Skip to content

Latest commit

 

History

History
109 lines (85 loc) · 4.22 KB

getting_started.md

File metadata and controls

109 lines (85 loc) · 4.22 KB
Raw

Using DeepLab2

In the following, we provide instructions on how to run DeepLab2.

Prerequisites

We assume DeepLab2 is successfully installed and the necessary datasets are configured.

Running DeepLab2

DeepLab2 contains several implementations of state-of-the-art methods. In the following, we discuss all steps from choosing a model, setting up the configuration to training and evaluating it.

Choosing a model

For this tutorial, we use Panoptic-DeepLab, however, running any other model follows the same steps. For each network architecture, we provide a guide that contains example configurations and (pretrained) checkpoints. You can find all guides here. For now, please checkout Panoptic-DeepLab.

We will use the Resnet50 model as an example for this guide. If you just want to run the network without training, please download the corresponding checkpoint trained by us. If you would like to train the network, please download the corresponding ImageNet pretrained checkpoint from here.

Defining a configuration

When you want to train or evaluate a network, DeepLab2 requires a corresponding configuration. This configuration contains information about the network architecture as well as all sorts of hyper-parameters. Fortunately, for almost all settings we provide default values and example configurations. The configuration of Panoptic-DeepLab with ResNet50 for the Cityscapes dataset can be found here.

Using our default parameters there are only a few things that needs to be defined:

  1. The name of the experiment experiment_name. The experiment name is used as a folder name to store all experiment related files in.
  2. The initial checkpoint initial_checkpoint, which can be an empty string for none or the path to a checkpoint (e.g., pretrained on ImageNet or fully trained by us.)
  3. The training dataset train_dataset_options.file_pattern, which should point to the TFRecords of the Cityscapes train set.
  4. The evaluation dataset eval_dataset_options.file_pattern, which should point to the TFRecords of the Cityscapes val set.
  5. If the custom CUDA kernel is successfully compiled, we recommend to set merge_semantic_and_instance_with_tf_op to true.

For a detailed explanation of all the parameters, we refer to the documented definitions of the proto files. A good starting place is the config.proto. The ExperimentOptions are a collection of all necessary configurations ranging from the model architecture to the training settings.

Training and Evaluating

We currently support four different modes to run DeepLab2:

  • Training: This will only train the network based on the provided configuration.
  • Evaluation: This will only evaluate the network based on the provided configuration.
  • Continuous Evaluation: This mode will constantly monitor a directory for newly saved checkpoints that will be evaluated until a timeout. This mode is useful when runing separate jobs for training and evaluation (e.g., a multi GPU job for training, and a single GPU job for evaluating).
  • Interleaved Training and Evaluation: In this mode, training and evaluation will run interleaved. This is not supported for multi GPU jobs.

Putting everything together

To run DeepLab2 on GPUs, the following command should be used:

python trainer/train.py \
    --config_file=${CONFIG_FILE} \
    --mode={train | eval | train_and_eval | continuous_eval} \
    --model_dir=${BASE_MODEL_DIRECTORY} \
    --num_gpus=${NUM_GPUS}

You can also launch DeepLab2 on TPUS. For this, the TPU address needs to be specified:

python trainer/train.py \
    --config_file=${CONFIG_FILE} \
    --mode={train | eval | train_and_eval | continuous_eval} \
    --model_dir=${BASE_MODEL_DIRECTORY} \
    --master=${TPU_ADDRESS}

For a detailed explanation of each option run:

python trainer/train.py --help