def buzzy_title_based_sim_dfs(
treat_strength,
con_strength,
noise_level,
setting="simple",
seed=0,
base_output_dir='../dat/sim/peerread_buzzytitle_based/'):
labeler = make_buzzy_based_simulated_labeler(treat_strength,
con_strength,
noise_level,
setting=setting,
seed=seed)
num_splits = 10
dev_splits = [0]
test_splits = [0]
# data_file = '../dat/reddit/proc.tf_record'
# vocab_file = "../../bert/pre-trained/uncased_L-12_H-768_A-12/vocab.txt"
tokenizer = tokenization.FullTokenizer(vocab_file=vocab_file,
do_lower_case=True)
input_dataset_from_filenames = make_input_fn_from_file(
data_file,
250,
num_splits,
dev_splits,
test_splits,
tokenizer,
is_training=False,
filter_test=False,
shuffle_buffer_size=25000,
seed=seed,
labeler=labeler)
output_df = dataset_fn_to_df(input_dataset_from_filenames)
output_df = output_df.rename(index=str,
columns={'theorem_referenced': 'treatment'})
output_dir = os.path.join(base_output_dir, "mode{}".format(setting))
os.makedirs(output_dir, exist_ok=True)
output_path = os.path.join(
output_dir,
"beta0{}.beta1{}.gamma{}.tsv".format(treat_strength, con_strength,
noise_level))
output_df.to_csv(output_path, '\t')
def main(_):
print("this program started")
tf.enable_eager_execution() # for debugging
tf.set_random_seed(FLAGS.seed)
tf.logging.set_verbosity(tf.logging.INFO)
if not FLAGS.do_train and not FLAGS.do_eval and not FLAGS.do_predict:
raise ValueError(
"At least one of `do_train`, `do_eval` or `do_predict' must be True.")
bert_config = modeling.BertConfig.from_json_file(FLAGS.bert_config_file)
if FLAGS.max_seq_length > bert_config.max_position_embeddings:
raise ValueError(
"Cannot use sequence length %d because the BERT model "
"was only trained up to sequence length %d" %
(FLAGS.max_seq_length, bert_config.max_position_embeddings))
tf.gfile.MakeDirs(FLAGS.output_dir)
tokenizer = tokenization.FullTokenizer(
vocab_file=FLAGS.vocab_file, do_lower_case=FLAGS.do_lower_case)
tpu_cluster_resolver = None
if FLAGS.use_tpu and FLAGS.tpu_name:
tpu_cluster_resolver = tf.contrib.cluster_resolver.TPUClusterResolver(
FLAGS.tpu_name, zone=FLAGS.tpu_zone, project=FLAGS.gcp_project)
is_per_host = tf.contrib.tpu.InputPipelineConfig.PER_HOST_V2
run_config = tf.contrib.tpu.RunConfig(
cluster=tpu_cluster_resolver,
master=FLAGS.master,
model_dir=FLAGS.output_dir,
save_checkpoints_steps=FLAGS.save_checkpoints_steps,
keep_checkpoint_max=FLAGS.keep_checkpoints,
# save_checkpoints_steps=None,
# save_checkpoints_secs=None,
save_summary_steps=10,
tpu_config=tf.contrib.tpu.TPUConfig(
iterations_per_loop=FLAGS.iterations_per_loop,
num_shards=FLAGS.num_tpu_cores,
per_host_input_for_training=is_per_host))
# Estimator and data pipeline setup
label_dict = \
{'accepted': 2,
'abstract_contains_deep': 2,
'abstract_contains_neural': 2,
'abstract_contains_embedding': 2,
'abstract_contains_outperform': 2,
'abstract_contains_novel': 2,
'abstract_contains_state_of_the_art': 2,
'abstract_contains_state-of-the-art': 2,
'title_contains_deep': 2,
'title_contains_neural': 2,
'title_contains_embedding': 2,
'contains_appendix': 2,
'theorem_referenced': 2,
'equation_referenced': 2,
'year': 11,
'venue': 9,
'arxiv': 3}
params = {'target_name': FLAGS.label, 'num_labels': label_dict[FLAGS.label]}
dev_splits = [int(s) for s in str.split(FLAGS.dev_splits)]
test_splits = [int(s) for s in str.split(FLAGS.test_splits)]
num_train_steps = FLAGS.num_train_steps
num_warmup_steps = FLAGS.num_warmup_steps
model_fn = multiclass_model_fn_builder(
bert_config=bert_config,
init_checkpoint=FLAGS.init_checkpoint,
learning_rate=FLAGS.learning_rate,
num_train_steps=num_train_steps,
num_warmup_steps=num_warmup_steps,
use_tpu=FLAGS.use_tpu,
use_one_hot_embeddings=FLAGS.use_tpu,
label_pred=FLAGS.label_pred,
unsupervised=FLAGS.unsupervised,
polyak=False)
