if step != 0 and step % args.training_steps == 0:
writer.add_scalar('Loss/train', running_loss / 100, step)
print(f"Training Loss: {running_loss/100}")
print("Getting Final Evaluation Results")
print("--------------------")
sorted_preds, indices = eval_util.calculate_predictions(
eval_loader,
model,
device,
args.target_publication,
version="Evaluation",
step=step,
writer=writer,
check_recall=True)
ranked_df = eval_util.create_ranked_results_list(
final_word_ids, args.word_embedding_type, sorted_preds, indices,
eval_data)
eval_util.save_ranked_df(output_path, "evaluation", ranked_df,
args.word_embedding_type)
print(
f"Ranked Data Saved to {output_path / 'results' / 'evaluation'}!")
check = False
break
# save model for easy future reloading
model_path = output_path / "model"
if not model_path.is_dir():
model_path.mkdir()
model_string = args.word_embedding_type + "-inner-product-model.pt"
model_path = model_path / model_string
torch.save(model.state_dict(), model_path)
abs_model_path = Path(args.model_path)
config_file = abs_model_path.parent / "config.json"
config = BertConfig.from_json_file(config_file)
model = BertForSequenceClassification(config)
model.load_state_dict(torch.load(abs_model_path))
model.to(device)
model.eval()
torch.no_grad()
print("Model Loaded")
print(model)
print("-------------------")
data_logit_list = []
for batch in tqdm(raw_loader):
current_logits = eval_util.calculate_batched_predictions(
batch, model, device, args.target_publication)
data_logit_list = data_logit_list + list(current_logits)
converted_list = np.array(data_logit_list)
sorted_preds = np.sort(converted_list)
indices = np.argsort(converted_list)
ranked_df = eval_util.create_ranked_results_list(final_word_ids, sorted_preds,
indices, raw_data)
eval_util.save_ranked_df(output_path, args.data_name, ranked_df)
print("Predictions Made")
print(
f"Ranked Data Saved to {output_path / 'results' / 'evaluation'} directory!"
)