Convert numpy arrays to floats using float constructor, not .item().

I plan on changing Keras to return Python floats instead of Numpy scalars, and this will break if .item() is called. Also, using the float constructor is cleaner, as it clearly communicates that a Python float is desired.

PiperOrigin-RevId: 297192289

official/nlp/transformer/misc.py

@@ -269,7 +269,7 @@ def build_stats(history, callbacks):
   if history and history.history:
     train_hist = history.history
     # Gets final loss from training.
-    stats['loss'] = train_hist['loss'][-1].item()
+    stats['loss'] = float(train_hist['loss'][-1])
 
   if not callbacks:
     return stats

official/r1/resnet/estimator_benchmark.py

@@ -94,12 +94,12 @@ class EstimatorBenchmark(tf.test.Benchmark):
     metrics = []
     if 'accuracy' in eval_results:
       metrics.append({'name': 'accuracy_top_1',
-                      'value': eval_results['accuracy'].item(),
+                      'value': float(eval_results['accuracy']),
                       'min_value': top_1_min,
                       'max_value': top_1_max})
     if 'accuracy_top_5' in eval_results:
       metrics.append({'name': 'accuracy_top_5',
-                      'value': eval_results['accuracy_top_5'].item()})
+                      'value': float(eval_results['accuracy_top_5'])})
 
     if examples_per_sec_hook:
       exp_per_second_list = examples_per_sec_hook.current_examples_per_sec_list

official/vision/image_classification/common.py

@@ -164,18 +164,18 @@ def build_stats(history, eval_output, callbacks):
   """
   stats = {}
   if eval_output:
-    stats['accuracy_top_1'] = eval_output[1].item()
-    stats['eval_loss'] = eval_output[0].item()
+    stats['accuracy_top_1'] = float(eval_output[1])
+    stats['eval_loss'] = float(eval_output[0])
 
   if history and history.history:
     train_hist = history.history
     # Gets final loss from training.
-    stats['loss'] = train_hist['loss'][-1].item()
+    stats['loss'] = float(train_hist['loss'][-1])
     # Gets top_1 training accuracy.
     if 'categorical_accuracy' in train_hist:
-      stats[TRAIN_TOP_1] = train_hist['categorical_accuracy'][-1].item()
+      stats[TRAIN_TOP_1] = float(train_hist['categorical_accuracy'][-1])
     elif 'sparse_categorical_accuracy' in train_hist:
-      stats[TRAIN_TOP_1] = train_hist['sparse_categorical_accuracy'][-1].item()
+      stats[TRAIN_TOP_1] = float(train_hist['sparse_categorical_accuracy'][-1])
 
   if not callbacks:
     return stats