Models · Semantic segmentation

mmseg CNN classics

The MMSegmentation authors' ADE20K checkpoints of FCN, PSPNet, DeepLabV3, DeepLabV3+ and HRNet, convolutional scene segmentation.

stable Fork of open-mmlab/mmsegmentation@c685fe6 · backends: portable, kernels

Checkpoint

Paper 2014 Added in 0.1.0 Evaluated with 0.2.0

Kernels, RTX 5060 Ti 16 GB 16.5 ms 1024x2048, fp16
vs the faster original 2.8× original fp16: 47.0 ms
Peak memory 288 MiB original: 1,925 MiB
Accuracy vs original fp32 −0.01 kernels mIoU %, ADE20K val (39.40 original)

Use it

pip install fastforks

Lean install with only what this model needs: pip install fastforks-core

from fastforks import segmentation

model = segmentation.load("fcn-r101-d8-ade20k")  # kernels on a supported GPU, else portable PyTorch
result = model(image)
result.logits
result.backend  # what ran: mode, device, dtype, kernels
fastforks bench fcn-r101-d8-ade20k

Times the model on this machine. load(…, mode="portable") or mode="kernels" picks a backend; usage guide.

Speed and memory

Every backend on one GPU (RTX 5060 Ti 16 GB) and the same input, the authors’ code included. Kernels runs fp16. Portable runs fp16 on CUDA.

BackenddtypeTime per call (ms)vs original fp16Peak memory
allocated / reserved
Load (s)First call (s)
Original fp32fp32 94.20.50×1,799 MiB/2,074 MiB––
Original fp16torch.autocast(fp16) around the forwardfp16 47.0–1,925 MiB/2,102 MiB––
Portable fp32fp32not measured ––––
Portablefp16 40.71.2×414 MiB/472 MiB1.40.04
Kernelsfp16 16.52.8×288 MiB/374 MiB4.30.02
Portable + torch.compile 35.61.3×414 MiB/472 MiB9 (build)–
TensorRT fp16fp16 27.41.7×882 MiBenginenot comparable22 (build)–

Measured on RTX 5060 Ti 16 GB, torch 2.14.1+cu130, 2026-11-03 to 2026-11-16. First call, 3 warm-up, peak reset, median of 3 rounds x median of 5. Ratios divide two times of this table (same GPU, same input). Original fp16 is the authors’ own precision option, else torch.autocast. torch.compile and TensorRT: build time instead of load time. TensorRT memory (engine, not comparable) is the GPU memory the engine took, read from the driver because TensorRT allocates outside PyTorch, so it can't be compared with the PyTorch peaks of the other rows.

Other devices

DeviceBackenddtypems / callPeak memoryOutput vs referenceInput
Apple M1 Pro 16 GB MPS, torch 2.14.1, 2026-11-12Originalfp32cannot run here–1024x2048
Portablefp16405 1,875 MiB processmatches1024x2048
Portablefp32468 2,978 MiB processmatches1024x2048
Kernelsfp16not implemented–1024x2048
NVIDIA L4 CUDA, torch 2.14.1+cu130, 2026-11-07Portablefp1640.3 414 MiB/472 MiB–1024x2048
Kernelsfp1621.8 288 MiB/374 MiB–1024x2048
NVIDIA L40S CUDA, torch 2.14.1+cu130, 2026-11-07Portablefp1612.6 414 MiB/472 MiB–1024x2048
Kernelsfp166.92 288 MiB/374 MiB–1024x2048
NVIDIA A10 CUDA, torch 2.14.1+cu130, 2026-11-07Portablefp1633.8 414 MiB/472 MiB–1024x2048
Kernelsfp1619.5 288 MiB/374 MiB–1024x2048
NVIDIA A100-SXM4-40GB CUDA, torch 2.14.1+cu130, 2026-11-07Portablefp1615.5 414 MiB/472 MiB–1024x2048
Kernelsfp1611.8 288 MiB/374 MiB–1024x2048
NVIDIA RTX PRO 6000 CUDA, torch 2.14.1+cu130, 2026-11-08Portablefp167.82 414 MiB/472 MiB–1024x2048
Kernelsfp165.21 288 MiB/374 MiB–1024x2048

CUDA memory is torch’s peak allocated / reserved; Apple has no such counter, so the Mac rows show the process footprint (weights, Python and torch included). “Output vs reference” compares the device’s output with the stored golden output under the card’s tolerance; – = not checked on that device. Kernels “not implemented” = no kernels for this model on that backend yet: it runs portable there.

