CoAtNet
CoAtNet (Dai et al. 2021; timm's `rw` networks nano to 2, ImageNet-1k, some pretrained on ImageNet-12k): convolution stages followed by attention stages. Apache-2.0.
experimental Fork of huggingface/pytorch-image-models@0df212b · backends: portable, kernels
- Paper CoAtNet: Marrying Convolution and Attention for All Data Sizes NeurIPS 2021 · first released 9 Jun 2021
Paper 2021 Added in 0.2.0 Evaluated with 0.2.0
Use it
pip install fastforks Lean install with only what this model needs: pip install "fastforks-core[timm]"
from fastforks import classification
model = classification.load("coatnet_nano_rw_224.sw_in1k") # kernels on a supported GPU, else portable PyTorch
result = model(image)
result.logits
result.backend # what ran: mode, device, dtype, kernels fastforks bench coatnet_nano_rw_224.sw_in1k 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.
Input: synthetic, 224x224 (the task’s standard size)
| Backend | dtype | Time per call (ms) | vs original fp32 | Peak memory allocated / reserved | Load (s) | First call (s) |
|---|---|---|---|---|---|---|
| Original fp32 | fp32 | 7.88 | – | 105 MiB/114 MiB | 1.6 | 0.36 |
| Original fp16torch.autocast(float16) around the call | fp16 | 10.6 | 0.75× | 124 MiB/146 MiB | 0.3 | 0.54 |
| Portable fp32 | fp32 | not measured | – | – | – | – |
| Portable | fp16 | 7.23 | 1.1× | 74 MiB/86 MiB | 1.4 | 0.42 |
| Kernels | not measured | – | – | – | – | |
| Portable + torch.compile | 4.71 | 1.7× | 96 MiB/126 MiB | 14 (build) | – | |
| TensorRT fp16 | fp16 | 0.709 | 11× | 38 MiBenginenot comparable | 32 (build) | – |
Measured on RTX 5060 Ti 16 GB, torch 2.14.1+cu130, 2026-11-11. 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
| Device | Backend | dtype | ms / call | Peak memory | Output vs reference | Input |
|---|---|---|---|---|---|---|
| Apple M1 Pro 16 GB MPS, torch 2.14.1, 2026-11-14 | Original | fp32 | 19.8 | 596 MiB process | – | 224x224 |
| Portable | fp16 | 19.6 | 572 MiB process | matches | 224x224 | |
| Portable | fp32 | 19.6 | 564 MiB process | matches | 224x224 | |
| Kernels | fp32 | not implemented | – | 224x224 | ||
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
- Weights
- Apache-2.0 source
- Training data
-
- ImageNet-1k (ILSVRC 2012 classification) noncommercial, non-commercial
Our reading of the licences, not legal advice.
Weights
coatnet_nano_rw_224.sw_in1k.safetensors 60.6 MB· hosted on Hugging Facefastforks downloads and checks these files on first use (fastforks download coatnet_nano_rw_224.sw_in1k).
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.
ImageNet-1k val
5000 random images, top-1 / top-5 %
| Run | dtype | top-1 % ↑ | top-5 % ↑ | Δ top-1 % vs fp32 original | images | ms / image |
|---|---|---|---|---|---|---|
| Original fp32 | fp32 | 82.58 | 95.82 | reference | 5000 | 1.59 |
| Original fp16 | fp16 | 82.58 | 95.80 | ±0.00 | 5000 | 0.970 |
| Portable fp32 | fp32 | 82.60 | 95.78 | +0.02 | 5000 | 2.05 |
| Portable | fp16 | 82.56 | 95.78 | −0.02 | 5000 | 0.912 |
| Kernels | not measured | |||||
fastforks: the model's own resize / crop on the GPU; original: the library's pretrained loader and its PIL eval transform. Original = the authors’ code, same checkpoint and protocol. Measured on original fp32, original fp16: RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-11; portable fp32: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 6f0a70e1, 2026-11-17; portable: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 7ba7307f, 2026-11-11.
