Code, data, and models for graspnet project v2. Please do not touch without first asking Bhargava or Joe.
- install caffe locally in this repository, under
caffe - copy relevant data in the correct subdirectories under origdata
- download pretrained models to correct subdirectories in models
- read
preprocess_data.py, make modifications as necess 5EC1 ary, and run - run
create_dataset.sh - run
make_dataset_mean.sh
After steps 1 and 2 above are done, the project structure should look like -
root/
caffe/ <- local install of caffe
models/ <- where model definitions and trained weight files go
tarballs/ <- zipped original images for 2 datasets, just in case
tensorflow/ <- local tensorflow install in virtualenv
origdata/
HandCam/
Images/
Anno_HandCam.json
ImageNet/ <- *curated imagenet*
Images/
Anno_ImageNet.json
DeepGrasping/
Images/
Anno_DeepGrasping.json
AllImages/ <- symlink all files under Images above here
trainingdata/ <- preprocessed data for training nets, with train and test text files
train/
val/
Using vgg{16,19} and resnet networks, both recent ilsrvc winners
Read this paper on fine tuning
on vgg and resnet networks and the original vgg paper
Will augment training sets with random flips potential problem - training images are nonsquare, net input is square, not optimal combo