Licence and weights

Commercial use allowed

Code
Apache-2.0 LICENSE
Weights
Apache-2.0 source
Training data

Our reading of the licences, not legal advice.

Weights

fcn-r101-d8-ade20k.pth 275 MB· hosted on OpenMMLab

fastforks downloads and checks these files on first use (fastforks download fcn-r101-d8-ade20k).

Accuracy

The test set through the authors’ code and through ours: same checkpoint, same protocol. The distance between the original’s fp32 and fp16 runs is what half precision alone costs; our runs should stay within it.

ADE20K val

every 4th image, mIoU / aAcc %

RundtypemIoU % ↑aAcc % ↑Δ mIoU % vs fp32 originalimagesms / image
Original fp32fp3239.4079.09 reference 500 57.8
Original fp16fp1639.4079.10 ±0.00 500 33.0
Portable fp32fp3239.4079.10 ±0.00 500 56.4
Portablefp1639.3979.10 −0.01 500 27.1
Kernelsfp1639.3979.09 −0.01 500 11.5

Each image at its own size, argmax of the logits. Original = the authors’ code, same checkpoint and protocol. Measured on original fp32, original fp16: RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-06; portable fp32: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 44640884, 2026-11-17; portable, kernels: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 939adfe8, 2026-11-06.

Closeness to the original

Beyond the test-set scores: the raw output of ours against the original’s fp32 output on the same input, pixel by pixel. CPU, float32: max |diff| <= 0.002 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.04 and p99 <= 0.15 (absolute error of the output logits).

CheckToleranceMeasured
CPU fp32, max |diff|≤ 0.002 not recorded
GPU default dtypes, mean |diff|≤ 0.04 0.00228
GPU default dtypes, p99 |diff|≤ 0.15 0.00975

Recorded for fcn-r101-d8-ade20k on RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-04, with these kernels: cudnn_hacccutlass_convigevseg_convseg_up

Details

The mmsegmentation ADE20K checkpoints of the CNN classics: FCN, PSPNet, DeepLabV3 and DeepLabV3+ on a dilated ResNet-101 (output stride 8), and HRNetV2-W48 with an FCN head; networks vendored from mmsegmentation (no mmseg / mmcv dependency).

Input / output
Input: one RGB image. Output: per-pixel class labels (logits with logits=True).
Input size
any H x W >= 1, resized or padded inside the model; tested up to 3000x4000 (7,407 MiB allocated)
Outputs
logits
Checkpoint
arch="fcn-r101-d8"
CUDA kernels
  • igevIGEV stereo: convolution epilogues, correlation lookups, GRU gates
  • cutlass_convCUTLASS convolutions with the recurrent update (GRU) fused in
  • cudnn_haccconvolutions with fp16 accumulation (cuDNN)
  • seg_convfused convolutions for segmentation CNNs
  • seg_upfused resize and argmax of segmentation scores
Apple (MPS)
fp32, fp16
Full reference, generated from the repository

The MMSegmentation authors' ADE20K checkpoints of FCN, PSPNet, DeepLabV3, DeepLabV3+ and HRNet, convolutional scene segmentation.

