It looks like the repo is set up to reproduce the bottom-left-most cell in Table 1 of the the MAML paper, which has 1-shot 5-way miniimagenet test accuracy of 48.70 ± 1.84%. Is that correct? When I run the code from the official MAML repo I get a similar number, but when I run this code I'm getting around 40%. Are there some settings I need to change? I like the simplicity of this code and would like to use it as long as I can reproduce the original results.