Name: ReferringRelationships
Owner: Stanford Vision and Learning Group
Description: null
Created: 2017-07-19 06:25:29.0
Updated: 2018-05-22 12:29:01.0
Pushed: 2018-03-28 01:07:17.0
Homepage: null
Size: 7127
Language: Python
GitHub Committers
| User | Most Recent Commit | # Commits |
|---|
Other Committers
| User | Most Recent Commit | # Commits |
|---|
This repository contains code used to produce the results in the following paper:
Ranjay Krishna†, Ines Chami†, Michael Bernstein, Li Fei-Fei
IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018
If you are using this repository, please use the following citation:
roceedings{krishna2018referring,
tle={Referring Relationships},
thor={Krishna, Ranjay and Chami, Ines and Bernstein, Michael and Fei-Fei, Li },
oktitle={IEEE Conference on Computer Vision and Pattern Recognition},
ar={2018}
You can clone the repository and install the requirements by running the following:
clone https://github.com/stanfordvl/ReferringRelationships.git
eferringRelationships
ualenv -p python3 env
ce env/bin/activate
install -r requirements.txt
To download the dataset used in the project, run:
ripts/download_data.sh
Note that we only distribute the annotations for the datasets. To download the images for these datasets, please use the following links:
To train the models, you will need to create an hdf5 dataset and then run
the following script to test and evaluate the model:
r the VRD dataset.
ripts/create_vrd_dataset.sh $LOCATION_OF_VRD_TRAIN_IMAGES $LOCATION_OF_VRD_TEST_IMAGES
ripts/train_vrd.sh
ripts/evaluate_vrd.sh
r the CLEVR dataset.
ripts/create_clevr_dataset.sh $LOCATION_OF_CLEVR_TRAIN_IMAGES $LOCATION_OF_CLEVR_VAL_IMAGES
ripts/train_clevr.sh $LOCATION_OF_MODEL
ripts/evaluate_clevr.sh $LOCATION_OF_MODEL
r the Visual Genome dataset.
ripts/create_visualgenome_dataset.sh $LOCATION_OF_VISUAL_GENOME_IMAGES
ripts/train_visualgenome.sh $LOCATION_OF_MODEL
ripts/evaluate_visualgenome.sh $LOCATION_OF_MODEL
This script will train the model and save the weights in the --save-dir
directory. It will also save the configuration parameters in a
params.json file and log events in train.log.
However, if you decide that you want more control over the training or evaluation scripts, check out the instructions below.
The script data.py will save masks for objects and subjects
in train/val/test directories that will be created in the directory
--save-dir. The script also saves numpy arrays for relationships.
The script has the following command line arguments to modify the dataset pre-processing:
, --help show this help message and exit
test When true, the data is not split into training and
validation sets
val-percent Fraction of images in validation split.
save-dir where to save the ground truth masks, this Location
where dataset should be saved.
img-dir Location where images are stored.
annotations Json with relationships for each image.
image-metadata Image metadata json file.
image-dim The size the images should be saved as.
output-dim The size the predictions should be saved as.
seed The random seed used to reproduce results.
num-images The random seed used to reproduce results.
save-images Use this flag to specify that the images should also
be saved.
max-rels-per-image Maximum number of relationships per image.
The model can be trained by calling python train.py with the following command
line arguments to modify your training:
onal arguments:
, --help Show this help message and exit
opt The optimizer used during training. Currently supports
rms, adam, adagrad and adadelta.
lr The learning rate for training.
lr_decay The learning rate decay.
batch-size The batch size used in training.
epochs The number of epochs to train.
seed The random seed used to reproduce results.
overwrite Train even if that folder already contains an existing
model.
save-dir The location to save the model and the results.
models-dir The location of the model weights
use-models-dir Indicates that new models can be saved in the models
directory set by --models-dir.
save-best-only Saves only the best model checkpoint.
use-subject Boolean indicating whether to use the subjects.
use-predicate Boolean indicating whether to use the predicates.
use-object Boolean indicating whether to use the objects.
embedding-dim Number of dimensions in our class embeddings.
hidden-dim Number of dimensions in the hidden unit.
feat-map-dim The size of the feature map extracted from the image.
input-dim Size of the input image.
num-predicates The number of predicates in the dataset.
num-objects The number of objects in the dataset.
dropout The dropout probability used in training.
train-data-dir Location of the training data.
val-data-dir Location of the validation data.
image-data-dir Location of the images.
heatmap-threshold The thresholds above which we consider a heatmap to
contain an object.
The evaluations can be run using python evaluate.py with the following options:
, --help show this help message and exit
batch-size The batch size used in training.
seed The random seed used to reproduce results.
workers Number workers used to load the data.
heatmap-threshold The thresholds above which we consider a heatmap to
contain an object.
model-checkpoint The model to evaluate.
data-dir Location of the data to evluate with.
The discovery based experiments can be run by setting the following flags
during training and using python evaluate_discovery.py when evaluating.
discovery Used when we run the discovery experinent where
objects are dropped during training.
always-drop-file Location of list of objects that should always be
dropped.
subject-droprate Rate at which subjects are dropped.
object-droprate Rate at which objects are dropped.
model-checkpoint The model to evaluate.
data-dir Location of the data to evluate with.
We welcome everyone to contribute to this reporsitory. Send us a pull request.
The code is under the MIT license. Check LICENSE for details.