Difference in reported result in paper and implementation for smallnorb

[sairaamVenkatraman]

Hi, this is not an issue as such. Your results posted for smallnorb using dynamic routing are reported as 99.56% - whereas those in the paper are 97.3% for dynamic routing. I haven't run your code as of yet and am more comfortable in pytorch than I am in tensorflow. My question is this: What new addition did you bring to get this result? This seems to be state of the art.

[j1schmid]

There is a bug in the function "extract_patch" in "DataReader.py" yielding in using the training images for both training and testing. The reported test error is actually the training error.

[jianghaojun]

Yes. Obviously, there is a bug in function "extract_patch".

def extract_patch(dataset):
    extracted = []
    for img in **train_dat**:
	img_ = img[0].reshape(96,96,1)
	extracted.append(img_)
	img_ = img[1].reshape(96,96,1)
	extracted.append(img_)
    return np.array(extracted)

[EscVM]

Hi!

Do you know what is the real test error?