Rename customized reset_parameters to init_weights (Fix #123)

CHANGELOG.md

@@ -4,6 +4,12 @@
 [Doing] Add support for saving pb file, exporting embeddings
 [Doing] Add support of multi-gpu training
 
+**FuxiCTR v2.3.4, 2024-11-05**
++ [Fix] Fix OOV token update ([#119](https://github.com/reczoo/FuxiCTR/issues/119))
++ [Fix] Speed up parquet dataset reading ([#121](https://github.com/reczoo/FuxiCTR/issues/121))
++ [Fix] Fix add_loss() isue that does not work after renaming to compute_loss() ([#122](https://github.com/reczoo/FuxiCTR/issues/122))
++ [Fix] Rename customized reset_parameters to init_weights ([#123](https://github.com/reczoo/FuxiCTR/issues/123))
+
 **FuxiCTR v2.3.3, 2024-10-14**
 + [Feature] Add EulerNet and DCNv3 models
 + [Feature] Add support to parquet as input, like csv format

fuxictr/pytorch/layers/embeddings/feature_embedding.py

@@ -112,7 +112,7 @@ def __init__(self,
                                                                       padding_idx=padding_idx)
                 elif feature_spec["type"] == "embedding":
                     self.embedding_layers[feature] = nn.Identity()
-        self.reset_parameters()
+        self.init_weights()
 
     def get_feature_encoder(self, encoder):
         try:
@@ -127,13 +127,13 @@ def get_feature_encoder(self, encoder):
         except:
             raise ValueError("feature_encoder={} is not supported.".format(encoder))
 
-    def reset_parameters(self):
+    def init_weights(self):
         embedding_initializer = get_initializer(self.embedding_initializer)
         for k, v in self.embedding_layers.items():
             if "share_embedding" in self._feature_map.features[k]:
                 continue
             if type(v) == PretrainedEmbedding: # skip pretrained
-                v.reset_parameters(embedding_initializer)
+                v.init_weights(embedding_initializer)
             elif type(v) == nn.Embedding:
                 if v.padding_idx is not None:
                     embedding_initializer(v.weight[1:, :]) # set padding_idx to zero

fuxictr/pytorch/layers/embeddings/pretrained_embedding.py

@@ -62,7 +62,7 @@ def __init__(self,
         if pretrain_usage == "concat":
             self.proj = nn.Linear(pretrain_dim + embedding_dim, embedding_dim, bias=False)
 
-    def reset_parameters(self, embedding_initializer):
+    def init_weights(self, embedding_initializer):
         if self.pretrain_usage in ["sum", "concat"]:
             nn.init.zeros_(self.id_embedding.weight) # set oov token embeddings to zeros
             embedding_initializer(self.id_embedding.weight[1:self.oov_idx, :])

fuxictr/pytorch/layers/interactions/bilinear_interaction.py

@@ -33,9 +33,9 @@ def __init__(self, num_fields, embedding_dim, bilinear_type="field_interaction")
             self.bilinear_W = nn.Parameter(torch.Tensor(self.interact_dim, embedding_dim, embedding_dim))
         else:
             raise NotImplementedError
-        self.reset_parameters()
+        self.init_weights()
 
-    def reset_parameters(self):
+    def init_weights(self):
         nn.init.xavier_normal_(self.bilinear_W)
 
     def forward(self, feature_emb):
@@ -70,9 +70,9 @@ def __init__(self, num_fields, embedding_dim, bilinear_type="field_interaction")
         else:
             raise NotImplementedError
         self.triu_index = nn.Parameter(torch.triu_indices(num_fields, num_fields, offset=1), requires_grad=False)
-        self.reset_parameters()
+        self.init_weights()
 
-    def reset_parameters(self):
+    def init_weights(self):
         nn.init.xavier_normal_(self.bilinear_W)
 
     def forward(self, feature_emb):

fuxictr/pytorch/models/rank_model.py

@@ -92,18 +92,18 @@ def compute_loss(self, return_dict, y_true):
         return loss
 
     def reset_parameters(self):
-        def reset_default_params(m):
+        def default_reset_params(m):
             # initialize nn.Linear/nn.Conv1d layers by default
             if type(m) in [nn.Linear, nn.Conv1d]:
                 nn.init.xavier_normal_(m.weight)
                 if m.bias is not None:
                     m.bias.data.fill_(0)
-        def reset_custom_params(m):
-            # initialize layers with customized reset_parameters
-            if hasattr(m, 'reset_custom_params'):
-                m.reset_custom_params()
-        self.apply(reset_default_params)
-        self.apply(reset_custom_params)
+        def custom_reset_params(m):
+            # initialize layers with customized init_weights()
+            if hasattr(m, 'init_weights'):
+                m.init_weights()
+        self.apply(default_reset_params)
+        self.apply(custom_reset_params)
 
     def get_inputs(self, inputs, feature_source=None):
         X_dict = dict()

model_zoo/APG/config/model_config.yaml

@@ -151,5 +151,3 @@ APG_DCNv2: # This is a config template
     seed: 2019
     monitor: {'AUC': 1, 'logloss': -1}
     monitor_mode: 'max'
-
-

model_zoo/FinalNet/src/FinalNet.py

@@ -121,7 +121,7 @@ def __init__(self, num_fields, gate_residual="concat"):
         assert gate_residual in ["concat", "sum"]
         self.gate_residual = gate_residual
 
-    def reset_custom_params(self):
+    def init_weights(self):
         nn.init.zeros_(self.linear.weight)
         nn.init.ones_(self.linear.bias)
 

tests/test_torch.sh

@@ -4,7 +4,7 @@ home="$(pwd)/../model_zoo"
 echo "=== Testing AFM ==="  && cd $home/AFM && python run_expid.py --expid AFM_test && \
 echo "=== Testing AFN ===" && cd $home/AFN && python run_expid.py --expid AFN_test && \
 echo "=== Testing AOANet ===" && cd $home/AOANet && python run_expid.py --expid AOANet_test && \
-echo "=== Testing APG ===" && cd $home/APG && python run_expid.py --expid APG_test && \
+echo "=== Testing APG ===" && cd $home/APG && python run_expid.py --expid APG_DeepFM_test && \
 echo "=== Testing AutoInt ===" && cd $home/AutoInt && python run_expid.py --expid AutoInt_test && \
 echo "=== Testing BST ===" && cd $home/BST && python run_expid.py --expid BST_test && \
 echo "=== Testing CCPM ===" && cd $home/CCPM && python run_expid.py --expid CCPM_test && \