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| <= 8.6e-06 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.0024 and p99 <= 0.011 (absolute error of the output logits).
| Check | Tolerance | Measured |
|---|---|---|
| CPU fp32, max |diff| | ≤ 0.0000086 | not recorded |
| GPU default dtypes, mean |diff| | ≤ 0.0024 | 0.00121 |
| GPU default dtypes, p99 |diff| | ≤ 0.011 | 0.0057 |
Recorded for coatnet_nano_rw_224.sw_in1k on RTX 5060 Ti, torch 2.10.0+cu130, 2026-11-17, with these kernels: cudnn_haccfvm_convfvm_extrafvm_mbfvm_r3
Details
CoAtNet (Dai et al. 2021) in timm's `rw` configurations: MBConv stages followed by transformer stages with relative position bias.
- Input / output
- Input: one RGB image. Output: ImageNet-1k class logits.
- Input size
- fixed 224x224: the input is resized (short side 248) and centre-cropped to it; tested up to 224x224 (74 MiB allocated)
- Outputs
logits- Checkpoint
arch="coatnet_nano_rw_224.sw_in1k"- CUDA kernels
cudnn_haccconvolutions with fp16 accumulation (cuDNN)fvm_convfused convolutions for image classifiersfvm_extrafused transformer row operations (rotary split, SwiGLU with LayerNorm)fvm_mbfused MBConv blocks (MobileNet and EfficientNet families)fvm_r3further fused classifier blocks
- Apple (MPS)
- fp32, fp16
Full reference, generated from the repository
CoAtNet (Dai et al. 2021; timm's rw networks nano to 2, ImageNet-1k, some pretrained on ImageNet-12k): convolution stages followed by attention stages. Apache-2.0.
Commercial use: yes.
| variant | input | backend | time (ms) | peak memory (MiB, allocated / reserved) |
|---|---|---|---|---|
coatnet_nano_rw_224.sw_in1k | synthetic (224x224) | original (fp32) original portable kernels portable + torch.compile TensorRT fp16 | 7.88 10.6 7.23 – 4.71 0.709 | 105 / 114 124 / 146 74 / 86 – 96 / 126 102 / 124 |
coatnet_rmlp_nano_rw_224.sw_in1k | synthetic (224x224) | original (fp32) original portable kernels portable + torch.compile TensorRT fp16 | 8.95 12.1 8.09 – 9.89 0.868 | 121 / 154 124 / 146 78 / 106 – 110 / 126 122 / 144 |
coatnet_0_rw_224.sw_in1k | synthetic (224x224) | original (fp32) original portable kernels portable + torch.compile TensorRT fp16 | 7.64 9.59 6.49 – 3.59 0.849 | 156 / 180 194 / 210 102 / 124 – 120 / 142 158 / 170 |
coatnet_1_rw_224.sw_in1k | synthetic (224x224) | original (fp32) original portable kernels portable + torch.compile TensorRT fp16 | 12.3 15.7 10.7 – 5.58 1.38 | 225 / 254 277 / 294 132 / 158 – 149 / 176 231 / 250 |
coatnet_rmlp_1_rw_224.sw_in1k | synthetic (224x224) | original (fp32) original portable kernels portable + torch.compile TensorRT fp16 | 14.0 18.4 12.4 – 6.54 1.47 | 213 / 234 278 / 296 131 / 156 – 155 / 176 243 / 262 |
coatnet_rmlp_1_rw2_224.sw_in12k_ft_in1k | synthetic (224x224) | original (fp32) original portable kernels portable + torch.compile TensorRT fp16 | 13.9 18.3 12.3 – 6.63 1.55 | 226 / 254 278 / 296 132 / 158 – 162 / 176 257 / 276 |
