Add document for loss weight learning

.readthedocs.yaml

@@ -1,11 +1,15 @@
 
 version: 2
 
+build:
+  os: ubuntu-22.04
+  tools:
+    python: "3.7"
+
 sphinx:
   configuration: docs/source/conf.py
 
 python:
-  version: "3.7"
   install:
   - requirements: requirements/runtime.txt
   - requirements: requirements/docs.txt

docs/source/train.md

@@ -236,31 +236,39 @@ EasyRec支持两种损失函数配置方式：1）使用单个损失函数；2
 多目标学习任务中，人工指定多个损失函数的静态权重通常不能获得最好的效果。EasyRec支持损失函数权重自适应学习，示例如下：
 
 ```protobuf
-      losses {
-        loss_type: CLASSIFICATION
-        learn_loss_weight: true
-      }
-      losses {
-        loss_type: BINARY_FOCAL_LOSS
-        learn_loss_weight: true
-        binary_focal_loss {
-          gamma: 2.0
-          alpha: 0.85
-        }
-      }
-      losses {
-        loss_type: PAIRWISE_FOCAL_LOSS
-        learn_loss_weight: true
-        pairwise_focal_loss {
-          session_name: "client_str"
-          hinge_margin: 1.0
-        }
-      }
+  loss_weight_strategy: Uncertainty
+  losses {
+    loss_type: CLASSIFICATION
+    learn_loss_weight: true
+  }
+  losses {
+    loss_type: BINARY_FOCAL_LOSS
+    learn_loss_weight: true
+    binary_focal_loss {
+      gamma: 2.0
+      alpha: 0.85
+    }
+  }
+  losses {
+    loss_type: PAIRWISE_FOCAL_LOSS
+    learn_loss_weight: true
+    pairwise_focal_loss {
+      session_name: "client_str"
+      hinge_margin: 1.0
+    }
+  }
 ```
 
 通过`learn_loss_weight`参数配置是否需要开启权重自适应学习，默认不开启。开启之后，`weight`参数不再生效。
 
-参考论文：《Multi-Task Learning Using Uncertainty to Weigh Losses for Scene Geometry and Semantics》
+- loss_weight_strategy: Uncertainty
+  - 表示通过不确定性来度量损失函数的权重；目前在`learn_loss_weight: true`时必须要设置该值
+- loss_weight_strategy: Random
+  - 表示损失函数的权重设定为归一化的随机数
+
+参考论文：
+- 《 Multi-Task Learning Using Uncertainty to Weigh Losses for Scene Geometry and Semantics 》
+- 《 [Reasonable Effectiveness of Random Weighting: A Litmus Test for Multi-Task Learning](https://arxiv.org/abs/2111.10603) 》
 
 ## 训练命令
 

easy_rec/python/layers/common_layers.py

@@ -171,15 +171,17 @@ def call(self, config, training):
           ln_name = self._group_name + 'f_%d' % i
           fea = layer_norm(fea, name=ln_name, reuse=self._reuse)
         if do_dropout and output_feature_list:
-          fea = self.dropout.apply(fea, training=training)
+          fea = self.dropout.call(fea, training=training)
         if do_feature_dropout:
           fea = tf.div(fea, keep_prob) * mask[i]
         feature_list[i] = fea
       if do_feature_dropout:
         features = tf.concat(feature_list, axis=-1)
 
     if do_dropout and not do_feature_dropout:
-      features = self.dropout.apply(features, training=training)
+      features = self.dropout.call(features, training=training)
+    if features.shape.ndims == 3 and int(features.shape[0]) == 1:
+      features = tf.squeeze(features, axis=0)
 
     if config.only_output_feature_list:
       return feature_list

easy_rec/python/model/multi_task_model.py

@@ -10,6 +10,7 @@
 from easy_rec.python.model.rank_model import RankModel
 from easy_rec.python.protos import tower_pb2
 from easy_rec.python.protos.loss_pb2 import LossType
+from easy_rec.python.protos.easy_rec_model_pb2 import EasyRecModel
 
 if tf.__version__ >= '2.0':
   tf = tf.compat.v1
@@ -188,7 +189,8 @@ def get_learnt_loss(self, loss_type, name, value):
       else:
         return tf.exp(-uncertainty) * value + 0.5 * uncertainty
     else:
-      raise ValueError('Unsupported loss weight strategy: ' + strategy.Name)
+      strategy_name = EasyRecModel.LossWeightStrategy.Name(strategy)
+      raise ValueError('Unsupported loss weight strategy: ' + strategy_name)
 
   def build_loss_graph(self):
     """Build loss graph for multi task model."""

easy_rec/python/tools/feature_selection.py

@@ -106,7 +106,11 @@ def _feature_dim_dropout_ratio(self):
       group_name = feature_group.group_name
 
       logit_p_name = 'logit_p' if group_name == 'all' else 'logit_p_%s' % group_name
-      logit_p = reader.get_tensor(logit_p_name)
+      try:
+        logit_p = reader.get_tensor(logit_p_name)
+      except Exception:
+        print('get `logit_p` failed, try to get `backbone/logit_p`')
+        logit_p = reader.get_tensor('backbone/' + logit_p_name)
       feature_dims_importance = tf.sigmoid(logit_p)
       with tf.Session() as sess:
         feature_dims_importance = feature_dims_importance.eval(session=sess)