GitHub - ml-lab/TensorBox

TensorBox is a simple framework for training neural networks to detect objects in images. Training requires a text file (see here, for example) of paths to images on disk and the corresponding object locations in each image. The basic model implements the simple and robust GoogLeNet-OverFeat algorithm. We additionally provide an implementation of the ReInspect algorithm, achieving state-of-the-art detection results on the TUD crossing and brainwash datasets.

Special thanks to Brett Kuprel of Sebastian Thrun's group for providing the initial code to hack into Google's pretrained ImageNet weights for finetuning.

OverFeat Installation & Training

First, install TensorFlow from source or pip

$ git clone http://github.com/russell91/tensorbox
$ cd tensorbox
$ ./download_data.sh
$ python train.py --hypes hypes/default.json --gpu 0 --logdir output

Note that running on your own dataset should only require modifying the hypes/default.json file. When finished training, you can use code from the provided ipython notebook to get results on your test set.

ReInspect Installation & Training

ReInspect, initially implemented in Caffe, is a neural network extension to Overfeat-GoogLeNet in Tensorflow. It is designed for high performance object detection in images with heavily overlapping instances. See the paper for details or the video for a demonstration.

$ git clone http://github.com/russell91/tensorbox
$ cd tensorbox
$ ./download_data.sh
$ # Install tensorflow from source
$ git clone --recurse-submodules https://github.com/tensorflow/tensorflow
$ # Add code for the custom hungarian layer user_op
$ cp /path/to/tensorbox/data/lstm/hungarian.cc /path/to/tensorflow/tensorflow/core/user_ops/
$ # Proceed with the GPU installation of tensorflow from source...
$ # (see https://www.tensorflow.org/versions/r0.7/get_started/os_setup.html#installing-from-sources)
$ cd /path/to/tensorbox/utils && make && cd ..
$ python train.py --hypes hypes/lstm.json --gpu 0 --logdir output

Tensorboard

You can visualize the progress of your experiments during training using Tensorboard.

$ cd /path/to/tensorbox
$ tensorboard --logdir output
$ # (optional, start an ssh tunnel if not experimenting locally)
$ ssh myserver -N -L localhost:6006:localhost:6006
$ # open localhost:6006 in your browser

For example, the following is a screenshot of a Tensorboard comparing two different experiments with learning rate decays that kick in at different points. The learning rate drops in half at 60k iterations for the green experiment and 300k iterations for red experiment.

TensorBox is a simple framework for training neural networks to detect objects in images. Training requires a text file (see here, for example) of paths to images on disk and the corresponding object locations in each image. The basic model implements the simple and robust GoogLeNet-OverFeat algorithm. We additionally provide an implementation of the ReInspect algorithm, achieving state-of-the-art detection results on the TUD crossing and brainwash datasets.

Special thanks to Brett Kuprel of Sebastian Thrun's group for providing the initial code to hack into Google's pretrained ImageNet weights for finetuning.

OverFeat Installation & Training

First, install TensorFlow from source or pip

$ git clone http://github.com/russell91/tensorbox
$ cd tensorbox
$ ./download_data.sh
$ python train.py --hypes hypes/default.json --gpu 0 --logdir output

Note that running on your own dataset should only require modifying the hypes/default.json file. When finished training, you can use code from the provided ipython notebook to get results on your test set.

ReInspect Installation & Training

ReInspect, initially implemented in Caffe, is a neural network extension to Overfeat-GoogLeNet in Tensorflow. It is designed for high performance object detection in images with heavily overlapping instances. See the paper for details or the video for a demonstration.

$ git clone http://github.com/russell91/tensorbox
$ cd tensorbox
$ ./download_data.sh
$ # Install tensorflow from source
$ git clone --recurse-submodules https://github.com/tensorflow/tensorflow
$ # Add code for the custom hungarian layer user_op
$ cp /path/to/tensorbox/data/lstm/hungarian.cc /path/to/tensorflow/tensorflow/core/user_ops/
$ # Proceed with the GPU installation of tensorflow from source...
$ # (see https://www.tensorflow.org/versions/r0.7/get_started/os_setup.html#installing-from-sources)
$ cd /path/to/tensorbox/utils && make && cd ..
$ python train.py --hypes hypes/lstm.json --gpu 0 --logdir output

Tensorboard

You can visualize the progress of your experiments during training using Tensorboard.

