v1_predictor

Training and evaluating the CNN for "Using deep learning to reveal the neural code for images in primary visual cortex" by William F. Kindel, Elijah D. Christensen and Joel Zylberberg https://arxiv.org/abs/1706.06208 (2017).

Overview

This program trains and evaluates a convolutional neural network (CNN) whose input is an image and whose output is the predicted firing rates of every neuron in a given data file. This CNN has two convolutional layers followed by a densily connected hidden layer with hyper-parameters described below. To quantify the performance of the predictor, we compare the network’s predicted firing rates to the neurons’ measured firing rates using a held-out evaluation set using the Pearson correlation coefficient.

Versions

Python 3.5.2
Tensorlow 1.0.1

Also see requirements.txt

Necessary Files

python files

runnet.py
buildnet.py

data files

stored in data/data.tar.gz

$ tar -xvzf data.tar.gz

data/
├── 01mean50ms_smallim_d2_crop.mat
├── 02mean50ms_smallim_d2_crop.mat
├── 03mean50ms_smallim_d2_crop.mat
├── 04mean50ms_smallim_d2_crop.mat
├── 05mean50ms_smallim_d2_crop.mat
├── 06mean50ms_smallim_d2_crop.mat
├── 07mean50ms_smallim_d2_crop.mat
├── 08mean50ms_smallim_d2_crop.mat
├── 09mean50ms_smallim_d2_crop.mat
└── 10mean50ms_smallim_d2_crop.mat

Examples

Building and Running the network

From a Unix terminal:

$ python3 runnet.py

Plotting Example

For example of plotting output files run:

examples/plot1516647610.py

$ cd examples
$ python3 plot1516647610.py

Which produces:

examples/training_manualsave_1516647610.npy
examples/persondata1516647610.npy

Network Outputs

Written to the manualsave folder located in the working directory. Folder created if not present.

All output files have the same ID siffix (denoted UNIXTIME) which is the Unix time when the training starts.

Outputs:

training_plot_UNIXTIME.png -- a plot of the training progress writting as the training goes on
training_manualsave_UNIXTIME.npy -- the data that does into the training plot. Load as: [FLAGS, trainlist, earlystoplist, evallist, rlist, step, lossbaseline, traintrials, earlystoptrials, evaltrials] = np.load(training_manualsave_UNIXTIME.npy)
personplotUNIXTIME.png -- a bar plot of the Pearson correlation coefficient over all neurons for the best nextowrk updated as the training goes on.
persondataUNIXTIME.npy -- the data that does into the Pearson plot. Load as: [rval,step,traintrials, earlystoptrials, evaltrials, FLAGS] = np.load(persondataUNIXTIME.npy)
network_manualsave_UNIXTIME.npy -- the trained weights/parameters of the CNN for the best nextowrk updated as the training goes on. Load as: [WC1, BC1, WC2, BC2, WH3, BH3, WL4, BL4, step] = np.load(network_manualsave_UNIXTIME.npy)

CNN Hyperparameters

The hyper-parameters of the are written as a FLAG which can be adjusted.

$ python3 runnet.py --help
usage: lnln_hyperopt.py [-h] [--data_dir DATA_DIR] [--save_dir SAVE_DIR]
                        [--save SAVE] [--savetraining SAVETRAINING]
                        [--savenetwork SAVENETWORK] [--fileindex FILEINDEX]
                        [--learning_rate LEARNING_RATE]
                        [--max_steps MAX_STEPS] [--batch_size BATCH_SIZE]
                        [--trainingset_frac TRAININGSET_FRAC]
                        [--earlystop_frac EARLYSTOP_FRAC] [--conv1 CONV1]
                        [--conv2 CONV2] [--conv1size CONV1SIZE] [--nk1 NK1]
                        [--nstride1 NSTRIDE1] [--conv2size CONV2SIZE]
                        [--nk2 NK2] [--nstride2 NSTRIDE2]
                        [--numconvlayer NUMCONVLAYER] [--hidden1 HIDDEN1]
                        [--hidden2 HIDDEN2] [--dropout DROPOUT]

optional arguments:
  -h, --help            show this help message and exit
  --data_dir DATA_DIR   Data directory containing mat files of the format:
                        0Nmean50ms_smallim_d2_crop
  --save_dir SAVE_DIR   Directory to save outputs
  --save SAVE           If true, save the results.
  --savetraining SAVETRAINING
                        If true, save the traing.
  --savenetwork SAVENETWORK
                        If true, save the network
  --fileindex FILEINDEX
                        index for which file to load
  --dropout DROPOUT     ...

training:
  --learning_rate LEARNING_RATE
                        Initial learning rate.
  --max_steps MAX_STEPS
                        Number of steps to run trainer.
  --batch_size BATCH_SIZE
                        Batch size.
  --trainingset_frac TRAININGSET_FRAC
                        Training set size (fraction of images).
  --earlystop_frac EARLYSTOP_FRAC
                        Early stop set size (fraction of images).

CNN params:
  --conv1 CONV1         Number of filters in conv 1.
  --conv2 CONV2         Number of filters in conv 2.
  --conv1size CONV1SIZE
                        Size (linear) of convolution kernel larer 1.
  --nk1 NK1             Size of max pool kernel layer 1.
  --nstride1 NSTRIDE1   Size of max pool stride layer 1.
  --conv2size CONV2SIZE
                        Size (linear) of convolution kernel larer 2.
  --nk2 NK2             Size of max pool kernel layer 2.
  --nstride2 NSTRIDE2   Size of max pool stride.
  --numconvlayer NUMCONVLAYER
                        number of convolutional layers

hidden layer:
  --hidden1 HIDDEN1     Number of units in hidden layer 1.
  --hidden2 HIDDEN2     Number of units in hidden layer 2. Not used.

FLAGS.learning_rate -- the learning rate
FLAGS.max_steps -- maximum number of steps to run trainer
FLAGS.dropout -- drop out keep rate during training
FLAGS.batch_size -- batch size for training
FLAGS.trainingset_frac -- training set size (fraction of images)
FLAGS.earlystop_frac -- early stop set size (fraction of images)
FLAGS.conv1 -- number of filters in conv 1.
FLAGS.conv2 -- number of filters in conv 2.
FLAGS.hidden1 -- number of units in hidden layer 1
FLAGS.conv1size -- size (linear) of convolution kernel larer 1
FLAGS.nk1 -- size of max pool kernel layer 1.
FLAGS.nstride1 -- size of max pool stride layer 1.
FLAGS.conv2size -- size (linear) of convolution kernel larer 2
FLAGS.nk2 -- size of max pool kernel layer 2
FLAGS.nstride2 -- size of max pool stride for conv 2
FLAGS.fileindex -- intiger index for which file to load
FLAGS.save -- if true, write output files

pfbalan/v1_predictor