DistriMIS Distributing Deep Learning Hyperparameter Tuning for 3D Medical Image Segmentation.
Quick Introduction • Setup • Data Preparation • Parallelism Modes
DistriMIS is the open implementation of the Data and Experiment Parallelism design for Multi-GPU and Supercomputing environments at the Barcelona Supercomputing Center (BSC), focused on Deep Learning for Medical Image Segmentation.
It provides a guide and methodology for deploying 3D U-Net-based Deep Learning neural networks in a distributed environment for scaling. Here we use TensorFlow and Ray frameworks as engine and experiment multi-node distribution. As an example and benchmark, the MSD Brain Tumor Segmentation dataset is used
Abstract
Most research on novel techniques for 3D Medical Image Segmentation (MIS) is currently done using Deep Learning with GPU accelerators. The principal challenge of such technique is that a single input can easily cope computing resources, and require prohibitive amounts of time to be processed. Distribution of deep learning and scalability over computing devices is an actual need for progressing on such research field. Conventional distribution of neural networks consist in "data parallelism", where data is scattered over resources (e.g., GPUs) to parallelize the training of the model. However, "experiment parallelism" is also an option, where different training processes (i.e., on a hyper-parameter search) are parallelized across resources. While the first option is much more common on 3D image segmentation, the second provides a pipeline design with less dependence among parallelized processes, allowing overhead reduction and more potential scalability. In this work we present a design for distributed deep learning training pipelines, focusing on multi-node and multi-GPU environments, where the two different distribution approaches are deployed and benchmarked. We take as proof of concept the 3D U-Net architecture, using the MSD Brain Tumor Segmentation dataset, a state-of-art problem in medical image segmentation with high computing and space requirements. Using the BSC MareNostrum supercomputer as benchmarking environment, we use TensorFlow and Ray as neural network training and experiment distribution platforms. We evaluate the experiment speed-up when parallelizing, showing the potential for scaling out on GPUs and nodes. Also comparing the different parallelism techniques, showing how experiment distribution leverages better such resources through scaling, e.g. by a speed-up factor from x12 to x14 using 32 GPUs. Finally, we provide the implementation of the design open to the community, and the non-trivial steps and methodology for adapting and deploying a MIS case as the here presented.
If you use our code for your research, please cite our preprint:
"TODO: Citar paper"
Just download the code, and execute the scripts inside the directory. You can clone the repository using:
foo@bar:~$ git clone https://github.com/HiEST/DistMIS
foo@bar:~$ cd DistMIS/
foo@bar:~$ python -m pip install -r requirements.txt
The scrips work with python >= 3.7.4 and uses the following packages:
tensorflow==2.3.0
tensorflow-addons==0.13.0
ray==1.4.1
numpy==1.18.4
imageio==2.6.1
matplotlib==3.4.3
nibabel==3.2.1
tqdm==4.46.1
The tfrecord script assumes that the Task01_BrainTumor.tar has been downloaded from its official webpage MSD Challenge and extracted to your local machine. The directory of the original dataset is slightly used in the next sections and it refers to the extracted folder. This folder must contain:
- dataset.json: Information about the dataset.
- imagesTr: Brain images for training.
- labelsTr: Label images for training.
Note that the folder imagesTs is not used because their respective labels are not provided. The training, validation and test sets are split using the data available (imagesTr and labelsTr).
The data is composed by 3D samples of shape (240,240,155,4) for the brain images and (240,240,155,1) for the ground truth segmentation masks. The data format is NIfTI, commonly used in medical imaging. A transposition of the channels is applied and the shape of the samples is reduced to (4,240,240,152) and (1, 240,240,152). Additionally, standardization is applied to the brain images and the 3 clases for the labels are joined to form a binary clasification problem: pixel is tumor (1) or is not (0).
In the following figure the 4 channels from the brain images and the ground truth are shown.
