling.
         :param kernel: gaussian or cauchy.
         :param reduction: using SUM reduction over batch axis,
+            this is for supporting multi-device training,
+            and the loss will be divided by global batch size,
             calling the loss like `loss(y_true, y_pred)` will return a scalar tensor.
         :param name: str, name of the loss.
         """
diff --git a/examples/custom_image_label_loss.py b/examples/custom_image_label_loss.py
@@ -22,6 +22,8 @@ def __init__(
         Init.
 
         :param reduction: using SUM reduction over batch axis,
+            this is for supporting multi-device training,
+            and the loss will be divided by global batch size,
             calling the loss like `loss(y_true, y_pred)` will return a scalar tensor.
         :param name: name of the loss
         """
diff --git a/examples/custom_parameterized_image_label_loss.py b/examples/custom_parameterized_image_label_loss.py
@@ -24,6 +24,8 @@ def __init__(
 
         :param p: order of the norm, 1 or 2.
         :param reduction: using SUM reduction over batch axis,
+            this is for supporting multi-device training,
+            and the loss will be divided by global batch size,
             calling the loss like `loss(y_true, y_pred)` will return a scalar tensor.
         :param name: name of the loss.
         """
