EfficientDet with TensorFlow and DALI

This is a modified version of original EfficientDet implementation google/automl. It has been changed to allow to use DALI data preprocessing.

To use DALI pipeline for data loading and preprocessing --pipeline dali_gpu or --pipeline dali_cpu, for original pipeline --pipeline tensorflow.

Preparing data files from COCO dataset

For creating TFrecords files:

python3 ./dataset/create_coco_tfrecord.py \
        --image_dir ./coco/train2017 \
        --object_annotations_file ./coco/annotations/instances_train2017.json \
        --output_file_prefix ./tfrecords/train

For creating TFrecord index files (necessary only for DALI pipelines):

python3 ./dataset/create_tfrecord_indexes.py \
        --tfrecord_file_pattern './tfrecords/*.tfrecord' \
        --tfrecord2idx_script ../../../../../tools/tfrecord2idx \

Training in Keras Fit/Compile mode

For the full training on all available GPUs with DALI gpu pipeline:

python3 train.py \
        --multi_gpu \
        --pipeline dali_gpu \
        --epochs 50 \
        --input_type tfrecord \
        --train_file_pattern './tfrecords/train*.tfrecord' \
        --batch_size 16 \
        --train_steps 2000 \
        --output_filename final_weights.h5

Evaluation in Keras Fit/Compile mode

For the evaluation with DALI gpu pipeline:

python3 eval.py \
        --pipeline dali_gpu \
        --input_type tfrecord \
        --eval_file_pattern './tfrecords/eval*.tfrecord' \
        --eval_steps 5000 \
        --weights final_weights.h5

Usage

usage: train.py [-h] [--initial_epoch INITIAL_EPOCH] [--epochs EPOCHS]
                --input_type {tfrecord,coco} [--images_path IMAGES_PATH]
                [--annotations_path ANNOTATIONS_PATH]
                [--train_file_pattern TRAIN_FILE_PATTERN]
                [--batch_size BATCH_SIZE] [--train_steps TRAIN_STEPS]
                [--eval_file_pattern EVAL_FILE_PATTERN]
                [--eval_steps EVAL_STEPS] [--eval_freq EVAL_FREQ]
                [--eval_during_training] [--eval_after_training]
                --pipeline_type {synthetic,tensorflow,dali_cpu,dali_gpu}
                [--multi_gpu [MULTI_GPU [MULTI_GPU ...]]] [--seed SEED]
                [--hparams HPARAMS] [--model_name MODEL_NAME]
                [--output_filename OUTPUT_FILENAME]
                [--start_weights START_WEIGHTS] [--log_dir LOG_DIR]
                [--ckpt_dir CKPT_DIR]

optional arguments:
  -h, --help            show this help message and exit
  --initial_epoch INITIAL_EPOCH
                        Epoch from which to start training.
  --epochs EPOCHS       Epoch on which training should finish.
  --input_type {tfrecord,coco}
                        Input type.
  --images_path IMAGES_PATH
                        Path to COCO images.
  --annotations_path ANNOTATIONS_PATH
                        Path to COCO annotations.
  --train_file_pattern TRAIN_FILE_PATTERN
                        TFrecord files glob pattern for files with training data.
  --batch_size BATCH_SIZE
  --train_steps TRAIN_STEPS
                        Number of steps (iterations) in each epoch.
  --eval_file_pattern EVAL_FILE_PATTERN
                        TFrecord files glob pattern for files with evaluation data,
                        defaults to `train_file_pattern` if not given.
  --eval_steps EVAL_STEPS
                        Number of examples to evaluate during each evaluation.
  --eval_freq EVAL_FREQ
                        During training evaluation frequency.
  --eval_during_training
                        Whether to run evaluation every `eval_freq` epochs.
  --eval_after_training
                        Whether to run evaluation after finished training.
  --pipeline_type {synthetic,tensorflow,dali_cpu,dali_gpu}
                        Pipeline type used while loading and preprocessing data.
                        One of: tensorflow – pipeline used in original
                        EfficientDet implementation on
                        https://github.com/google/automl/tree/master/efficientdet
                        synthetic – like `tensorflow` pipeline type but repeats
                        one batch endlessly dali_gpu – pipeline which uses
                        Nvidia Data Loading Library (DALI) to run part of data
                        preprocessing on GPUs to improve efficiency
                        dali_cpu – like `dali_gpu` pipeline type but restricted
                        to run only on CPU
  --multi_gpu [MULTI_GPU [MULTI_GPU ...]]
                        List of GPUs to use, if empty defaults to all visible GPUs.
  --seed SEED
  --hparams HPARAMS     String or filename with parameters.
  --model_name MODEL_NAME
  --output_filename OUTPUT_FILENAME
                        Filename for final weights to save.
  --start_weights START_WEIGHTS
  --log_dir LOG_DIR     Directory for tensorboard logs.
  --ckpt_dir CKPT_DIR   Directory for saving weights each step.

usage: eval.py [-h] --input_type {tfrecord,coco} [--images_path IMAGES_PATH]
               [--annotations_path ANNOTATIONS_PATH]
               [--eval_file_pattern EVAL_FILE_PATTERN]
               [--eval_steps EVAL_STEPS]
               --pipeline_type {synthetic,tensorflow,dali_cpu,dali_gpu}
               [--weights WEIGHTS] [--model_name MODEL_NAME] [--hparams HPARAMS]

optional arguments:
  -h, --help            show this help message and exit
  --input_type {tfrecord,coco}
                        Input type.
  --images_path IMAGES_PATH
                        Path to COCO images.
  --annotations_path ANNOTATIONS_PATH
                        Path to COCO annotations.
  --eval_file_pattern EVAL_FILE_PATTERN
                        TFrecord files glob pattern for files with evaluation data.
  --eval_steps EVAL_STEPS
                        Number of examples to evaluate.
  --pipeline_type {synthetic,tensorflow,dali_cpu,dali_gpu}
                        Pipeline type used while loading and preprocessing data.
                        One of: tensorflow – pipeline used in original
                        EfficientDet implementation on
                        https://github.com/google/automl/tree/master/efficientdet
                        synthetic – like `tensorflow` pipeline type but repeats
                        one batch endlessly dali_gpu – pipeline which uses
                        Nvidia Data Loading Library (DALI) to run part of data
                        preprocessing on GPUs to improve efficiency dali_cpu –
                        like `dali_gpu` pipeline type but restricted to run
                        only on CPU
  --weights WEIGHTS     Name of the file with model weights.
  --model_name MODEL_NAME
  --hparams HPARAMS     String or filename with parameters.

Requirements

pip install -r requirements.txt