Why do I get mse=0.91 for the MSLR dataset using DANets?

[ghost]

I download the code+date from your github, but get mse=0.91, not 0.55. There is some codes I debug: (1) in data_util.py, add '.txt' def svm2pkl(source, save_path): before: X_train, y_train = load_svmlight_file(os.path.join(source, 'train')) X_valid, y_valid = load_svmlight_file(os.path.join(source, 'vali')) X_test, y_test = load_svmlight_file(os.path.join(source, 'test')) after: def svm2pkl(source, save_path): X_train, y_train = load_svmlight_file(os.path.join(source, 'train.txt')) X_valid, y_valid = load_svmlight_file(os.path.join(source, 'vali.txt')) X_test, y_test = load_svmlight_file(os.path.join(source, 'test.txt')) (2)in deault.py, add 'cfg.fit.weight_decay = 1e-5','cfg.fit.schedule_step = 20'. As I do regression task, when I run main.py, the Keyerror: weight_decay ,Keyerror: schedule_step .I can't find these two parameters define in deault.py ,MSLR.yaml, main.py, so I add 'cfg.fit.weight_decay = 1e-5','cfg.fit.schedule_step = 20' in deault.py. (3)Run codes by using python predict.py -d [dataset_name] -m [model_file_path] -g [gpu_id] where the best.pth path is already specified. I don't get the mean: Replace the resume_dir path with the file path containing your trained model/weight. Does it mean set the resume_dir path to the best.pth path?

[WhatAShot]

Thank you for your interests. The results may fluctuate slightly, but not by such a large margin.

Maybe you forgot to pre-process the data?

Maybe you did not train the model before inference?

You may check this two aspects first.