FPN | Feature Pyramid Networks for Object Detection | Computer Vision library

I believe that model.fit_generator is deprecated, in TensorFlow 2.2 and higher you can just use model.fit because this now supports generators.

https://www.tensorflow.org/api_docs/python/tf/keras/Model#fit_generator

No pandas needed.

Looking at the conversion script seems like dets is a combo of boxes plus score

It's a bug of the library as reported at this link. COCO metrics don't have this problem, so use it to evaluate your model. The problem is not fixed yet. If you want to follow updates made to the code(they work fine) please follow the previous link and also this link

With some changes to the train_loop function in model_lib.py, you can alternate between training and evaluation in the same application. See the example below.

From what I understand, the Tensorflow Object Detection API is developed with a focus on distributed learning and if you were using mulitple GPUs/TPUs then you could have some devices doing training and other devices doing evaluation. So I suspect the way model_lib.py is implemented currently does not fully support doing training and evaluation on the same device.

I'm not certain the root cause of the error you are seeing, typically I have seen Tensorflow throw OOM errors when there is a memory issue. It may be that how Tensorflow is using CUDA does not support two applications using the same device.

Regarding your second question, I followed the advice here on the same thread and this worked for me. Duplicating the code in the third code block below. Initially, this did not appear to work for me because I naively updated the file in the Object Detection repository I created, but your application may be using the Object Detection API that is installed in your site-libs, so I would recommend confirming that the file you are changing is the same one being loaded in your import statements.

--

This is outside of the training loop

If tf.session can understand your BalloonConfig, it suffices to just add the property to your config:

Ah, the age old question, "does size matter"? Yes, it does. A model that is trained on 640x640 images expects to receive only 640x640 images. If you have an image of 1915x1080, it is your job to reduce that to 640x640 for classification. Often, the biggest part of a classification app is munging varied input into the standard form the model expects.

YOU will have to decide how to shrink the image. Do you need to maintain aspect ratio? Do you leave black bars top and bottom?

And do not assume that the 1024 model necessarily does a better job than the 640 model. It is very often the case that more detail simply confuses the classification.

The problem in your case is that your input tensor shape is of the form (1,1219,1920,4), more precisely the 4 is problematic.

The first element, 1, stands for the batch size (added in input_tensor[tf.newaxis, ...]).

You get that part right, but where you actually read the images, the problem takes place, because there are 4 channels (assuming you read RGB-A?) not 3 (typical RGB) or 1 (grayscale).

I recommend that you check your images and to force the conversion to RGB, i.e. Image.open(path).convert('RGB')

These problems get rewritten into Boolean logic and fp.mul is simply hard to analyze, just like multiplication in many other theories. You're basically at the mercy of the SAT solver, which uses various kinds of heuristics to solve many problems quickly, but on any given problem, it's not easy to predict it's behavior. It's very common for tiny changes to the input to result in huge changes in solving time. You can run Z3 with -v:10 to see which tactics get applied and to see some of the statistics from the SAT solver.