search

ultralytics / yolov5

YOLOv5 🚀 in PyTorch > ONNX > CoreML > TFLite
https://docs.ultralytics.com
GNU Affero General Public License v3.0
51.17k stars 16.43k forks source link

Small performance gain after export to TensorRT #8769

Closed GiorgosBetsos closed 2 years ago

GiorgosBetsos commented 2 years ago

Search before asking

Question

My system specs are the following:

OS: Ubuntu 20.04.4 LTS GPU: Nvidia GTX 1080 with 8GB RAM CPU: Core I5 11500 RAM: 32 GB DDR4

I have cloned the yolov5 repository and installed all requirements, including the ones necessary for onnx and TernsorRT export, in a new conda environment. I then followed the intructions from TFLite, ONNX, CoreML, TensorRT Export guide and managed to successfully produce yolov5s.engine file.

I then used detect.py to perform inference on about 20 jpeg images. The results I got are far from the YOLOv5 Export Benchmarks for GPU.

Part of the output I got using yolov5s.engine file is as follows:

python detect.py --source data/images/test --weights yolov5s.engine detect: weights=['yolov5s.engine'], source=data/images/test, data=data/coco128.yaml, imgsz=[640, 640], conf_thres=0.25, iou_thres=0.45, max_det=1000, device=, view_img=False, save_txt=False, save_conf=False, save_crop=False, nosave=False, classes=None, agnostic_nms=False, augment=False, visualize=False, update=False, project=runs/detect, name=exp, exist_ok=False, line_thickness=3, hide_labels=False, hide_conf=False, half=False, dnn=False YOLOv5 🚀 v6.1-324-g0b5ac22 Python-3.8.13 torch-1.12.0+cu102 CUDA:0 (NVIDIA GeForce GTX 1080, 8111MiB)

Loading yolov5s.engine for TensorRT inference... [07/29/2022-00:12:08] [TRT] [I] [MemUsageChange] Init CUDA: CPU +194, GPU +0, now: CPU 299, GPU 753 (MiB) [07/29/2022-00:12:08] [TRT] [I] Loaded engine size: 31 MiB [07/29/2022-00:12:08] [TRT] [I] [MemUsageChange] TensorRT-managed allocation in engine deserialization: CPU +0, GPU +32, now: CPU 0, GPU 32 (MiB) [07/29/2022-00:12:09] [TRT] [I] [MemUsageChange] TensorRT-managed allocation in IExecutionContext creation: CPU +0, GPU +34, now: CPU 0, GPU 66 (MiB) image 1/21 /home/giorgos/repos/third-party/yolov5/data/images/test/bus (3rd copy).jpg: 640x640 4 persons, 1 bus, Done. (0.005s) image 2/21 /home/giorgos/repos/third-party/yolov5/data/images/test/bus (4th copy).jpg: 640x640 4 persons, 1 bus, Done. (0.005s) image 21/21 /home/giorgos/repos/third-party/yolov5/data/images/test/zidane.jpg: 640x640 2 persons, 2 ties, Done. (0.005s) Speed: 0.4ms pre-process, 5.3ms inference, 0.9ms NMS per image at shape (1, 3, 640, 640)

And this is the output from using yolov5s.pt

python detect.py --source data/images/test --weights yolov5s.pt --half detect: weights=['yolov5s.pt'], source=data/images/test, data=data/coco128.yaml, imgsz=[640, 640], conf_thres=0.25, iou_thres=0.45, max_det=1000, device=, view_img=False, save_txt=False, save_conf=False, save_crop=False, nosave=False, classes=None, agnostic_nms=False, augment=False, visualize=False, update=False, project=runs/detect, name=exp, exist_ok=False, line_thickness=3, hide_labels=False, hide_conf=False, half=True, dnn=False YOLOv5 🚀 v6.1-324-g0b5ac22 Python-3.8.13 torch-1.12.0+cu102 CUDA:0 (NVIDIA GeForce GTX 1080, 8111MiB)

Fusing layers... YOLOv5s summary: 213 layers, 7225885 parameters, 0 gradients image 1/21 /home/giorgos/repos/third-party/yolov5/data/images/test/bus (3rd copy).jpg: 640x480 4 persons, 1 bus, Done. (0.007s) image 2/21 /home/giorgos/repos/third-party/yolov5/data/images/test/bus (4th copy).jpg: 640x480 4 persons, 1 bus, Done. (0.007s) ... image 20/21 /home/giorgos/repos/third-party/yolov5/data/images/test/zidane (copy).jpg: 384x640 2 persons, 2 ties, Done. (0.006s) image 21/21 /home/giorgos/repos/third-party/yolov5/data/images/test/zidane.jpg: 384x640 2 persons, 2 ties, Done. (0.006s) Speed: 0.3ms pre-process, 6.9ms inference, 0.7ms NMS per image at shape (1, 3, 640, 640)

As you can see the speedup in inference time is nowhere near the one reported in the link at the beginning of this post: using the .engine file I got 5.3ms inference time vs 6.9ms obtained using pytorch model.

