模型上线

train/train_service_sk_tf2_distibute.py

@@ -236,7 +236,7 @@ early_stopping = tf.estimator.experimental.stop_if_no_increase_hook(wideAndDeepM
 
 hooks = [early_stopping]
 
-train_spec = tf.estimator.TrainSpec(input_fn = lambda: input_fn(DATA_DIR + 'train_samples.csv', 3, True, 2048), hooks = hooks)
+train_spec = tf.estimator.TrainSpec(input_fn = lambda: input_fn(DATA_DIR + 'train_samples.csv', 1, True, 2048), hooks = hooks)
 
 serving_feature_spec = tf.feature_column.make_parse_example_spec(
   linear_feature_columns + dnn_feature_columns)
@@ -248,7 +248,7 @@ exporter = tf.estimator.BestExporter(
   name = "best_exporter",
   compare_fn = lambda best_eval_result, current_eval_result: current_eval_result['auc'] > best_eval_result['auc'],
   serving_input_receiver_fn = serving_input_receiver_fn,
-  exports_to_keep = 3)
+  exports_to_keep = 1)
 
 eval_spec = tf.estimator.EvalSpec(input_fn = lambda: input_fn(DATA_DIR + 'eval_samples.csv', 1, False, 2 ** 15), steps = None, throttle_secs = 120, exporters = exporter)
 
@@ -257,20 +257,21 @@ eval_spec = tf.estimator.EvalSpec(input_fn = lambda: input_fn(DATA_DIR + 'eval_s
 #                                                             summation_method='careful_interpolation')}
 
 # wideAndDeepModel = tf.contrib.estimator.add_metrics(wideAndDeepModel, my_auc)
-
+# /data/files/wideAndDeep/checkpoints/export/best_exporter
 tf.estimator.train_and_evaluate(wideAndDeepModel, train_spec, eval_spec)
 
-wideAndDeepModel.evaluate(lambda: input_fn(DATA_DIR + 'eval_samples.csv', 1, False, 2 ** 15))
+# wideAndDeepModel.evaluate(lambda: input_fn(DATA_DIR + 'eval_samples.csv', 1, False, 2 ** 15))
 
-wideAndDeepModel.export_saved_model(MODEL_BASE_DIR , serving_input_receiver_fn, as_text = False)
+# wideAndDeepModel.export_saved_model(MODEL_BASE_DIR , serving_input_receiver_fn, as_text = False)
 
-filenames = os.listdir(MODEL_BASE_DIR)
+BEST_EXPORT_DIR = CHECKPOINTS_DIR + "export/best_exporter"
+filenames = os.listdir(BEST_EXPORT_DIR)
 
 filename = filenames[0]
 
 from datetime import date
 filename_ = date.fromtimestamp(int(filename)).strftime('%Y%m%d')
 
-shutil.move(os.path.join(MODEL_BASE_DIR, filename), os.path.join(MODEL_BASE_DIR, filename_))
-
+shutil.move(os.path.join(BEST_EXPORT_DIR, filename), os.path.join(BEST_EXPORT_DIR, filename_))
+shutil.move(os.path.join(BEST_EXPORT_DIR, filename_), os.path.join(MODEL_BASE_DIR, filename_))
 print("训练耗时: {}s".format(time.time() - start))
\ No newline at end of file