Skip to content

Latest commit

 

History

History
96 lines (76 loc) · 4.78 KB

GETTING_STARTED.md

File metadata and controls

96 lines (76 loc) · 4.78 KB
Raw

Getting Started

This document provides basic tutorials for the usage of MMAction. For installation, please refer to INSTALL.md. For data deployment, please refer to DATASET.md.

An example on UCF-101

We first give an example of testing and training action recognition models on UCF101.

Prepare data

First of all, please follow PREPARING_UCF101.md for data preparation.

Test a reference model

Reference models are stored in MODEL_ZOO.md. We download a reference model spatial stream BN-Inception at $MMACTION/modelzoo using:

wget -c https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmaction/models/ucf101/tsn_2d_rgb_bninception_seg3_f1s1_b32_g8-98160339.pth -P ./modelzoo/

Then, together with provided configs files, we run the following code to test with multiple GPUs:

./tools/dist_test_recognizer.sh test_configs/TSN/ucf101/tsn_rgb_bninception.py tsn_2d_rgb_bninception_seg3_f1s1_b32_g8-98160339.pth 8

When testing 3D ConvNets, the oversample we used is 10 clips x 3 crops by default. For some extremely large models, it might be difficult for so many samples to fit on 1 GPU. When it happens, you can use commands for heavy test instead:

./tools/dist_test_recognizer_heavy.sh test_configs/CSN\ircsn_kinetics400_se_rgb_r152_seg1_32x2.py ircsn_kinetics400_se_rgb_r152_f32s2_ig65m_fbai-9d6ed879.pth 8 --batch_size=5

Train a model with multiple GPUs

To reproduce the model, we provide training scripts as follows:

./tools/dist_train_recognizer.sh configs/TSN/ucf101/tsn_rgb_bninception.py 8 --validate
  • --validate: performs evaluation every k (default=1) epochs during the training, which help diagnose training process.

More examples

The procedure is not limited to action recognition in UCF101. To perform spatial-temporal detection on AVA, we can train a baseline model by running

./tools/dist_train_detector.sh configs/ava/ava_fast_rcnn_nl_r50_c4_1x_kinetics_pretrain_crop.py 8 --validate

and evaluate a reference model by running

wget -c wget -c https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmaction/models/ava/fast_rcnn_ava2.1_nl_r50_c4_1x_f32s2_kin-e2495b48.pth -P modelzoo/
python tools/test_detector.py ava_fast_rcnn_nl_r50_c4_1x_kinetics_pretrain_crop.py modelzoo/fast_rcnn_ava2.1_nl_r50_c4_1x_f32s2_kin-e2495b48.pth --out ava_fast_rcnn_nl_r50_multiscale.pkl --gpus 8 --eval bbox

To perform temporal action detection on THUMOS14, we can training a baseline model by running

./tools/dist_train_localizer.sh configs/thumos14/ssn_thumos14_rgb_bn_inception.py 8

and evaluate a reference model by running

wget -c wget -c https://open-mmlab.s3.ap-northeast-2.amazonaws.com/mmaction/models/thumos14/ssn_thumos14_rgb_bn_inception_tag-dac9ddb0.pth -P modelzoo/
python tools/test_detector.py configs/thumos14/ssn_thumos14_rgb_bn_inception.py modelzoo/ssn_thumos14_rgb_bn_inception_tag-dac9ddb0.pth --gpus 8 --out ssn_thumos14_rgb_bn_inception.pkl --eval thumos14

More Abstract Usage

Inference with pretrained models

We provide testing scripts to evaluate a whole dataset.

Test a dataset

python tools/test_${ARCH}.py ${CONFIG_FILE} ${CHECKPOINT_FILE} [--out ${RESULT_FILE}] [other task-specific arguments]

Arguments:

  • ${ARCH} could be
    • "recognizer" for action recognition (TSN, I3D, SlowFast, R(2+1)D, CSN, ...)
    • "localizer" for temporal action detection/localization (SSN)
    • "detector" for spatial-temporal action detection (a re-implmented Fast-RCNN baseline)
  • ${CONFIG_FILE} is the config file stored in $MMACTION/test_configs.
  • ${CHECKPOINT_FILE} is the checkpoint file. Please refer to MODEL_ZOO.md for more details.

Train a model

MMAction implements distributed training and non-distributed training, powered by the same engine of mmdetection.

Train with multiple GPUs (Recommended)

Training with multiple GPUs follows the rules below:

./tools/dist_train_${ARCH}.sh ${CONFIG_FILE} ${GPU_NUM} [optional arguments]
  • ${ARCH} could be
    • "recognizer" for action recognition (TSN, I3D, ...)
    • "localizer" for temporal action detection/localization (SSN)
    • "detector" for spatial-temporal action detection (a re-implmented Fast-RCNN baseline)
  • ${CONFIG_FILE} is the config file stored in $MMACTION/configs.
  • ${GPU_NUM} is the number of GPU (default: 8). If you are using number other than 8, please adjust the learning rate in the config file linearly.