Is there sth I could be doing wrong?

Additional

No response

github-actions[bot] commented 2 years ago

👋 Hello @GiorgosBetsos, thank you for your interest in YOLOv5 🚀! Please visit our ⭐️ Tutorials to get started, where you can find quickstart guides for simple tasks like Custom Data Training all the way to advanced concepts like Hyperparameter Evolution.

If this is a 🐛 Bug Report, please provide screenshots and minimum viable code to reproduce your issue, otherwise we can not help you.

If this is a custom training ❓ Question, please provide as much information as possible, including dataset images, training logs, screenshots, and a public link to online W&B logging if available.

For business inquiries or professional support requests please visit https://ultralytics.com or email support@ultralytics.com.

Requirements

Python>=3.7.0 with all requirements.txt installed including PyTorch>=1.7. To get started:

git clone https://github.com/ultralytics/yolov5  # clone
cd yolov5
pip install -r requirements.txt  # install

Environments

YOLOv5 may be run in any of the following up-to-date verified environments (with all dependencies including CUDA/CUDNN, Python and PyTorch preinstalled):

Status

CI CPU testing

If this badge is green, all YOLOv5 GitHub Actions Continuous Integration (CI) tests are currently passing. CI tests verify correct operation of YOLOv5 training (train.py), validation (val.py), inference (detect.py) and export (export.py) on macOS, Windows, and Ubuntu every 24 hours and on every commit.

glenn-jocher commented 2 years ago

@GiorgosBetsos 👋 Hello! Thanks for asking about benchmarks. YOLOv5 🚀 inference is officially supported in 11 formats:

💡 ProTip: Export to ONNX or OpenVINO for up to 3x CPU speedup. See CPU Benchmarks. 💡 ProTip: Export to TensorRT for up to 5x GPU speedup. See GPU Benchmarks.

Format export.py --include Model
PyTorch - yolov5s.pt
TorchScript torchscript yolov5s.torchscript
ONNX onnx yolov5s.onnx
OpenVINO openvino yolov5s_openvino_model/
TensorRT engine yolov5s.engine
CoreML coreml yolov5s.mlmodel
TensorFlow SavedModel saved_model yolov5s_saved_model/
TensorFlow GraphDef pb yolov5s.pb
TensorFlow Lite tflite yolov5s.tflite
TensorFlow Edge TPU edgetpu yolov5s_edgetpu.tflite
TensorFlow.js tfjs yolov5s_web_model/

Benchmarks

Benchmarks below run on a Colab Pro with the YOLOv5 tutorial notebook Open In Colab. To reproduce:

python utils/benchmarks.py --weights yolov5s.pt --imgsz 640 --device 0

Colab Pro V100 GPU

benchmarks: weights=/content/yolov5/yolov5s.pt, imgsz=640, batch_size=1, data=/content/yolov5/data/coco128.yaml, device=0, half=False, test=False
Checking setup...
YOLOv5 🚀 v6.1-135-g7926afc torch 1.10.0+cu111 CUDA:0 (Tesla V100-SXM2-16GB, 16160MiB)
Setup complete ✅ (8 CPUs, 51.0 GB RAM, 46.7/166.8 GB disk)

Benchmarks complete (458.07s)
                   Format  mAP@0.5:0.95  Inference time (ms)
0                 PyTorch        0.4623                10.19
1             TorchScript        0.4623                 6.85
2                    ONNX        0.4623                14.63
3                OpenVINO           NaN                  NaN
4                TensorRT        0.4617                 1.89
5                  CoreML           NaN                  NaN
6   TensorFlow SavedModel        0.4623                21.28
7     TensorFlow GraphDef        0.4623                21.22
8         TensorFlow Lite           NaN                  NaN
9     TensorFlow Edge TPU           NaN                  NaN
10          TensorFlow.js           NaN                  NaN