$ cd /path/to/tensorbox
$ tensorboard --logdir output
$ # (optional, start an ssh tunnel if not experimenting locally)
$ ssh myserver -N -L localhost:6006:localhost:6006
$ # open localhost:6006 in your browser

For example, the following is a screenshot of a Tensorboard comparing two different experiments with learning rate decays that kick in at different points. The learning rate drops in half at 60k iterations for the green experiment and 300k iterations for red experiment.

TensorBox is a simple framework for training neural networks to detect objects in images. Training requires a text file (see here, for example) of paths to images on disk and the corresponding object locations in each image. The basic model implements the simple and robust GoogLeNet-OverFeat algorithm. We additionally provide an implementation of the ReInspect algorithm, achieving state-of-the-art detection results on the TUD crossing and brainwash datasets.

Special thanks to Brett Kuprel of Sebastian Thrun's group for providing the initial code to hack into Google's pretrained ImageNet weights for finetuning.

OverFeat Installation & Training

First, install TensorFlow from source or pip

$ git clone http://github.com/russell91/tensorbox
$ cd tensorbox
$ ./download_data.sh
$ python train.py --hypes hypes/default.json --gpu 0 --logdir output

Note that running on your own dataset should only require modifying the hypes/default.json file. When finished training, you can use code from the provided ipython notebook to get results on your test set.

ReInspect Installation & Training

ReInspect, initially implemented in Caffe, is a neural network extension to Overfeat-GoogLeNet in Tensorflow. It is designed for high performance object detection in images with heavily overlapping instances. See the paper for details or the video for a demonstration.

$ git clone http://github.com/russell91/tensorbox
$ cd tensorbox
$ ./download_data.sh
$ # Install tensorflow from source
$ git clone --recurse-submodules https://github.com/tensorflow/tensorflow
$ # Add code for the custom hungarian layer user_op
$ cp /path/to/tensorbox/data/lstm/hungarian.cc /path/to/tensorflow/tensorflow/core/user_ops/
$ # Proceed with the GPU installation of tensorflow from source...
$ # (see https://www.tensorflow.org/versions/r0.7/get_started/os_setup.html#installing-from-sources)
$ cd /path/to/tensorbox/utils && make && cd ..
$ python train.py --hypes hypes/lstm.json --gpu 0 --logdir output

Tensorboard

You can visualize the progress of your experiments during training using Tensorboard.

$ cd /path/to/tensorbox
$ tensorboard --logdir output
$ # (optional, start an ssh tunnel if not experimenting locally)
$ ssh myserver -N -L localhost:6006:localhost:6006
$ # open localhost:6006 in your browser

For example, the following is a screenshot of a Tensorboard comparing two different experiments with learning rate decays that kick in at different points. The learning rate drops in half at 60k iterations for the green experiment and 300k iterations for red experiment.

TensorBox is a simple framework for training neural networks to detect objects in images. Training requires a text file (see here, for example) of paths to images on disk and the corresponding object locations in each image. The basic model implements the simple and robust GoogLeNet-OverFeat algorithm. We additionally provide an implementation of the ReInspect algorithm, achieving state-of-the-art detection results on the TUD crossing and brainwash datasets.

Special thanks to Brett Kuprel of Sebastian Thrun's group for providing the initial code to hack into Google's pretrained ImageNet weights for finetuning.

OverFeat Installation & Training

First, install TensorFlow from source or pip

$ git clone http://github.com/russell91/tensorbox
$ cd tensorbox
$ ./download_data.sh
$ python train.py --hypes hypes/default.json --gpu 0 --logdir output

Note that running on your own dataset should only require modifying the hypes/default.json file. When finished training, you can use code from the provided ipython notebook to get results on your test set.

ReInspect Installation & Training

ReInspect, initially implemented in Caffe, is a neural network extension to Overfeat-GoogLeNet in Tensorflow. It is designed for high performance object detection in images with heavily overlapping instances. See the paper for details or the video for a demonstration.

