Can not pass test_train_cifar10 in which my accuracy and loss did not match golden standard values

[rocinant]

My result of test_train_cifar10 is (np.float64(0.1015625), np.float32(3.6889665)) which is different from [0.09375, 3.5892258], which mean my acc and loss are both little higher. And I just can not figure out why. I have passed all previous tests in hw4 until this train_cifar10 test. I looked back and forth and the code runs just fine, it is just the result can not match up. Any idea is appreciated.

[rocinant]

While, I finally located the reason why my implementation's result is different from golden standard. In CIFAR10Dataset when loading the file from data_batch_1 to data_batch_5 and packing them together, you need to pack them one by one sequentially. But I use some os.walk to find all related files whose order is not from 1 to 5 thus the traning data is kind of shuffled. While the test of test_cifar10 just get passed even the data is shuffled. I think this should be pointed out in the document, otherwise others will waste lots of time trying to debug this.