Colab Pro CPU

benchmarks: weights=/content/yolov5/yolov5s.pt, imgsz=640, batch_size=1, data=/content/yolov5/data/coco128.yaml, device=cpu, half=False, test=False
Checking setup...
YOLOv5 🚀 v6.1-135-g7926afc torch 1.10.0+cu111 CPU
Setup complete ✅ (8 CPUs, 51.0 GB RAM, 41.5/166.8 GB disk)

Benchmarks complete (241.20s)
                   Format  mAP@0.5:0.95  Inference time (ms)
0                 PyTorch        0.4623               127.61
1             TorchScript        0.4623               131.23
2                    ONNX        0.4623                69.34
3                OpenVINO        0.4623                66.52
4                TensorRT           NaN                  NaN
5                  CoreML           NaN                  NaN
6   TensorFlow SavedModel        0.4623               123.79
7     TensorFlow GraphDef        0.4623               121.57
8         TensorFlow Lite        0.4623               316.61
9     TensorFlow Edge TPU           NaN                  NaN
10          TensorFlow.js           NaN                  NaN

Good luck 🍀 and let us know if you have any other questions!

GiorgosBetsos commented 2 years ago

I have seen these benchmarks, this is the reason why I opened this issue. According to the benchmarks TensorRT inference time is 1.89ms vs 10.19ms for PyTorch inference. As opposed to this I get 5.3ms vs 6.9ms.

This is what I get when I execute:

python utils/benchmarks.py --weights yolov5s.pt --imgsz 640 --device 0

on my box:

benchmarks: weights=yolov5s.pt, imgsz=640, batch_size=1, data=/home/giorgos/repos/third-party/yolov5/data/coco128.yaml, device=0, half=False, test=False, pt_only=False, hard_fail=False
Checking setup...
YOLOv5 🚀 v6.1-324-g0b5ac22 Python-3.8.13 torch-1.12.0+cu102 CUDA:0 (NVIDIA GeForce GTX 1080, 8111MiB)
Setup complete ✅ (12 CPUs, 31.2 GB RAM, 90.1/227.7 GB disk)

Benchmarks complete (139.15s)
                   Format  Size (MB)  mAP@0.5:0.95  Inference time (ms)
0                 PyTorch       14.1        0.4716                 6.77
1             TorchScript       28.1        0.4716                 7.14
2                    ONNX       28.0        0.4716                66.00
3                OpenVINO        NaN           NaN                  NaN
4                TensorRT       31.9        0.4716                 4.62
glenn-jocher commented 2 years ago

@GiorgosBetsos sure, results will vary with hardware, software, firmware etc. The results I provided are run on a reproducible environments, Colab Pro.

lennartmoritz commented 2 years ago

Hey @GiorgosBetsos, I face the same issue. I wanted to speed up inference with the new TensorRT format in my application. But in the benchmarks the improvements between the PyTorch and TensorRT format are only valid for batch sizes of 1. For a batch size of 15, as i would use in my application, the TensorRT format was detrimental to the inference time.

Why would higher batch sizes be detrimental for TensorRT?

Could this be unintended behaviour that can be fixed by adjusting export parameters @glenn-jocher?

I used the benchmark features on a local PC with an RTX 2060 and a server with a GTX 1080 Ti.

Worse TensorRT results with RTX 2060

I benchmarked with an image size of 320 and batch size of [1, 15].

benchmarks: weights=models/yolov5/yolov5s.pt, imgsz=320, batch_size=1, data=/usr/src/app/detector/yolov5_61/data/coco128.yaml, device=0, half=False, test=False, pt_only=False, hard_fail=False
Checking setup...
YOLOv5 🚀 2022-8-5 Python-3.7.7 torch-1.11.0 CUDA:0 (NVIDIA GeForce RTX 2060, 5932MiB)
Setup complete ✅ (6 CPUs, 15.5 GB RAM, 86.0/93.6 GB disk)

