[tfdbg2] A few fixes and improvements to example debug_mnist_v2

1. Change the default `--dump_tensor_debug_mode` flag value to
   `FULL_HEALTH`, a mode more suitable for numerical instability
   debugging for the particular bug in this example than the previous
   default value `NO_TENSOR`.
2. Change the default `--dump_circular_buffer_size` value to -1, to
   accommodate the possibility of longer runs where user would
   want to see the debug data in its entirety.
3. Rename a few weight variables. They were previously named in a
   confusing way.
4. Change "logits" to "probs", as they are generated by a `softmax`
   operation and hence are more accurately described as probability
   scores.

PiperOrigin-RevId: 316473164
Change-Id: I4eb13f9581a4d4e550b3b3a5cd132eeffc7dd043

tensorflow/python/debug/examples/v2/debug_mnist_v2.py

@@ -98,16 +98,18 @@ def parse_args():
   parser.add_argument(
       "--dump_tensor_debug_mode",
       type=str,
-      default="NO_TENSOR",
+      default="FULL_HEALTH",
       help="Mode for dumping tensor values. Options: NO_TENSOR, CURT_HEALTH, "
-      "CONCISE_HEALTH, SHAPE, FULL_TENSOR. This is relevant only when "
+      "CONCISE_HEALTH, SHAPE, FULL_HEALTH. This is relevant only when "
       "--dump_dir is set.")
   # TODO(cais): Add more tensor debug mode strings once they are supported.
   parser.add_argument(
       "--dump_circular_buffer_size",
       type=int,
-      default=1000,
+      default=-1,
       help="Size of the circular buffer used to dump execution events. "
+      "A value <= 0 disables the circular-buffer behavior and causes "
+      "all instrumented tensor values to be dumped. "
       "This is relevant only when --dump_dir is set.")
   parser.add_argument(
       "--use_random_config_path",
@@ -178,9 +180,9 @@ def main(_):
     return activations
 
   # init model
-  hidden = get_dense_weights(IMAGE_SIZE**2, HIDDEN_SIZE)
-  logits = get_dense_weights(HIDDEN_SIZE, NUM_LABELS)
-  variables = hidden + logits
+  hidden_weights = get_dense_weights(IMAGE_SIZE**2, HIDDEN_SIZE)
+  output_weights = get_dense_weights(HIDDEN_SIZE, NUM_LABELS)
+  variables = hidden_weights + output_weights
 
   @tf.function
   def model(x):
@@ -193,15 +195,25 @@ def main(_):
     Returns:
       A (?, 10) tensor containing the class scores for each example.
     """
-    hidden_act = dense_layer(hidden, x)
-    logits_act = dense_layer(logits, hidden_act, tf.identity)
+    hidden_act = dense_layer(hidden_weights, x)
+    logits_act = dense_layer(output_weights, hidden_act, tf.identity)
     y = tf.nn.softmax(logits_act)
     return y
 
   @tf.function
-  def loss(logits, labels):
-    """Calculates cross entropy loss."""
-    diff = -(labels * tf.math.log(logits))
+  def loss(probs, labels):
+    """Calculates cross entropy loss.
+
+    Args:
+      probs: Class probabilities predicted by the model. The shape is expected
+        to be (?, 10).
+      labels: Truth labels for the classes, as one-hot encoded vectors. The
+        shape is expected to be the same as `probs`.
+
+    Returns:
+      A scalar loss tensor.
+    """
+    diff = -labels * tf.math.log(probs)
     loss = tf.reduce_mean(diff)
     return loss