The framework is composed by 4 main scripts: tfrecord, visualize (basic preparation), also data_parallel and exp_parallel (parallelism modes).
The tfrecord script reads the dataset.json file located inside the dataset folder, and creates a tf.data.Dataset from the imagesTr and labelsTr samples. It requires the orignal dataset directory path as argument, e.g. /home/Task01_BrainTumor/. If the target directory is provided, the tf.data.Dataset is serialized and saved in TFRecord format into that directory. Running this script is the first step in order to run the other scripts. Serializing the dataset the memory footprint will be larger but it has good benefits:
- Perform offline preprocessing and data augmentation.
- Optimize data reading, online preprocessing and data augmentation.
It can seem the benefits are poor, but in Deep Learning applications with this type of data, these optimization techniques save hours of training.
foo@bar:~$ python tfrecord.py --help
usage: tfrecord.py [-h] --source-dir SOURCE_DIR [--target-dir TARGET_DIR] [--split SPLIT] [--reshape RESHAPE]
optional arguments:
-h, --help Show this help message and exit
--source-dir SOURCE_DIR Path: Source data directory. The directory must contain the dataset.json file,
and the two folders imagesTr, labelsTr.
--target-dir TARGET_DIR Path: Target data directory. It must exist. It is where the TFRecord data will be
written to.
--split SPLIT Tuple: Ratios into which the dataset will be divided (all sum up to 1). Train,
validation and test set. Default=(0.7, 0.15, 0.15).
--reshape RESHAPE Tuple: Shape of the written data. Default=(240, 240, 152) which is its original
shape. If a different shape is provided, a trilinear resize will be applied to
the data.
Creating the target directory
foo@bar:~$ mkdir dataset
Creating the tfrecord dataset into the created directory dataset.
foo@bar:~$ python tfrecord.py --source-dir /home/Task01_BrainTumor/ --target-source /home/dataset/
Creating a tfrecord dataset with smaller size data, and different split sets.
foo@bar:~$ python tfrecord.py --source-dir /home/Task01_BrainTumor/ --target-source /home/dataset/ --reshape (120, 120, 152) --split (0.8, 0.1, 0.1)
The visualize script is just an auxiliary script for visualizing the data after doing the tfrecord and other possible transformations, e.g. offline data_augmentation. It is also useful for debugging purposes, e.g. testing some transformation or preprocessing functions, before deploying.
foo@bar:~$ python visualize.py --help
usage: visualize.py [-h] --dataset-dir DATASET_DIR [--sample SAMPLE] [--data-shape DATA_SHAPE]
[--no-screen NO_SCREEN]
optional arguments:
-h, --help Show this help message and exit
--dataset-dir DATASET_DIR Path: TFRecord dataset directory.
--sample SAMPLE Int: Sample to visualize. Default=0. It has to be
0 <= sample <= size_dataset.
--data-shape DATA_SHAPE Tuple: Shape of the data in the dataset path provided.
Default=(240, 240, 152) which is the orginal data shape.
--no-screen No X session (graphical Linux desktop) mode. If it used, a
GIF file will be saved in the current directory ('./')
containing the plotted image.
Visualizing the written data in TFRecord format, the default sample is the number 0.
foo@bar:~$ python visualize.py --dataset-dir /home/dataset/
Visualizing the sample number 350, since we are working via ssh with no x session, we enable --no-screeen flag to save a GIF file.
foo@bar:~$ python visualize.py --dataset-dir /home/dataset/ --sample 350 --no-screen
The data_parallel script is the first approach presented in the paper, given a model in tensorflow and a TFRecord dataset it performs data parallelism. Data parallelism consists in, given n GPUs, the model is replicated n times and each replica is sent to a GPU. After that, the data is split into n chunks, i.e. the batch size is diveded by n, these chunks are distributed across the GPUs, where each chunk is assigned to a GPU. If we are training m models and m >= n, then we proceed sequentially for each model. Since we are using more than 1 GPU we are speeding-up the training of each model, and therefore, the m models. Our cluster has 4 GPUs per node, so if the number of GPUs used is less than 4, i.e. we are using only one node, tf.MirroredStrategy is used. For multi-node, i.e. >= 4 GPUs, we use ray.cluster which handles all the comunications between nodes and ray.sgd which is a wrapper around tf.MultiWorkerMirroredStrategy. Both tf.MirroredStrategy and tf.MultiWorkerMirroredStrategy are built-in functions from tensorflow distributed API.
