def build_dlrm(emb_dict):
if 'user' in emb_dict or 'item' in emb_dict:
emb_list = [emb for _, emb in emb_dict.items()]
output = MLPInteraction(num_layers=2)(emb_list)
else:
sparse_feat_mlp_output = [MLPInteraction()( [emb_dict['sparse']] )] if 'sparse' in emb_dict else []
dense_feat_mlp_output = [MLPInteraction()( [emb_dict['dense']] )] if 'dense' in emb_dict else []
output = MLPInteraction(num_layers=2)(sparse_feat_mlp_output + dense_feat_mlp_output)
return output
def test_input_missing():
input_node1 = Input()
input_node2 = Input()
output_node1 = MLPInteraction()(input_node1)
output_node2 = MLPInteraction()(input_node2)
output_node = ConcatenateInteraction()([output_node1, output_node2])
output_node = RatingPredictionOptimizer()(output_node)
with pytest.raises(ValueError) as info:
graph_module.HyperGraph(input_node1, output_node)
assert 'A required input is missing for HyperModel' in str(info.value)
def build_autoint(emb_dict):
if 'user' in emb_dict or 'item' in emb_dict:
emb_list = [emb for _, emb in emb_dict.items()]
fm_output = [SelfAttentionInteraction()(emb_list)]
bottom_mlp_output = [MLPInteraction(num_layers=2)(emb_list)]
output = MLPInteraction(num_layers=2)(fm_output + bottom_mlp_output)
else:
fm_output = [SelfAttentionInteraction()( [emb_dict['sparse']] )] if 'sparse' in emb_dict else []
bottom_mlp_output = [MLPInteraction()( [emb_dict['dense']] )] if 'dense' in emb_dict else []
output = MLPInteraction(num_layers=2)(fm_output + bottom_mlp_output)
return output
def test_input_output_disconnect():
input_node1 = Input()
output_node = input_node1
_ = MLPInteraction()(output_node)
input_node = Input()
output_node = input_node
output_node = MLPInteraction()(output_node)
output_node = RatingPredictionOptimizer()(output_node)
with pytest.raises(ValueError) as info:
graph_module.HyperGraph(input_node1, output_node)
assert 'Inputs and outputs not connected.' in str(info.value)
def build_neumf(emb_dict):
emb_list = [emb for _, emb in emb_dict.items()]
innerproduct_output = [
ElementwiseInteraction(elementwise_type="innerporduct")(emb_list)
]
mlp_output = [MLPInteraction(num_layers=2)(emb_list)]
output = innerproduct_output + mlp_output
return output
def test_graph_basics():
input_node = Input(shape=(30, ))
output_node = input_node
output_node = MLPInteraction()(output_node)
output_node = RatingPredictionOptimizer()(output_node)
graph = graph_module.PlainGraph(input_node, output_node)
model = graph.build_keras_graph().build(hp_module.HyperParameters())
assert model.input_shape == (None, 30)
assert model.output_shape == (None, )
def build_mlp(user_num, item_num):
input = Input(shape=[2])
user_emb_mlp = LatentFactorMapper(feat_column_id=0,
id_num=user_num,
embedding_dim=64)(input)
item_emb_mlp = LatentFactorMapper(feat_column_id=1,
id_num=user_num,
embedding_dim=64)(input)
output = MLPInteraction()([user_emb_mlp, item_emb_mlp])
output = RatingPredictionOptimizer()(output)
model = RPRecommender(inputs=input, outputs=output)
return model
def build_neumf(user_num, item_num):
input = Input(shape=[2])
user_emb_gmf = LatentFactorMapper(feat_column_id=0,
id_num=user_num,
embedding_dim=64)(input)
item_emb_gmf = LatentFactorMapper(feat_column_id=1,
id_num=item_num,
embedding_dim=64)(input)
innerproduct_output = ElementwiseInteraction(
elementwise_type="innerporduct")([user_emb_gmf, item_emb_gmf])
user_emb_mlp = LatentFactorMapper(feat_column_id=0,
id_num=user_num,
embedding_dim=64)(input)
item_emb_mlp = LatentFactorMapper(feat_column_id=1,
id_num=item_num,
embedding_dim=64)(input)
mlp_output = MLPInteraction()([user_emb_mlp, item_emb_mlp])
output = RatingPredictionOptimizer()([innerproduct_output, mlp_output])
model = RPRecommender(inputs=input, outputs=output)
return model
def build_neumf(user_num, item_num):
input = Input(shape=[2])
user_emb_gmf = LatentFactorMapper(column_id=0,
num_of_entities=user_num,
embedding_dim=64)(input)
item_emb_gmf = LatentFactorMapper(column_id=1,
num_of_entities=item_num,
embedding_dim=64)(input)
innerproduct_output = InnerProductInteraction()(
[user_emb_gmf, item_emb_gmf])
user_emb_mlp = LatentFactorMapper(column_id=0,
num_of_entities=user_num,
embedding_dim=64)(input)
item_emb_mlp = LatentFactorMapper(column_id=1,
num_of_entities=item_num,
embedding_dim=64)(input)
mlp_output = MLPInteraction()([user_emb_mlp, item_emb_mlp])
