Quickstart for Model Developers¶
This quickstart uses the MNIST dataset to demonstrate basic Determined functionality and walks you through the steps needed to install Determined, run training jobs, and visualize experiment results in your browser. Three examples show the scalability and enhanced functionality gained from simple configuration setting changes:
Train on a local, single CPU or GPU.
Run a distributed training job on multiple GPUs.
Use hyperparameter tuning.
An experiment is a training job that consists of one or more variations, or trials, on the same model. By calling Determined API functions from your training loops, you automatically get metric frequency output, plots, and checkpointing for every experiment without writing extra code. You can use the WebUI to view model information, configuration settings, output logs, and training metrics for all of your experiments.
Each of these quickstart examples uses the same model code and example dataset, differing only in their configuration settings. For a list of all experiment configuration settings and more detailed information about each, see Experiment Configuration Reference.
Determined agent and master nodes must be configured with Ubuntu 16.04 or higher, CentOS 7, or macOS 10.13 or higher.
Agent nodes must have Docker installed.
To run jobs with GPUs, install Nvidia drivers, version 384.81 or higher, on each agent. The drivers can be installed as part of a CUDA installation but the rest of the CUDA toolkit is not required.
At least 4-CPU cores, Intel Broadwell or later. The master node does not require GPUs.
200GB of free disk space.
At least 2-CPU cores, Intel Broadwell or later.
If you are using GPUs, Nvidia GPUs with compute capability 3.7 or greater are required: K80, P100, V100, A100, GTX 1080, GTX 1080 Ti, TITAN, or TITAN XP.
50GB of free disk space.
Quickstart Training Examples¶
Download and extract the files used in this quickstart to a local directory:
Extract the configuration and model files:
tar xzvf mnist_pytorch.tgz
You should see the following files in the
adaptive.yaml const.yaml data.py distributed.yaml layers.py model_def.py README.md
Each of the YAML-formatted configuration files corresponds to one of the following example experiments:
Train a single model on a single GPU/CPU, with constant hyperparameter values.
Train a single model using multiple, distributed GPUs.
Perform a hyperparameter search using the Determined adaptive hyperparameter tuning algorithm.
Model and Pipeline Definition¶
Although the Python model and data pipeline definition files are not explained in this quickstart, you might want to review them to see how to call the Determined API from your code:
Model data loading and preparation code.
Convolutional layers used by the model.
Model definition and training/validation loops.
After gaining basic familiarity with Determined tools and operations, you can replacing these files with your model data and code, and setting configuration parameters for the kind of experiments you want to run.
Run a Local Single CPU/GPU Training Job¶
This exercise trains a single model for a fixed number of batches, using constant values for all hyperparameters on a single slot. A slot is a CPU or GPU computing device, which the master schedules to run.
To install the Determined library and start a cluster locally, enter:
pip install determined det deploy local cluster-up
If your local machine does not have a supported Nvidia GPU, include the
pip install determined det deploy local cluster-up --no-gpu
mnist_pytorchdirectory, create an experiment specifying the
det experiment create const.yaml .
The last dot (.) argument uploads all of the files in the current directory as the context directory for your model. Determined copies the model context directory contents to the trial container working directory.
You should receive confirmation that the experiment is created:
Preparing files (.../mnist_pytorch) to send to master... 8.6KB and 7 files Created experiment 1
To automatically stream log messages for the first trial in an experiment to
stdout, specifying the configuration file and context directory, enter:
det e create const.yaml . -f
-foption is the short form of
To view the uncategorized experiments, enter your username and password. If you deployed locally, accept the default determined username with no password and click Sign In:
Enter the cluster address in the browser address bar to view experiment progress in the WebUI. If you installed locally using the
det deploy localcommand, the URL is
http://localhost:8080/. Accept the default
determinedusername and click Sign In. No password is required.
The figure shows two experiments. Experiment 3 has COMPLETED and experiment 4 is still ACTIVE. Your experiment number and status can differ depending on how many times you run the examples.
While an experiment is in the ACTIVE, training state, click the experiment name to see the Metrics graph update for your currently defined metrics:
In this example, the graph displays the loss.
After the experiment completes, click the experiment name to view the trial page:
With this fundamental understanding of Determined, you are ready to scale to distributed training in the next example.
Run a Remote Distributed Training Job¶
In the distributed training example, a Determined cluster comprises a master and one or more agents. The master provides centralized management of the agent resources.
This example requires a Determined cluster with multiple GPUs and, while it does not fully demonstrate the benefits of distributed training, it does show how to work with added hardware resources.
distributed.yaml configuration file for this example is the same as the
in the previous example, except that a
resources.slots_per_trial field is defined and set to a
resources: slots_per_trial: 8
This is the number of available GPU resources. The
slots_per_trial value must be divisible by
the number of GPUs per machine. You can change the value to match your hardware configuration.
To connect to a Determined master running on a remote instance, set the remote IP address and port number in the
Create and run the experiment:
det experiment create distributed.yaml .
You can also use the
-moption to specify a remote master IP address:
det -m http://<ipAddress>:8080 experiment create distributed.yaml .
To view the WebUI dashboard, enter the cluster address in your browser address bar, accept the default
determinedusername, and click Sign In. A password is not required.
Click the Experiment name to view the experiment’s trial display. The loss curve is similar to the single-GPU experiment in the previous exercise but the time to complete the trial is reduced by about half.
Run a Hyperparameter Tuning Job¶
This example demonstrates hyperparameter search. The example uses the
configuration file, which is similar to the
const.yaml file in the first example but includes
additional hyperparameter settings:
hyperparameters: global_batch_size: 64 learning_rate: type: double minval: .0001 maxval: 1.0 n_filters1: type: int minval: 8 maxval: 64 n_filters2: type: int minval: 8 maxval: 72 dropout1: type: double minval: .2 maxval: .8 dropout2: type: double minval: .2 maxval: .8
Hyperparameter searches involve multiple trials or model variations per experiment. The configuration settings tell the search algorithm the ranges to explore for each hyperparameter.
adaptive_asha search method and maximum number of trials, max_trials` are also specified:
searcher: name: adaptive_asha metric: validation_loss smaller_is_better: true max_trials: 16 max_length: batches: 937
This example uses a fixed batch size and searches on dropout size, filters, and learning rate. The
max_trials setting of
16 indicates how many model configurations to explore.
Create and run the experiment:
det experiment create adaptive.yaml .
To view the WebUI dashboard, enter your cluster address in the browser address bar, accept the default determined username, and click Sign In. No password is required.
The experiment can take some time to complete. You can monitor progress in the WebUI Dashboard by clicking the Experiment name. Notice that more trials have started:
Determined runs the number of
max_trialstrials and automatically starts new trials as resources become available. For 16 trials, it should take about 10 minutes to train with at least one trial performing at about 98 percent validation accuracy. The hyperparameter search halts poorly performing trials.
For detailed information on administrator tasks and how to install Determined on different platforms, see Basic Setup.
In the Example Solutions documentation, you can find machine learning models that have been ported to the Determined APIs. Each example includes a model definition and one or more experiment configuration files, and instructions on how to run the example.
To learn more about the hyperparameter search algorithm, see the Hyperparameter Tuning section.
For faster, less structured ways to run a Determined cluster without writing a model, see: