Codes for reproducing results in paper: https://arxiv.org/abs/1706.07555. This README will introduce about how to compile and run whisker model, how to generate dataset, and how to train deep neural networks mentioned in the paper. If you have any problems reproducing the results, please feel free to start an issue under this repo or contact Chengxu directly.
Codes for building whisker model and generating datasets include codes in folders bullet_demos_extracted/ and cmd_gen_mp4/.
For compiling the whisker model, you need to have:
cmakeboost, for parsing the config files. Local installation is supported.Bullet, we use it as physic engine. We also require that the build of bullet is done inbullet_buildunder bullet source code directory. Local installation is supported.hdf5, hdf5 is needed to generate the dataset. Only global installation is supported.
Currently, we only support Linux and Mac OS systems. Windows are not supported now.
Our whisker model is based on demos in Bullet. Therefore, all requirements of Bullet demos also apply here.
As the basic framework and some useful tools are borrowed from these demos, building the whisker model requires building all the demos in folder bullet_demos_extracted/, which includes demos extracted by us.
Specifically, we implement our whisker model in demo TestHingeTorque and the actual code is in bullet_demos_extracted/examples/Constraints/TestHingeTorque.cpp.
Make a new directory in any place you want, then run following command for compiling:
cmake -D BULLET_PHYSICS_SOURCE_DIR:SPRING=/path/to/your/bullet/code/repo/ /path/to/this/repo/bullet_demos_extracted/examples/
The reason we need path of bullet source code as input is that we need one library file compiled during building Bullet, which is also why we require that the build of bullet is done in bullet_build under bullet source code directory.
If you install bullet and boost locally, then you also need to specify the BOOST_ROOT and BULLET_ROOT in cmake command by inserting -D BOOST_ROOT:SPRING=/path/to/your/boost/installation/ and -D BULLET_ROOT:SPRING=/path/to/your/bullet/installation/ before /path/to/this/repo/bullet_demos_extracted/examples/.
After successfully running cmake command, you could run make under that directory to actually make the model.
Once make is done, change directory to /path/to/your/build/ExampleBrowser/ and then run ./App_ExampleBrowser --start_demo_name=TestHingeTorque.
If everything is correct, you will see one single whisker behaving unnaturalistically, as we need to provide correct config file to the program using following commands.
To take a quick view of whisker model used in our experiment, run following command under folder cmd_gen_mp4/:
python cmd_gen.py --mp4flag 0 --testmode 1 --pathexe /path/to/your/build/ExampleBrowser/App_ExampleBrowser --fromcfg /path/to/this/repo/cmd_gen_mp4/opt_results/para_ --indxend 31
Like interacting in Bullet demos, you could use "Ctrl + Pressing left mouse + rotate" within the window to rotate the view and "Pressing left mouse + move" to try to apply forces to the whiskers.
Within the command we use, mp4flag determines whether the program will generate a video in mp4 format (ffmpeg required) and fromcfg specifies the location of specific parameters for each whiskers, which we got from behavior optimization mentioned in the paper.
You can also read the source code in cmd_gen.py to understand more of those parameters.
Especially, if you want to use the whisker model in your way, you could read each parameter in config_dict in cmd_gen.py, as these parameters are sent to the whisker model to modify its behavior.
If you are also interested in reproducing the behavior optimization results by yourself, please contact Chengxu. The code is also provided here, but reproduction is complex and may not be of general interest.
We use objects in ShapeNet to generate the dataset.
After downloading the 3D models, you need to use v-hacd to process all the models to transfer each object into a set of "near" convex parts.
The parameter we used in v-hacd is --resolution 500000 --maxNumVerticesPerCH 64.
After processing the models, the path of one model should be organized as following example: /path/to/your/models/02691156/a6693555a4c0bd47434e905131c8d6c6/a6693555a4c0bd47434e905131c8d6c6.obj.
We sampled 9981 objects from ShapeNet to get a balanced distribution in categories (see our paper for details, the actual code is in get_obj_list.py).