# If TPU is not available, this will fall back to normal Estimator on CPU
# or GPU.
estimator = tf.contrib.tpu.TPUEstimator(
use_tpu=FLAGS.use_tpu,
model_fn=model_fn,
config=run_config,
train_batch_size=FLAGS.train_batch_size,
eval_batch_size=FLAGS.eval_batch_size,
predict_batch_size=FLAGS.predict_batch_size,
params=params)
if FLAGS.do_train:
input_files_or_glob = FLAGS.input_files_or_glob
tf.logging.info("***** Running training *****")
tf.logging.info(" Batch size = %d", FLAGS.train_batch_size)
tf.logging.info(" Num steps = %d", num_train_steps)
# subsample and process the data
with tf.name_scope("training_data"):
train_input_fn = make_input_fn_from_file(
input_files_or_glob=input_files_or_glob,
seq_length=FLAGS.max_seq_length,
num_splits=FLAGS.num_splits,
dev_splits=dev_splits,
test_splits=test_splits,
tokenizer=tokenizer,
is_training=True,
shuffle_buffer_size=25000, # note: bert hardcoded this, and I'm following suit
seed=FLAGS.seed)
# additional logging
hooks = []
if FLAGS.label_pred:
hooks += [
tf.train.LoggingTensorHook({
# 'token_ids': 'token_ids',
# 'token_mask': 'token_mask',
# 'label_ids': 'label_ids',
# 'pred_in': 'summary/in_split/predictions',
# 'pred_out': 'summary/out_split/predictions',
# 'ra_in': 'summary/in_split/labels/kappa/batch_random_agreement/random_agreement',
# 'ra_out': 'summary/out_split/labels/kappa/batch_random_agreement/random_agreement',
},
every_n_iter=1000)
]
estimator.train(input_fn=train_input_fn, max_steps=num_train_steps, hooks=hooks)
if FLAGS.do_train and (FLAGS.do_eval or FLAGS.do_predict):
# reload the model to get rid of dropout and input token masking
trained_model_checkpoint = tf.train.latest_checkpoint(FLAGS.output_dir)
model_fn = multiclass_model_fn_builder(
bert_config=bert_config,
init_checkpoint=trained_model_checkpoint,
learning_rate=FLAGS.learning_rate,
num_train_steps=num_train_steps,
num_warmup_steps=num_warmup_steps,
use_tpu=FLAGS.use_tpu,
use_one_hot_embeddings=FLAGS.use_tpu,
label_pred=True,
unsupervised=False,
polyak=False
)
estimator = tf.contrib.tpu.TPUEstimator(
use_tpu=FLAGS.use_tpu,
model_fn=model_fn,
config=run_config,
train_batch_size=FLAGS.train_batch_size,
eval_batch_size=FLAGS.eval_batch_size,
predict_batch_size=FLAGS.predict_batch_size,
params=params)
if FLAGS.do_eval:
tf.logging.info("***** Running evaluation *****")
# tf.logging.info(" Num examples = %d", len(eval_examples))
tf.logging.info(" Batch size = %d", FLAGS.eval_batch_size)
# This tells the estimator to run through the entire set.
eval_steps = None
# However, if running eval on the TPU, you will need to specify the
# number of steps.
if FLAGS.use_tpu:
# Eval will be slightly WRONG on the TPU because it will truncate
# the last batch.
pass
# eval_steps = int(len(eval_examples) / FLAGS.eval_batch_size)
eval_drop_remainder = True if FLAGS.use_tpu else False
eval_input_fn = make_input_fn_from_file(
input_files_or_glob=FLAGS.input_files_or_glob,
seq_length=FLAGS.max_seq_length,
num_splits=FLAGS.num_splits,
dev_splits=dev_splits,
test_splits=test_splits,
tokenizer=tokenizer,
is_training=False,
filter_test=False,
seed=FLAGS.seed)
result = estimator.evaluate(input_fn=eval_input_fn, steps=eval_steps)
output_eval_file = os.path.join(FLAGS.output_dir, "eval_results.txt")
with tf.gfile.GFile(output_eval_file, "w") as writer:
tf.logging.info("***** Eval results *****")
for key in sorted(result.keys()):
tf.logging.info(" %s = %s", key, str(result[key]))
writer.write("%s = %s\n" % (key, str(result[key])))
if FLAGS.do_predict:
tf.logging.info("***** Running prediction*****")
if FLAGS.use_tpu:
# Warning: According to tpu_estimator.py Prediction on TPU is an
# experimental feature and hence not supported here
raise ValueError("Prediction in TPU not supported")
predict_input_fn = make_input_fn_from_file(
input_files_or_glob=FLAGS.input_files_or_glob,
seq_length=FLAGS.max_seq_length,
num_splits=FLAGS.num_splits,
dev_splits=dev_splits,
test_splits=test_splits,
tokenizer=tokenizer,
is_training=False,
filter_test=False,
seed=FLAGS.seed)
result = estimator.predict(input_fn=predict_input_fn)
output_predict_file = os.path.join(FLAGS.output_dir, "test_results.tsv")
with tf.gfile.GFile(output_predict_file, "w") as writer:
tf.logging.info("***** Predict results *****")
attribute_names = ['in_test',
'treatment_probability',
'expected_outcome_st_treatment', 'expected_outcome_st_no_treatment',
'outcome', 'treatment']
header = "\t".join(
attribute_name for attribute_name in attribute_names) + "\n"
writer.write(header)
for prediction in result:
output_line = "\t".join(
str(prediction[attribute_name]) for attribute_name in attribute_names) + "\n"
writer.write(output_line)
def propensity_sim_ground_truth_and_naive_from_dataset(
beta0=0.25,
beta1=0.0,
gamma=0.0,
exogenous_confounding=0.0,
setting='simple',
tfrecord_file='../dat/PeerRead/proc/arxiv-all.tf_record',
vocab_file="../../bert/pre-trained/uncased_L-12_H-768_A-12/vocab.txt",
base_propensities_path='logs/peerread/arxiv/joint_dragon/seed0/o_accepted_t_buzzy_title/split0/predict/test_results.tsv',
seed=0):
"""
This is a helper function that computes the ground truth and naive estimates
by creating a tf dataset and 'sampling' the required observed data.