Benchmarks complete (95.81s)
                   Format  Size (MB)  mAP@0.5:0.95  Inference time (ms)
0                 PyTorch       14.1        0.4114                 6.21
1             TorchScript       27.9        0.4114                 4.15
2                    ONNX       27.7        0.4114                34.92
3                OpenVINO        NaN           NaN                  NaN
4                TensorRT       35.0        0.4114                 2.71
5                  CoreML        NaN           NaN                  NaN
6   TensorFlow SavedModel        NaN           NaN                  NaN
7     TensorFlow GraphDef        NaN           NaN                  NaN
8         TensorFlow Lite        NaN           NaN                  NaN
9     TensorFlow Edge TPU        NaN           NaN                  NaN
10          TensorFlow.js        NaN           NaN                  NaN

benchmarks: weights=models/yolov5/yolov5s.pt, imgsz=320, batch_size=15, data=/usr/src/app/detector/yolov5_61/data/coco128.yaml, device=0, half=False, test=False, pt_only=False, hard_fail=False
Checking setup...
YOLOv5 🚀 2022-8-5 Python-3.7.7 torch-1.11.0 CUDA:0 (NVIDIA GeForce RTX 2060, 5932MiB)
Setup complete ✅ (6 CPUs, 15.5 GB RAM, 86.0/93.6 GB disk)

Benchmarks complete (139.08s)
                   Format  Size (MB)  mAP@0.5:0.95  Inference time (ms)
0                 PyTorch       14.1        0.4114                 2.07
1             TorchScript       27.9        0.4114                 1.45
2                    ONNX       27.7        0.4114                30.42
3                OpenVINO        NaN           NaN                  NaN
4                TensorRT       35.0        0.4114                 2.68
5                  CoreML        NaN           NaN                  NaN
6   TensorFlow SavedModel        NaN           NaN                  NaN
7     TensorFlow GraphDef        NaN           NaN                  NaN
8         TensorFlow Lite        NaN           NaN                  NaN
9     TensorFlow Edge TPU        NaN           NaN                  NaN
10          TensorFlow.js        NaN           NaN                  NaN

Benchmark with GTX 1080 Ti

I benchmarked with an image size of [320, 620] and batch size of [1, 15].

benchmarks: weights=models/yolov5/yolov5s.pt, imgsz=640, batch_size=15, data=/usr/src/app/detector/yolov5_61/data/coco128.yaml, device=0, half=False, test=False, pt_only=False, hard_fail=False
Checking setup...
YOLOv5 🚀 2022-8-17 Python-3.8.12 torch-1.11.0a0+17540c5 CUDA:0 (NVIDIA GeForce GTX 1080 Ti, 11179MiB)
Setup complete ✅ (40 CPUs, 62.8 GB RAM, 754.7/915.8 GB disk)

Benchmarks complete (119.54s)
                   Format  Size (MB)  mAP@0.5:0.95  Inference time (ms)
0                 PyTorch       14.1        0.4716                 3.87
1             TorchScript       28.0        0.4716                 3.60
2                    ONNX       28.0        0.4716                12.60
3                OpenVINO        NaN           NaN                  NaN
4                TensorRT       42.7        0.4716                 4.07
5                  CoreML        NaN           NaN                  NaN
6   TensorFlow SavedModel        NaN           NaN                  NaN
7     TensorFlow GraphDef        NaN           NaN                  NaN
8         TensorFlow Lite        NaN           NaN                  NaN
9     TensorFlow Edge TPU        NaN           NaN                  NaN
10          TensorFlow.js        NaN           NaN                  NaN

benchmarks: weights=models/yolov5/yolov5s.pt, imgsz=640, batch_size=1, data=/usr/src/app/detector/yolov5_61/data/coco128.yaml, device=0, half=False, test=False, pt_only=False, hard_fail=False
Checking setup...
YOLOv5 🚀 2022-8-17 Python-3.8.12 torch-1.11.0a0+17540c5 CUDA:0 (NVIDIA GeForce GTX 1080 Ti, 11179MiB)
Setup complete ✅ (40 CPUs, 62.8 GB RAM, 754.7/915.8 GB disk)

Benchmarks complete (107.81s)
                   Format  Size (MB)  mAP@0.5:0.95  Inference time (ms)
0                 PyTorch       14.1        0.4716                12.91
1             TorchScript       28.0        0.4716                 6.93
2                    ONNX       28.0        0.4716                13.30
3                OpenVINO        NaN           NaN                  NaN
4                TensorRT       42.7        0.4716                 4.05
5                  CoreML        NaN           NaN                  NaN
6   TensorFlow SavedModel        NaN           NaN                  NaN
7     TensorFlow GraphDef        NaN           NaN                  NaN
8         TensorFlow Lite        NaN           NaN                  NaN
9     TensorFlow Edge TPU        NaN           NaN                  NaN
10          TensorFlow.js        NaN           NaN                  NaN

benchmarks: weights=models/yolov5/yolov5s.pt, imgsz=320, batch_size=15, data=/usr/src/app/detector/yolov5_61/data/coco128.yaml, device=0, half=False, test=False, pt_only=False, hard_fail=False
Checking setup...
YOLOv5 🚀 2022-8-17 Python-3.8.12 torch-1.11.0a0+17540c5 CUDA:0 (NVIDIA GeForce GTX 1080 Ti, 11179MiB)
Setup complete ✅ (40 CPUs, 62.8 GB RAM, 754.7/915.8 GB disk)

