Standalone IBM DVS128 Gesture Dataset on PyTorch

Standalone IBM DVS128 Gesture Dataset on PyTorch. Most codes in this repository are extracted from Spiking Jelly, which is a neuromorphic simulator repository. This is intended for anyone to experiment with the IBM DVS128 Gesture dataset without solving all dependencies of Spiking Jelly that could be daunting sometimes.

Different from Tonic - another neuromorphic dataset library, the dataset class in Spiking Jelly and this repository directly extracts sample from the original IBM dataset. Tonic has already pre-processed the original dataset and remove some samples, thus the number of samples in Tonic are slightly smaller.

Example of dataset usage without any pre-process. This allows one to directly access events corresponding to each sample.

from custom_dataset import DVS128Gesture
import torch
from tqdm import tqdm

dataset_dir = '<ENTER PATH OF YOU DATASET HERE>'
# If dataset doesn't exist, the dataset will be download to the specified location

dataset = DVS128Gesture(root=dataset_dir, train=True, data_type='event')
data_loader = torch.utils.data.DataLoader(dataset, batch_size=1, shuffle=True, num_workers=1)
for raw_events, target in tqdm(data_loader, desc='Loading training data'):
    pass
    # Do something
    
# Example: print all components of raw events corresponding to the last sample in dataloader
print(raw_events[0]['t'])
print(raw_events[0]['x'])
print(raw_events[0]['y'])
print(raw_events[0]['p'])
# Then, print target corresponding to the sample. This is number between 0-10 as there are 11 classes of actions in the dataset.
print(target)

Example of usage for training that pre-processes events from each sample into frames; each of which has the same number of events

# Test loading DVS 128 gesture dataset and spliting each sample into N frames
# such that each frame has about the same number of events
print("Loading data - Example mode 1")
dataset_train = DVS128Gesture(root=dataset_dir, train=True, data_type='frame', frames_number=16, split_by='number')
dataset_test = DVS128Gesture(root=dataset_dir, train=False, data_type='frame', frames_number=16s, split_by='number')
print(f'dataset_train:{dataset_train.__len__()}, dataset_test:{dataset_test.__len__()}')

print("Creating data loaders")
data_loader = torch.utils.data.DataLoader(
    dataset_train, batch_size=16,
    shuffle=True, num_workers=4, pin_memory=False)
data_loader_test = torch.utils.data.DataLoader(
    dataset_test, batch_size=16,
    shuffle=False, num_workers=4, pin_memory=False)

for event_reprs, target in tqdm(data_loader, desc='Loading training data'):
    pass
    # Do something
print(event_reprs.shape, target.shape)

for event_reprs, target in tqdm(data_loader_test, desc='Loading testing data'):
    pass
    # Do something
print(event_reprs.shape, target.shape)

Another example of dataset usage for training that pre-processes events from each sample into frames, but each frame with equal duration

# Test loading DVS 128 gesture dataset and spliting each sample into abritrary number of frames
# such that each frame has about the same duration for instance 3e5 micro second
print("Loading data - Example mode 2")
dataset_train = DVS128Gesture(root=dataset_dir, train=True, data_type='frame', split_by='frame_duration', frame_duration=300000)
dataset_test = DVS128Gesture(root=dataset_dir, train=False, data_type='frame', split_by='frame_duration', frame_duration=300000)
print(f'dataset_train:{dataset_train.__len__()}, dataset_test:{dataset_test.__len__()}')

print("Creating data loaders")
# Collate function is needed because each sample may have a different size
data_loader = torch.utils.data.DataLoader(
    dataset_train, batch_size=16, collate_fn=base_dataset.pad_seq,
    shuffle=True, num_workers=4, pin_memory=False)
data_loader_test = torch.utils.data.DataLoader(
    dataset_test, batch_size=16, collate_fn=base_dataset.pad_seq,
    shuffle=False, num_workers=4, pin_memory=False)

# Suppose we want to measure length of event representation 
train_repr_lens = []
for event_reprs, repr_lens, target in tqdm(data_loader, desc='Loading training data'):
    event_reprs = event_reprs
    target = target
    # Collecting length of event representation when splitting by this method
    train_repr_lens.extend(list(repr_lens))
train_repr_lens = torch.as_tensor(train_repr_lens)
# Print statistic of the event representation length 
print(torch.min(train_repr_lens), torch.max(train_repr_lens), torch.mean(train_repr_lens.float()), torch.std(train_repr_lens.float()))

# Repeat the same thing with test set 
test_repr_lens = []
for event_reprs, repr_lens, target in tqdm(data_loader_test, desc='Loading testing data'):
    # Do something
    event_reprs = event_reprs
    target = target
    event_reprs = event_reprs.float()
    test_repr_lens.extend(list(repr_lens))
test_repr_lens = torch.as_tensor(test_repr_lens)
# Print statistic of the event representation length 
print(torch.min(test_repr_lens), torch.max(test_repr_lens), torch.mean(test_repr_lens.float()), torch.std(test_repr_lens.float()))

There are other modes for loading dataset. See base_dataset.py for details.

QianpengLi577 / dvs128_gesture_pytorch

Standalone IBM DVS128 Gesture Dataset on PyTorch

About

Languages