torchevent 0.0.3

A PyTorch library for event-based data processing

Torchevent: Spiking Neural Network Framework

Torchevent is a PyTorch-based framework for Spiking Neural Networks (SNNs). It supports training and inference for event-based datasets like NMNIST, offering custom models, loss functions, and transformations optimized for SNN workflows.

---

References

This project draws inspiration from the following works:

1. Paper:
   TSSL-BP: Temporal-Spike-Sequence Learning via Backpropagation for Spiking Neural Networks
   Proceedings of the 34th Conference on Neural Information Processing Systems (NeurIPS), 2020.
   Link to Paper

2. GitHub Repository:
   TSSL-BP: Temporal Spike-Sequence Learning Framework
   GitHub Repository

We thank the authors of these works for providing valuable insights into spiking neural network research and implementation.

---

Features

1. SNN Models (TSSL-BP)

• Models like NMNISTNet and NCARSNet are specifically designed for event-based datasets.
• Easily configurable for various spiking network architectures and time-step dynamics.

2. Event Data Transformations

• Transformations tailored for event-based data processing:
  • RandomTemporalCrop: Randomly crops events based on a given time window.
  • TemporalCrop: Sequentially crops events within a fixed time window.
  • ToFrameAuto: Converts events into frames with dynamic configurations.

3. Loss Functions

• Loss functions designed for SNN-specific requirements:
  • SpikeKernelLoss: Computes the loss using Post-Synaptic Potentials (PSP).
  • SpikeCountLoss: Optimizes models to match desired spike counts.
  • SpikeSoftmaxLoss: Combines spike data with softmax and cross-entropy for classification tasks.

---

Installation

To install torchevent manually:

git clone https://github.com/devcow85/torchevent.git
cd torchevent
pip install .

---

Usage

The following script demonstrates training the NMNISTNet model using the NMNIST dataset from tonic api:

import tonic
import tonic.transforms as transforms
import torch
from torch.utils.data import DataLoader

from torchevent.utils import set_seed, spike2data
from torchevent.transforms import RandomTemporalCrop, TemporalCrop
from torchevent import models, loss

# Set seed for reproducibility
set_seed(7)

# Prepare the dataset
transform = transforms.Compose([
    RandomTemporalCrop(time_window=99000),
    transforms.ToFrame(sensor_size=tonic.datasets.NMNIST.sensor_size, n_time_bins=5),
])

train_ds = tonic.datasets.NMNIST(save_to="data", train=True, transform=transform)
val_ds = tonic.datasets.NMNIST(save_to="data", train=False, transform=transform)

# Create data loaders
train_loader = DataLoader(train_ds, shuffle=True, batch_size=32, num_workers=8, pin_memory=True)
val_loader = DataLoader(val_ds, shuffle=False, batch_size=32, num_workers=8, pin_memory=True)

# Initialize model, optimizer, and loss function
model = models.NMNISTNet(5, 1, n_steps=5).to("cuda")
optimizer = torch.optim.AdamW(model.parameters(), lr=0.0005)
criterion = loss.SpikeCountLoss(desired_count=4, undesired_count=1)

# Training loop
for epoch in range(3):
    model.train()
    for data, targets in train_loader:
        data, targets = data.to("cuda", non_blocking=True), targets.to("cuda", non_blocking=True)
        optimizer.zero_grad()
        spikes = model(data.to(torch.float32))
        spike_loss = criterion(spikes, targets)
        spike_loss.backward()
        torch.nn.utils.clip_grad_norm_(model.parameters(), 1)
        optimizer.step()
        print(f"Epoch [{epoch+1}], Loss: {spike_loss.item():.4f}")

Expected Result

Epoch [1/3], Step [10/1875], Loss: 40.6000, Elapsed Time: 0.13s
...
Epoch [1/3] completed. Average Loss: 22.7644, Elapsed Time: 163.37s
...
Epoch [2/3], Step [1870/1875], Loss: 20.1000, Elapsed Time: 0.06s
Epoch [2/3] completed. Average Loss: 18.2996, Elapsed Time: 107.86s
...
Epoch [3/3], Step [1870/1875], Loss: 10.9000, Elapsed Time: 0.06s
Epoch [3/3] completed. Average Loss: 15.9984, Elapsed Time: 108.05s
Validation Loss: 15.2796, Accuracy: 91.01%, Elapsed Time: 5.16s

Contact

For quenstions, suggestions, or support, please contact Seokhun Jeon ([e-mail])