coatnet_2_rw_224.sw_in12k_ft_in1k | synthetic (224x224) | original (fp32) original portable kernels portable + torch.compile TensorRT fp16 | 14.6 15.7 11.9 – 5.53 1.73 | 376 / 410 466 / 470 207 / 240 – 211 / 240 374 / 416 |
coatnet_rmlp_2_rw_224.sw_in1k | synthetic (224x224) | original (fp32) original portable kernels portable + torch.compile TensorRT fp16 | 15.3 18.6 15.1 – 6.34 1.92 | 376 / 410 466 / 472 207 / 240 – 238 / 260 404 / 446 |
coatnet_rmlp_2_rw_224.sw_in12k_ft_in1k | synthetic (224x224) | original (fp32) original portable kernels portable + torch.compile TensorRT fp16 | 15.3 18.6 12.8 – 6.52 1.87 | 376 / 410 466 / 470 207 / 240 – 238 / 260 402 / 444 |
coatnet_rmlp_2_rw_384.sw_in12k_ft_in1k | synthetic (384x384) | original (fp32) original portable kernels portable + torch.compile TensorRT fp16 | 27.9 17.9 13.9 – 6.92 3.69 | 520 / 566 516 / 528 296 / 330 – 326 / 370 464 / 500 |
Details
- Original code: https://github.com/huggingface/pytorch-image-models at
0df212b369a5 - Code licence: Apache-2.0 (permissive); files:
- Weights licence: Apache-2.0
- Licence evidence: https://huggingface.co/timm/coatnet_nano_rw_224.sw_in1k/blob/89fb9ccadc876a973c6ff954b089eb6723174c7a/README.md
- Training data: ImageNet-1k (ILSVRC 2012 classification)
coatnet_rmlp_1_rw2_224.sw_in12k_ft_in1k: ImageNet-21k / ImageNet-22k (full ImageNet, fall 2011 release), ImageNet-1k (ILSVRC 2012 classification)coatnet_2_rw_224.sw_in12k_ft_in1k: ImageNet-21k / ImageNet-22k (full ImageNet, fall 2011 release), ImageNet-1k (ILSVRC 2012 classification)coatnet_rmlp_2_rw_224.sw_in12k_ft_in1k: ImageNet-21k / ImageNet-22k (full ImageNet, fall 2011 release), ImageNet-1k (ILSVRC 2012 classification)coatnet_rmlp_2_rw_384.sw_in12k_ft_in1k: ImageNet-21k / ImageNet-22k (full ImageNet, fall 2011 release), ImageNet-1k (ILSVRC 2012 classification)
- Kernel libraries:
cudnn_hacc,fvm_conv,fvm_extra,fvm_mb,fvm_r3 - Variants:
coatnet_nano_rw_224.sw_in1k(default ofcoatnet_nano_rw_224),coatnet_rmlp_nano_rw_224.sw_in1k(default ofcoatnet_rmlp_nano_rw_224),coatnet_0_rw_224.sw_in1k(default ofcoatnet_0_rw_224),coatnet_1_rw_224.sw_in1k(default ofcoatnet_1_rw_224),coatnet_rmlp_1_rw_224.sw_in1k(default ofcoatnet_rmlp_1_rw_224),coatnet_rmlp_1_rw2_224.sw_in12k_ft_in1k(default ofcoatnet_rmlp_1_rw2_224),coatnet_2_rw_224.sw_in12k_ft_in1k(default ofcoatnet_2_rw_224),coatnet_rmlp_2_rw_224.sw_in1k,coatnet_rmlp_2_rw_224.sw_in12k_ft_in1k(default ofcoatnet_rmlp_2_rw_224),coatnet_rmlp_2_rw_384.sw_in12k_ft_in1k(default ofcoatnet_rmlp_2_rw_384) - Outputs: logits; label_set:
imagenet-1k - Input size:
coatnet_nano_rw_224.sw_in1k,coatnet_rmlp_nano_rw_224.sw_in1k: fixed 224x224: the input is resized (short side 248) and centre-cropped to itcoatnet_0_rw_224.sw_in1k,coatnet_1_rw_224.sw_in1k,coatnet_rmlp_1_rw_224.sw_in1k,coatnet_rmlp_1_rw2_224.sw_in12k_ft_in1k,coatnet_2_rw_224.sw_in12k_ft_in1k,coatnet_rmlp_2_rw_224.sw_in1k,coatnet_rmlp_2_rw_224.sw_in12k_ft_in1k: fixed 224x224: the input is resized (short side 235) and centre-cropped to itcoatnet_rmlp_2_rw_384.sw_in12k_ft_in1k: fixed 384x384: the input is resized (short side 384) and centre-cropped to it
