memflow / README.md
Flor's picture
card: update from fastforks 1cbf1671
51f667d verified
|
Raw History Blame Contribute Delete
7.76 kB
metadata
library_name: fastforks
pipeline_tag: image-to-image
license: apache-2.0
tags:
  - flow
  - optical-flow
  - motion-estimation
  - cuda
  - pytorch
  - fastforks
  - fastforks-card

MemFlow (fastforks)

MemFlow (Dong & Fu, CVPR 2024): optical flow with a memory of past motion. A RAFT-style network at 1/8 resolution: BasicEncoder feature and context networks, an all-pairs correlation pyramid, a motion encoder and update block of depthwise-separable large-kernel blocks (SKFlow) instead of a GRU, and an attention readout of the motion features of memory frames; convex 8x upsampling. The authors' released weights; the network is ported from DQiaole/MemFlow. A pair (image1, image2) runs as upstream's first frame of a video (empty memory: the readout attends over the pair's own motion features). Paper (video with memory, val_decoder_depth 15): Sintel test EPE 1.05 (clean) / 1.91 (final), KITTI 2015 test Fl-all 4.10 %.

This repo is the model card only: it holds no weights. fastforks downloads the authors' weights from the links below and checks them against the SHA-256 in the model card. Code: https://github.com/kleinicke/fastforks. Kernels for the kernels backend: https://huggingface.co/fastforks/kernels. Original code: https://github.com/DQiaole/MemFlow at 5486fb88a55b.

Task: flow. Tier: experimental. Outputs: flow. Backends: portable, kernels.

Input: two RGB images (consecutive frames). Output: per-pixel motion from the first to the second (dx, dy in pixels).

Variants

Weights are the authors' files (not hosted here). SHA-256 shortened to 12 hex digits; the full hashes are in card.toml.

variant weights (author link) SHA-256 weights licence commercial use
memflow@1 MemFlowNet_things.pth 694cc33c1ac5 Apache-2.0 yes
memflow-kitti@1 MemFlowNet_kitti.pth c3ecadb24486 Apache-2.0 yes
memflow-sintel@1 MemFlowNet_sintel.pth 47c0ecd56761 Apache-2.0 yes

Licence

Original-code strings, recorded as they are (research-use package; check them before any other use). Code: Apache-2.0. Licence files: LICENSE.

Training data: FlyingChairs (Freiburg), Freiburg SceneFlow (FlyingThings3D, Driving, Monkaa) (FlyingThings3D (part of SceneFlow), starting from SKFlow's FlyingChairs + FlyingThings3D checkpoint, as upstream's train.sh).

Licence evidence: https://github.com/DQiaole/MemFlow/blob/5486fb88a55b04c2e2b292b3abaa7641d9988831/LICENSE

Input size

any H x W >= 128, padded to a multiple of 8 Tested up to (portable, fp16, RTX 5060 Ti 16 GB): memflow 1024x2048 (8.2 GiB peak allocated, 9.2 reserved); memflow-kitti 1024x2048 (8.2 GiB peak allocated, 9.2 reserved); memflow-sintel 1024x2048 (8.2 GiB peak allocated, 9.2 reserved)

Speed and memory

Flow, synthetic

model input time (ms): original fp32 original portable kernels kernels speed-up: vs original vs portable peak memory (MiB, allocated / reserved): original portable kernels
memflow 1080x1920 – 3116 964 865 3.6x 1.1x 15061 / 15594 8265 / 8382 968 / 1022
memflow-kitti 1080x1920 – 3247 965 865 3.8x 1.1x 15061 / 15594 8265 / 8382 968 / 1022
memflow-sintel 1080x1920 – 3214 964 865 3.7x 1.1x 15061 / 15634 8265 / 8382 968 / 1022

Startup and first call

model load (s): original portable kernels first call (s): original portable kernels
memflow 0.16 0.06 1.18 3.82 0.97 0.88
memflow-kitti 0.14 0.14 0.97 3.06 0.98 0.88
memflow-sintel 0.14 0.10 0.95 3.06 0.98 0.88

torch.compile and TensorRT

model input time: portable compile TensorRT fp16 kernels preparation: compile (cold first call) ONNX + engine build peak MiB (alloc / reserved): portable compile kernels TensorRT (engine + torch)
memflow 1080x1920 964 ms 1642 ms fails: TensorRT engine build failed 865 ms 78.9 s – 8265 / 8382 6813 / 6946 968 / 1022 –
memflow-kitti 1080x1920 965 ms 1644 ms fails: TensorRT engine build failed 865 ms 72.6 s – 8265 / 8382 6813 / 6946 968 / 1022 –
memflow-sintel 1080x1920 964 ms 1640 ms fails: TensorRT engine build failed 865 ms 76.0 s – 8265 / 8382 6813 / 6946 968 / 1022 –

Fixed-size cells: time (ms) / peak allocated / peak reserved memory (MiB).

Hardware, protocol and caveats: benchmarks (RTX 5060 Ti 16 GB, sm_120). Original = the original code named above at the precision of portable (fp16, or (fp32) where fp16 is unsupported or not finite); original fp32 = its default fp32 run. Peak memory is the torch allocator's.

Accuracy

On public datasets

Frames _10 -> _11 at full size against the flow_occ ground truth; EPE = mean of the per-pair mean end-point error, Fl-all = share of all valid pixels with an error > 3 px and > 5 % of the GT magnitude. Original = the original code named under Details, same checkpoint and protocol.

variant dataset original fp32 original fp16 portable kernels note
memflow KITTI 2015 train, EPE / Fl-all % 4.108 / 14.43 4.107 / 14.43 4.109 / 14.44 4.103 / 14.42
memflow-kitti KITTI 2015 train, EPE / Fl-all % 0.600 / 1.21 0.600 / 1.21 0.600 / 1.21 0.600 / 1.21 trained on this dataset's public images; not a zero-shot result

Measured on NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks db9e79d5+uncommitted, 2026-11-13.

Parity with the original code

Every variant is checked against reference outputs of the original code at its pinned commit. CPU, float32: max |diff| <= 4.2e-05 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.0018 and p99 <= 0.0065 (absolute error of the output flow).

Limitations

  • Pairs only: the video memory (the keys and values of earlier frames that upstream's inference keeps) is not exposed; each pair runs as the first frame of a video.
  • The MemFlow-T checkpoints (Twins-SVT encoders) and MemFlow-P (flow prediction) are not ported.

Usage

from fastforks import flow

model = flow.load("memflow")  # mode="auto": the preferred backend from recorded measurements
result = model("frame1.png", "frame2.png")
result.flow  # [2, H, W] float32 tensor: dx, dy in input pixels
result.backend  # what ran, e.g. Backend(mode='portable', device='cuda:0', dtype='float16')

Install with pip install fastforks (see https://github.com/kleinicke/fastforks). Other variants: memflow, memflow-kitti, memflow-sintel.

Kernels backend

flow.load("memflow", mode="kernels") runs our CUDA kernels (raft_flow). It needs a prebuilt kernel for your GPU architecture (downloaded from https://huggingface.co/fastforks/kernels on first use) and a recent NVIDIA driver; if there is none it raises BackendUnavailableError with the reason and never falls back silently. mode="auto" picks kernels when it can run and logs the choice.