Benchmarks complete (101.88s)
                   Format  Size (MB)  mAP@0.5:0.95  Inference time (ms)
0                 PyTorch       14.1        0.4114                 1.36
1             TorchScript       27.9        0.4114                 1.14
2                    ONNX       27.7        0.4114                 6.87
3                OpenVINO        NaN           NaN                  NaN
4                TensorRT       42.3        0.4114                 2.16
5                  CoreML        NaN           NaN                  NaN
6   TensorFlow SavedModel        NaN           NaN                  NaN
7     TensorFlow GraphDef        NaN           NaN                  NaN
8         TensorFlow Lite        NaN           NaN                  NaN
9     TensorFlow Edge TPU        NaN           NaN                  NaN
10          TensorFlow.js        NaN           NaN                  NaN

benchmarks: weights=models/yolov5/yolov5s.pt, imgsz=320, batch_size=1, data=/usr/src/app/detector/yolov5_61/data/coco128.yaml, device=0, half=False, test=False, pt_only=False, hard_fail=False
Checking setup...
YOLOv5 🚀 2022-8-17 Python-3.8.12 torch-1.11.0a0+17540c5 CUDA:0 (NVIDIA GeForce GTX 1080 Ti, 11179MiB)
Setup complete ✅ (40 CPUs, 62.8 GB RAM, 754.7/915.8 GB disk)

Benchmarks complete (103.19s)
                   Format  Size (MB)  mAP@0.5:0.95  Inference time (ms)
0                 PyTorch       14.1        0.4114                10.98
1             TorchScript       27.9        0.4114                 7.27
2                    ONNX       27.7        0.4114                 5.91
3                OpenVINO        NaN           NaN                  NaN
4                TensorRT       42.3        0.4114                 2.13
5                  CoreML        NaN           NaN                  NaN
6   TensorFlow SavedModel        NaN           NaN                  NaN
7     TensorFlow GraphDef        NaN           NaN                  NaN
8         TensorFlow Lite        NaN           NaN                  NaN
9     TensorFlow Edge TPU        NaN           NaN                  NaN
10          TensorFlow.js        NaN           NaN                  NaN
glenn-jocher commented 2 years ago

@lennartmoritz torch inference will normally see speedups at larger batch sizes, see https://community.ultralytics.com/t/yolov5-study-batch-size-vs-speed/31, not sure about TRT as I only have batch size 1 experience there.

BTW you'll get significantly better speedup at FP16 with TRT.

glenn-jocher commented 2 years ago

@lennartmoritz ah I just realized val.py is forcing batch-size 1 for most formats on L149. You might want to debug expanding this to TRT dynamic and higher batch size models (but not batch-size 1 models) and then submit a PR with your updates. https://github.com/ultralytics/yolov5/blob/4a8ab3bc42d32f3e2e9c026b87dc29fba6143064/val.py#L139-L151

lennartmoritz commented 2 years ago

Thank you for the hint. I've created a workaround for the benchmark but I don't have the time right now to look into dynamic batch sizes and add correct "fixed batch size" handling needed in a PR.

With the workaround the TensorRT model "engine" is built with the selected batch size, but the batch size must be divisible without remainder by the amount of images in the benchmark (128 Images for the default COCO images) to avoid an error since batch sizes smaller than what the ThesorRT model is built for are not supported. Custom handling to add padding "placeholder images" to non-full input batches would be required.

By adding elif f == "engine" case like this:

# Export
if f == "-":
    w = weights  # PyTorch format
elif f == "engine":
    w = export.run(weights=weights, imgsz=[imgsz], batch_size=batch_size, include=[f], device=device, half=half)[-1]  # Set correct batch
else:
    w = export.run(weights=weights, imgsz=[imgsz], include=[f], device=device, half=half)[-1]  # all others
assert suffix in str(w), "export failed"

To the original code: https://github.com/ultralytics/yolov5/blob/4a8ab3bc42d32f3e2e9c026b87dc29fba6143064/utils/benchmarks.py#L71-L76

glenn-jocher commented 2 years ago

@lennartmoritz ah got it. You probably just want --dynamic TRT export then for the last batch problem.

