Loading train.py +5 −0 Original line number Diff line number Diff line Loading @@ -290,6 +290,8 @@ if __name__ == "__main__": ## Preprocess training sequences data_np = ary start_time = datetime.now() ## Prepare the lev1, lev2 training data, for lev2, we have different strategies based on the mapping_mode if config['mapping_mode'] == 0: # single rank for level 2 lev2_data = np.apply_over_axes(np.sum, data_np, [1, 2, 3]) # 1D array Loading Loading @@ -754,6 +756,9 @@ if __name__ == "__main__": ### extract the individual tensor (sequence) with reversed index (for lower level ranks) rank_npz_dic = {} end_time = datetime.now() print0(f"Execution time: {(end_time - start_time).total_seconds()} seconds") if rank == gid * (int(scale_lev1) + 1): ## roof rank pred_te_seq = te_raw_predictions['prediction'][0, :, 3] ## single sequence for roof rank (for now) rank_npz_dic['roof'] = pred_te_seq.cpu().numpy() ### temp key for this roof rank Loading Loading
train.py +5 −0 Original line number Diff line number Diff line Loading @@ -290,6 +290,8 @@ if __name__ == "__main__": ## Preprocess training sequences data_np = ary start_time = datetime.now() ## Prepare the lev1, lev2 training data, for lev2, we have different strategies based on the mapping_mode if config['mapping_mode'] == 0: # single rank for level 2 lev2_data = np.apply_over_axes(np.sum, data_np, [1, 2, 3]) # 1D array Loading Loading @@ -754,6 +756,9 @@ if __name__ == "__main__": ### extract the individual tensor (sequence) with reversed index (for lower level ranks) rank_npz_dic = {} end_time = datetime.now() print0(f"Execution time: {(end_time - start_time).total_seconds()} seconds") if rank == gid * (int(scale_lev1) + 1): ## roof rank pred_te_seq = te_raw_predictions['prediction'][0, :, 3] ## single sequence for roof rank (for now) rank_npz_dic['roof'] = pred_te_seq.cpu().numpy() ### temp key for this roof rank Loading