support session_config in estimator and refine evalute (intel-analyti…

…cs#2665)

python/orca/src/bigdl/orca/tfpark/estimator.py

@@ -115,7 +115,7 @@ def _call_model_fn(self, features, labels, mode, config):
 
         return model_fn_results
 
-    def train(self, input_fn, steps=None):
+    def train(self, input_fn, steps=None, session_config=None):
         """Trains a model given training data `input_fn`.
 
         :param input_fn: A function that constructs the input data for evaluation. The
@@ -163,7 +163,8 @@ def train(self, input_fn, steps=None):
                                                     spec.loss,
                                                     sess=sess,
                                                     dataset=result,
-                                                    model_dir=zoo_ckpt_path)
+                                                    model_dir=zoo_ckpt_path,
+                                                    session_config=session_config)
 
                     opt.optimize(MaxIteration(steps))
                     sess.run(assign_step, feed_dict={add_step_input: steps})
@@ -223,9 +224,17 @@ def evaluate(self, input_fn, eval_methods, steps=None, checkpoint_path=None):
                         sess.run(tf.global_variables_initializer())
                     if isinstance(spec.predictions, dict):
                         if "mae" in eval_methods:
+                            key = prediction_keys.PredictionKeys.PREDICTIONS
+                            msg = "{} is required for evaluating mse,".format(key) + \
+                                  " please add it in your model_fn predictions"
+                            assert key in spec.prediction, msg
                             outputs = [
                                 spec.predictions[prediction_keys.PredictionKeys.PREDICTIONS]]
                         else:
+                            key = prediction_keys.PredictionKeys.LOGITS
+                            msg = "{} is required in for evaluating,".format(key) + \
+                                  " please add it in your model_fn predictions"
+                            assert key in spec.predictions, msg
                             outputs = [
                                 spec.predictions[prediction_keys.PredictionKeys.LOGITS]]
                     else: