/GEDI

Primary LanguagePython

The GEDI CNN project

Your first step is to install requirements. (Use sudo if you must).

python install -r requirements.txt.

Next, download the following model ``weight'' files from the below links

  1. https://mega.nz/#!YU1FWJrA!O1ywiCS2IiOlUCtCpI6HTJOMrneN-Qdv3ywQP5poecM
  2. https://mega.nz/#F!iY0zCaBQ!AzJHMIIEnYALG5xxnAWOFg
    • This is the tracking model.
  3. https://mega.nz/#F!hbo1GAxT!Fj7eeSE7qWQBJBkhuXwQqQ
    • This is a newer tracking model trained on full rotation augmentations
  4. https://mega.nz/#F!6ZEQyCiR!ldTgD9CqGyXX7bkiUPyZnw
    • A tracking model trained with rotation, contrast, and brightness augs.
  5. https://mega.nz/#!yQNRkJqI!Gv9qSWoGqbdV1MjdJ9WrrQ3OMx3qogMhWPLTrVVuSME
    • This is a repackaging of the GEDI model into a .npy. Download this and add it to your config as self.gedi_weight_path.

Last but not least, copy the 'gedi_config.py.template' file as 'gedi_config.py' and adjust it to match your system.

If you're using this project you will want to do one of four things (advanced operations, like ratio and time-course prediction will be added below):

  1. Convert data into a CNN-friendly format.
  2. Train a model.
  3. Test data on a model.
  4. Visualize why a model made the decisions it made.

I will walk you through each of these steps. But first, some basic familiarization with the project.

  • gedi_config.py contains all of the settings for the project.
    • You will need to set self.home_dir and self.src_dir to folders on your machine.
    • You will need to put your GEDI TIFF images into the directory pointed to by self.original_image_dir.
    • Set self.which_dataset to the image panel you're testing -- GFP/masked-GFP/GEDI/ratio.
    • Set self.experiment_image_set to match whatever the structure of the image folders you are using. It is expected that folders take the format {XXX}_{YYY} where {XXX} is the image category (e.g. Live or Dead) and {YYY} is the stem indicated by self.experiment_image_set. An example is Live_rat_gedi_images.
    • Take a look at the other settings, but if you're using a typical set-up is unlikely you'll have to change them.

1. Convert data into a CNN-friendly format.

To combat bottlenecks in data transfer related to python and CPU memory swapping, we use a memory-mapped'' formate with our CNNs. Since this is tensorflow, we use tf-records''. The following will walk you through how to create these.

  • Option A: I want to convert data into a tf-records to use on a model that's already been trained.

    • Edit line 34 in gedi_config.py. Set self.easy_analysis = True. Now whatever image folders you point the script to in self.experiment_image_set will be captured into a single tf-records file, that can then be passed through a testing script.

  • Option B: I want to convert data into a tf-records for training/validation (no seperate testing data).

    • Edit line 34 in gedi_config.py. Set self.easy_analysis = False. Set test_set = False.

  • Option C: I want to convert data into a tf-records for training/validation with a seperate testing set.

    • Edit line 34 in gedi_config.py. Set self.easy_analysis = False. Set test_set = True.

  • Option D: I want to package a ``blinded'' dataset, without its class labels.

    • Edit line 34 in gedi_config.py. Set self.easy_analysis = False. Set blinded = True.

After choosing your appropriate settings, run sh run_preprocessing_scipts.sh

2. Train a model.

After preparing data in a CNN-friendly format, you want to train a model. Look in gedi_config.py for settings that you can adjust for CNN training.

Run sh train_models.sh X where X is the GPU you wish to run training on.

3. Test data on a model.

  • Option A: Pass your tiff images through a pretrained model. Put all of your images in one folder ('image_folder' in the command below) and all of your model checkpoint files that you downloaded above (the second mega.nz link above) in a seperate folder ('model_dir' in the command below). python training_and_eval/test_vgg16_placeholder.py --image_dir=/home/to/my_images_for_cnn --model_file=/path/to/GEDI_trained_model/model_58600.ckpt-58600

  • Option B: Test a model on any dataset (assuming similar charactaristics of both, e.g. both are Rat neurons). (model_dir is your model, validation_data is the tf-records file you want to test on.)

python training_and_eval/test_vgg16.py --model_dir=/media/data/GEDI/drew_images/project_files/train_checkpoint/gfp_2017_05_27_13_56_55 --validation_data=/media/data/GEDI/drew_images/project_files/tfrecords/all_rh_analysis_rat_gfp/val.tfrecords --selected_ckpts=32

  • Option C: Train an SVM on a model to optimize the transfer of its predictions to a new dataset, e.g. trained on rat and tested on human.

python training_and_eval/test_vgg16_tf_svm.py --model_dir=/media/data/GEDI/drew_images/project_files/train_checkpoint/gfp_2017_05_27_13_56_55 --validation_data=/media/data/GEDI/drew_images/project_files/tfrecords/all_rh_analysis_rat_gfp/test.tfrecords --selected_ckpts=32

4. Visualize model decisions in pixel space -- why did the model make the decisions it made?

  • Option A: Visualize decisions on any dataset with the gradient image method. (model_dir is your model, validation_data is the tf-records file you want to test on.)

python visualization/smooth_gradient_image_placeholder.py --live_ims=/path/to/Live_rat --dead_ims=/path/to/Dead_rat model_file=/path/to/GEDI_trained_model/model_58600.ckpt-58600 --output_folder=/path/to/output_visualization_folder

Common errors:

  • Traceback (most recent call last):   File "training_and_eval/test_vgg16_placeholder.py", line 13, in <module>     from gedi_config import GEDIconfig ImportError: No module named gedi_config Setup did not successfully add the project directory to your pythonpath, so you have to do this by hand: export PYTHONPATH=$PYTHONPATH:/my/path/with/gedi_project or export PYTHONPATH=$PYTHONPATH:$(pwd)

  • Segmentation fault. This was raised when running the test_vgg16_placeholder.py script on the CPU. It did not repeat. Try running the script again.