The object information we used is stored in obj_choice_2.txt under cmd_gen_mp4/.
To generate the dataset using those 9981 objects, we use non-interactive whisker model, which should be at /path/to/your/build/Constraints/App_TestHinge. Starting App_TestHinge will not start a winodw and the simulation will be done in the same way as starting App_ExampleBrowser. Besides, another folder (config_folder) needs to be created to hold all the configs generated during dataset generation. The command to generate the whole dataset under cmd_gen_mp4/ is as following:
python cmd_dataset.py --objsta 0 --objlen 9981 --bigsamnum 24 --pathexe /path/to/your/build/Constraints/App_TestHinge --fromcfg /path/to/this/repo/cmd_gen_mp4/opt_results/para_ --pathconfig /path/to/your/config_folder --savedir /path/to/store/hdf5s --loaddir /path/to/your/models --seedbas 0
This command will generate 9981*24 hdf5s in /path/to/store/hdf5s.
You can parallel it by running multiple commands with different --objsta (starting index of objects for generation, between 0 and 9980) and --objlen (number of objects for generation).
And here --bigsamnum 24 means 24 independent settings will be generated for each object.
We will use 1/13 of those hdf5s as validation dataset.
With --seedbas 10000 --bigsamnum 2 and different folder for hdf5s, we can generate validation dataset as well. (Here we set seedbas to be 10000 as in the program, as seedbas + objIndex will be used as random seed for each object)
After all the hdf5s have been generated, we use cmd_to_tfr_bycat.py under cmd_gen_mp4/ to generate tfrecords needed to train the models using tensorflow.
Of course, you need to install tensorflow.
The command is as following:
python cmd_to_tfr_bycat.py --catsta 0 --catlen 117 --seedbas 0 --loaddir /path/to/store/hdf5s --savedir /path/to/store/tfrecords --suffix strain --bigsamnum 24
Here, we are generating tfrecords by each category. There are overall 117 categories.
Parameter catsta indicates the starting index of this generation and catlen is the number of categories this generation will cover.
Parameter suffix is just the suffix of tfrecord names, which we will use later to distinguish train/val split.
In order to generate tfrecords for validation, the command need to be modified as following:
python cmd_to_tfr_bycat.py --catsta 0 --catlen 117 --seedbas 10000 --loaddir /path/to/store/validation/hdf5s --savedir /path/to/store/tfrecords --suffix sval --bigsamnum 2
If nothing is wrong, you will be able to see three folders under /path/to/store/tfrecords: Data_force, Data_torque, category.
Each of these folders stores tfrecords for force, torque, and label respectively.
Before training, you need to run script make_meta.py under cmd_gen_mp4/ to create some help files under each tfrecord folders by:
python make_meta.py --dir /path/to/store/tfrecords
Codes for training deep neural networks reported in paper are in folder train_barrel_net/. We use tensorflow and tfutils.
Tfutils is a repo helping the use of tensorflow for model training, validating, saving, resuming, and logging. It will store everything in a MongoDB. Therefore, you need to start your own MongoDB. Please check tfutils for more tutorials about using it. As tfutils is actively updating its master branch, please checkout to multiple_net branch and install it or put the folder under your PYTHONPATH.
After launching your MongoDB at port your_port, you can train deep neural networks through running following commands under folder train_barrel_net/:
For network S_rand:
python train_catenet.py --gpu your_gpu --expId catenet_newdata_n1_fix_1 --cacheDirPrefix /path/to/your/cache --whichopt 2 --initlr 0.001 --newdata 1 --fixweights 1 --newdataprefix /path/to/store/tfrecords
Here, --gpu parameter will be sent to system environment variable CUDA_VISIBLE_DEVICES to control gpus you will use. Usually, only one number is enough. If you want to run the training on multiple gpus, you could set two or more gpus as --gpu 0,1,2 and you also need to add --parallel 1 in your command.
