add HEART vs HAT_M speed benchmark (PyTorch/ONNX/TensorRT)
Browse files
README.md
CHANGED
|
@@ -304,3 +304,46 @@ HEART is a re-implementation/simplification inspired by:
|
|
| 304 |
|
| 305 |
The architecture file is original code; verify attribution/compatibility
|
| 306 |
against those papers before commercial redistribution.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 304 |
|
| 305 |
The architecture file is original code; verify attribution/compatibility
|
| 306 |
against those papers before commercial redistribution.
|
| 307 |
+
|
| 308 |
+
## Update — 2026-09-30: speed benchmark (HEART vs HAT_M)
|
| 309 |
+
|
| 310 |
+
Measured on an **RTX 3060**, input `1x3x256x256` (a typical SR tile), fp16,
|
| 311 |
+
warmup 5, mean of 20 runs. ONNX is static-shape; TensorRT uses the same builder
|
| 312 |
+
options as NERVE's published config (`builderOptimizationLevel=5` + CUDA graphs).
|
| 313 |
+
|
| 314 |
+
| model | params | PyTorch fp16 | ONNX fp16 | **TensorRT fp16** | peak VRAM |
|
| 315 |
+
|---|---|---|---|---|---|
|
| 316 |
+
| **HEART 4x** | 16.7M | 816 ms | 745 ms | **266 ms** (208 FPS) | **679 MB** |
|
| 317 |
+
| HEART 4x OTF GAN | 16.7M | 809 ms | 753 ms | **270 ms** | 679 MB |
|
| 318 |
+
| HAT_M 4x | 20.8M | 1485 ms | 849 ms | 598 ms (107 FPS) | 2428 MB |
|
| 319 |
+
|
| 320 |
+
**Verdict.** At a 256px tile HEART is **~1.8x faster than HAT_M** in PyTorch and
|
| 321 |
+
uses **~3.6x less VRAM** (679 MB vs 2.4 GB) — a bigger margin than our earlier
|
| 322 |
+
30k-iteration note suggested. Under **TensorRT** both get dramatically faster,
|
| 323 |
+
and HEART keeps a **~2.2x** lead (266 ms vs 598 ms).
|
| 324 |
+
|
| 325 |
+
**TensorRT is a big win for HEART** (3.1x over PyTorch), much larger than for
|
| 326 |
+
NERVE (1.05x) — a bigger model gives the builder more to optimize. Use it:
|
| 327 |
+
|
| 328 |
+
```python
|
| 329 |
+
import onnxruntime as ort
|
| 330 |
+
|
| 331 |
+
sess = ort.InferenceSession(
|
| 332 |
+
"4x_HEART_fp16.onnx", # static shape, e.g. 1x3x256x256
|
| 333 |
+
providers=[
|
| 334 |
+
("TensorrtExecutionProvider", {
|
| 335 |
+
"trt_fp16_enable": True,
|
| 336 |
+
"trt_builder_optimization_level": 5,
|
| 337 |
+
"trt_cuda_graph_enable": True,
|
| 338 |
+
"trt_engine_cache_enable": True,
|
| 339 |
+
"trt_engine_cache_path": "trt_cache",
|
| 340 |
+
"trt_timing_cache_enable": True,
|
| 341 |
+
}),
|
| 342 |
+
"CUDAExecutionProvider",
|
| 343 |
+
],
|
| 344 |
+
)
|
| 345 |
+
```
|
| 346 |
+
|
| 347 |
+
Caveats: this is a **single size (256px)** on one GPU, and ONNX/TensorRT need a
|
| 348 |
+
**static shape** — export one per size you use. HEART is **ONNX-only** (no ncnn;
|
| 349 |
+
its attention ops don't convert cleanly).
|