$ git clone http://github.com/russell91/tensorbox
$ cd tensorbox
$ ./download_data.sh
$ # Install tensorflow from source
$ git clone --recurse-submodules https://github.com/tensorflow/tensorflow
$ # Add code for the custom hungarian layer user_op
$ cp /path/to/tensorbox/data/lstm/hungarian.cc /path/to/tensorflow/tensorflow/core/user_ops/
$ # Proceed with the GPU installation of tensorflow from source...
$ # (see https://www.tensorflow.org/versions/r0.7/get_started/os_setup.html#installing-from-sources)
$ cd /path/to/tensorbox/utils && make && cd ..
$ python train.py --hypes hypes/lstm.json --gpu 0 --logdir output

Tensorboard

You can visualize the progress of your experiments during training using Tensorboard.

$ cd /path/to/tensorbox
$ tensorboard --logdir output
$ # (optional, start an ssh tunnel if not experimenting locally)
$ ssh myserver -N -L localhost:6006:localhost:6006
$ # open localhost:6006 in your browser

For example, the following is a screenshot of a Tensorboard comparing two different experiments with learning rate decays that kick in at different points. The learning rate drops in half at 60k iterations for the green experiment and 300k iterations for red experiment.

TensorBox is a simple framework for training neural networks to detect objects in images. Training requires a text file (see here, for example) of paths to images on disk and the corresponding object locations in each image. The basic model implements the simple and robust GoogLeNet-OverFeat algorithm. We additionally provide an implementation of the ReInspect algorithm, achieving state-of-the-art detection results on the TUD crossing and brainwash datasets.

Special thanks to Brett Kuprel of Sebastian Thrun's group for providing the initial code to hack into Google's pretrained ImageNet weights for finetuning.

OverFeat Installation & Training

First, install TensorFlow from source or pip

$ git clone http://github.com/russell91/tensorbox
$ cd tensorbox
$ ./download_data.sh
$ python train.py --hypes hypes/default.json --gpu 0 --logdir output

Note that running on your own dataset should only require modifying the hypes/default.json file. When finished training, you can use code from the provided ipython notebook to get results on your test set.

ReInspect Installation & Training

ReInspect, initially implemented in Caffe, is a neural network extension to Overfeat-GoogLeNet in Tensorflow. It is designed for high performance object detection in images with heavily overlapping instances. See the paper for details or the video for a demonstration.

$ git clone http://github.com/russell91/tensorbox
$ cd tensorbox
$ ./download_data.sh
$ # Install tensorflow from source
$ git clone --recurse-submodules https://github.com/tensorflow/tensorflow
$ # Add code for the custom hungarian layer user_op
$ cp /path/to/tensorbox/data/lstm/hungarian.cc /path/to/tensorflow/tensorflow/core/user_ops/
$ # Proceed with the GPU installation of tensorflow from source...
$ # (see https://www.tensorflow.org/versions/r0.7/get_started/os_setup.html#installing-from-sources)
$ cd /path/to/tensorbox/utils && make && cd ..
$ python train.py --hypes hypes/lstm.json --gpu 0 --logdir output

Tensorboard

You can visualize the progress of your experiments during training using Tensorboard.

$ cd /path/to/tensorbox
$ tensorboard --logdir output
$ # (optional, start an ssh tunnel if not experimenting locally)
$ ssh myserver -N -L localhost:6006:localhost:6006
$ # open localhost:6006 in your browser

For example, the following is a screenshot of a Tensorboard comparing two different experiments with learning rate decays that kick in at different points. The learning rate drops in half at 60k iterations for the green experiment and 300k iterations for red experiment.

TensorBox is a simple framework for training neural networks to detect objects in images. Training requires a text file (see here, for example) of paths to images on disk and the corresponding object locations in each image. The basic model implements the simple and robust GoogLeNet-OverFeat algorithm. We additionally provide an implementation of the ReInspect algorithm, achieving state-of-the-art detection results on the TUD crossing and brainwash datasets.

Special thanks to Brett Kuprel of Sebastian Thrun's group for providing the initial code to hack into Google's pretrained ImageNet weights for finetuning.

OverFeat Installation & Training

First, install TensorFlow from source or pip

$ git clone http://github.com/russell91/tensorbox
$ cd tensorbox
$ ./download_data.sh
$ python train.py --hypes hypes/default.json --gpu 0 --logdir output

Note that running on your own dataset should only require modifying the hypes/default.json file. When finished training, you can use code from the provided ipython notebook to get results on your test set.