For network S_2c0f:
python train_catenet.py --gpu your_gpu --expId catenet_newdata_shallow4 --cacheDirPrefix /path/to/your/cache --whichopt 2 --initlr 0.003 --newdata 1 --pathconfig catenet_shallow4_config.cfg --newdataprefix /path/to/store/tfrecords
For network S_1c0f:
python train_catenet.py --gpu your_gpu --expId catenet_nd_shal2 --cacheDirPrefix /path/to/your/cache --whichopt 2 --initlr 0.003 --newdata 1 --parallel 1 --pathconfig catenet_shallow2_config.cfg --newdataprefix /path/to/store/tfrecords
For network S_1c2f:
python train_catenet.py --gpu your_gpu --expId catenet_newdata_sm_1conv --cacheDirPrefix /path/to/your/cache --whichopt 2 --initlr 0.003 --newdata 1 --pathconfig catenet_sm2_1conv_config.cfg --newdataprefix /path/to/store/tfrecords
For network S_2c1f:
python train_catenet.py --gpu your_gpu --expId catenet_newdata_sm_2c1f --cacheDirPrefix /path/to/your/cache --whichopt 2 --initlr 0.003 --newdata 1 --pathconfig catenet_sm2_2c1f_config.cfg --newdataprefix /path/to/store/tfrecords
For network S_3c1f:
python train_catenet.py --gpu your_gpu --expId catenet_newdata_sm_3c1f --cacheDirPrefix /path/to/your/cache --whichopt 2 --initlr 0.003 --newdata 1 --pathconfig catenet_sm2_3c1f_config.cfg --newdataprefix /path/to/store/tfrecords
For network S_few:
python train_catenet.py --gpu your_gpu --expId catenet_nd_shal3 --cacheDirPrefix /path/to/your/cache --whichopt 2 --initlr 0.003 --newdata 1 --pathconfig catenet_shallow3_config.cfg --newdataprefix /path/to/store/tfrecords
For network S_2c2f:
python train_catenet.py --gpu your_gpu --expId catenet_newdata_sm_2conv --cacheDirPrefix /path/to/your/cache --whichopt 2 --initlr 0.003 --newdata 1 --pathconfig catenet_sm2_2conv_config.cfg --newdataprefix /path/to/store/tfrecords
For network S_3D:
python train_catenet.py --gpu your_gpu --newdataprefix /path/to/store/tfrecords --expId catenet_nd_3d --cacheDirPrefix /path/to/your/cache --whichopt 2 --initlr 0.003 --pathconfig catenet_3d_config.cfg --namefunc catenet_3d_tfutils --expand 1 --newdata 1
For network S_3c2f:
python train_catenet.py --gpu your_gpu --newdataprefix /path/to/store/tfrecords --expId catenet_newdata_sm_3conv --cacheDirPrefix /path/to/your/cache --whichopt 2 --initlr 0.003 --newdata 1 --parallel 1 --pathconfig catenet_sm2_3conv_config.cfg
For network S_4c2f:
python train_catenet.py --gpu your_gpu --newdataprefix /path/to/store/tfrecords --expId catenet_newdata_sm_low2_2 --cacheDirPrefix /path/to/your/cache --whichopt 2 --initlr 0.006 --newdata 1 --parallel 1 --pathconfig catenet_sm_low2_config.cfg
For network S:
python train_catenet.py --gpu your_gpu --newdataprefix /path/to/store/tfrecords --expId catenet_newdata_sm2 --cacheDirPrefix /path/to/your/cache --whichopt 2 --initlr 0.01 --newdata 1 --pathconfig catenet_sm2_config.cfg
For network S_more:
python train_catenet.py --gpu your_gpu --newdataprefix /path/to/store/tfrecords --expId catenet_newdata_n1 --cacheDirPrefix /path/to/your/cache --whichopt 2 --initlr 0.01 --newdata 1
For network S_deep:
