Soda Bottle Identification Challenge

If you already have an account with DeepCognition. There is no separate registration required to participate.

Register to participate is now only visible to users who does not have account with Deep Cognition or not logged in.

error message when i train:

Training process exited abnormly.

what happen?

i also have the same problem even with shallow net

Hi, @matt

  1. Fix path to images files in the train.csv file.
    Image Label
    ./0.jpg P.Zero

Image Label
./P.Zero/0.jpg P.Zero

  1. Copy whole folder to the folder with public datasets

  2. Add meta.json file, see mnist’s file, there are specification of names inputs and outputs

studio will be able to load your dataset, but currently I am not guarantying it will connect with the model - I am still struggling…

I believe when I am logged-in I should not get Register to Participate. if you check the snapshot I attached, that clearly shows I am already logged in.

Please let me know once it is fixed.

Yes, we have fixed it. Register to participate is not visible now If you are logged in.

I am facing problem in training my model. I used the autoML feature to build my model. But when I start training, the system free memory goes from 14GB to 0 and my laptop hangs. I tried with the public datasets and everything is working fine. What is the matter with this ? Please help.

If you are using “Full Dataset” in memory option under data tab, you can try batch by batch setting.

I have tried the same earlier. But it did not work. Btw the data for this challenge is very less and my ram is 16 GB. Also, a bigger dataset like that of MNIST is working fine. What should I do now?

In comparison to MNIST, each image is 1175 times bigger (640 * 480 * 3/(28*28)). So even being smaller dataset, it needs around 100 times more memory than MNIST to store all the images in RAM.

If you have tried batch by batch setting, then try reducing the batch size to small value. Try with value 1 first and then increase it from there.

Thanks for the clarification. I resized all the images and now it’s working fine. I am getting nearly 85% accuracy.

I am not able to start training. I press ’ start training ’ button. Then it shows ’ connected to compute server ’ and just stays there for all types of instances.

@ManikSoni, we have checked the issue. It seems browser is blocking non-HTTPS connections to compute server.

If you are using Chrome, there is option on the right side in address bar to allow unsafe connections. Please click on that and choose “Load unsafe scripts” as a workaround for now.


1 Like

yeah after I allowed that, I was able to start training but then I got this error.

simple_bind error. Arguments: Input_0: (32, 3, 480, 640) [12:48:20] src/operator/./cudnn_convolution-inl.h:556: Check failed: e == CUDNN_STATUS_SUCCESS (9 vs. 0) cuDNN: CUDNN_STATUS_NOT_SUPPORTED Stack trace returned 10 entries: [bt] (0) /usr/local/lib/python3.5/dist-packages/mxnet/ [0x7f1464dcac5c] [bt] (1) /usr/local/lib/python3.5/dist-packages/mxnet/ [0x7f146816307d] [bt] (2) /usr/local/lib/python3.5/dist-packages/mxnet/ [0x7f1468159add] [bt] (3) /usr/local/lib/python3.5/dist-packages/mxnet/ [0x7f1466ebe5b4] [bt] (4) /usr/local/lib/python3.5/dist-packages/mxnet/ [0x7f1466e35668] [bt] (5) /usr/local/lib/python3.5/dist-packages/mxnet/ [0x7f1464e98a25] [bt] (6) /usr/local/lib/python3.5/dist-packages/mxnet/ [0x7f1466c13bd2] [bt] (7) /usr/local/lib/python3.5/dist-packages/mxnet/ [0x7f1466c32a1b] [bt] (8) /usr/local/lib/python3.5/dist-packages/mxnet/ [0x7f1466c36b06] [bt] (9) /usr/local/lib/python3.5/dist-packages/mxnet/ [0x7f1466c37174]

This error could be result of too big network to fit on GPU/System. Can you try reducing batch size and/or reducing the image size ?

Can we participate without actually using DeepCognition desktop app, or its cloud services ?

Model needs to be developed using Deep Learning Studio (DLS). Submission process will accept model trained only with DLS.

1 Like

Did you resize the images within deep cognition studio? If yes, could you tell me how to do it within DLS?

Hi, @snehalvartak

First option:
Go to Dataset tab -> click on Images label button, it will show options on the right panel -> choose Resize option -> write target size, click enter. DLS will process images with OpenCV’s resize method and will update Input size in the model viewport

Second option:
Use MaxPooling2D or AveragePooling2D layer, that can be viewed (roughly) as the integer scaling of the image

1 Like

Could you, please, provide intermediate version of DLS with bug-fixes for pre-trained nets? So state-of-the-art nets can be used for fine tuning. With ‘include top layers’ option equal to false, fully non-trainable (Trainable=0) or fully trainable (Trainable=100). Include-top is not critical, as for me, because can be constructed manually, intermediate states for option ‘Trainable’ for me are hard to interpret what to learn and what not to learn.