spacenetchallenge / roaddetector Goto Github PK

Hi,

After running pre-trained model for albu solution (batch size 1 and only one image(#230) is placed as data set), the docker memory rapidly grows and ultimately system is hanged. After some time, the process is auto killed by OS.
Please advise what is the resolution of this issue. I have my Ubuntu machine 18.04 with 64G physical memory.

Thanks,
Ajmal

Hi,

I have been trying to download the training data for the SpaceNet Road Dataset without success.

I tried commands like aws s3 cp s3://spacenet-dataset/SpaceNet_Roads_Competition/SpaceNet_Roads_Sample.tar.gz .
in my Ubuntu 18.04, with cli but received error every time I did.

I want to use this network for road detection on 512*512 images. I have binary images as ground truth, so I skipped create_mask, and directly copy training images & labels into /wdata/train

I can train the model successfully, but when it do prediction, it reports several errors after I change line 63 in train_eval.py to rows, cols = 556, 556 (512 + 44 padding). I kept other parts intact.

Traceback (most recent call last):
File "train_eval.py", line 77, in
eval_roads()
File "train_eval.py", line 72, in eval_roads
keval.predict(fold, e)
File "/opt/app/src/pytorch_utils/eval.py", line 92, in predict
predicted = predict(model, samples, flips=self.flips)
File "/opt/app/src/pytorch_utils/eval.py", line 41, in predict
pred1 = F.sigmoid(model(batch))
File "/miniconda/lib/python3.6/site-packages/torch/nn/modules/module.py", line 325, in call
result = self.forward(*input, **kwargs)
File "/opt/app/src/pytorch_zoo/abstract_model.py", line 103, in forward
x = bottleneck(x, enc_results[rev_idx])
File "/miniconda/lib/python3.6/site-packages/torch/nn/modules/module.py", line 325, in call
result = self.forward(*input, **kwargs)
File "/opt/app/src/pytorch_zoo/abstract_model.py", line 24, in forward
x = torch.cat([dec, enc], dim=1)
RuntimeError: inconsistent tensor sizes at /pytorch/torch/lib/THC/generic/THCTensorMath.cu:157

If I instead use rows, cols = 512, 512, It can run, but the outputs have issues near the boundary. Please let me know how to fix it, thanks!

Hi,

Where can I find pre-trained nn models? they don't exist at https://s3.us-east-2.amazonaws.com/vdurnov/nn_models.zip

Thanks,
Ajmal

How to download s3://spacenet-dataset/SpaceNet_Roads_Competition/Pretrained_Models/01-Albu/weights/

It required an account with credit card details. So I need to pay to download this? Please clarify and let me know how can I get it.

Thanks,
Muhammad Ajmal Siddiqui