-1

I am trying to convert a frozen graph of a resnet-50 model to onnx model and then to tensorRT. I want to make sure the floating point precision at each conversion.

Shruthi Manian
  • 1
  • 1
  • Please edit the question to limit it to a specific problem with enough detail to identify an adequate answer. – Community Sep 11 '21 at 19:11

1 Answers1

1

Suppose you are passing an image x to the model like model(x) or model.forward(x), then you can check the datatype of x. You can use dtype property to get the type of a tensorflow variable.

> x = tf.Variable(tf.random_normal([256, 100]))
> x.dtype

<dtype: 'float32_ref'>

You can use as_numpy_dtype property of dtype to convert from tf.dtype to numpy dtype.

> x = tf.Variable(tf.random_normal([256, 100]))
> x.dtype.as_numpy_dtype
<class 'numpy.float32'>

For Onnx, you can import the onnx/graphsurgeon library to perform various operations. But the easiest way would be to use netron.

  1. pip install netron
  2. open https://localhost:8080
  3. Click the input node
  4. The information panel on the right would give a lot of information including data-type.
romil
  • 68
  • 1
  • 8