Commercial use: yes.

variantinputbackendtime (ms)peak memory (MiB, allocated / reserved)
fcn-r101-d8-ade20ksynthetic (1024x2048)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
94.2
47.0
40.7
16.5
35.6
27.4
1799 / 2074
1925 / 2102
414 / 472
288 / 374
414 / 472
1656 / 1734
pspnet-r101-d8-ade20ksynthetic (1024x2048)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
92.5
47.2
48.7
24.9
41.3
29.8
1797 / 2074
1922 / 2222
414 / 472
335 / 394
414 / 472
1859 / 1898
deeplabv3-r101-d8-ade20ksynthetic (1024x2048)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
108
54.8
48.0
20.5
42.1
30.6
1870 / 2134
2031 / 2162
450 / 492
356 / 414
450 / 512
1764 / 1826
deeplabv3plus-r101-d8-ade20ksynthetic (1024x2048)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
104
53.4
47.1
17.9
44.6
30.9
1776 / 2036
1891 / 2024
410 / 452
355 / 436
410 / 432
1683 / 1732
hrnet-w48-fcn-ade20ksynthetic (1024x2048)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
48.1
54.8
27.3
11.2
23.9
20.9
1789 / 1870
1915 / 2010
421 / 470
265 / 344
421 / 448
1662 / 1712
Details
  • Original code: https://github.com/open-mmlab/mmsegmentation at c685fe6767c4
  • Code licence: Apache-2.0 (permissive); files: LICENSE
  • Weights licence: Apache-2.0
  • Licence evidence: https://github.com/open-mmlab/mmsegmentation/blob/c685fe6767c4cadf6b051983ca6208f1b9d1ccb8/README.md#license
  • Training data: ImageNet-1k (ILSVRC 2012 classification), ADE20K (SceneParsing, 150 classes)
  • Kernel libraries: igev, cutlass_conv, cudnn_hacc, seg_conv, seg_up
  • Variants: fcn-r101-d8-ade20k, pspnet-r101-d8-ade20k, deeplabv3-r101-d8-ade20k, deeplabv3plus-r101-d8-ade20k, hrnet-w48-fcn-ade20k
  • Outputs: logits; label_set: ade20k-150
  • Input size: any H x W >= 1, resized or padded inside the model
  • Tested up to (portable, fp16, RTX 5060 Ti 16 GB): fcn-r101-d8-ade20k 3000x4000 (7.2 GiB peak allocated, 7.5 reserved); pspnet-r101-d8-ade20k 3000x4000 (7.2 GiB peak allocated, 7.5 reserved); deeplabv3-r101-d8-ade20k 3000x4000 (7.3 GiB peak allocated, 7.5 reserved); deeplabv3plus-r101-d8-ade20k 3000x4000 (7.2 GiB peak allocated, 7.5 reserved); hrnet-w48-fcn-ade20k 3000x4000 (7.2 GiB peak allocated, 7.5 reserved)
  • fcn-r101-d8-ade20k 1024x2048, first call: portable + torch.compile 9 s (cold Inductor cache), TensorRT fp16 22 s (ONNX + engine).
  • pspnet-r101-d8-ade20k 1024x2048, first call: portable + torch.compile 11 s (cold Inductor cache), TensorRT fp16 44 s (ONNX + engine).
  • deeplabv3-r101-d8-ade20k 1024x2048, first call: portable + torch.compile 10 s (cold Inductor cache), TensorRT fp16 26 s (ONNX + engine).
  • deeplabv3plus-r101-d8-ade20k 1024x2048, first call: portable + torch.compile 11 s (cold Inductor cache), TensorRT fp16 23 s (ONNX + engine).
  • hrnet-w48-fcn-ade20k 1024x2048, first call: portable + torch.compile 36 s (cold Inductor cache), TensorRT fp16 30 s (ONNX + engine).
  • fcn-r101-d8-ade20k original 1024x2048: torch.autocast(fp16) around the forward
  • pspnet-r101-d8-ade20k original 1024x2048: torch.autocast(fp16) around the forward
  • deeplabv3-r101-d8-ade20k original 1024x2048: torch.autocast(fp16) around the forward
  • deeplabv3plus-r101-d8-ade20k original 1024x2048: torch.autocast(fp16) around the forward
  • hrnet-w48-fcn-ade20k original 1024x2048: torch.autocast(fp16) around the forward
  • fcn-r101-d8-ade20k load / first call: original 4.72 s / 0.393 s (512x512), portable 1.42 s / 0.102 s (512x512), kernels 4.35 s / 0.02 s (512x512)
  • pspnet-r101-d8-ade20k load / first call: original 1.35 s / 0.092 s (512x512), portable 1.42 s / 0.296 s (512x512), kernels 1.59 s / 0.025 s
  • deeplabv3-r101-d8-ade20k load / first call: original 1.64 s / 0.031 s (512x512), portable 2.28 s / 0.112 s (512x512), kernels 1.92 s / 0.033 s (512x512)
  • deeplabv3plus-r101-d8-ade20k load / first call: original 4.72 s / 0.662 s (512x512), portable 1.11 s / 0.111 s (512x512), kernels 1.15 s / 0.022 s (512x512)
  • hrnet-w48-fcn-ade20k load / first call: original 1.48 s / 0.079 s (512x512), portable 1.18 s / 0.107 s (512x512), kernels 1.52 s / 0.025 s (512x512)
  • Protocol, first call, 3 warm-up, peak reset, median of 3 rounds x median of 5