First of all a configuration JSON file is required to execute the script. This configuration file has some parameters which are required shown below:
- num_replicas
(int) Number of GPUs to train each model. The model will be replicated this number of times.
- batch_size_per_replica
(int) Batch size handled by each replica, i.e. GPU. In data parallelism the total_batch_size is batch_size_per_replica * num_replicas.
- num_epochs:
(int) Number of epochs each model will train for.
- debug:
(bool) Mode debug. If true, no tensorboard files will be saved and the training verbosity is set for every step. Otherwise the training verbosity is set for epoch and the tensorboard files will be saved.
Once the configuration file is ready, if we are using a single node we can just call the script.
foo@bar:~$ python data_parallel.py --help
usage: data_parallel.py [-h] --config CONFIG
optional arguments:
-h, --help Show this help message and exit
--config CONFIG Path: Json file configuration
If we are using multi node, we first need to initialize a ray cluster and then execute the script as above. Please refer to the section Multi-node Ray Cluster.
So first, let's define our config JSON file named config.json.
{
"num_replicas": 4,
"batch_size_per_replica": 2,
"num_epochs": 20,
"debug": false
}
Afterwards, we can simply call the script with our config file.
foo@bar:~$ python exp_parallel.py --config ./config.json
The exp_parallel script is the second approach presented in the paper, given a model in tensorflow and a TFRecord dataset it performs experiment parallelism using ray.tune which manages all the low level parallelism implementaion. Experiment parallelism consists in, given n GPUs, m models and m >= n, assigning a model to each GPU available. Hence, we are training n models at the same time, speeding-up the computations of all the m models. As mentioned above our cluster has 4 GPU per node, so if the number of GPUs used is less than 4, i.e. we are using only one node, ray.tune is used. For multi-node, i.e. >= 4 GPUs, we use ray.cluster which handles all the comunications between nodes and ray.tune.
First of all a configuration JSON file is required to execute the script. This configuration file has some parameters which are required shown below:
- batch_size_per_replica
(int) Batch size handled by each replica, i.e. GPU. The total_batch_size is batch_size_per_replica * num_replicas. Since in experiment parallelism we are not applying data parallelism and we train each model with a GPU, num_replicas = 1 and total_batch_size = batch_size_per_replica.
- num_epochs:
(int) Number of epochs each model will train for.
- debug:
(bool) Mode debug. If true, no tensorboard files will be saved and the training verbosity is set for every step. Otherwise the training verbosity is set for epoch and the tensorboard files will be saved.
In order to execute the script first we need to start a ray.cluster with the required resources, i.e. we want to use NUM_GPUS and NUM_CPUS. If we are using a single node then we can type the following command. If we are using multi-node, please ignore this command and refer to the to the section Multi-node ray cluster.
foo@bar:~$ ray start --head --num-cpus=NUM_CPUS --num-gpus=NUM_GPUS
Once the ray cluster is started, we can call our script with our configuration json file.
foo@bar:~$ python exp_parallel.py --help
usage: exp_parallel.py [-h] --config CONFIG
optional arguments:
-h, --help Show this help message and exit
--config CONFIG Path: Json file configuration
It is a well practice to shutdown the ray cluster when the work is done.
foo@bar:~$ ray stop
First we define out config as JSON file named config.json and afterwards we initialize a ray cluster with 20 CPUs and 2 GPUs.
