upgrade zero inflated lognormal loss, support export structure path, upgrade pre-commit hooks

Author: yangxudong

.pre-commit-config.yaml

@@ -10,8 +10,8 @@ repos:
     rev: 5.12.0
     hooks:
       - id: isort
-  - repo: https://github.com/google/yapf
-    rev: v0.43.0
+  - repo: https://github.com/pre-commit/mirrors-yapf
+    rev: v0.32.0
     hooks:
       - id: yapf
   - repo: https://github.com/pre-commit/pre-commit-hooks

easy_rec/python/compat/embedding_parallel_saver.py

@@ -171,8 +171,8 @@ def _load_embed(
           embed_ids_o = part_id_o + embed_ids_o * len(embed_files)
           sel_ids = np.where(
             np.logical_and(
-              (embed_ids_o % part_num) == part_id, embed_ids_o
-              < embed_part_size * part_num
+              (embed_ids_o % part_num) == part_id,
+              embed_ids_o < embed_part_size * part_num
             )
           )[0]
           part_update_cnt += len(sel_ids)

easy_rec/python/compat/optimizers.py

@@ -497,20 +497,21 @@ def _get_grad_norm(grads_and_vars, embedding_parallel=False):
       sparse_norms.append(gen_nn_ops.l2_loss(grad.values))
     else:
       dense_norms.append(gen_nn_ops.l2_loss(grad))
-  reduced_norms = hvd.grouped_allreduce(
-    part_norms, op=hvd.Sum, compression=hvd.compression.NoneCompressor
-  )
-  sparse_norms = sparse_norms + reduced_norms
-  all_norms = reduced_norms + dense_norms
+  if hvd is not None and part_norms:
+    reduced_norms = hvd.grouped_allreduce(
+      part_norms, op=hvd.Sum, compression=hvd.compression.NoneCompressor
+    )
+    sparse_norms = sparse_norms + reduced_norms
+  all_norms = sparse_norms + dense_norms
   sparse_norm = math_ops.sqrt(
     math_ops.reduce_sum(array_ops.stack(sparse_norms) * 2.0)
-  )
+  ) if sparse_norms else tf.constant(0.0)
   dense_norm = math_ops.sqrt(
     math_ops.reduce_sum(array_ops.stack(dense_norms) * 2.0)
-  )
+  ) if dense_norms else tf.constant(0.0)
   grad_norm = math_ops.sqrt(
     math_ops.reduce_sum(array_ops.stack(all_norms)) * 2.0
-  )
+  ) if all_norms else tf.constant(0.0)
   return sparse_norm, dense_norm, grad_norm
 
 

easy_rec/python/main.py

@@ -209,13 +209,13 @@ def _metric_cmp_fn(best_eval_result, current_eval_result):
       )
       if export_config.metric_bigger:
         return (
-          best_eval_result[export_config.best_exporter_metric]
-          < current_eval_result[export_config.best_exporter_metric]
+          best_eval_result[export_config.best_exporter_metric] <
+          current_eval_result[export_config.best_exporter_metric]
         )
       else:
         return (
-          best_eval_result[export_config.best_exporter_metric]
-          > current_eval_result[export_config.best_exporter_metric]
+          best_eval_result[export_config.best_exporter_metric] >
+          current_eval_result[export_config.best_exporter_metric]
         )
 
     exporters = [

easy_rec/python/model/deepfm.py

@@ -1,5 +1,6 @@
 # -*- encoding:utf-8 -*-
 # Copyright (c) Alibaba, Inc. and its affiliates.
+import logging
 import tensorflow as tf
 
 from easy_rec.python.layers import dnn, fm
@@ -47,6 +48,10 @@ def build_input_layer(self, model_config, feature_configs):
     if not has_final:
       assert model_config.deepfm.wide_output_dim == model_config.num_class
     self._wide_output_dim = model_config.deepfm.wide_output_dim
+    if self._wide_output_dim != 1:
+      logging.warning(
+        'wide_output_dim not equal to 1, it is not a standard model'
+      )
     super(DeepFM, self).build_input_layer(model_config, feature_configs)
 
   def build_predict_graph(self):

easy_rec/python/model/easy_rec_estimator.py

@@ -322,8 +322,7 @@ def _train_model_fn(self, features, labels, run_config):
       'embedding_learning_rate_multiplier'
     ):
       gradient_multipliers = {
-        var:
-        self.train_config.optimizer_config[0].
+        var: self.train_config.optimizer_config[0].
         embedding_learning_rate_multiplier
         for var in tf.trainable_variables() if 'embedding_weights:' in var.name
         or '/embedding_weights/part_' in var.name

easy_rec/python/utils/shape_utils.py

@@ -175,9 +175,8 @@ def pad_or_clip_nd(tensor, output_shape):
   """
   tensor_shape = tf.shape(tensor)
   clip_size = [
-    tf.where(tensor_shape[i] -
-             shape > 0, shape, -1) if shape is not None else -1
-    for i, shape in enumerate(output_shape)
+    tf.where(tensor_shape[i] - shape > 0, shape, -1)
+    if shape is not None else -1 for i, shape in enumerate(output_shape)
   ]
   clipped_tensor = tf.slice(
     tensor, begin=tf.zeros(len(clip_size), dtype=tf.int32), size=clip_size