Skip to content

Latest commit

 

History

History
129 lines (106 loc) · 6.2 KB

GETTING_STARTED.md

File metadata and controls

129 lines (106 loc) · 6.2 KB
Raw

Getting started

This document describes how to train and test models with our repository. In the following we consider AVA with R50-I3D-NL backbone as an example. Other datasets or backbones are trained/tested in a similar way.

Each experiment is defined by a configuration (.yaml) file. The list of all yaml files are available in README.md.

Training

Download pre-trained weights

We pre-train all models on ImageNet-1K and Kinetics-400. The pre-trained model weights are provided. Please download (R50-I3D-NL, R101-I3D-NL) and put/symlink them in [video-lfb root]/pretrained_weights folder.

Training a baseline model

Run train_net.py with a corresponding yaml file.

python2 tools/train_net.py \
  --config_file configs/ava_r50_baseline.yaml \
  CHECKPOINT.DIR /tmp/baseline-output

Training an LFB model

It involves 2 steps:

  1. Train a baseline model that will be used to infer LFB.
python2 tools/train_net.py \
  --config_file configs/ava_r50_baseline.yaml \
  CHECKPOINT.DIR /tmp/lfb-nl-step1-output
  1. Train an LFB model.
python2 tools/train_net.py \
  --config_file configs/ava_r50_lfb_nl.yaml \
  LFB.MODEL_PARAMS_FILE [path to baseline model weight from step1] \
  LFB.WRITE_LFB True \
  CHECKPOINT.DIR /tmp/lfb-nl-step2-output

This command will first construct an LFB using the model weight LFB.MODEL_PARAMS_FILE, and then train the LFB-NL model. Here we set LFB.WRITE_LFB as True to store the inferred LFB, so that we can use LFB.LOAD_LFB and LFB.LOAD_LFB_PATH to reuse it next time. LFB will be stored in training job output folder (CHECKPOINT.DIR). Storing LFB takes about 3.3 GB of space for AVA, 1.1 GB for EPIC-Kitchens verbs, and 4.4 GB for Charades.

Linear scaling rule

All of our released configs make use of 8 GPUs. If you will be using different number of GPUs for training, please modify training schedules and hyperparameters according to the linear scaling rule. See Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour paper for details.

Testing

To test/evaluate a model, use the config (yaml) file you used for training, and run tools/test_net.py with the model weight to be evaluated.

Testing a baseline model:

python2 tools/test_net.py \
  --config_file configs/ava_r50_baseline.yaml \
  TEST.PARAMS_FILE [path to model weight]

Testing an LFB model:

python2 tools/test_net.py \
  --config_file configs/ava_r50_lfb_nl.yaml \
  LFB.LOAD_LFB True \
  LFB.LOAD_LFB_PATH [path to folder that stores LFB] \
  TEST.PARAMS_FILE [path to model weight]

Here we assume we set LFB.WRITE_LFB as True during training, so we can directly load the LFB without re-constructing it. In this case, [path to folder that stores LFB] is the output folder of training job.

For AVA models, you can set AVA.TEST_MULTI_CROP as True to perform multi-scale, 2-flip, and 3-crop testing.

Dataset specific details

AVA Person Detector

To use our predicted person boxes for experiments, please download (train, val, test) and put them in data/ava/annotations/.

To train the person detector yourself, please use Detectron with e2e_keypoint_rcnn_X-101-32x8d-FPN_s1x_ava.yaml. The model weight of our trained detector is available [here]. The model obtains 93.9 AP@50 on validation set.

EPIC-Kitchens Noun LFB

Our pre-computed "Noun LFB" is available for download (train, val). It contains the features of the top-10 objects extracted at 1 FPS from an object detector. To train/test an LFB model on EPIC-Kitchens "Nouns", please download and put (or symlink) them in data/epic/noun_lfb.

To train the detector yourself, please use Detectron with epic_object_detector.yaml. This yaml file makes use of Visual Genome pre-trained model weights, which is available [here]). The final model weight is available [here].

Evaluating EPIC-Kitchens "action" predictions

After training a verb/noun model, a epic_predictions_final.pkl file will be automatically generated. Please run

python2 tools/evaluate_actions.py --verb_file [path to verb predictions] --noun_file [path to noun predictions]

to combine "verb" and "noun" predictions and evaluate their "action" accuracy.

Charades two-stage training

Note that we train LFB models for Charades in two stages: We first train the backbone without LFB (i.e., like training a baseline model), and then add LFB and continue training for half of the schedule.

Stage 1:

python2 tools/train_net.py \
  --config_file configs/charades_r50_baseline.yaml \
  CHECKPOINT.DIR /tmp/lfb-nl-stage1-output

Stage 2:

python2 tools/train_net.py \
  --config_file configs/charades_r50_lfb_nl.yaml \
  TRAIN.PARAMS_FILE [path to model weight from Stage 1] \
  LFB.MODEL_PARAMS_FILE [path to model weight for LFB inference] \
  LFB.WRITE_LFB True \
  CHECKPOINT.DIR /tmp/lfb-nl-stage2-output