def save(self, path: RichPath) -> None:
variables_to_save = list(set(self.__sess.graph.get_collection(tf.GraphKeys.GLOBAL_VARIABLES)))
weights_to_save = self.__sess.run(variables_to_save)
weights_to_save = {var.name: value
for (var, value) in zip(variables_to_save, weights_to_save)}
data_to_save = {
"model_type": type(self).__name__,
"hyperparameters": self.hyperparameters,
"metadata": self.__metadata,
"weights": weights_to_save,
"run_name": self.__run_name,
}
path.save_as_compressed_file(data_to_save)
def run_experiment(queries: List[Dict[str,
Any]], best_index: int, n_trials: int,
dfs: Dict[str, pd.DataFrame], dataset_params: Dict[str,
Any],
optimizer_params: Dict[str, Any], output_file: RichPath):
times: List[float] = []
total_time: float = 0.0
best_count: float = 0.0
avg_times: List[float] = []
best_frac: List[float] = []
selected_index: List[int] = []
optimizer = get_optimizer(optimizer_params['name'], queries,
**optimizer_params['params'])
for t in range(1, n_trials + 1):
start = time.time()
query_index, q = optimizer.get_query(time=t)
df = merge_frames(q['from'], q['on'], options=[['home', 'away']])
elapsed = time.time() - start
optimizer.update(query_index, reward=-elapsed)
df.copy()
# Avoid time measurements on the first iteration due to caching
selected_index.append(query_index)
if t > 1:
times.append(elapsed)
total_time += elapsed
avg_times.append(total_time / (t - 1))
best_count += float(query_index == best_index)
best_frac.append(best_count / (t - 1))
if t % 100 == 0:
print(f'Completed {t} trials')
# Collect and write the metrics
metrics = dict(times=times,
avg_times=avg_times,
best_frac=best_frac,
selected_index=selected_index)
output_file.save_as_compressed_file(metrics)
def run_predict(model_path: RichPath, test_data_path: RichPath,
output_file: RichPath):
test_run_id = "_".join(
[time.strftime("%Y-%m-%d-%H-%M-%S"),
str(os.getpid())])
test_hyper_overrides = {
'run_id': test_run_id,
"dropout_keep_rate": 1.0,
}
test_data_chunks = test_data_path.get_filtered_files_in_dir('*.jsonl.gz')
# Restore model
model = model_restore_helper.restore(model_path,
is_train=False,
hyper_overrides=test_hyper_overrides)
def predictions():
for annotation in model.annotate(test_data_chunks):
if ignore_annotation(annotation.original_annotation):
continue
ordered_annotation_predictions = sorted(
annotation.predicted_annotation_logprob_dist,
key=lambda x: -annotation.predicted_annotation_logprob_dist[
x])[:10]
annotation_dict = annotation._asdict()
logprobs = annotation_dict['predicted_annotation_logprob_dist']
filtered_logprobs = []
for annot in ordered_annotation_predictions:
logprob = float(logprobs[annot])
if annot == '%UNK%' or annot == '%UNKNOWN%':
annot = 'typing.Any'
filtered_logprobs.append((annot, logprob))
annotation_dict[
'predicted_annotation_logprob_dist'] = filtered_logprobs
yield annotation_dict
output_file.save_as_compressed_file(predictions())
