| import os |
| import sys |
| import time |
| import traceback |
| import pandas as pd |
|
|
| |
| ROOT = os.path.dirname(__file__) |
| if ROOT not in sys.path: |
| sys.path.insert(0, ROOT) |
|
|
| import run_inference as ri |
|
|
|
|
| def predict_and_time(input_csv: str, output_csv: str): |
| df = pd.read_csv(input_csv) |
| results = [] |
|
|
| for idx, row in df.iterrows(): |
| q = row.get('question') |
| ground = row.get('action_label') |
| try: |
| start = time.perf_counter() |
| |
| with ri.torch.no_grad(): |
| emb = ri.embedder.encode(q, convert_to_tensor=True).to(ri.DEVICE) |
| action, log_prob, _ = ri.model.get_action(emb) |
| pred_idx = int(action.item()) |
| end = time.perf_counter() |
| elapsed = end - start |
| except Exception: |
| pred_idx = None |
| elapsed = None |
| traceback.print_exc() |
|
|
| results.append({ |
| 'question': q, |
| 'ground_truth': ground, |
| 'predicted_label': pred_idx, |
| 'time_took': elapsed, |
| }) |
|
|
| out_df = pd.DataFrame(results) |
| out_df.to_csv(output_csv, index=False) |
| print(f"Wrote results to {output_csv} ({len(out_df)} rows)") |
|
|
|
|
| if __name__ == '__main__': |
| inp = os.path.join(ROOT, 'eval_queries.csv') |
| out = os.path.join(ROOT, 'eval_with_timing_results.csv') |
| print(f"Input: {inp}\nOutput: {out}\nDevice: {ri.DEVICE}") |
| predict_and_time(inp, out) |
|
|
