vdurnov / xview2_1st_place_solution Goto Github PK

Hi, I have a question about the validate function of classification stage.
The codes are in def validate(net,data_loader), train50_cls_cce.py.
`
for j in range(msks.shape[0]): # msks.shape[0] is batch size
tp[4] += np.logical_and(msks[j, 0] > 0, msk_pred[j] > 0).sum()
fn[4] += np.logical_and(msks[j, 0] < 1, msk_pred[j] > 0).sum()
fp[4] += np.logical_and(msks[j, 0] > 0, msk_pred[j] < 1).sum()

            targ = lbl_msk[j][msks[j, 0] > 0]
            pred = msk_damage_pred[j].argmax(axis=0)
            pred = pred * (msk_pred[j] > _thr)
            pred = pred[msks[j, 0] > 0]
            for c in range(4):
                tp[c] += np.logical_and(pred == c, targ == c).sum()
                fn[c] += np.logical_and(pred != c, targ == c).sum()
                fp[c] += np.logical_and(pred == c, targ != c).sum()

`
I was wondering why there is 'pred == c, targ == c' , when the values of c ranges from 0 to 3 while the values of targ ranges from 1 to 4?
Or did I get it wrong? Please explain to me. Thanks a million!

In the individual model output folders, I'm getting the correct outputs, i.e. there are buildings having different colors associated with their damage values.

But the submissions folder being created has all black images. I tried reading them using cv2 and all pixel values are indeed 0. Not sure how to proceed.

Can you please give insight into the need for all kinds of randomized transformation on images? And how it is going to help in model? Thanks in advance

In the part of prediction, for example predict50cls.py, the script would find files with 'pre' and 'post' in their names. But when I only have one post disaster image, what should I do to use your model?

`

    msk[..., 1][msk[..., 2:].max(axis=2)] = False
    msk[..., 3][msk[..., 2]] = False
    msk[..., 4][msk[..., 2]] = False
    msk[..., 4][msk[..., 3]] = False
    msk[..., 0][msk[..., 1:].max(axis=2)] = False

`
I have a question, what does the above code do？ The code is in ‘train50_cls_cce.py’ , about at 235 line. I've been trying to figure out what this means all day...

se154_cls_cce_2_1_best model is missing in the downloaded model file.

In predict script such as pred92cls.py, there are two images generated saying part1 and part2. Are these image created to use cv2 which write 3 bands only and then there is overlapping of damage classes in both part?

Thanks for your great job. May I ask for the latest Trained mode wights link? Thanks.

Hi

I am trying to use your model but the link in README.md to trained model weights https://vdurnov.s3.amazonaws.com/xview2_1st_weights.zip is not working and returns access denied.

response:

<Error>
<Code>AccessDenied</Code>
<Message>Access Denied</Message>
<RequestId>S1VMK858D0ATDQT5</RequestId>
<HostId>fgY53YOWp206iRsWX9lTUqV3sy35GN5gEeRuq2Hgv8/yZZRaEwPepJLPsO4yqcmC9DzqOGneeA0=</HostId>
</Error>

can you update this link?

Thanks to your excellent work!
looks like you use several models separately (ResNet34, SeNet154...). May I ask which one has the best results? i.e., from which model does the 1st place score on xview2 leaderboard resulted from?