- Tested up to (portable, fp16, RTX 5060 Ti 16 GB):
coatnet_nano_rw_224.sw_in1k224x224 (0.1 GiB peak allocated, 0.1 reserved);coatnet_rmlp_nano_rw_224.sw_in1k224x224 (0.1 GiB peak allocated, 0.1 reserved);coatnet_0_rw_224.sw_in1k224x224 (0.1 GiB peak allocated, 0.1 reserved);coatnet_1_rw_224.sw_in1k224x224 (0.1 GiB peak allocated, 0.2 reserved);coatnet_rmlp_1_rw_224.sw_in1k224x224 (0.1 GiB peak allocated, 0.2 reserved);coatnet_rmlp_1_rw2_224.sw_in12k_ft_in1k224x224 (0.1 GiB peak allocated, 0.2 reserved);coatnet_2_rw_224.sw_in12k_ft_in1k224x224 (0.2 GiB peak allocated, 0.2 reserved);coatnet_rmlp_2_rw_224.sw_in1k224x224 (0.2 GiB peak allocated, 0.2 reserved);coatnet_rmlp_2_rw_224.sw_in12k_ft_in1k224x224 (0.2 GiB peak allocated, 0.2 reserved);coatnet_rmlp_2_rw_384.sw_in12k_ft_in1k384x384 (0.3 GiB peak allocated, 0.3 reserved) coatnet_nano_rw_224.sw_in1k224x224, first call: portable + torch.compile 14 s (cold Inductor cache), TensorRT fp16 32 s (ONNX + engine).coatnet_rmlp_nano_rw_224.sw_in1k224x224, first call: portable + torch.compile 14 s (cold Inductor cache), TensorRT fp16 37 s (ONNX + engine).coatnet_0_rw_224.sw_in1k224x224, first call: portable + torch.compile 12 s (cold Inductor cache), TensorRT fp16 38 s (ONNX + engine).coatnet_1_rw_224.sw_in1k224x224, first call: portable + torch.compile 16 s (cold Inductor cache), TensorRT fp16 42 s (ONNX + engine).coatnet_rmlp_1_rw_224.sw_in1k224x224, first call: portable + torch.compile 19 s (cold Inductor cache), TensorRT fp16 44 s (ONNX + engine).coatnet_rmlp_1_rw2_224.sw_in12k_ft_in1k224x224, first call: portable + torch.compile 18 s (cold Inductor cache), TensorRT fp16 43 s (ONNX + engine).coatnet_2_rw_224.sw_in12k_ft_in1k224x224, first call: portable + torch.compile 16 s (cold Inductor cache), TensorRT fp16 36 s (ONNX + engine).coatnet_rmlp_2_rw_224.sw_in1k224x224, first call: portable + torch.compile 18 s (cold Inductor cache), TensorRT fp16 46 s (ONNX + engine).coatnet_rmlp_2_rw_224.sw_in12k_ft_in1k224x224, first call: portable + torch.compile 18 s (cold Inductor cache), TensorRT fp16 41 s (ONNX + engine).coatnet_rmlp_2_rw_384.sw_in12k_ft_in1k384x384, first call: portable + torch.compile 19 s (cold Inductor cache), TensorRT fp16 42 s (ONNX + engine).coatnet_nano_rw_224.sw_in1koriginal 224x224: torch.autocast(float16) around the callcoatnet_rmlp_nano_rw_224.sw_in1koriginal 224x224: torch.autocast(float16) around the callcoatnet_0_rw_224.sw_in1koriginal 224x224: torch.autocast(float16) around the callcoatnet_1_rw_224.sw_in1koriginal 224x224: torch.autocast(float16) around the callcoatnet_rmlp_1_rw_224.sw_in1koriginal 224x224: torch.autocast(float16) around