ReInspect Installation & Training

ReInspect, initially implemented in Caffe, is a neural network extension to Overfeat-GoogLeNet in Tensorflow. It is designed for high performance object detection in images with heavily overlapping instances. See the paper for details or the video for a demonstration.

$ git clone http://github.com/russell91/tensorbox
$ cd tensorbox
$ ./download_data.sh
$ # Install tensorflow from source
$ git clone --recurse-submodules https://github.com/tensorflow/tensorflow
$ # Add code for the custom hungarian layer user_op
$ cp /path/to/tensorbox/data/lstm/hungarian.cc /path/to/tensorflow/tensorflow/core/user_ops/
$ # Proceed with the GPU installation of tensorflow from source...
$ # (see https://www.tensorflow.org/versions/r0.7/get_started/os_setup.html#installing-from-sources)
$ cd /path/to/tensorbox/utils && make && cd ..
$ python train.py --hypes hypes/lstm.json --gpu 0 --logdir output

Tensorboard

You can visualize the progress of your experiments during training using Tensorboard.

$ cd /path/to/tensorbox
$ tensorboard --logdir output
$ # (optional, start an ssh tunnel if not experimenting locally)
$ ssh myserver -N -L localhost:6006:localhost:6006
$ # open localhost:6006 in your browser

For example, the following is a screenshot of a Tensorboard comparing two different experiments with learning rate decays that kick in at different points. The learning rate drops in half at 60k iterations for the green experiment and 300k iterations for red experiment.

TensorBox is a simple framework for training neural networks to detect objects in images. Training requires a text file (see here, for example) of paths to images on disk and the corresponding object locations in each image. The basic model implements the simple and robust GoogLeNet-OverFeat algorithm. We additionally provide an implementation of the ReInspect algorithm, achieving state-of-the-art detection results on the TUD crossing and brainwash datasets.

Special thanks to Brett Kuprel of Sebastian Thrun's group for providing the initial code to hack into Google's pretrained ImageNet weights for finetuning.

OverFeat Installation & Training

First, install TensorFlow from source or pip

$ git clone http://github.com/russell91/tensorbox
$ cd tensorbox
$ ./download_data.sh
$ python train.py --hypes hypes/default.json --gpu 0 --logdir output

Note that running on your own dataset should only require modifying the hypes/default.json file. When finished training, you can use code from the provided ipython notebook to get results on your test set.

ReInspect Installation & Training

ReInspect, initially implemented in Caffe, is a neural network extension to Overfeat-GoogLeNet in Tensorflow. It is designed for high performance object detection in images with heavily overlapping instances. See the paper for details or the video for a demonstration.

$ git clone http://github.com/russell91/tensorbox
$ cd tensorbox
$ ./download_data.sh
$ # Install tensorflow from source
$ git clone --recurse-submodules https://github.com/tensorflow/tensorflow
$ # Add code for the custom hungarian layer user_op
$ cp /path/to/tensorbox/data/lstm/hungarian.cc /path/to/tensorflow/tensorflow/core/user_ops/
$ # Proceed with the GPU installation of tensorflow from source...
$ # (see https://www.tensorflow.org/versions/r0.7/get_started/os_setup.html#installing-from-sources)
$ cd /path/to/tensorbox/utils && make && cd ..
$ python train.py --hypes hypes/lstm.json --gpu 0 --logdir output

Tensorboard

You can visualize the progress of your experiments during training using Tensorboard.

$ cd /path/to/tensorbox
$ tensorboard --logdir output
$ # (optional, start an ssh tunnel if not experimenting locally)
$ ssh myserver -N -L localhost:6006:localhost:6006
$ # open localhost:6006 in your browser

For example, the following is a screenshot of a Tensorboard comparing two different experiments with learning rate decays that kick in at different points. The learning rate drops in half at 60k iterations for the green experiment and 300k iterations for red experiment.

TensorBox is a simple framework for training neural networks to detect objects in images. Training requires a text file (see here, for example) of paths to images on disk and the corresponding object locations in each image. The basic model implements the simple and robust GoogLeNet-OverFeat algorithm. We additionally provide an implementation of the ReInspect algorithm, achieving state-of-the-art detection results on the TUD crossing and brainwash datasets.

