Condition Error in the 'evaluate' function.

[cifang-Kasit]

Dear author,

Thank you for bring us the nice work.

As the picture shown, in the evaluate function, the way that extract the information from the condition will lead to the information loss of the 'pos_w'.

Would you mind take a quick look and see if this is an issue or my misunderstand?

Thank you so much.

[wenhaomin]

Hi, thank you for the question. Currently, we do not use the pos_w. We leave the interface here, feel free to use them make your model design!

[cifang-Kasit]

Thank you for your quick reply. I used the hyperparameter that provided by the code with the PEMS08 Dataset.

'T_h': 12, 'T_p': 12, 'batch_size': 8, 'data': {'feature_file': 'c:\Users\dell\Documents\GPU_directories\Ten\DiffSTG-main\DiffSTG-main/data/dataset/PEMS08/flow.npy', 'name': 'PEMS08', 'num_features': 1, 'num_recent': 1, 'num_vertices': 170, 'path': 'c:\Users\dell\Documents\GPU_directories\Ten\DiffSTG-main\DiffSTG-main/data/dataset/', 'points_per_hour': 12, 'spatial': 'c:\Users\dell\Documents\GPU_directories\Ten\DiffSTG-main\DiffSTG-main/data/dataset/PEMS08/adj.npy', 'test_start_idx': 14284, 'val_start_idx': 10713}, 'device': device(type='cuda', index=1), 'early_stop': 10, 'epoch': 300, 'gpu_id': 1, 'is_test': False, 'load_model': True, 'load_model_path': './DiffSTG-main/output/model/UGnet+32+100+quad+0.1+100+ddim_multi+12+8+True+PEMS08+0.0+False+False+0.002+8N-100+T_h-12+T_p-12+epsilon_theta-UGnet.dm4stg', 'log_path': 'c:\Users\dell\Documents\GPU_directories\Ten\DiffSTG-main\DiffSTG-main/output/log//UGnet+32+200+quad+0.1+200+ddim_multi+12+8+True+PEMS08+0.0+False+False+0.002+8+True+./DiffSTG-main/output/model/UGnet+32+100+quad+0.1+100+ddim_multi+12+8+True+PEMS08+0.0+False+False+0.002+8N-100+T_h-12+T_p-12+epsilon_theta-UGnet.dm4stg.log', 'logger': <utils.common_utils.Logger object at 0x00000152396EE520>, 'lr': 0.002, 'mask_ratio': 0.0, 'model': {'C': 32, 'F': 1, 'N': 200, 'T_h': 12, 'T_p': 12, 'V': 170, 'beta_end': 0.1, 'beta_schedule': 'quad', 'channel_multipliers': [1, 2], 'd_h': 32, 'day_len': 288, 'device': device(type='cuda', index=1), 'epsilon_theta': 'UGnet', 'is_label_condition': True, 'n_channels': 32, 'sample_steps': 200, 'sample_strategy': 'ddim_multi', 'supports_len': 2, 'week_len': 7}, 'model_name': 'DiffSTG', 'n_samples': 8, 'nni': False, 'optimizer': 'adam', 'start_epoch': 0, 'trial_name': 'UGnet+32+200+quad+0.1+200+ddim_multi+12+8+True+PEMS08+0.0+False+False+0.002+8+True+./DiffSTG-main/output/model/UGnet+32+100+quad+0.1+100+ddim_multi+12+8+True+PEMS08+0.0+False+False+0.002+8N-100+T_h-12+T_p-12+epsilon_theta-UGnet.dm4stg', 'wd': 1e-05} model loaded from: ./DiffSTG-main/output/model/UGnet+32+100+quad+0.1+100+ddim_multi+12+8+True+PEMS08+0.0+False+False+0.002+8N-100+T_h-12+T_p-12+epsilon_theta-UGnet.dm4stg sample num: 10690 sample num: 3548 sample num: 3549

Sorry for directly plot the log. However, I can not achieve the performance as the paper presented.

May I kindly ask some suggestion on the hyperparameter that you performed? For example, I am wondering if the early step is too small.

Since the training really takes time, may I ask if the weight is possible to share or if it's possible to achieve the similar performance from my device.