把训练时l2正则化改成l1正则化

eda/esmm/Model_pipline/train.py

@@ -71,6 +71,7 @@ def input_fn(filenames, batch_size=32, num_epochs=1, perform_shuffle=False):
 
     # Extract lines from input files using the Dataset API, can pass one filename or filename list
     # dataset = tf.data.TFRecordDataset(filenames).map(_parse_fn, num_parallel_calls=8).prefetch(500000)    # multi-thread pre-process then prefetch
+
     # Randomizes input using a window of 256 elements (read into memory)
     # if perform_shuffle:
     #     dataset = dataset.shuffle(buffer_size=256)
@@ -172,7 +173,7 @@ def model_fn(features, labels, mode, params):
         x_cvr = x_concat
         for i in range(len(layers)):
             x_cvr = tf.contrib.layers.fully_connected(inputs=x_cvr, num_outputs=layers[i], \
-                weights_regularizer=tf.contrib.layers.l1_regularizer(l2_reg), scope='cvr_mlp%d' % i)
+                weights_regularizer=tf.contrib.layers.l2_regularizer(l2_reg), scope='cvr_mlp%d' % i)
 
             if FLAGS.batch_norm:
                 x_cvr = batch_norm_layer(x_cvr, train_phase=train_phase, scope_bn='cvr_bn_%d' %i)       #放在RELU之后 https://github.com/ducha-aiki/caffenet-benchmark/blob/master/batchnorm.md#bn----before-or-after-relu
@@ -180,7 +181,7 @@ def model_fn(features, labels, mode, params):
                 x_cvr = tf.nn.dropout(x_cvr, keep_prob=dropout[i])                                  #Apply Dropout after all BN layers and set dropout=0.8(drop_ratio=0.2)
 
         y_cvr = tf.contrib.layers.fully_connected(inputs=x_cvr, num_outputs=1, activation_fn=tf.identity, \
-            weights_regularizer=tf.contrib.layers.l1_regularizer(l2_reg), scope='cvr_out')
+            weights_regularizer=tf.contrib.layers.l2_regularizer(l2_reg), scope='cvr_out')
         y_cvr = tf.reshape(y_cvr,shape=[-1])
 
     with tf.name_scope("CTR_Task"):
@@ -192,7 +193,7 @@ def model_fn(features, labels, mode, params):
         x_ctr = x_concat
         for i in range(len(layers)):
             x_ctr = tf.contrib.layers.fully_connected(inputs=x_ctr, num_outputs=layers[i], \
-                weights_regularizer=tf.contrib.layers.l1_regularizer(l2_reg), scope='ctr_mlp%d' % i)
+                weights_regularizer=tf.contrib.layers.l2_regularizer(l2_reg), scope='ctr_mlp%d' % i)
 
             if FLAGS.batch_norm:
                 x_ctr = batch_norm_layer(x_ctr, train_phase=train_phase, scope_bn='ctr_bn_%d' %i)       #放在RELU之后 https://github.com/ducha-aiki/caffenet-benchmark/blob/master/batchnorm.md#bn----before-or-after-relu
@@ -200,7 +201,7 @@ def model_fn(features, labels, mode, params):
                 x_ctr = tf.nn.dropout(x_ctr, keep_prob=dropout[i])                                  #Apply Dropout after all BN layers and set dropout=0.8(drop_ratio=0.2)
 
         y_ctr = tf.contrib.layers.fully_connected(inputs=x_ctr, num_outputs=1, activation_fn=tf.identity, \
-            weights_regularizer=tf.contrib.layers.l1_regularizer(l2_reg), scope='ctr_out')
+            weights_regularizer=tf.contrib.layers.l2_regularizer(l2_reg), scope='ctr_out')
         y_ctr = tf.reshape(y_ctr,shape=[-1])
 
     with tf.variable_scope("MTL-Layer"):
@@ -222,7 +223,7 @@ def model_fn(features, labels, mode, params):
         ctr_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=y_ctr, labels=y))
         #cvr_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=y_ctcvr, labels=z))
         cvr_loss = tf.reduce_mean(tf.losses.log_loss(predictions=pctcvr, labels=z))
-        loss = ctr_task_wgt * ctr_loss + (1  -ctr_task_wgt) * cvr_loss + l2_reg * tf.nn.l1_loss(Feat_Emb)
+        loss = ctr_task_wgt * ctr_loss + (1  -ctr_task_wgt) * cvr_loss + l2_reg * tf.nn.l2_loss(Feat_Emb)
 
         tf.summary.scalar('ctr_loss', ctr_loss)
         tf.summary.scalar('cvr_loss', cvr_loss)
@@ -410,7 +411,7 @@ def update_or_insert(df2,queue_name):
 
 
 if __name__ == "__main__":
-    print("开始训练")
+
     b = time.time()
     path = "hdfs://172.16.32.4:8020/strategy/esmm/"
     tf.logging.set_verbosity(tf.logging.INFO)