the callcoatnet_rmlp_1_rw2_224.sw_in12k_ft_in1koriginal 224x224: torch.autocast(float16) around the callcoatnet_2_rw_224.sw_in12k_ft_in1koriginal 224x224: torch.autocast(float16) around the callcoatnet_rmlp_2_rw_224.sw_in1koriginal 224x224: torch.autocast(float16) around the callcoatnet_rmlp_2_rw_224.sw_in12k_ft_in1koriginal 224x224: torch.autocast(float16) around the callcoatnet_rmlp_2_rw_384.sw_in12k_ft_in1koriginal 384x384: torch.autocast(float16) around the callcoatnet_nano_rw_224.sw_in1kload / first call: original 0.28 s / 0.539 s, portable 1.37 s / 0.422 scoatnet_rmlp_nano_rw_224.sw_in1kload / first call: original 0.26 s / 0.187 s, portable 1.36 s / 0.393 scoatnet_0_rw_224.sw_in1kload / first call: original 0.38 s / 0.312 s, portable 1.37 s / 0.378 scoatnet_1_rw_224.sw_in1kload / first call: original 0.47 s / 0.15 s, portable 1.44 s / 0.366 scoatnet_rmlp_1_rw_224.sw_in1kload / first call: original 0.47 s / 0.022 s, portable 1.47 s / 0.373 scoatnet_rmlp_1_rw2_224.sw_in12k_ft_in1kload / first call: original 0.47 s / 0.022 s, portable 1.49 s / 0.366 scoatnet_2_rw_224.sw_in12k_ft_in1kload / first call: original 0.74 s / 0.314 s, portable 1.52 s / 0.505 scoatnet_rmlp_2_rw_224.sw_in1kload / first call: original 0.71 s / 0.022 s, portable 1.54 s / 0.506 scoatnet_rmlp_2_rw_224.sw_in12k_ft_in1kload / first call: original 0.71 s / 0.022 s, portable 1.8 s / 0.504 scoatnet_rmlp_2_rw_384.sw_in12k_ft_in1kload / first call: original 0.74 s / 0.053 s (384x384), portable 1.52 s / 0.435 s (384x384)- Protocol, first call, 3 warm-up, peak reset, median of 3 rounds x median of 5
Measured
fastforks: the model's own resize / crop on the GPU; original: the library's pretrained loader and its PIL eval transform. Original = the original code named under Details, same checkpoint and protocol.
| variant | dataset | original fp32 | original fp16 | portable | portable fp32 | note |
|---|---|---|---|---|---|---|
coatnet_nano_rw_224.sw_in1k | ImageNet-1k val, 5000 random images, top-1 / top-5 % | 82.58 / 95.82 | 82.58 / 95.80 | 82.56 / 95.78 | 82.60 / 95.78 | |
coatnet_rmlp_nano_rw_224.sw_in1k | ImageNet-1k val, 5000 random images, top-1 / top-5 % | 82.48 / 95.98 | 82.54 / 95.98 | 82.44 / 95.96 | 82.40 / 95.96 | |
coatnet_0_rw_224.sw_in1k | ImageNet-1k val, 5000 random images, top-1 / top-5 % | 83.34 / 96.30 | 83.32 / 96.30 | 83.38 / 96.30 | 83.40 / 96.28 | |
coatnet_1_rw_224.sw_in1k | ImageNet-1k val, 5000 random images, top-1 / top-5 % | 84.32 / 96.54 | 84.38 / 96.56 | 84.44 / 96.56 | 84.46 / 96.56 | |
coatnet_rmlp_1_rw_224.sw_in1k | ImageNet-1k val, 5000 random images, top-1 / top-5 % | 84.58 / 96.74 | 84.58 / 96.74 | 84.46 / 96.56 | 84.50 / 96.56 | |
coatnet_rmlp_1_rw2_224.sw_in12k_ft_in1k | ImageNet-1k val, 5000 random images, top-1 / top-5 % | 85.34 / 97.02 | 85.36 / 97.02 | 85.30 / 97.04 | 85.30 / 97.02 | |