{
"batch_size_per_replica": 2,
"num_epochs": 20,
"debug": false
}
foo@bar:~$ ray start --head --num-cpus=20 --num-gpus=2
Finally we can call the script with our config file.
foo@bar:~$ python exp_parallel.py --config ./config.json
In our case we are using a cluster with 4 GPUs per node, so given n GPUs for n >= 4, we are using multi-node. If you are using multi-node, you need to start a ray cluster in a different way from what is shown in the previous sections. Once the cluster is initialized you can run the script as usual. Here we show an example, a bash script to start a ray cluster using SLURM.
#!/bin/bash
# shellcheck disable=SC2206
# THIS FILE IS GENERATED BY AUTOMATION SCRIPT! PLEASE REFER TO ORIGINAL SCRIPT!
# THIS FILE IS MODIFIED AUTOMATICALLY FROM TEMPLATE AND SHOULD BE RUNNABLE!
#SBATCH --job-name 16_gpus
#SBATCH -D .
#SBATCH --output=./%j.out
#SBATCH --error=./%j.err
### This script works for any number of nodes, Ray will find and manage all resources
#SBATCH --nodes=4
### Give all resources to a single Ray task, ray can manage the resources internally
#SBATCH --ntasks-per-node=1
#SBATCH --cpus-per-task=160
#SBATCH --gres='gpu:4'
#SBATCH --time 48:00:00
#SBATCH --exclusive
# Load modules or your own conda environment here
module purge; module load gcc/8.3.0 cuda/10.2 cudnn/7.6.4 nccl/2.4.8 tensorrt/6.0.1 \
openmpi/4.0.1 atlas/3.10.3 scalapack/2.0.2 fftw/3.3.8 szip/2.1.1 ffmpeg/4.2.1 opencv/4.1.1 \
python/3.7.4_ML ray/1.4.1
# ===== DO NOT CHANGE THINGS HERE UNLESS YOU KNOW WHAT YOU ARE DOING =====
# This script is a modification to the implementation suggest by gregSchwartz18 here:
# https://github.com/ray-project/ray/issues/826#issuecomment-522116599
redis_password=$(uuidgen)
export redis_password
nodes=$(scontrol show hostnames "$SLURM_JOB_NODELIST") # Getting the node names
nodes_array=($nodes)
node_1=${nodes_array[0]}
ip=$(srun --nodes=1 --ntasks=1 -w "$node_1" hostname --ip-address) # making redis-address
# if we detect a space character in the head node IP, we'll
# convert it to an ipv4 address. This step is optional.
if [[ "$ip" == *" "* ]]; then
IFS=' ' read -ra ADDR <<< "$ip"
if [[ ${#ADDR[0]} -gt 16 ]]; then
ip=${ADDR[1]}
else
ip=${ADDR[0]}
fi
echo "IPV6 address detected. We split the IPV4 address as $ip"
fi
port=6379
ip_head=$ip:$port
export ip_head
echo "IP Head: $ip_head"
echo "STARTING HEAD at $node_1"
srun --nodes=1 --ntasks=1 -w "$node_1" \
ray start --head --num-cpus 40 --node-ip-address="$ip" --port=$port --redis-password="$redis_password" --block &
sleep 30
worker_num=$((SLURM_JOB_NUM_NODES - 1)) #number of nodes other than the head node
for ((i = 1; i <= worker_num; i++)); do
node_i=${nodes_array[$i]}
echo "STARTING WORKER $i at $node_i"
srun --nodes=1 --ntasks=1 -w "$node_i" ray start --num-cpus 40 --address "$ip_head" --redis-password="$redis_password" --block &
sleep 5
done
# ===== Call your code below =====
With the first part of the script we have requested the resources via slurm and we have started a ray cluster. The script continues with the following lines where you can call the script you wish.
export PYTHONUNBUFFERED=1
python data_parallelism.py --config ./config.json
Additionally to the Tensorflow based open implementation of the Data and Experiment Parallelism design for Supercomputing environments, focused on Deep Learning for Image Segmentation, we have also provided an extension using Pytorch. In this addendum we provide the scripts and configurations needed to replicate the experimentation that has been done, comparing via N experiments the scalability of data and experiment parallelism with Pytorch.
