The program always be killed about avsr-tf1 HOT 10 CLOSED

ok, I understand, I will try first segment the pretrain set and than use it. Thank you so much.

from avsr-tf1.

Hallo @wentaoxandry

I remember seeing a similar error whenever my system ran out of RAM.
Could you monitor the RAM usage after launching the experiment ?

Also, try reducing the buffer size of the function that shuffles the batches in memory:
https://github.com/georgesterpu/Sigmedia-AVSR/blob/master/avsr/io_utils.py#L89

from avsr-tf1.

Hallo @georgesterpu ,

Thank you for your help, I have try change buffer_size from 5000 to 100, but it getting better but still be killed, before change it can only do the first 3 batch, now after first 6 batch, it still be killed.

The memory will always increase after each batch,

from avsr-tf1.

What batch size are you using ?
And what is the average length (in seconds) of a sentence in your dataset ?

from avsr-tf1.

the batch size is still (48, 64), but I have also tried (24,32), still be killed.
the average length is 7s.

from avsr-tf1.

What video resolution did you use when writing the video tfrecord files ?

from avsr-tf1.

I used output_resolution = (36,36)

from avsr-tf1.

I believe that, for large scale experiments, we may need to redesign and optimise the input pipeline, but this won't happen until April.

Can you please confirm that your script gets killed when the system runs out of memory ?

from avsr-tf1.

Actually I also use lrs2 database, but I combine pretrain and train set and use the whole set as the training set, for test is still test set.

I can‘t find the specific script , but only know it is in avsr.py, the train function,

in this while loop.
try:
while True:
out = self._train_session.run([self._train_model.model.train_op,
self._train_model.model.batch_loss,
], **self.sess_opts)

                    if self._hparams.profiling is True:
                    self.profiler.add_step(batches, self.run_meta)

                    from tensorflow.python.profiler import option_builder

                    self.profiler.profile_name_scope(options=(option_builder.ProfileOptionBuilder
                                                              .trainable_variables_parameter()))

                    opts = option_builder.ProfileOptionBuilder.time_and_memory()
                    self.profiler.profile_operations(options=opts)

                    opts = (option_builder.ProfileOptionBuilder(
                        option_builder.ProfileOptionBuilder.time_and_memory())
                            .with_step(batches)
                            .with_timeline_output('/tmp/timelines/').build())

                    self.profiler.profile_graph(options=opts)

                sum_loss += out[1]
                print('batch: {}'.format(batches))
                batches += 1

from avsr-tf1.

I see, LRS2 can have some very long sentences. Sigmedia-AVSR doesn't work well with those, as it fills up your memory quite fast.

from avsr-tf1.