Created
November 20, 2019 13:14
-
-
Save sayakpaul/e595b12b09dbb40e452db0289a6aa76c to your computer and use it in GitHub Desktop.
Thanks for passing this along @tlkh. Will incorporate your suggestions and report back.
@sayakpaul can you also test with larger images? CIFAR's 32x32 is "too small".
@tlkh will do! Your materials here (https://github.com/NVAITC/pycon-sg19-tensorflow-tutorial) are just fantastic. Do you mind if I used them in my decks with a proper citation?
@sayakpaul no problem! Please feel free to modify and use for your own purposes.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Note that just calling
opt = tf.keras.mixed_precision.experimental.LossScaleOptimizer(opt, "dynamic")
alone does not activate mixed precision. Moving forward, you should calltf.keras.mixed_precision.experimental.set_policy('mixed_float16')
before constructing the model. To fix the current "dtype" error, set the dtype of the softmax layer totf.float32
manually.