RuntimeError during the training example with OIM about open-reid HOT 21 CLOSED

@GBJim, @Cysu I guess that Dataparallel of pytorch doesn't work well with Nvidia-docker. Or maybe it is caused by pytorch pytorch forum

from open-reid.

I wonder if it is fine to run the official mnist example?

from open-reid.

Hi @Cysu
After going through the MNIST example. No errors happen.

I also tried to train the inception net in example: python examples/inception.py -d viper -b 64 -j 2 --loss xentropy --logs-dir logs/inception-viper-xentropy
No errors happen as well.

The interesting thing is that I tried to train ResNet again:
The training process froze like the following:, but no errors.

Files already downloaded and verified
VIPeR dataset loaded
subset | # ids | # images

train | 216 | 432
val | 100 | 200
trainval | 316 | 632
query | 316 | 632
gallery | 316 | 632
Epoch: [0][1/7] Time 160.275 (160.275) Data 0.446 (0.446) Loss 5.375 (5.375) Prec 0.00% (0.00%)
Epoch: [0][2/7] Time 0.563 (80.419) Data 0.001 (0.223) Loss 10.057 (7.716) Prec 0.00% (0.00%)

Is this caused by the GPU resource usage?
Currently, some Caffe process is also using my GPUs.

from open-reid.

I'm not sure if it is caused by some deadlocks between pytorch and caffe, especially when both are using NCCL. You may try to run it again when the caffe experiments are finished.

from open-reid.

Hi @Cysu
Sorry for late response.
I tried it again after my Caffe process is terminated.

The training will be frozen when the -j (worker) argument is set to be bigger than 1.
If the -j argument is set to be 1, I get error: [Errno 111] Connection refused

from open-reid.

@GBJim Could you please change the num_workers in the official mnist example and see if it has the same problem?

from open-reid.

@Cysu:

I tested the MNIST example with 16 workers. Everything is correct

from open-reid.

Sorry but currently I have no idea why it happened. There should be no much difference between our data loader with the mnist ones. I'm not sure if it is related to using root instead of normal user on Linux.

from open-reid.

Thanks @Cysu
I will try to figure it out!

from open-reid.

@GBJim any luck on this?

from open-reid.

Hi @Cysu
I've built a new environment for open RE-ID and cloned the latest commit.
But it seems like the resnet.py and inception.py are removed from the example folder.

Is there new tutorial of how to do a training or testing?
Thanks!

from open-reid.

It seems like the codes are re-organized into oim_loss.py, softmax_loss.py and, triplet_loss.py
Let me check if my these scripts can work

from open-reid.

@Cysu

I tried these commands: python examples/oim_loss.py -d viper or python examples/softmax_loss.py -d viper and python examples/triplet_loss.py -d viper as well.
Tthe following output is prompted and then the process was frozen. I need to use ctrl+z to exit for the process

root@e50f76502ce4:~/open-reid# python examples/oim_loss.py -d viper
Files already downloaded and verified
VIPeR dataset loaded
  subset   | # ids | # images
  ---------------------------
  train    |   216 |      432
  val      |   100 |      200
  trainval |   316 |      632
  query    |   316 |      632
  gallery  |   316 |      632

from open-reid.

@GBJim Oh, I forgot to update the tutorials. Just finished. Please check here.

Does the previous error still occur when -j 1 is use?

from open-reid.

@Cysu

The process is still frozen when I set to single job. (Maybe I should wait for the process for longer time)

I set job to 1 and tried the following combinations:

OIM + ResNet --> Frozen

OIM + Inception --> RuntimeError: The expanded size of the tensor (128) must match the existing size (64) at non-singleton dimension 1. at /root/pytorch/torch/lib/THC/generic/THCTensor.c:323

SOFTMAX + ResNet --> Frozen

SOFTMAX + Inception --> Works Normally

And thank you for updating the documentation!

from open-reid.

That's weird... What's the script for OIM + Inception?

from open-reid.

@Cysu
python examples/oim_loss.py -d viper -a inception -j 1

from open-reid.

I meet the same issue. The problems that @GBJim had happen to me as well. Particularly, this, inception.py has nothing wrong, but resnet.py is Frozen.

from open-reid.

@lzj322 Do you use Nvidia-docker to host the environment?

from open-reid.

@GBJim yes. Would that be a problem? I don't know much about it. I asked the administrator to reset the docker. Now it has normal results. But we don't know why.
I am afraid that this issue could happen someday again.

from open-reid.

@lzj322 Yeah, two programs cannot run on the same device if using NCCL.

from open-reid.