Python code for the Larisch, Berger, Hamker (2023) publication.
- René Larisch (rene.larisch@informatik.tu-chemnitz.de)
- Lucien Berger (lucien.berger@informatik.tu-chemnitz.de)
- Python v3.9
- numpy v1.22.3
- matplotlib v3.5.1
- ANNarchy v4.7.1.4 (https://annarchy.github.io/)
- Tonic v1.1.2 (https://tonic.readthedocs.io)
- tqdm v4.62.2
The Train directory contains the python script to learn the V1 layer 4 models on the event-based MNIST (N-MNIST) dataset. The dataset is provided by the Tonic package.
Start the learning with:
python ann_simple.py
or with four cores:
python ann_simple.py -j4
Please be aware, that depending on your hardware, training the full
The Evaluation directory contains five subdirectories, one for each of the five tested corruptions. Please perform the following steps to evaluate the finally trained network on one of the corruptions:
-
Copy the weights matrices from the output directory, created by the ann_simple.py, to the network_data directory. Please remove the number from the file name. The names of the files should be:
- InhibW.txt: feedforward weights from the LGN/input population to the inhibitory population
- INLat.txt: lateral inhibitory weights between all inhibitory neurons
- INtoV1.txt: inhibitory feedback weights from the inhibitory population to the excitatory simple cell population
- V1toIN.txt: feedforward weights from the simple cell population to the inhibitory population
- V1weight.txt: feedforward weights from the LGN/input population to the excititatory simple cell population
-
To present the complete N-MNIST training set and the test set with the corresponding corruption use (* depends on the corruption):
python nmnist_run_*.py
The corruption is presented on different levels to the network, including a corruption strength of zero (no corruption). An output_svm directory is created, where the activity on the training set is saved, and in corresponding subdirectories the activities on the test set with a certain level of corruption.
- To evaluate the classification robustness with respect to a certain corruption use start the svm.py script. It first fits a support-vector machine with the network activities on the N-MNIST training set and then predicts the classes based on the network activities, observed on the test set with the corruption. For each level of corruption strength, it creates one *.txt file, containing the accuracy score. Additionally, in each subdirectory in the output_svm directory, the image of a confusion matrix is created.
-
Train
- ann_simple.py: contains the network defintion and training loop
- Make_lilarray.py: script to process the event list from Tonic to create a list of lists for ANNarchy
- plots.py: some additional plots to monitor the network activity
- rezeptiv.py: creates receptive fields of the final trained network
-
Evaluation
- directories for the five different corruption which are tested
- each directory contains the network definition (net.py), a script to run the support vector machine for classification (svm.py), and a script to present the N-MNIST dataset with the corresponding corruption
- svm_L_Transform_DropArea: Present the normal N-MNIST dataset and the drop area corruption for different area sizes (nmnist_run_droparea.py)
- svm_L_Transform_DropEvent: Present the normal N-MNIST dataset and the drop event corruption for levels of drop rates for single events (nmnist_run_drop_event.py)
- svm_L_Transform_UniformNoise_custom: Present the normal N-MNIST dataset and the uniform noise corruption for different number of additional events (nmnist_run_UniformNoise.py)
- svm_L_Transform_SpatialJitter: Present the normal N-MNIST dataset and the spatial jitter corruption whith different levels of shifts along the space dimensions (nmnist_run_Jitter.py)
- svm_L_Transform_TimeJitter: Present the normal N-MNIST dataset and the spatial jitter corruption whith different levels of shifts along the time dimension (nmnist_run_Jitter.py)