Similarly to the Tensorflow proposal, we have provided a data parallel script that replicates the model on each GPU and divides the data in chunks, each sent to one device. To do so, we have used Ray.SGD and the corresponding TorchTrainer object. The TorchTrainer object is a wrapper around torch.distributed.launch that automatically replicates the training components across different machines so that training can be executed in parallel. One of the main benefits of Ray.SGD is that we can scale up the number of workers seamlessly across multiple nodes without making any change to the code.
First, a configuration JSON file must be defined to execute the script. This configuration file requires the following parameters:
- lr
(int) Hyperparameter that defines the step size at each iteration while moving toward a minimum of a loss function.
- epochs
(int) Number of epochs each model will train for.
- verbose
(int) Number of epochs each model will train for.
- nodes
(int) Number of nodes.
- gpus
(int) Number of GPUs per node.
- batch_size
(int) Batch size handled by each replica, i.e. GPU.
- num_workers
(int) Number of nodes * number of GPUs per node.
- use_gpu
(bool) Boolean that indicates if the train is going to be done using GPU resources.
Once the configuration file is ready, we can replicate the experimentation with one node as follows. First, we initialize the ray cluster, specifying the number of gpus and cpus per gpu.
foo@bar:~$ ray start --head --num-cpus=20 --num-gpus=2
Under the hood, TorchTrainer will create replicas of your model, each of which is managed by a Ray actor connected to the Ray cluster. Thus, we can now execute the training script, defining with flag g the number of gpus.
foo@bar:~$ python multiexperiment.py -g 2
If we are using multi node, we first need to initialize ray with a more complex bash script. Please refer to the section Multi-node Ray Cluster.
Similarly to the Experiment Parallelism approach presented in Tensorflow, we used the ray.tune library for experiment execution and hyperparameter tuning at any scale. Given an object trainable and the number of samples (experiments) that we want to make, the function tune.run executes the hyperparameter tuning. This function manages the experiment and provides many features such as logging, checkpointing, and early stopping. As happens in Data Parallelism, in a multinode environment we have to deal with more complex ray initializations, so the bash script gets harder to deal with (Multi-node Ray Cluster).
First, a configuration JSON file must be defined to execute the script. This configuration file requires the following parameters:
-
lr
(int) Hyperparameter that defines the step size at each iteration while moving toward a minimum of a loss function.
-
epochs
(int) Number of epochs each model will train for.
-
verbose
(int) Verbose.
-
nodes
(int) Number of nodes.
-
gpus
(int) Number of GPUs per node.
-
batch_size
(int) Batch size handled by each replica, i.e. GPU.
-
num_workers
(int) Number of nodes * number of GPUs per node.
-
use_gpu
(bool) Boolean that indicates if the train is going to be done using GPU resources.
-
multinode
(bool) Boolean that indicates if the train is going to be done in a multinode environment.
-
cpu_per_trial
(int) Number of CPUs assigned to each trial.
-
gpu_per_trial
(int) Number of GPUs assigned to each trial.
-
num_samples
(int) Number of experiments.
In order to execute the experiment parallelism script we first need to start a ray.cluster with the required resources. If we are using a single node then we can type the following command with the given cpus and gpus.
foo@bar:~$ ray start --head --num-cpus=20 --num-gpus=2
Once the ray cluster is started, we can call our script with our configuration json file.
foo@bar:~$ python train.py -c ../config/2gpu/config.json
It is a good practice to shutdown the ray cluster when the work is done.
foo@bar:~$ ray stop