4

I train different models with tensor object detection (TFOD) API and I would like to know how many parameters are trained for a given model.

I run faster RCNN, SSD, RFCN and also with different image resolution, I would like to have a way to know how many parameters are trained. Is there a way to do that?

I have tried answers found here How to count total number of trainable parameters in a tensorflow model? with no luck.

Here is the code I added line 103 of model_main.py:

print("Training {} parameters".format(np.sum([np.prod(v.get_shape().as_list()) for v in tf.trainable_variables()]))

I think the problem is that I do not access the tf.Session() the TFOD is running, hence my code always return 0.0 parameters (although training strats just fine and train, hopefully, millions of parameters), and I don't know how to solve that issue.

CDuvert
  • 367
  • 1
  • 3
  • 14

2 Answers2

1

When using the export_inference_graph.py, the script also analyzes your model, and counts parameters and FLOPS (if possible). If looks like this:

_TFProfRoot (--/# total params)
  FeatureExtractor (--/# params)
  ...
  WeightSharedConvolutionalBoxPredictor (--/# params)
  ...
netanel-sam
  • 1,862
  • 9
  • 14
0

TFOD API used tf.estimator.Estimator to train and evaluate. The Estimator object provided function to get all the variables, Estimator.get_variable_names() (reference).

You can add this line print(estimator.get_variable_names()) after estimator.train_and_evaluate() (here).

You will see all variable names printed after the training is completed. To see the results faster, you can train for just 1 step.

Danny Fang
  • 3,843
  • 1
  • 19
  • 25