Example of inference of model about gazeml HOT 6 CLOSED

Hi:
I have created a DMS demo that use GazeML, here is the code that show how GazeML was used, https://github.com/parai/dms/blob/master/models/gaze.py.

from gazeml.

Hi many apologies for taking so long to reply to this...

To be honest it's because I don't have a straight-forward solution for a simple way to do inference. What I recommend is to copy the model code to a simpler code-base for yourself, then load the checkpoints.

I use tf.Saver to load model parameters. See https://github.com/swook/GazeML/blob/master/src/core/checkpoint_manager.py#L47

The restoring occurs based on the tf.variable_scope definitions and thus naming of various tensors. So I recommend keeping the neural net spec. as it is.

from gazeml.

Of course, in tensorflow, it could be:

with tf.Session(...) as sess:
    tf.saved_model.loader.load(sess, ...)
    sess.run('myOutput:0',
               feed_dict={'myInput:0': ...

Can i first decode image into tensor:

img_tensor = tf.image.decode_image(img_raw)

Then feed the tensor to the model?

from gazeml.

I found the code in inverence_generator.py

data_source = next(iter(self._train_data.values()))
while True:
      fetches = dict(self.output_tensors['train'], **data_source.output_tensors)
      start_time = time.time()
      outputs = self._tensorflow_session.run(
      fetches=fetches,
       feed_dict={
            self.is_training: False,
            self.use_batch_statistics: True,
        },
       )
     outputs['inference_time'] = 1e3*(time.time() - start_time)
      yield outputs

This fetch from datasource, right? How about for a single image? could we inference with out put image into datasource?

Thank you

from gazeml.

According to your paper and output or training, there are 2 models, first is hourglass, the other is densenet. Now I already trained the models. How to load and run?

Thank you for your help.

from gazeml.

Hi! I am also facing with same problem. How do you fix it?

from gazeml.