AttributeError: 'numpy.ndarray' object has no attribute 'pop' about ssd_keras HOT 7 CLOSED

I'm having trouble reproducing this issue. Yes, pop() is a method of lists, not arrays, but batch_y is a list. In order to reproduce this I've changed the generator to always pop the last batch item and it works without issues. Are you using one of the Jupyter notebooks to train? Did you make any changes to the code, either in the notebook or elsewhere?

from ssd_keras.

Hi, strange ...
Here's what I changed in the code:

in [3]: weights_path = '/home/ubuntu/ssd_keras/pretrained_weights/vgg-16_ssd-fcn_ILSVRC-CLS-LOC.h5'

in [5]: # changed the beginning of the paths to /home/ubuntu/ssd_keras/... 

That's it...

in ssd_batch_generator.py, Ln 676, the scripts is turning batch_y into an array.

Could I maybe comment-out this whole for-loop which is causing the error and deactivate random cropping?

Thanks!

from ssd_keras.

Just noticed that random cropping is "False" by default. Why is this happening after all?

from ssd_keras.

not sure if thats important, but Im also getting this warnings:

/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/tensor_util.py:509: DeprecationWarning: The binary mode of fromstring is deprecated, as it behaves surprisingly on unicode inputs. Use frombuffer instead
  return np.fromstring(tensor.tensor_content, dtype=dtype).reshape(shape)

from ssd_keras.

Well, for starters, this implementation requires Python 3 as stated in the dependencies. I see that you are using Python 2. It makes little sense to try to spot the problem as long as something as basic as the Python version is incorrect. I would also recommend to install things like TensorFlow, Keras, and other libraries within virtual environments (e.g. Anaconda environments) rather than directly so that you can maintain multiple different environment configurations.

As for your comments above:

in ssd_batch_generator.py, Ln 676, the scripts is turning batch_y into an array.

No, it only turns the i-th element of batch_y into an array:

batch_y[i] = np.array(batch_y[i])

batch_y is still a list.

Just noticed that random cropping is "False" by default. Why is this happening after all?

Because max_crop_and_resize and full_crop_and_resize ultimately use random_crop to perform the actual cropping.

I wouldn't know why this issue would be related to the Python version (after all, a list is a list regardless of the Python version), but since I cannot reproduce the issue in Python 3, I can only suggest that you use a Python 3 environment.

from ssd_keras.

ah damnit. didn't see that.

I commented it out and training is running now. will see if the results are still fine, but afterwards I will follow your advice with the virtualenv. Since no one else seems to have this error, this might be the issue. Thanks anyway, will close this now.

from ssd_keras.

Yeah, commenting it out is actually a good makeshift solution. The only effect will be that some batch items might have no ground truth boxes in them after the random cropping, but this should happen very rarely and even so it likely won't impact the training at all.

from ssd_keras.