Tensor shape error while training starts about keras-maskrcnn HOT 27 OPEN

Sorry for the delay, we've been quite busy lately. Is this still an issue for you? If so, could you share a single image and the corresponding COCO style dataset file? I could maybe have a look sometime to see what the problem is.

from keras-maskrcnn.

yes I am still facing the issue. I am having around 40k training images of size 500x500. Dataset is in coco format. Do i have to resize the images to any specific size ? what is the acceptable input size to the network ?

from keras-maskrcnn.

The network should accept any input size. There are some lower limits (I believe 224x224) and some upper limit but that is restricted by your hardware. I meant if you can share a minimal example of a dataset that produces this error, it gives me something to experiment with.

from keras-maskrcnn.

Thanks @hgaiser . My dataset is in coco format. as a subset can i provide you the image and corresponding masks ?

from keras-maskrcnn.

Hi @hgaiser , would you like to take a look at the dataset to figure out whats happening with the dataset. I know its something with the dataset or generator. could you guide me what specific should i look for in data or any debugging steps. i have used debug.py to view the annotations and anchors but it seems fine.

Any help is appreciated.

Thanks in advance

from keras-maskrcnn.

Hey @amardeepjaiman . Yes if you can share a minimal dataset somewhere with this issue, then I can have a look.

from keras-maskrcnn.

Hi @hgaiser , I have uploaded a sample dataset on following location. Please have a look.

https://www.dropbox.com/s/2o4364568bf7vj2/sample_dataset.zip?dl=0

from keras-maskrcnn.

Do you happen to have the annotations as well? These are just the images and the masks, correct?

from keras-maskrcnn.

yes, but i have annotations for full dataset in coco format. do you want me provide annotaion for sample too. I can generate and provide to you.

from keras-maskrcnn.

Yes, the smaller the dataset the better.

from keras-maskrcnn.

ok. Please access this zip file for small dataset with coco annotations.

https://www.dropbox.com/s/ijzz9ippm8mjvvy/sample_dataset_with_anno%28coco%29.zip?dl=0

from keras-maskrcnn.

hey @hgaiser , did you get a chance to look at the dataset ?

from keras-maskrcnn.

I just trained a while on the training set (had to move some files around to different dirs to get it to accept the dataset). What version of tensorflow are you running?

from keras-maskrcnn.

I have tried running it on different version of tensorflow like 1.10 and currently 1.13. Any pointers what could be wrong with the dataset ? or whats your observation ?

from keras-maskrcnn.

@hgaiser , any update on this issue ?

from keras-maskrcnn.

I don't know, as I said the dataset seems to be training fine for me. Would you like me to re-upload your modified dataset and give you the command I used for training?

from keras-maskrcnn.

that would be great i guess. could you please highlight the modification you did. and also the training command.

also do you have any DockerFile of this environment you are using. So that i can replicate this here?

from keras-maskrcnn.

Hi @hgaiser ,

any update ?Did you upload the dataset somewhere ?

Thanks,
Amardeep

from keras-maskrcnn.

python keras_maskrcnn/bin/train.py --weights snapshots/resnet50_coco_v0.2.0.h5 --no-evaluation coco ~/crash_dataset/train

The modifications I made were just moving some files around so that train.py knew how to parse it as a generator.

crash_dataset.zip

from keras-maskrcnn.

Thanks @hgaiser . I will try this. is there any specific tensorflow version installed on your environment ?

from keras-maskrcnn.

from keras-maskrcnn.

Hi @hgaiser , I have narrowed down the problem now. If i use the --batch-size argument then only my problem occurs. If i leave to default i.e. 1, then problem dont occur.

python keras_maskrcnn/bin/train.py --weights snapshots/resnet50_coco_v0.2.0.h5 --batch-size=2 --no-evaluation coco ~/crash_dataset/train

its giving same error. Could you please check ? However I have started the training without giving batch size.

from keras-maskrcnn.

from keras-maskrcnn.

I just ran the same command, but with --batch-size 2:

python keras_maskrcnn/bin/train.py --weights snapshots/resnet50_coco_v0.2.0.h5 --no-evaluation --batch-size 2 coco ~/crash_dataset/train

It didn't crash for me. Is this also what you ran?

from keras-maskrcnn.

yes. Giving batch size crashing for me.

from keras-maskrcnn.

Then I've run out of ideas for things to test =\

from keras-maskrcnn.

Hi @hgaiser , I have narrowed down the problem now. If i use the --batch-size argument then only my problem occurs. If i leave to default i.e. 1, then problem dont occur.

python keras_maskrcnn/bin/train.py --weights snapshots/resnet50_coco_v0.2.0.h5 --batch-size=2 --no-evaluation coco ~/crash_dataset/train

its giving same error. Could you please check ? However I have started the training without giving batch size.

@amardeepjaiman not sure if this might be the problem but please make sure you are not using "=" sign after "--batch-size" and rather a space before "2".

from keras-maskrcnn.