TensorFlow-Slim Multi-GPU training

Question

I'm using TensorFlow-Slim. My aim is to run a given standard script (located in /models/slim/scripts) in multi-GPU mode. I've tested the finetune_resnet_v1_50_on_flowers.sh script (cloned on 12.04.2017). I've just added --num_clones=2 at the end of the training part (inspired by /slim/deployment/model_deploy_test.py and previous StackOverflow answers):

python train_image_classifier.py \
  --train_dir=${TRAIN_DIR} \
  --dataset_name=flowers \
  --dataset_split_name=train \
  --dataset_dir=${DATASET_DIR} \
  --model_name=resnet_v1_50 \
  --checkpoint_path=${PRETRAINED_CHECKPOINT_DIR}/resnet_v1_50.ckpt \
  --checkpoint_exclude_scopes=resnet_v1_50/logits \
  --trainable_scopes=resnet_v1_50/logits \
  --max_number_of_steps=3000 \
  --batch_size=32 \
  --learning_rate=0.01 \
  --save_interval_secs=60 \
  --save_summaries_secs=60 \
  --log_every_n_steps=100 \
  --optimizer=rmsprop \
  --weight_decay=0.00004 \
  --num_clones=2

Code from deployment/model_deploy_test.py:

def testMultiGPU(self):
    deploy_config = model_deploy.DeploymentConfig(num_clones=2)

I've got one warning ('Ignoring device specification'):

I tensorflow/core/common_runtime/gpu/gpu_device.cc:975] Creating TensorFlow device (/gpu:0) -> (device: 0, name: Tesla P100-SXM2-16GB, pci bus id: 0000:85:00.0)
I tensorflow/core/common_runtime/gpu/gpu_device.cc:975] Creating TensorFlow device (/gpu:1) -> (device: 1, name: Tesla P100-SXM2-16GB, pci bus id: 0000:86:00.0)
I tensorflow/core/common_runtime/simple_placer.cc:669] Ignoring device specification /GPU:1 for node 'clone_1/fifo_queue_Dequeue' because the input edge from 'prefetch_queue/fifo_queue' is a reference connection and already has a device field set to /CPU:0
I tensorflow/core/common_runtime/simple_placer.cc:669] Ignoring device specification /GPU:0 for node 'clone_0/fifo_queue_Dequeue' because the input edge from 'prefetch_queue/fifo_queue' is a reference connection and already has a device field set to /CPU:0

The GPUs run normally (memory usage and GPU utilization), but the training is not faster compared with a single GPU training.

This issue might be related to: https://github.com/tensorflow/tensorflow/issues/8061

I would be very pleased to receive your answers, opinions or concrete proposals to this problem.

CUDA version: release 8.0, V8.0.53

TensorFlow installed from binary, tested versions: 1.0.1 and 1.1.0rc

GPU: NVIDIA Tesla P100 (SXM2)

Vallum Vallum · Accepted Answer · 2017-11-21T06:07:52

Please follow this document https://github.com/tensorflow/tensorflow/issues/12689 To make sure that the Variables are stored in CPU, we need to use context manager with slim.arg_scope([slim.model_variable, slim.variable], device='/cpu:0'):

It solved my problem.

TensorFlow-Slim Multi-GPU training

2 Answers