output = RatingPredictionOptimizer()([innerproduct_output, mlp_output])
model = RPRecommender(inputs=input, outputs=output)
return model
user_emb_gmf = LatentFactorMapper(column_id=0,
num_of_entities=user_num,
embedding_dim=64)(input)
item_emb_gmf = LatentFactorMapper(column_id=1,
num_of_entities=item_num,
embedding_dim=64)(input)
user_emb_mlp = LatentFactorMapper(column_id=0,
num_of_entities=user_num,
embedding_dim=64)(input)
item_emb_mlp = LatentFactorMapper(column_id=1,
num_of_entities=item_num,
embedding_dim=64)(input)
# Step 2.2: Setup interactors to handle models
innerproduct_output = InnerProductInteraction()([user_emb_gmf, item_emb_gmf])
mlp_output = MLPInteraction()([user_emb_mlp, item_emb_mlp])
# Step 2.3: Setup optimizer to handle the target task
output = RatingPredictionOptimizer()([innerproduct_output, mlp_output])
model = RPRecommender(inputs=input, outputs=output)
# Step 3: Build the searcher, which provides search algorithm
searcher = Search(model=model,
tuner='greedy', # random, greedy
tuner_params={"max_trials": 5, 'overwrite': True}
)
# Step 4: Use the searcher to search the recommender
searcher.search(x=[train_X_categorical],
y=train_y,
x_val=[val_X_categorical],
hash_size = criteo.get_hash_size()
# Step 2: Build the recommender, which provides search space
# Step 2.1: Setup mappers to handle inputs
dense_input_node = Input(shape=[numerical_count])
sparse_input_node = Input(shape=[categorical_count])
dense_feat_emb = DenseFeatureMapper(num_of_fields=numerical_count,
embedding_dim=2)(dense_input_node)
sparse_feat_emb = SparseFeatureMapper(num_of_fields=categorical_count,
hash_size=hash_size,
embedding_dim=2)(sparse_input_node)
# Step 2.2: Setup interactors to handle models
attention_output = SelfAttentionInteraction()(
[dense_feat_emb, sparse_feat_emb])
bottom_mlp_output = MLPInteraction()([dense_feat_emb])
top_mlp_output = MLPInteraction()([attention_output, bottom_mlp_output])
# Step 2.3: Setup optimizer to handle the target task
output = CTRPredictionOptimizer()(top_mlp_output)
model = CTRRecommender(inputs=[dense_input_node, sparse_input_node],
outputs=output)
# Step 3: Build the searcher, which provides search algorithm
searcher = Search(
model=model,
tuner='random',
tuner_params={
'max_trials': 2,
'overwrite': True
},
dense_feat_emb = DenseFeatureMapper(num_of_fields=13,
embedding_dim=2)(dense_input_node)
# TODO: preprocess data to get sparse hash_size
sparse_feat_emb = SparseFeatureMapper(num_of_fields=26,
hash_size=[
1444, 555, 175781, 128509, 306, 19,
11931, 630, 4, 93146, 5161, 174835,
3176, 28, 11255, 165206, 11, 4606,
2017, 4, 172322, 18, 16, 56456, 86,
43356
],
embedding_dim=2)(sparse_input_node)
crossnet_output = CrossNetInteraction()([dense_feat_emb, sparse_feat_emb])
bottom_mlp_output = MLPInteraction()([dense_feat_emb])
top_mlp_output = MLPInteraction()([crossnet_output, bottom_mlp_output])
output = PointWiseOptimizer()(top_mlp_output)
model = CTRRecommender(inputs=[dense_input_node, sparse_input_node],
outputs=output)
# AutoML search and predict.
searcher = Search(
model=model,
tuner='random',
tuner_params={
'max_trials': 2,
'overwrite': True
},
)
num_of_fields=13,
embedding_dim=16)(dense_input_node)
# TODO: preprocess data to get sparse hash_size
sparse_feat_emb = SparseFeatureMapper(
num_of_fields=26,
hash_size=[
1444, 555, 175781, 128509, 306, 19,
11931, 630, 4, 93146, 5161, 174835,
3176, 28, 11255, 165206, 11, 4606,
2017, 4, 172322, 18, 16, 56456,
86, 43356
],
embedding_dim=16)(sparse_input_node)
sparse_feat_mlp_output = MLPInteraction()([sparse_feat_emb])
dense_feat_mlp_output = MLPInteraction()([dense_feat_emb])
top_mlp_output = MLPInteraction(num_layers=2)([sparse_feat_mlp_output, dense_feat_mlp_output])
output = PointWiseOptimizer()(top_mlp_output)
model = CTRRecommender(inputs=[dense_input_node, sparse_input_node], outputs=output)
# AutoML search and predict.
searcher = Search(model=model,
tuner='random',
tuner_params={'max_trials': 2, 'overwrite': True},
)
searcher.search(x=train_X,
y=train_y,
x_val=val_X,
y_val=val_y,
def build_neumf(emb_dict):
emb_list = [emb for _, emb in emb_dict.items()]
innerproduct_output = [InnerProductInteraction()(emb_list)]
mlp_output = [MLPInteraction(num_layers=2)(emb_list)]
output = innerproduct_output + mlp_output
return output