在用机器学习PAI attention的时候没有效果，有没有什么建议呀？

在用机器学习PAI attention的时候没有效果，大佬们有没有什么建议呀？截取了部分配置如下：

features {
input_names: "clk_pscate_seq"
feature_type: SequenceFeature
embedding_dim: 8
hash_bucket_size: 1000000
separator: ""
max_seq_len: 25
}
features {
input_names: "clk_event_seq"
feature_type: SequenceFeature
embedding_dim: 8
hash_bucket_size: 1000000
separator: ""
max_seq_len: 25
}
}
model_config {
model_name: "BST"
model_class: "RankModel"
feature_groups {
group_name: "normal"
feature_names: "code_seat_id"
feature_names: "imp_pkg_list"
feature_names: "clk_pkg_list"
feature_names: "interest_list"
feature_names: "ad_creative_id"
feature_names: "ad_group_id"
feature_names: "ad_plan_id"
feature_names: "advertiser_id"
feature_names: "creativity_style_id"
feature_names: "image_scale"
feature_names: "image_width"
feature_names: "image_height"
feature_names: "package_name"
feature_names: "ps_package_name"
feature_names: "bill_type"
feature_names: "bill"
feature_names: "purpose_type"
wide_deep: DEEP
}
feature_groups {
group_name: "imp_seq"
feature_names: "imp_c_seq"
feature_names: "imp_g_seq"
feature_names: "imp_p_seq"
feature_names: "imp_a_seq"
feature_names: "imp_csi_seq"
feature_names: "imp_pkg_seq"
feature_names: "imp_pscate_seq"

feature_names: "ad_creative_id"
feature_names: "ad_group_id"
feature_names: "ad_plan_id"
feature_names: "advertiser_id"
feature_names: "code_seat_id"
feature_names: "ps_package_name"
feature_names: "ps_cate"
wide_deep: DEEP

}
feature_groups {
group_name: "clk_seq"
feature_names: "clk_c_seq"
feature_names: "clk_g_seq"
feature_names: "clk_p_seq"
feature_names: "clk_a_seq"
feature_names: "clk_csi_seq"
feature_names: "clk_pkg_seq"
feature_names: "clk_pscate_seq"
feature_names: "ad_creative_id"
feature_names: "ad_group_id"
feature_names: "ad_plan_id"
feature_names: "advertiser_id"
feature_names: "code_seat_id"
feature_names: "ps_package_name"
feature_names: "ps_cate"
wide_deep: DEEP
}
backbone {
blocks {
name: "deep"
inputs {
feature_group_name: 'normal'
}
keras_layer {
class_name: 'MLP'
mlp {
hidden_units: [256, 128]
}
}
}
blocks {
name: "imp_seq_input"
inputs {
feature_group_name: "imp_seq"
}
input_layer {
output_seq_and_normal_feature: true
}
}
blocks {
name: "imp_seq_bst"
inputs {
block_name: "imp_seq_input"
}
keras_layer {
class_name: 'BST'
bst {
hidden_size: 56
num_attention_heads: 4
num_hidden_layers: 2
intermediate_size: 32
hidden_act: 'gelu'
max_position_embeddings: 25
hidden_dropout_prob: 0.1
attention_probs_dropout_prob: 0
}
}
}
blocks {
name: "clk_seq_input"
inputs {
feature_group_name: "clk_seq"
}
input_layer {
output_seq_and_normal_feature: true
}
}
blocks {
name: "clk_seq_bst"
inputs {
block_name: "clk_seq_input"
}
keras_layer {
class_name: 'BST'
bst {
hidden_size: 56
num_attention_heads: 4
num_hidden_layers: 2
intermediate_size: 32
hidden_act: 'gelu'
max_position_embeddings: 25
hidden_dropout_prob: 0.1
attention_probs_dropout_prob: 0
}
}
}
top_mlp {
hidden_units: [64,32]
}
}
model_params {
l2_regularization: 1e-04
}
embedding_regularization: 1e-04
}
export_config {
multi_placeholder: true
}