tensorflow API检测框和评估

Question

I am trying to use the API detection on tensorflow following this tutorial https://towardsdatascience.com/how-to-train-your-own-object-detector-with-tensorflows-object-detector-api-bec72ecfe1d9 . But there are some details I am not understanding.

First, I do not understand some parameters for the evaluation in the config file. "num readers" and "max_evals" parameters. "Max evals" seems to be the number of evaluation on the dataset but why it is not 1 by default ? Because we need to test only one time the checkpoint (or I am wrong ?). Regarding the training, are the data shuffle automatically ?

Second, I am wondering if we can use tensorboard so as to display the box in an image during the training WITH the API detection. If yes, what are the steps to obtain it ?

Answer 1

• num_readers : I' not entirely sure, but it probably has something to do with how many processes you'll need to read the eval input. You could try and change it if you see that reading the data is really where you're losing time (it's usually not the case though, inference takes more time).

• max_evals : says how many checkpoints you want to evaluate. While training, the trianing job will save checkpoints regularly (so that if you stop the training, you don't lose all the previous work, for instance; or so that you can choose between the networks created respectively by n and m iterations). You'll usually start an evaluation job at the same time, which will regularly evaluate the last created checkpoint, up to max_evals checkpoints. If you don't fill this value, it will evaluate all created checkpoints, thus providing you an evaluation performance regularly during training.

• about shuffling the training data: you need to add shuffle: True to the train input reader (in the config file) for that, but apparently it may not work on the first epoch, so you should probably also make sure your dataset is not sorted beforehand.

• to dispay images with boxes during training: yes you can. You'll need to run the evaluation job for that. You can set how many images will be saved in the config file. To see the evolution of the detections for these images, you'll need shuffle: False in your eval input reader. Otherwise it will show you different images at every step, so you won't really see the progress. Then you just run Tensorboard, the images will be in the "images" tab.