Images being used with .predict() don't appear to be normalized the same way as the images pulled using get_x from the datasets (for feeding into predict_array).

[nwertzberger]

Hello,

I've been trying to track down why I get a different answer for:

learner.predict_array(dataset.val_dl.dataset.get_x(0)[None])

vs what i see in preds from running (learner.predict()) using the same dataset. I noticed the following:

• When I look up predict_with_targs, it seems that it turns the dataloader into an iterator, and then shoves those values into the model via this:

    def predict_with_targs(m, dl):
        m.eval()
        if hasattr(m, 'reset'): m.reset()
        res = []
        for *x,y in iter(dl): res.append([get_prediction(m(*VV(x))),y])
        preda,targa = zip(*res)
        return to_np(torch.cat(preda)), to_np(torch.cat(targa))
  

• When I pull out some data using this method, I notice that the resulting tensor is NOT normalized, but the dimensions are 3x128x128 (what my model expects):

    x = list(iter(data2.val_dl))[0]
    to_np(x[0][0,:,:,:])
    # This shows a tensor with lots of values bigger than 1 and smaller than -1
  

• When I pull from the dataset using get_x, it IS normalized, but the dimensions are 128x128x3

Given that the other interfaces return normalized images, i think this might not be on purpose, but I'm no expert on what should or should not be normalized. I know that it not being normalized makes lime integration difficult, and, because the dimensions of the images pulled in each case is different, I have to transpose them as well, which seems awkward.

[jph00]

You should never call get_x directly. Instead just use the standard dataset indexer: dataset.val_ds[i]

[nwertzberger]

Is it intentional that the resulting tensor is not normalized between 0 and 1?

…

On Wed, Jun 6, 2018 at 3:12 PM Jeremy Howard ***@***.***> wrote: You should never call get_x directly. Instead just use the standard dataset indexer: dataset.val_ds[i] — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#535 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAV_m2qo2WER1Y4UBKjpr7bTQRNr3hFwks5t6DfJgaJpZM4UX3Y0> .

[yanneta]

Here is the answer to your question https://discuss.pytorch.org/t/whats-the-range-of-the-input-value-desired-to-use-pretrained-resnet152-and-vgg19/1683 On Wed, Jun 6, 2018 at 1:53 PM, Nick Wertzberger <notifications@github.com> wrote:

…

Is it intentional that the resulting tensor is not normalized between 0 and 1? On Wed, Jun 6, 2018 at 3:12 PM Jeremy Howard ***@***.***> wrote: > You should never call get_x directly. Instead just use the standard > dataset indexer: dataset.val_ds[i] > > — > You are receiving this because you authored the thread. > Reply to this email directly, view it on GitHub > <#535 (comment)>, or mute > the thread > <https://github.com/notifications/unsubscribe-auth/AAV_ m2qo2WER1Y4UBKjpr7bTQRNr3hFwks5t6DfJgaJpZM4UX3Y0> > . > — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#535 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AC3CZSrApH80DGGkqt8jiZHnl5GWEBf_ks5t6EFCgaJpZM4UX3Y0> .

[nwertzberger]

I’m sorry, I am not quite understanding the answer here. In one interface, I see the images are normalized from 0 to 1, and that makes sense. I expect to see normalized images as mentioned In Torchvision. On the iterator interface, I see tensors that do not appear to be normalized. Also, when I give a normalized image to my network, it gives significantly different predictions than my non normalized input, making me think that nothing is auto-normalizing the input, at least through predict_array.

…

On Wed, Jun 6, 2018 at 4:05 PM Yannet ***@***.***> wrote: Here is the answer to your question https://discuss.pytorch.org/t/whats-the-range-of-the-input-value-desired-to-use-pretrained-resnet152-and-vgg19/1683 On Wed, Jun 6, 2018 at 1:53 PM, Nick Wertzberger ***@***.*** > wrote: > Is it intentional that the resulting tensor is not normalized between 0 and > 1? > On Wed, Jun 6, 2018 at 3:12 PM Jeremy Howard ***@***.***> > wrote: > > > You should never call get_x directly. Instead just use the standard > > dataset indexer: dataset.val_ds[i] > > > > — > > You are receiving this because you authored the thread. > > Reply to this email directly, view it on GitHub > > <#535 (comment)>, > or mute > > the thread > > <https://github.com/notifications/unsubscribe-auth/AAV_ > m2qo2WER1Y4UBKjpr7bTQRNr3hFwks5t6DfJgaJpZM4UX3Y0> > > . > > > > — > You are receiving this because you are subscribed to this thread. > Reply to this email directly, view it on GitHub > <#535 (comment)>, or mute > the thread > < https://github.com/notifications/unsubscribe-auth/AC3CZSrApH80DGGkqt8jiZHnl5GWEBf_ks5t6EFCgaJpZM4UX3Y0 > > . > — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#535 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAV_mzDoF4iqg_xWDOY0f2i_0Y2zvry1ks5t6EQTgaJpZM4UX3Y0> .