coatnet_2_rw_224.sw_in12k_ft_in1k | ImageNet-1k val, 5000 random images, top-1 / top-5 % | 87.04 / 98.16 | 87.04 / 98.14 | 87.18 / 98.16 | 87.18 / 98.16 | |
coatnet_rmlp_2_rw_224.sw_in1k | ImageNet-1k val, 5000 random images, top-1 / top-5 % | 85.12 / 96.86 | 85.08 / 96.86 | 85.16 / 96.80 | 85.16 / 96.78 | |
coatnet_rmlp_2_rw_224.sw_in12k_ft_in1k | ImageNet-1k val, 5000 random images, top-1 / top-5 % | 87.04 / 98.18 | 87.02 / 98.18 | 87.02 / 98.12 | 87.02 / 98.12 | |
coatnet_rmlp_2_rw_384.sw_in12k_ft_in1k | ImageNet-1k val, 5000 random images, top-1 / top-5 % | 87.72 / 98.64 | 87.72 / 98.68 | 87.84 / 98.54 | 87.82 / 98.54 |
Measured on portable: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 7ba7307f, 2026-11-11; portable32: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 6f0a70e1, 2026-11-17.
Parity
CPU, float32: max |diff| <= 8.6e-06 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.0024 and p99 <= 0.011 (absolute error of the output logits).
| variant | kernels error vs original, mean / p99 | GPU, torch | kernels used |
|---|---|---|---|
coatnet_nano_rw_224.sw_in1k | 0.00121 / 0.0057 | RTX 5060 Ti, torch 2.10.0+cu130 | cudnn_hacc@cudnn91501, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120 |
coatnet_rmlp_nano_rw_224.sw_in1k | 0.00112 / 0.00488 | RTX 5060 Ti, torch 2.10.0+cu130 | cudnn_hacc@cudnn91501, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120 |
coatnet_0_rw_224.sw_in1k | 0.00169 / 0.00689 | RTX 5060 Ti, torch 2.10.0+cu130 | cudnn_hacc@cudnn91501, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120 |
coatnet_1_rw_224.sw_in1k | 0.00075 / 0.00309 | RTX 5060 Ti, torch 2.10.0+cu130 | cudnn_hacc@cudnn91501, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120 |
coatnet_rmlp_1_rw_224.sw_in1k | 0.00159 / 0.00642 | RTX 5060 Ti, torch 2.10.0+cu130 | cudnn_hacc@cudnn91501, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120 |
coatnet_rmlp_1_rw2_224.sw_in12k_ft_in1k | 0.00143 / 0.00522 | RTX 5060 Ti, torch 2.10.0+cu130 | cudnn_hacc@cudnn91501, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120 |
coatnet_2_rw_224.sw_in12k_ft_in1k | 0.00148 / 0.00543 | RTX 5060 Ti, torch 2.10.0+cu130 | cudnn_hacc@cudnn91501, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120 |
coatnet_rmlp_2_rw_224.sw_in1k | 0.00134 / 0.00632 | RTX 5060 Ti, torch 2.10.0+cu130 | cudnn_hacc@cudnn91501, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120 |
coatnet_rmlp_2_rw_224.sw_in12k_ft_in1k | 0.00144 / 0.0055 | RTX 5060 Ti, torch 2.10.0+cu130 | cudnn_hacc@cudnn91501, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120 |
coatnet_rmlp_2_rw_384.sw_in12k_ft_in1k | 0.00128 / 0.00505 | RTX 5060 Ti, torch 2.10.0+cu130 | cudnn_hacc@cudnn91501, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120 |
Example output
One fixed input through this checkpoint, the same for every model of the task (portable · float16 · NVIDIA GeForce RTX 5060 Ti).
Top 5
Input: Chelsea the cat, Stéfan van der Walt (CC0).