Apply tf_upgrade_v2 for metrics (#913)

blueoil/metrics/mean_average_precision.py

@@ -113,10 +113,10 @@ def _streaming_tp_fp_array(
             shape=[], dtype=tf.int64, name="num_gt_boxes_value")
 
         # Update operations.
-        tp_op = tf.assign(tp_value, tf.concat([tp_value, tp], axis=0), validate_shape=False)
-        fp_op = tf.assign(fp_value, tf.concat([fp_value, fp], axis=0), validate_shape=False)
-        scores_op = tf.assign(scores_value, tf.concat([scores_value, scores], axis=0), validate_shape=False)
-        num_gt_boxes_op = tf.assign_add(num_gt_boxes_value, num_gt_boxes)
+        tp_op = tf.compat.v1.assign(tp_value, tf.concat([tp_value, tp], axis=0), validate_shape=False)
+        fp_op = tf.compat.v1.assign(fp_value, tf.concat([fp_value, fp], axis=0), validate_shape=False)
+        scores_op = tf.compat.v1.assign(scores_value, tf.concat([scores_value, scores], axis=0), validate_shape=False)
+        num_gt_boxes_op = tf.compat.v1.assign_add(num_gt_boxes_value, num_gt_boxes)
 
         # Value and update ops.
         values = (tp_value, fp_value, scores_value, num_gt_boxes_value)
@@ -136,7 +136,7 @@ def _average_precision(precision, recall, name=None):
     The implementation follows Pascal 2012 and ILSVRC guidelines.
     See also: https://sanchom.wordpress.com/tag/average-precision/
     """
-    with tf.name_scope(name, 'average_precision', [precision, recall]):
+    with tf.compat.v1.name_scope(name, 'average_precision', [precision, recall]):
         # Convert to float64 to decrease error on Riemann sums.
         precision = tf.cast(precision, dtype=tf.float64)
         recall = tf.cast(recall, dtype=tf.float64)
@@ -164,7 +164,7 @@ def _safe_div_zeros(numerator, denominator, name):
     Returns:
       0 if `denominator` <= 0, else `numerator` / `denominator`
     """
-    return tf.where(
+    return tf.compat.v1.where(
         tf.greater(denominator, 0),
         tf.divide(numerator, denominator),
         tf.zeros_like(numerator),
@@ -180,7 +180,7 @@ def _safe_div_ones(numerator, denominator, name):
     Returns:
       1 if `denominator` <= 0, else `numerator` / `denominator`
     """
-    return tf.where(
+    return tf.compat.v1.where(
         tf.greater(denominator, 0),
         tf.divide(numerator, denominator),
         tf.ones_like(numerator),
@@ -200,7 +200,7 @@ def _precision_recall(
     """
     default_name = 'precision_recall_{}'.format(class_name)
     # Sort by score.
-    with tf.name_scope(scope, default_name, [num_gt_boxes, tp, fp, scores]):
+    with tf.compat.v1.name_scope(scope, default_name, [num_gt_boxes, tp, fp, scores]):
         num_detections = tf.size(scores)
         # Sort detections by score.
         scores, idxes = tf.nn.top_k(scores, k=num_detections, sorted=True)
@@ -215,15 +215,15 @@ def _precision_recall(
 
         scalar_precision = tf.cond(
             tf.equal(tf.size(precision), 0),
-            lambda: tf.constant(0, dtype=dtype),
-            lambda: precision[-1],
+            true_fn=lambda: tf.constant(0, dtype=dtype),
+            false_fn=lambda: precision[-1],
             name="scalar_precision"
         )
 
         scalar_recall = tf.cond(
             tf.equal(tf.size(recall), 0),
-            lambda: tf.constant(0, dtype=dtype),
-            lambda: recall[-1],
+            true_fn=lambda: tf.constant(0, dtype=dtype),
+            false_fn=lambda: recall[-1],
             name="scalar_recall"
         )
 
@@ -244,7 +244,7 @@ def _cummax(x, reverse=False, name=None):
     Returns:
     A `Tensor`. Has the same type as `x`.
     """
-    with tf.name_scope(name, "Cummax", [x]) as name:
+    with tf.compat.v1.name_scope(name, "Cummax", [x]) as name:
         # x = tf.convert_to_tensor(x, name="x")
         # Not very optimal: should directly integrate reverse into tf.scan.
         if reverse:

blueoil/metrics/metrics.py

@@ -30,13 +30,13 @@ def tp_tn_fp_fn_for_each(output, labels, threshold=0.5):
 
     """
     predicted = tf.greater_equal(output, threshold)
-    gt_positive = tf.reduce_sum(tf.cast(labels, tf.int32), 0, keep_dims=True)
-    gt_negative = tf.reduce_sum(tf.cast(tf.logical_not(labels), tf.int32), 0, keep_dims=True)
+    gt_positive = tf.reduce_sum(tf.cast(labels, tf.int32), axis=0, keepdims=True)
+    gt_negative = tf.reduce_sum(tf.cast(tf.logical_not(labels), tf.int32), axis=0, keepdims=True)
     true_positive = tf.math.logical_and(predicted, labels)
-    true_positive = tf.reduce_sum(tf.cast(true_positive, tf.int32), 0, keep_dims=True)
+    true_positive = tf.reduce_sum(tf.cast(true_positive, tf.int32), axis=0, keepdims=True)
 
     true_negative = tf.math.logical_and(tf.logical_not(predicted), tf.math.logical_not(labels))
-    true_negative = tf.reduce_sum(tf.cast(true_negative, tf.int32), 0, keep_dims=True)
+    true_negative = tf.reduce_sum(tf.cast(true_negative, tf.int32), axis=0, keepdims=True)
     false_negative = gt_positive - true_positive
     false_positive = gt_negative - true_negative