Density Estimation by Fully Convolutional Networks (DEFCoN) - A fluorescent spot counter for single molecule localization microscopy.
DEFCoN is a fluorescent spot counter implemented as a fully convolutional neural network. It is designed with the following criteria in mind:
- Fast enough for real-time analysis (~10 ms per frame)
- Parameter free predictions
- Trainable on custom datasets
- Simple API
Please see the setup.py file for the full list of required packages. Specifically, the trained DEFCoN model requires a combination of Python 3.6, TensorFlow 1.3.0, and Keras 2.0.8.
A combination of Python 3.5, TensorFlow 1.4.1, and Keras 2.1.1 is known to work for training new DEFCoN networks.
To install DEFCoN, clone this repository and then use the following command from inside DEFCoN's parent directory (the one above the cloned repository):
pip install DEFCoN
If you wish to develop DEFCoN, we recommend installing as a development package:
pip install -e DEFCoN
The steps required to set up TensorFlow to use the GPU vary widely depending on your system and your Python environment. We recommend reading the following directions for installing TensorFlow to ensure that it is properly configured to use your system's GPU in your particular environment: https://www.tensorflow.org/install/install_linux
In particular, we use the following commands from within a Python virtual environment after we have already installed DEFCoN.
pip uninstall tensorflow
pip install --ignore-installed --upgrade tensorflow-gpu==1.3.0
For us, uninstalling the CPU version of TensorFlow before installing the GPU version is necessary to successfully detect the GPU.
You may verify that TensorFlow successfully detects your GPU after setup by running the following commands from within your Python interpreter:
from tensorflow.python.client import device_lib
print(device_lib.list_local_devices())If you see a string containing something similar to name: "/device:GPU:0" device_type: "GPU" that is printed to the screen, then you should be good to go.
import pkg_resources
from defcon.networks import FCN
# Obtain the filename for the saved DEFCoN model and load it
defcon_model = pkg_resources.resource_filename('defcon.resources', 'defcon_tf13.h5')
model = FCN.from_file(defcon_model)
# Predict from a TIF image stack
tif_file = 'path/to/tif/stack.tif'
y_pred = model.predict_tiff(tif_file)More detailed descriptions on how to perform density map estimations and how to train your own DEFCoN network are found in the examples folder.
- How to use DEFCoN: https://gitter.im/leb_epfl/DEFCoN
- Bug reports: https://github.com/LEB-EPFL/DEFCoN/issues
- Feature requests: https://github.com/LEB-EPFL/DEFCoN/issues
- Developer questions: https://gitter.im/leb_epfl/DEFCoN
DEFCoN was written by Baptiste Ottino as a Master's thesis project under the guidance of Kyle M. Douglass and Suliana Manley in the Laboratory of Experimental Biophysics.