RAFT/core/raft.py

145 lines
4.8 KiB
Python
Raw Normal View History

2020-03-27 04:19:08 +01:00
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
2020-07-26 01:36:17 +02:00
from update import BasicUpdateBlock, SmallUpdateBlock
from extractor import BasicEncoder, SmallEncoder
from corr import CorrBlock, AlternateCorrBlock
2020-03-27 04:19:08 +01:00
from utils.utils import bilinear_sampler, coords_grid, upflow8
2020-07-26 01:36:17 +02:00
try:
autocast = torch.cuda.amp.autocast
except:
# dummy autocast for PyTorch < 1.6
class autocast:
def __init__(self, enabled):
pass
def __enter__(self):
pass
def __exit__(self, *args):
pass
2020-03-27 04:19:08 +01:00
class RAFT(nn.Module):
def __init__(self, args):
super(RAFT, self).__init__()
self.args = args
if args.small:
self.hidden_dim = hdim = 96
self.context_dim = cdim = 64
args.corr_levels = 4
args.corr_radius = 3
else:
self.hidden_dim = hdim = 128
self.context_dim = cdim = 128
args.corr_levels = 4
args.corr_radius = 4
2020-08-29 01:18:41 +02:00
if 'dropout' not in self.args:
self.args.dropout = 0
2020-03-27 04:19:08 +01:00
2020-08-29 01:18:41 +02:00
if 'alternate_corr' not in self.args:
self.args.alternate_corr = False
2020-03-27 04:19:08 +01:00
# feature network, context network, and update block
if args.small:
self.fnet = SmallEncoder(output_dim=128, norm_fn='instance', dropout=args.dropout)
self.cnet = SmallEncoder(output_dim=hdim+cdim, norm_fn='none', dropout=args.dropout)
self.update_block = SmallUpdateBlock(self.args, hidden_dim=hdim)
else:
self.fnet = BasicEncoder(output_dim=256, norm_fn='instance', dropout=args.dropout)
self.cnet = BasicEncoder(output_dim=hdim+cdim, norm_fn='batch', dropout=args.dropout)
self.update_block = BasicUpdateBlock(self.args, hidden_dim=hdim)
def freeze_bn(self):
for m in self.modules():
if isinstance(m, nn.BatchNorm2d):
m.eval()
def initialize_flow(self, img):
""" Flow is represented as difference between two coordinate grids flow = coords1 - coords0"""
N, C, H, W = img.shape
2021-09-16 16:34:37 +02:00
coords0 = coords_grid(N, H//8, W//8, device=img.device)
coords1 = coords_grid(N, H//8, W//8, device=img.device)
2020-03-27 04:19:08 +01:00
# optical flow computed as difference: flow = coords1 - coords0
return coords0, coords1
2020-07-26 01:36:17 +02:00
def upsample_flow(self, flow, mask):
""" Upsample flow field [H/8, W/8, 2] -> [H, W, 2] using convex combination """
N, _, H, W = flow.shape
mask = mask.view(N, 1, 9, 8, 8, H, W)
mask = torch.softmax(mask, dim=2)
up_flow = F.unfold(8 * flow, [3,3], padding=1)
up_flow = up_flow.view(N, 2, 9, 1, 1, H, W)
up_flow = torch.sum(mask * up_flow, dim=2)
up_flow = up_flow.permute(0, 1, 4, 2, 5, 3)
return up_flow.reshape(N, 2, 8*H, 8*W)
def forward(self, image1, image2, iters=12, flow_init=None, upsample=True, test_mode=False):
2020-03-27 04:19:08 +01:00
""" Estimate optical flow between pair of frames """
image1 = 2 * (image1 / 255.0) - 1.0
image2 = 2 * (image2 / 255.0) - 1.0
2020-07-26 01:36:17 +02:00
image1 = image1.contiguous()
image2 = image2.contiguous()
2020-03-27 04:19:08 +01:00
hdim = self.hidden_dim
cdim = self.context_dim
# run the feature network
2020-07-26 01:36:17 +02:00
with autocast(enabled=self.args.mixed_precision):
fmap1, fmap2 = self.fnet([image1, image2])
fmap1 = fmap1.float()
fmap2 = fmap2.float()
if self.args.alternate_corr:
2020-08-29 01:18:41 +02:00
corr_fn = AlternateCorrBlock(fmap1, fmap2, radius=self.args.corr_radius)
else:
corr_fn = CorrBlock(fmap1, fmap2, radius=self.args.corr_radius)
2020-03-27 04:19:08 +01:00
# run the context network
2020-07-26 01:36:17 +02:00
with autocast(enabled=self.args.mixed_precision):
cnet = self.cnet(image1)
net, inp = torch.split(cnet, [hdim, cdim], dim=1)
net = torch.tanh(net)
inp = torch.relu(inp)
2020-03-27 04:19:08 +01:00
coords0, coords1 = self.initialize_flow(image1)
2020-07-26 01:36:17 +02:00
if flow_init is not None:
coords1 = coords1 + flow_init
2020-03-27 04:19:08 +01:00
flow_predictions = []
for itr in range(iters):
coords1 = coords1.detach()
corr = corr_fn(coords1) # index correlation volume
flow = coords1 - coords0
2020-07-26 01:36:17 +02:00
with autocast(enabled=self.args.mixed_precision):
net, up_mask, delta_flow = self.update_block(net, inp, corr, flow)
2020-03-27 04:19:08 +01:00
# F(t+1) = F(t) + \Delta(t)
coords1 = coords1 + delta_flow
2020-07-26 01:36:17 +02:00
# upsample predictions
if up_mask is None:
2020-03-27 04:19:08 +01:00
flow_up = upflow8(coords1 - coords0)
else:
2020-07-26 01:36:17 +02:00
flow_up = self.upsample_flow(coords1 - coords0, up_mask)
flow_predictions.append(flow_up)
2020-03-27 04:19:08 +01:00
2020-07-26 01:36:17 +02:00
if test_mode:
return coords1 - coords0, flow_up
2020-03-27 04:19:08 +01:00
return flow_predictions