Add TPU and evaluation saving for GLUE finetuning #234
Conversation
|
Thanks for your pull request! It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA). View this failed invocation of the CLA check for more information. For the most up to date status, view the checks section at the bottom of the pull request. |
chenmoneygithub
changed the title
examples/bert/bert_finetune_glue.py
Outdated
| @@ -161,12 +179,14 @@ def __init__( | |||
| kernel_initializer=initializer, | |||
| name="logits", | |||
| ) | |||
| self._drop_out = tf.keras.layers.Dropout(dropout) | |||
examples/bert/bert_finetune_glue.py
Outdated
| model = keras.models.load_model(FLAGS.saved_model_input, compile=False) | ||
| # model = keras.models.load_model(FLAGS.saved_model_input, compile=False) | ||
| model = keras.models.load_model( | ||
| "gs://chenmoney-testing-east/" + FLAGS.saved_model_input, |
There was a problem hiding this comment.
I think this needs an update!
| @@ -270,6 +306,58 @@ def preprocess_data(inputs, labels): | |||
| f"The best hyperparameters found are:\nLearning Rate: {best_hp['lr']}" | |||
| ) | |||
|
|
|||
| if FLAGS.saved_evaluations_output: | |||
| filenames = { | |||
| "cola": "CoLA.tsv", | |||
There was a problem hiding this comment.
are these the exact filename needed for GLUE submission?
There was a problem hiding this comment.
yes, they require the exact name to match
examples/bert/bert_finetune_glue.py
Outdated
| @@ -53,6 +53,12 @@ | |||
| "The directory to save the finetuned model.", | |||
| ) | |||
|
|
|||
| flags.DEFINE_string( | |||
| "saved_evaluations_output", | |||
There was a problem hiding this comment.
maybe something a little more directly named here? tsv_prediction_output?
| with tf.io.gfile.GFile(filename, "w") as f: | ||
| # Write the required headline for GLUE. | ||
| f.write("index\tprediction\n") | ||
| for i in range(test_ds.cardinality()): |
There was a problem hiding this comment.
seems like we should be able to do this with a more readable tf.data loop here. for idx, x in enumerate(dataset): works directly, let's avoid calling iter and range and next ourselves
| ) | ||
|
|
||
| @tf.function | ||
| def eval_step(iterator): |
There was a problem hiding this comment.
can we just the actual keras.evaluate here?
There was a problem hiding this comment.
We need the predicted label to write to the file. evaluate only returns metrics?
chenmoneygithub
force-pushed
the
bert-finetuning
branch
from
d1abb4a
to
b0dd210
Compare
examples/bert/bert_finetune_glue.py
Outdated
| @@ -161,12 +173,14 @@ def __init__( | |||
| kernel_initializer=initializer, | |||
| name="logits", | |||
| ) | |||
| self._drop_out_layer = tf.keras.layers.Dropout(dropout) | |||
examples/bert/bert_finetune_glue.py
Outdated
| @@ -178,6 +192,7 @@ def __init__(self, model_config): | |||
|
|
|||
| def build(self, hp): | |||
| model = keras.models.load_model(FLAGS.saved_model_input, compile=False) | |||
| model = model.bert_model | |||
There was a problem hiding this comment.
I don't think this will work right? this seems to be implying we save the whole model in bert_train.py, which we do not
examples/bert/bert_finetune_glue.py
Outdated
|
|
||
| labelname = labelnames.get(FLAGS.task_name) | ||
| with tf.io.gfile.GFile(filename, "w") as f: | ||
| # Write the required headline for GLUE. |
There was a problem hiding this comment.
I think you can use the builtin csv with a different delimiter to write tsvs
https://docs.python.org/3/library/csv.html
https://stackoverflow.com/q/29895602
We should prefer that over doing this writing outself
No description provided.