Fix shared embeddings and string sparse validation

Author: shenweichen

deepctr/feature_column.py

@@ -14,6 +14,23 @@
 DEFAULT_GROUP_NAME = "default_group"
 
 
+def _is_string_dtype(dtype):
+    try:
+        return tf.as_dtype(dtype) == tf.string
+    except TypeError:
+        return dtype == "string"
+
+
+def _check_sparse_feature_dtype(fc):
+    if _is_string_dtype(fc.dtype) and not fc.use_hash:
+        raise ValueError(
+            "SparseFeat(name='{}', dtype='string') requires use_hash=True "
+            "so string ids can be converted before embedding lookup. "
+            "Alternatively, encode the feature values to integer ids before "
+            "passing them to DeepCTR.".format(fc.name)
+        )
+
+
 class SparseFeat(namedtuple('SparseFeat',
                             ['name', 'vocabulary_size', 'embedding_dim', 'use_hash', 'vocabulary_path', 'dtype', 'embeddings_initializer',
                              'embedding_name',
@@ -129,12 +146,14 @@ def build_input_features(feature_columns, prefix=''):
     input_features = OrderedDict()
     for fc in feature_columns:
         if isinstance(fc, SparseFeat):
+            _check_sparse_feature_dtype(fc)
             input_features[fc.name] = Input(
                 shape=(1,), name=prefix + fc.name, dtype=fc.dtype)
         elif isinstance(fc, DenseFeat):
             input_features[fc.name] = Input(
                 shape=(fc.dimension,), name=prefix + fc.name, dtype=fc.dtype)
         elif isinstance(fc, VarLenSparseFeat):
+            _check_sparse_feature_dtype(fc)
             input_features[fc.name] = Input(shape=(fc.maxlen,), name=prefix + fc.name,
                                             dtype=fc.dtype)
             if fc.weight_name is not None:

deepctr/inputs.py

@@ -16,32 +16,58 @@
 from .layers.utils import Hash
 
 
+def _create_embedding_layer(feat, l2_reg, prefix, name_suffix, mask_zero=False):
+    emb = Embedding(feat.vocabulary_size, feat.embedding_dim,
+                    embeddings_initializer=feat.embeddings_initializer,
+                    embeddings_regularizer=l2(l2_reg),
+                    name=prefix + '_' + name_suffix + '_' + feat.embedding_name,
+                    mask_zero=mask_zero)
+    emb.trainable = feat.trainable
+    return emb
+
+
+def _check_embedding_compatible(embedding_name, existing_feat, feat):
+    for attr in ('vocabulary_size', 'embedding_dim', 'trainable'):
+        if getattr(existing_feat, attr) != getattr(feat, attr):
+            raise ValueError(
+                "Feature columns with the same embedding_name must share the same "
+                "{}. embedding_name='{}' has {} and {}.".format(
+                    attr, embedding_name, getattr(existing_feat, attr), getattr(feat, attr)
+                )
+            )
+
+
 def get_inputs_list(inputs):
     return list(chain(*list(map(lambda x: x.values(), filter(lambda x: x is not None, inputs)))))
 
 
 def create_embedding_dict(sparse_feature_columns, varlen_sparse_feature_columns, seed, l2_reg,
                           prefix='sparse_', seq_mask_zero=True):
     sparse_embedding = {}
+    embedding_feature_dict = {}
+    varlen_embedding_names = set(
+        feat.embedding_name for feat in varlen_sparse_feature_columns
+    ) if varlen_sparse_feature_columns else set()
+
     for feat in sparse_feature_columns:
-        emb = Embedding(feat.vocabulary_size, feat.embedding_dim,
-                        embeddings_initializer=feat.embeddings_initializer,
-                        embeddings_regularizer=l2(l2_reg),
-                        name=prefix + '_emb_' + feat.embedding_name)
-        emb.trainable = feat.trainable
-        sparse_embedding[feat.embedding_name] = emb
+        embedding_name = feat.embedding_name
+        if embedding_name in sparse_embedding:
+            _check_embedding_compatible(embedding_name, embedding_feature_dict[embedding_name], feat)
+            continue
+        mask_zero = seq_mask_zero and feat.embedding_name in varlen_embedding_names
+        emb = _create_embedding_layer(feat, l2_reg, prefix, 'emb', mask_zero)
+        sparse_embedding[embedding_name] = emb
+        embedding_feature_dict[embedding_name] = feat
 
