add support for TensorRT conversion and inference

README.md

@@ -86,6 +86,31 @@ The mean dice score over 3200 epochs (the bright curve is smoothed, and the dark
 
 ![Validation mean dice score over 3200 epochs (the bright curve is smoothed, and the dark one is the actual curve)](https://developer.download.nvidia.com/assets/Clara/Images/clara_pt_net_arch_search_segmentation_validation_4-3.png)
 
+#### TensorRT speedup
+This bundle supports acceleration with TensorRT. The table below displays the speedup ratios observed on an A100 80G GPU.
+
+| method | torch_fp32(ms) | torch_amp(ms) | trt_fp32(ms) | trt_fp16(ms) | speedup amp | speedup fp32 | speedup fp16 | amp vs fp16|
+| :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
+| model computation | 54611.72 | 19240.66 | 16104.8 | 11443.57 | 2.84 | 3.39 | 4.77 | 1.68 |
+| end2end | 133.93 | 43.41 | 35.65 | 26.63 | 3.09 | 3.76 | 5.03 | 1.63 |
+
+Where:
+- `model computation` means the speedup ratio of model's inference with a random input without preprocessing and postprocessing
+- `end2end` means run the bundle end-to-end with the TensorRT based model.
+- `torch_fp32` and `torch_amp` are for the PyTorch models with or without `amp` mode.
+- `trt_fp32` and `trt_fp16` are for the TensorRT based models converted in corresponding precision.
+- `speedup amp`, `speedup fp32` and `speedup fp16` are the speedup ratios of corresponding models versus the PyTorch float32 model
+- `amp vs fp16` is the speedup ratio between the PyTorch amp model and the TensorRT float16 based model.
+
+This result is benchmarked under:
+ - TensorRT: 8.6.1+cuda12.0
+ - Torch-TensorRT Version: 1.4.0
+ - CPU Architecture: x86-64
+ - OS: ubuntu 20.04
+ - Python version:3.8.10
+ - CUDA version: 12.1
+ - GPU models and configuration: A100 80G
+
 ### Searched Architecture Visualization
 Users can install Graphviz for visualization of searched architectures (needed in [decode_plot.py](https://github.com/Project-MONAI/tutorials/blob/main/automl/DiNTS/decode_plot.py)). The edges between nodes indicate global structure, and numbers next to edges represent different operations in the cell searching space. An example of searched architecture is shown as follows:
 