This is necessary because we simulate outcomes on the fly using tf's dataset abstraction.
:return:
"""
output = pd.read_csv(base_propensities_path, '\t')
base_propensity_scores = output['treatment_probability'].values
# print(base_propensity_scores)
example_indices = output['index'].values
labeler = make_propensity_based_simulated_labeler(
treat_strength=beta0,
con_strength=beta1,
noise_level=gamma,
base_propensity_scores=base_propensity_scores,
example_indices=example_indices,
exogeneous_con=exogenous_confounding,
setting=setting,
seed=seed)
tokenizer = tokenization.FullTokenizer(vocab_file=vocab_file,
do_lower_case=True)
input_fn = make_input_fn_from_file(tfrecord_file,
250,
num_splits=10,
dev_splits=[0],
test_splits=[0],
tokenizer=tokenizer,
is_training=True,
filter_test=False,
shuffle_buffer_size=25000,
labeler=labeler,
seed=0)
# input_fn = make_input_fn_from_tfrecord(tokenizer=tokenizer, tfrecord=tfrecord_file)
params = {'batch_size': 4096}
dataset = input_fn(params)
sampler = dataset.make_one_shot_iterator()
sample = sampler.get_next()
with tf.Session() as sess:
# confounding = sess.run(sample['confounding'])
y, y0, y1, t = sess.run([
sample['outcome'], sample['y0'], sample['y1'], sample['treatment']
])
tf.reset_default_graph()
ground_truth = y1.mean() - y0.mean()
very_naive = y1[t == 1].mean() - y0[t == 0].mean()
return {'ground_truth': ground_truth, 'noiseless_very_naive': very_naive}
def buzzy_sim_ground_truth_and_naive_from_dataset(
beta0=0.25,
beta1=0.0,
gamma=0.0,
setting='simple',
tfrecord_file='../dat/PeerRead/proc/arxiv-all.tf_record',
vocab_file="../../bert/pre-trained/uncased_L-12_H-768_A-12/vocab.txt",
seed=0):
"""
This is a helper function that computes the ground truth and naive estimates
by creating a tf dataset and 'sampling' the required observed data.
This is necessary because we simulate outcomes on the fly using tf's dataset abstraction.
:return:
"""
tokenizer = tokenization.FullTokenizer(vocab_file=vocab_file,
do_lower_case=True)
labeler = make_buzzy_based_simulated_labeler(treat_strength=beta0,
con_strength=beta1,
noise_level=gamma,
setting=setting,
seed=seed)
input_fn = make_input_fn_from_file(tfrecord_file,
250,
num_splits=10,
dev_splits=[0],
test_splits=[0],
tokenizer=tokenizer,
is_training=True,
filter_test=False,
shuffle_buffer_size=25000,
labeler=labeler,
seed=0)
# input_fn = make_input_fn_from_tfrecord(tokenizer=tokenizer, tfrecord=tfrecord_file)
params = {'batch_size': 4096}
dataset = input_fn(params)
sampler = dataset.make_one_shot_iterator()
sample = sampler.get_next()
with tf.Session() as sess:
# confounding = sess.run(sample['confounding'])
y, y0, y1, t = sess.run([
sample['outcome'], sample['y0'], sample['y1'],
sample['theorem_referenced']
])
tf.reset_default_graph()
# print(y0[t==0].mean())
# print(y1[t==1].mean())
ground_truth = y1.mean() - y0.mean()
noiseless_naive = y1[t == 1].mean() - y0[t == 0].mean()
very_naive = y[t == 1].mean() - y[t == 0].mean()
return {
'ground_truth': ground_truth,
'very_naive': very_naive,
'noiseless_naive': noiseless_naive
}