python train_catenet.py --gpu your_gpu --newdataprefix /path/to/store/tfrecords --expId catenet_nd_sm2_deep --cacheDirPrefix /path/to/your/cache --whichopt 2 --initlr 0.01 --newdata 1 --parallel 1 --pathconfig catenet_sm2_deep_config.cfg
For network TS_few:
python train_catenet.py --gpu your_gpu --newdataprefix /path/to/store/tfrecords --expId catenet_nd_tempspa_less --cacheDirPrefix /path/to/your/cache --whichopt 2 --initlr 0.003 --pathconfig catenet_temp_spa_evenless_config.cfg --namefunc catenet_temp_spa_tfutils --newdata 1
For network TS:
python train_catenet.py --gpu your_gpu --newdataprefix /path/to/store/tfrecords --expId catenet_nd_tempspa --cacheDirPrefix /path/to/your/cache --whichopt 2 --initlr 0.003 --pathconfig catenet_temp_spa_less_config.cfg --namefunc catenet_temp_spa_tfutils --newdata 1
For network ST:
python train_catenet.py --gpu your_gpu --newdataprefix /path/to/store/tfrecords --expId catenet_nd_st_sp22_compare --cacheDirPrefix /path/to/your/cache --whichopt 2 --initlr 0.003 --pathconfig catenet_spa_temp_sp22_more_config.cfg --namefunc catenet_spa_temp_tfutils --expand 1 --newdata 1
For RNN class networks, you also need to install repo tnn. It's a repo helping the build of recurrent neural networks in tensorflow.
For network RNN_byp:
python train_catenet.py --gpu your_gpu --newdataprefix /path/to/store/tfrecords --expId catenet_nd_tnn_byp_more --cacheDirPrefix /path/to/your/cache --whichopt 2 --initlr 0.003 --pathconfig catenet_tnn_byp_more_config.cfg --namefunc catenet_tnn_tfutils --expand 1 --tnn 1 --tnndecay 0.1 --decaytrain 1 --newdata 1
For network RNN_lstm:
python train_catenet.py --gpu your_gpu --newdataprefix /path/to/store/tfrecords --expId catenet_nd_tnn_lstm_sep22_relu --cacheDirPrefix /path/to/your/cache --whichopt 2 --initlr 0.003 --pathconfig catenet_tnn_lstm_sep22_relu_config.cfg --namefunc catenet_tnn_tfutils --expand 1 --tnn 1 --tnndecay 0.1 --newdata 1
For network RNN_gru:
python train_catenet.py --gpu your_gpu --newdataprefix /path/to/store/tfrecords --expId catenet_nd_tnn_grusep22_relu --cacheDirPrefix /path/to/your/cache --whichopt 2 --initlr 0.003 --pathconfig catenet_tnn_gru_sep22_relu_config.cfg --namefunc catenet_tnn_tfutils --expand 1 --tnn 1 --tnndecay 0.1 --newdata 1
For network RNN:
python train_catenet.py --gpu your_gpu --newdataprefix /path/to/store/tfrecords --expId catenet_newdata_tnn_more --cacheDirPrefix /path/to/your/cache --whichopt 2 --initlr 0.003 --pathconfig catenet_tnn_sep22_more_config.cfg --namefunc catenet_tnn_tfutils --expand 1 --tnn 1 --tnndecay 0.1 --decaytrain 1 --newdata 1
For network RNN_fdb:
python train_catenet.py --gpu your_gpu --newdataprefix /path/to/store/tfrecords --expId catenet_newdata_tnn_fdb --cacheDirPrefix /path/to/your/cache --whichopt 2 --initlr 0.003 --pathconfig catenet_tnn_fdb_more_config.cfg --namefunc catenet_tnn_tfutils --expand 1 --tnn 1 --tnndecay 0.1 --decaytrain 1 --newdata 1
For all the network trained, we will manually change learning rate to half of that when performance on validation dataset saturates, through stopping the command and restarting it with a lower --initlr.