     if varlen_sparse_feature_columns and len(varlen_sparse_feature_columns) > 0:
         for feat in varlen_sparse_feature_columns:
-            # if feat.name not in sparse_embedding:
-            emb = Embedding(feat.vocabulary_size, feat.embedding_dim,
-                            embeddings_initializer=feat.embeddings_initializer,
-                            embeddings_regularizer=l2(
-                                l2_reg),
-                            name=prefix + '_seq_emb_' + feat.name,
-                            mask_zero=seq_mask_zero)
-            emb.trainable = feat.trainable
+            embedding_name = feat.embedding_name
+            if embedding_name in sparse_embedding:
+                _check_embedding_compatible(embedding_name, embedding_feature_dict[embedding_name], feat)
+                continue
+            emb = _create_embedding_layer(feat, l2_reg, prefix, 'seq_emb', seq_mask_zero)
             sparse_embedding[feat.embedding_name] = emb
+            embedding_feature_dict[feat.embedding_name] = feat
     return sparse_embedding
 
 

tests/feature_test.py

@@ -1,6 +1,8 @@
 from deepctr.models import DeepFM
 from deepctr.feature_column import SparseFeat, DenseFeat, VarLenSparseFeat, get_feature_names
+from deepctr.inputs import create_embedding_matrix
 import numpy as np
+import pytest
 
 
 def test_long_dense_vector():
@@ -28,3 +30,28 @@ def test_feature_column_sparsefeat_vocabulary_path():
     vlsf = VarLenSparseFeat(sf, 6)
     if vlsf.vocabulary_path != vocab_path:
         raise ValueError("vlsf.vocabulary_path is invalid")
+
+
+def test_create_embedding_matrix_reuses_same_embedding_name():
+    feature_columns = [
+        SparseFeat('item_id', 4, embedding_dim=8),
+        SparseFeat('item_id_copy', 4, embedding_dim=8, embedding_name='item_id'),
+        VarLenSparseFeat(SparseFeat('hist_item_id', 4, embedding_dim=8, embedding_name='item_id'), maxlen=3),
+        VarLenSparseFeat(SparseFeat('neg_hist_item_id', 4, embedding_dim=8, embedding_name='item_id'), maxlen=3),
+    ]
+
+    embedding_dict = create_embedding_matrix(feature_columns, l2_reg=0, seed=1024)
+
+    assert list(embedding_dict.keys()) == ['item_id']
+    assert embedding_dict['item_id'].name == 'sparse_emb_item_id'
+    assert embedding_dict['item_id'].mask_zero is True
+
+
+def test_create_embedding_matrix_rejects_inconsistent_shared_embedding():
+    feature_columns = [
+        SparseFeat('item_id', 4, embedding_dim=8),
+        VarLenSparseFeat(SparseFeat('hist_item_id', 5, embedding_dim=8, embedding_name='item_id'), maxlen=3),
+    ]
+
+    with pytest.raises(ValueError, match="same embedding_name"):
+        create_embedding_matrix(feature_columns, l2_reg=0, seed=1024)

tests/models/MTL_test.py

@@ -1,6 +1,7 @@
 import pytest
 import tensorflow as tf
 
+from deepctr.feature_column import SparseFeat
 from deepctr.models.multitask import SharedBottom, ESMM, MMOE, PLE
 from ..utils_mtl import get_mtl_test_data, check_mtl_model
 
@@ -27,6 +28,16 @@ def test_ESMM():
     check_mtl_model(model, model_name, x, y_list, task_types=['binary', 'binary'])
 
 
+def test_ESMM_string_sparse_requires_hash():
+    with pytest.raises(ValueError, match="use_hash=True"):
+        ESMM([SparseFeat('user_id', 10, dtype='string')], tower_dnn_hidden_units=(8,))
+
+
+def test_ESMM_string_sparse_with_hash():
+    model = ESMM([SparseFeat('user_id', 10, use_hash=True, dtype='string')], tower_dnn_hidden_units=(8,))
+    assert len(model.outputs) == 2
+
+
 def test_MMOE():
     if tf.__version__ == "1.15.0":  # slow in tf 1.15
         return