Special thanks to Brett Kuprel of Sebastian Thrun's group for providing the initial code to hack into Google's pretrained ImageNet weights for finetuning.

OverFeat Installation & Training

First, install TensorFlow from source or pip

$ git clone http://github.com/russell91/tensorbox
$ cd tensorbox
$ ./download_data.sh
$ python train.py --hypes hypes/default.json --gpu 0 --logdir output

Note that running on your own dataset should only require modifying the hypes/default.json file. When finished training, you can use code from the provided ipython notebook to get results on your test set.

ReInspect Installation & Training

ReInspect, initially implemented in Caffe, is a neural network extension to Overfeat-GoogLeNet in Tensorflow. It is designed for high performance object detection in images with heavily overlapping instances. See the paper for details or the video for a demonstration.

$ git clone http://github.com/russell91/tensorbox
$ cd tensorbox
$ ./download_data.sh
$ # Install tensorflow from source
$ git clone --recurse-submodules https://github.com/tensorflow/tensorflow
$ # Add code for the custom hungarian layer user_op
$ cp /path/to/tensorbox/data/lstm/hungarian.cc /path/to/tensorflow/tensorflow/core/user_ops/
$ # Proceed with the GPU installation of tensorflow from source...
$ # (see https://www.tensorflow.org/versions/r0.7/get_started/os_setup.html#installing-from-sources)
$ cd /path/to/tensorbox/utils && make && cd ..
$ python train.py --hypes hypes/lstm.json --gpu 0 --logdir output

Tensorboard

You can visualize the progress of your experiments during training using Tensorboard.

$ cd /path/to/tensorbox
$ tensorboard --logdir output
$ # (optional, start an ssh tunnel if not experimenting locally)
$ ssh myserver -N -L localhost:6006:localhost:6006
$ # open localhost:6006 in your browser

For example, the following is a screenshot of a Tensorboard comparing two different experiments with learning rate decays that kick in at different points. The learning rate drops in half at 60k iterations for the green experiment and 300k iterations for red experiment.

Name		Name	Last commit message	Last commit date
hypes		hypes
license		license
utils		utils
.gitignore		.gitignore
README.md		README.md
download_data.sh		download_data.sh
evaluate.ipynb		evaluate.ipynb
train.py		train.py

Name		Name	Last commit message	Last commit date
hypes		hypes
license		license
utils		utils
.gitignore		.gitignore
README.md		README.md
download_data.sh		download_data.sh
evaluate.ipynb		evaluate.ipynb
train.py		train.py

Name		Name	Last commit message	Last commit date
hypes		hypes
license		license
utils		utils
.gitignore		.gitignore
README.md		README.md
download_data.sh		download_data.sh
evaluate.ipynb		evaluate.ipynb
train.py		train.py

Name		Name	Last commit message	Last commit date
hypes		hypes
license		license
utils		utils
.gitignore		.gitignore
README.md		README.md
download_data.sh		download_data.sh
evaluate.ipynb		evaluate.ipynb
train.py		train.py

Name		Name	Last commit message	Last commit date
hypes		hypes
license		license
utils		utils
.gitignore		.gitignore
README.md		README.md
download_data.sh		download_data.sh
evaluate.ipynb		evaluate.ipynb
train.py		train.py

Uh oh!

Latest commit

History

Folders and files

Repository files navigation

OverFeat Installation & Training

ReInspect Installation & Training

Tensorboard

About

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages

Uh oh!

Latest commit

History

Folders and files

Repository files navigation

OverFeat Installation & Training

ReInspect Installation & Training

Tensorboard

About

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages

Uh oh!

Latest commit

History

Folders and files

Repository files navigation

OverFeat Installation & Training

ReInspect Installation & Training

Tensorboard

About

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages

Uh oh!

Latest commit

History

Folders and files

Repository files navigation

OverFeat Installation & Training

ReInspect Installation & Training

Tensorboard

About

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages

Uh oh!

Latest commit

History

Folders and files

Repository files navigation

OverFeat Installation & Training

ReInspect Installation & Training

Tensorboard

About

Resources

Stars

Watchers