Measured

Each image at its own size, argmax of the logits. Original = the original code named under Details, same checkpoint and protocol.

variantdatasetoriginal fp32original fp16portableportable fp32kernelsnote
fcn-r101-d8-ade20kADE20K val, every 4th image, mIoU / aAcc %39.40 / 79.0939.40 / 79.1039.39 / 79.1039.40 / 79.1039.39 / 79.09
pspnet-r101-d8-ade20kADE20K val, every 4th image, mIoU / aAcc %44.44 / 80.9744.44 / 80.9744.43 / 80.9744.44 / 80.9744.44 / 80.97
deeplabv3-r101-d8-ade20kADE20K val, every 4th image, mIoU / aAcc %43.62 / 80.8043.62 / 80.8043.61 / 80.8043.62 / 80.8043.62 / 80.80
deeplabv3plus-r101-d8-ade20kADE20K val, every 4th image, mIoU / aAcc %44.36 / 81.0144.36 / 81.0144.35 / 81.0144.36 / 81.0144.36 / 81.01
hrnet-w48-fcn-ade20kADE20K val, every 4th image, mIoU / aAcc %40.11 / 79.3440.11 / 79.3540.10 / 79.3540.11 / 79.3440.11 / 79.34

Measured on portable: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 939adfe8, 2026-11-06; portable32: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 44640884, 2026-11-17; kernels: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 939adfe8, 2026-11-06.

Parity

CPU, float32: max |diff| <= 0.002 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.04 and p99 <= 0.15 (absolute error of the output logits).

variantkernels error vs original, mean / p99GPU, torchkernels used
fcn-r101-d8-ade20k0.00228 / 0.00975RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120, seg_conv@8eec95783e28:sm_120, seg_up@3e44e71200cd:sm_120
pspnet-r101-d8-ade20k0.00713 / 0.02575RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120, seg_conv@8eec95783e28:sm_120, seg_up@3e44e71200cd:sm_120
deeplabv3-r101-d8-ade20k0.00223 / 0.00945RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120, seg_conv@8eec95783e28:sm_120, seg_up@3e44e71200cd:sm_120
deeplabv3plus-r101-d8-ade20k0.00313 / 0.01297RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120, seg_conv@8eec95783e28:sm_120, seg_up@3e44e71200cd:sm_120
hrnet-w48-fcn-ade20k0.00175 / 0.00723RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120, seg_conv@8eec95783e28:sm_120, seg_up@3e44e71200cd:sm_120

Example output

One fixed input through this checkpoint, the same for every model of the task (kernels · float16 · NVIDIA GeForce RTX 5060 Ti).

Input
Output
  • floor 26%
  • table 25%
  • sink 24%
  • tray 18%
  • plate 5.6%
  • wall 1.2%
A cup of coffee, input
Input · 100%

Input: A cup of coffee, Rachel Michetti (CC0).