This is the official PyTorch implementation of Unified Recurrence Modeling for Video Action Anticipation.
import torch
from mpnnel import MPNNEL
model = MPNNEL(
input_channels=256,
mpnn_tokens=8*8,
hidden_channels=512,
)
input = torch.randn(1, 8, 256, 8, 8) # (Batch, Timesteps, Channels, tokens)
out = model(input) # (1, 8, 512) -- (Batch, Timesteps, hiden_channels)
import torch
from mpnnel_ctp import MPNNELCTP
model = MPNNELCTP(
input_channels=256,
mpnn_tokens=8*8,
hidden_channels=512,
)
input = torch.randn(1, 8, 256, 8, 8) # (Batch, Timesteps, Channels, tokens)
out, noun, verb = model(input) # (1, 8, 512), (1, 8, 512), (1, 8, 512) -- (Batch, Timesteps, hiden_channels) for out, noun, verb
import torch
from mpnnel_tb import MPNNELTB
model = MPNNELTB(
input_channels=256,
mpnn_tokens=8*8,
hidden_channels=512,
tb_size=512
)
input = torch.randn(1, 8, 256, 8, 8) # (Batch, Timesteps, Channels, tokens)
out = model(input) # (1, 8, 512) -- (Batch, Timesteps, hiden_channels)
@article{tai2022unified,
title={Unified Recurrence Modeling for Video Action Anticipation},
author={Tai, Tsung-Ming and Fiameni, Giuseppe and Lee, Cheng-Kuang and See, Simon and Lanz, Oswald},
journal={arXiv:2206.01009},
year={2022}
}