config.txt

# config.txt
# This file is an example configuration file for the basicbert class.

# Written by David Stein (david@djstein.com).
# See https://www.djstein.com/basicbert/ for more info.
# Source: https://github.com/neuron-whisperer/basicbert


### Parameters About File and Folder Locations

# data_dir: The input data directory containing the .tsv files.
data_dir = input/

# output_dir: The output directory where the model checkpoints will be saved.
output_dir = output/

# bert_config_file: The config.json file for the pre-trained BERT model.
bert_config_file = bert_base/bert_config.json

# vocab_file: The vocabulary file that the BERT model was trained on.
vocab_file = bert_base/vocab.txt

# init_checkpoint: The checkpoint of the pretrained BERT model.
init_checkpoint = bert_base/bert_model.ckpt

# exported_model_dir: A folder containing a saved_model.pb and variables/.
# If this parameter is omitted (such as remaining commented out), BERT will
# use the output directory to export models (and to look for exported models).
# exported_model_dir = /path/to/exported/model/folder/

# labels_file: A file containing the labels, which is saved during training
# and read during prediction. If omitted or nonexistent, BERT will attempt
# to read the labels from the training file.
labels_file = output/labels.txt

# tf_output_file: A file that receives complete output from TensorFlow. If
# omitted (such as remaining commented out), BERT will discard all TensorFlow
# output.
# tf_output_file = output/tf_out.txt

### Parameters About Input Data

# do_lower_case: Whether input should be lower-cased (for uncased BERT model).
do_lower_case = True

# max_seq_length: Input length. Longer input is truncated; shorter input is
# padded.
max_seq_length = 256


### Parameters About Training

# warmup_proportion: Training inputs to use for linear learning warmup period.
# Explanation of warmup: https://stackoverflow.com/a/55942518/1564195
warmup_proportion = 0.05

# learning_rate: Initial learning rate for the Adam optimizer.
learning_rate = 5e-5

# num_train_epochs: Number of training epochs to run.
num_train_epochs = 10

# train_batch_size: Number of batches per training step.
train_batch_size = 25

# save_checkpoint_steps: Frequency of checkpoints.
save_checkpoint_steps = 10000