han-jd / gru-d Goto Github PK

I find that you use a default batch_size = 1 when you initialize data loader. But when I assign other value to batch_size, it will raise an error which shows that dimension mismatch. I wonder should I reset parameters' shape, if I want to use arbitrary batch_size?

1. In your code, it seems that you are just using 0 for missing value.
2. The decay equation seems not correct

When running the code with only three modifications:

1. Download Outcomes-a.txt from https://physionet.org/content/challenge-2012/1.0.0/ since the file is not in input/.
2. Removing the # from #np.save('./input/inputs', inputs), so the code can save and load inputs.
3. Adding allow_pickle=True to all np.load().

The value of AUC obtained is around 0.5304 instead of 0.8431. Any of these changes or perhaps a different version of a module could be the reason for this lower result? What is the setup to run this code and obtained the correct result?

Solved. Because os.listdir() doesn't read files in order, which makes features and labels mismatch.
Just change "for filename in os.listdir(inputpath):" to this:
filenames = os.listdir(inputpath)
filenames.sort(key=lambda x: int(x[:-4]))
for filename in filenames:

Aren't you calculating mean over the whole dataset? When you should be doing it over only train?

.

I hope you can introduce the data processing part of PhysioNet in detail, how it is feed into the network, thank you very much!!!