Yang Feng

@HYPJUDY You are correct. It doesn't speed up a single eval. I may provide an eval file you wanted in the future. Sometimes, the CIDEr increases and the captioning quality...

start_id is the line number of \ and end_id is the line number of \ in the dictionary file.

tf.contrib has been moved to other packages such as tf_addons, tf_slim. I will provide a tf2 compatible training code in the future.

Are you using the MSCOCO dataset? Have you tried the data files in https://drive.google.com/drive/folders/1ol8gLj6hYgluldvdj9XFKm16TCqOr7EE?

Did you generate the training input by yourself? You may try to use the input files from google drive.

Could you first sample length uniformly and then sample the start position?

If the length of the reference is L, sample the target length in the range [1, L] uniformly.

@tanghaoyu258 group_by_window is used to improve running speed. It is safe to remove it. dynamic_rnn will be run n steps, where n is the length of the longest sequence in...

A test code ``` import tensorflow as tf import tensorflow_addons as tfa im = tf.zeros([100, 100, 3]) mat = tf.constant([1, 0, 0, 0, 1, 0, 0, 0], dtype=tf.float32) with tf.GradientTape()...

@inio I find the function you mentioned in https://github.com/google-ar/arcore-android-sdk/blob/master/samples/hello_ar_java/app/src/main/java/com/google/ar/core/examples/java/common/rendering/BackgroundRenderer.java#L133. However, c-a and d-b are 0 or 1e-8 when I run the code on Google Pixel. Is it normal?