b1bfb8b4159dca7e057c3253dbd519984865b893,examples/transformer_standalone.py,,,#,41

Before Change

                        "is the global batch size and not the per-shard batch.")
tf.flags.DEFINE_float("dropout", 0.1, "dropout")
tf.flags.DEFINE_float("label_smoothing", 0.1, "label smoothing")
tf.flags.DEFINE_integer(
    "train_steps", 100000, "Total number of training steps.")

// DISTRIBUTED LAYOUT
// When running on TPU, make sure that the size of the mesh
// (the product of all dimension sizes) equals the number of TPU cores.

After Change

    ,
    "data_dir for TensorFlow Datasets"
)
tf.flags.DEFINE_boolean(
    "text2self",
    False,
    "Whether to train a language model (True) or encoder-decoder text-to-text "
    "model (False)."
)

// MODEL HYPERPARAMETERS
tf.flags.DEFINE_integer("max_length", 256,
                        "maximum sequence length (checkpoints depend on this)")

Instances