Does the TRT batched inference show speedup?

lennartmoritz commented 2 years ago

@glenn-jocher yes, the batched inference speed has improved notably compared to the ealier benchmarks. I'll add benchmarks with the RTX 2060 below. I agree that a dynamic TRT export should solve the problem. Since the benchmark.py offers no --dynamic flag for the ArgumentParser, I could add a hardcoded dynamic=true to the elif case from my earlier comment.

For some reason I could not replicate any improvements of using half precision as in the below benchmarks in my own application yet, but I can probably figure that out on my own. Since I only compared total detection times including input resizing, halving, upload to & download from GPU this might as well be a "me-problem".

RTX 2060, img_size = 320, batch = 16, half = false

benchmarks: weights=models/yolov5/yolov5s.pt, imgsz=320, batch_size=16, data=/usr/src/app/detector/yolov5_61/data/coco128.yaml, device=0, half=False, test=False, pt_only=False, hard_fail=False
Checking setup...
YOLOv5 🚀 2022-8-5 Python-3.7.7 torch-1.11.0 CUDA:0 (NVIDIA GeForce RTX 2060, 5932MiB)
Setup complete ✅ (6 CPUs, 15.5 GB RAM, 85.8/93.6 GB disk)

Benchmarks complete (87.86s)
                   Format  Size (MB)  mAP@0.5:0.95  Inference time (ms)
0                 PyTorch       14.1        0.4114                 1.51
1             TorchScript       27.9        0.4114                 1.41
2                    ONNX        NaN           NaN                  NaN
3                OpenVINO        NaN           NaN                  NaN
4                TensorRT       30.2        0.4114                 1.31
5                  CoreML        NaN           NaN                  NaN
6   TensorFlow SavedModel        NaN           NaN                  NaN
7     TensorFlow GraphDef        NaN           NaN                  NaN
8         TensorFlow Lite        NaN           NaN                  NaN
9     TensorFlow Edge TPU        NaN           NaN                  NaN
10          TensorFlow.js        NaN           NaN                  NaN

RTX 2060, img_size = 320, batch = 16, half = true

benchmarks: weights=models/yolov5/yolov5s.pt, imgsz=320, batch_size=16, data=/usr/src/app/detector/yolov5_61/data/coco128.yaml, device=0, half=True, test=False, pt_only=False, hard_fail=False
Checking setup...
YOLOv5 🚀 2022-8-5 Python-3.7.7 torch-1.11.0 CUDA:0 (NVIDIA GeForce RTX 2060, 5932MiB)
Setup complete ✅ (6 CPUs, 15.5 GB RAM, 85.8/93.6 GB disk)

Benchmarks complete (248.03s)
                   Format  Size (MB)  mAP@0.5:0.95  Inference time (ms)
0                 PyTorch       14.1        0.4116                 1.29
1             TorchScript       14.1        0.4117                 0.99
2                    ONNX        NaN           NaN                  NaN
3                OpenVINO        NaN           NaN                  NaN
4                TensorRT       15.6        0.4111                 0.43
5                  CoreML        NaN           NaN                  NaN
6   TensorFlow SavedModel        NaN           NaN                  NaN
7     TensorFlow GraphDef        NaN           NaN                  NaN
8         TensorFlow Lite        NaN           NaN                  NaN
9     TensorFlow Edge TPU        NaN           NaN                  NaN
10          TensorFlow.js        NaN           NaN                  NaN
glenn-jocher commented 2 years ago

@lennartmoritz this looks good. I looked at the val.py code again and it seems like TensorRT models at different batch sizes should be handled automatically, though as you mentined the last batch will produce an error if not dynamic. If you export TRT --dynamic can you run benchmarks at different batch sizes without any code modifications?

On L144: https://github.com/ultralytics/yolov5/blob/fc8758a49bd30526fb21d0683359e86be3a292a8/val.py#L139-L151

github-actions[bot] commented 2 years ago

👋 Hello, this issue has been automatically marked as stale because it has not had recent activity. Please note it will be closed if no further activity occurs.

Access additional YOLOv5 🚀 resources:

Access additional Ultralytics ⚡ resources:

Feel free to inform us of any other issues you discover or feature requests that come to mind in the future. Pull Requests (PRs) are also always welcomed!

Thank you for your contributions to YOLOv5 🚀 and Vision AI ⭐!