@@ -144,6 +169,18 @@ python -m monai.bundle run --config_file configs/inference.yaml
 python -m monai.bundle ckpt_export network_def --filepath models/model.ts --ckpt_file models/model.pt --meta_file configs/metadata.json --config_file configs/inference.yaml
 ```
 
+#### Export checkpoint to TensorRT based models with fp32 or fp16 precision:
+
+```
+python -m monai.bundle trt_export --net_id network_def --filepath models/model_trt.ts --ckpt_file models/model.pt --meta_file configs/metadata.json --config_file configs/inference.yaml  --precision <fp32/fp16> --use_trace "True" --dynamic_batchsize "[1, 4, 8]" --converter_kwargs "{'truncate_long_and_double':True, 'torch_executed_ops': ['aten::upsample_trilinear3d']}"
+```
+
+#### Execute inference with the TensorRT model:
+
+```
+python -m monai.bundle run --config_file "['configs/inference.yaml', 'configs/inference_trt.yaml']"
+```
+
 # References
 
 [1] He, Y., Yang, D., Roth, H., Zhao, C. and Xu, D., 2021. Dints: Differentiable neural network topology search for 3d medical image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 5841-5850).

configs/inference_trt.yaml

@@ -0,0 +1,8 @@
+---
+imports:
+- "$import glob"
+- "$import os"
+- "$import torch_tensorrt"
+handlers#0#_disabled_: true
+network_def: "$torch.jit.load(@bundle_root + '/models/model_trt.ts')"
+evaluator#amp: false

configs/metadata.json

@@ -1,7 +1,8 @@
 {
     "schema": "https://github.com/Project-MONAI/MONAI-extra-test-data/releases/download/0.8.1/meta_schema_20220324.json",
-    "version": "0.4.2",
+    "version": "0.4.3",
     "changelog": {
+        "0.4.3": "add support for TensorRT conversion and inference",
         "0.4.2": "update search function to match monai 1.2",
         "0.4.1": "fix the wrong GPU index issue of multi-node",
         "0.4.0": "remove error dollar symbol in readme",

docs/README.md

@@ -79,6 +79,31 @@ The mean dice score over 3200 epochs (the bright curve is smoothed, and the dark
 
 ![Validation mean dice score over 3200 epochs (the bright curve is smoothed, and the dark one is the actual curve)](https://developer.download.nvidia.com/assets/Clara/Images/clara_pt_net_arch_search_segmentation_validation_4-3.png)
 
+#### TensorRT speedup
+This bundle supports acceleration with TensorRT. The table below displays the speedup ratios observed on an A100 80G GPU.
+
+| method | torch_fp32(ms) | torch_amp(ms) | trt_fp32(ms) | trt_fp16(ms) | speedup amp | speedup fp32 | speedup fp16 | amp vs fp16|
+| :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
+| model computation | 54611.72 | 19240.66 | 16104.8 | 11443.57 | 2.84 | 3.39 | 4.77 | 1.68 |
+| end2end | 133.93 | 43.41 | 35.65 | 26.63 | 3.09 | 3.76 | 5.03 | 1.63 |
+
+Where:
+- `model computation` means the speedup ratio of model's inference with a random input without preprocessing and postprocessing
+- `end2end` means run the bundle end-to-end with the TensorRT based model.
+- `torch_fp32` and `torch_amp` are for the PyTorch models with or without `amp` mode.
+- `trt_fp32` and `trt_fp16` are for the TensorRT based models converted in corresponding precision.
+- `speedup amp`, `speedup fp32` and `speedup fp16` are the speedup ratios of corresponding models versus the PyTorch float32 model
+- `amp vs fp16` is the speedup ratio between the PyTorch amp model and the TensorRT float16 based model.
+
+This result is benchmarked under:
+ - TensorRT: 8.6.1+cuda12.0
+ - Torch-TensorRT Version: 1.4.0
+ - CPU Architecture: x86-64
+ - OS: ubuntu 20.04
+ - Python version:3.8.10
+ - CUDA version: 12.1
+ - GPU models and configuration: A100 80G
+
 ### Searched Architecture Visualization
 Users can install Graphviz for visualization of searched architectures (needed in [decode_plot.py](https://github.com/Project-MONAI/tutorials/blob/main/automl/DiNTS/decode_plot.py)). The edges between nodes indicate global structure, and numbers next to edges represent different operations in the cell searching space. An example of searched architecture is shown as follows:
 
@@ -137,6 +162,18 @@ python -m monai.bundle run --config_file configs/inference.yaml
 python -m monai.bundle ckpt_export network_def --filepath models/model.ts --ckpt_file models/model.pt --meta_file configs/metadata.json --config_file configs/inference.yaml
 ```
 
+#### Export checkpoint to TensorRT based models with fp32 or fp16 precision:
+
+```
+python -m monai.bundle trt_export --net_id network_def --filepath models/model_trt.ts --ckpt_file models/model.pt --meta_file configs/metadata.json --config_file configs/inference.yaml  --precision <fp32/fp16> --use_trace "True" --dynamic_batchsize "[1, 4, 8]" --converter_kwargs "{'truncate_long_and_double':True, 'torch_executed_ops': ['aten::upsample_trilinear3d']}"
+```
+
+#### Execute inference with the TensorRT model:
+
+```
+python -m monai.bundle run --config_file "['configs/inference.yaml', 'configs/inference_trt.yaml']"
+```
+
 # References
 
 [1] He, Y., Yang, D., Roth, H., Zhao, C. and Xu, D., 2021. Dints: Differentiable neural network topology search for 3d medical image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 5841-5850).