Chen Qian
Chen Qian
@susnato Thanks for your interest! To clarify - we finetuned BERT earlier on SST2 to make `bert_tiny_uncased_en_sst2`, however, dating back to that time our `BertClassifier` did not have the dropout...
@shivance Could you check your TF version on the notebook? The error looks like a stale one which XLA team has fixed.
I am actually very confused about the error, which appears to be from an optimizer line. But we have never heard XLA is broken on optimizer so far. Could you...
``` mirrored_strategy = tf.distribute.MirroredStrategy() with mirrored_strategy.scope(): model = tf.keras.Sequential([tf.keras.layers.Dense(1, input_shape=(1,))]) model.compile(loss='mse', optimizer='sgd', jit_compile=True) dataset = tf.data.Dataset.from_tensors(([1.], [1.])).repeat(100).batch(10) model.fit(dataset, epochs=2) ``` This works well, weird.
Thanks for checking! It's a problem with MirroredStrategy + XLA + sparse gradients, I am talking to TF teams for a solution.
@harupy @BenWilson2 This is overlapping with rendering notebook in artifact section as we discussed yesterday, do we currently support rendering an html doc?
This proposal is pretty valid to me, is this currently supported? @dbczumar
@sampathweb Thanks for the clarification!
@SuryanarayanaY Yes, by the way could you make sure this information is included somewhere? thanks!