A pre-trained model for the endoscopic tool segmentation task, trained using a flexible unet structure with an efficientnet-b2 [1] as the backbone and a UNet architecture [2] as the decoder. Datasets use private samples from Activ Surgical.
The PyTorch model and torchscript model are shared in google drive. Details can be found in large_files.yml
file. Modify the "bundle_root" parameter specified in configs/train.json and configs/inference.json to reflect where models are downloaded. Expected directory path to place downloaded models is "models/" under "bundle_root".
A pre-trained encoder weights would benefit the model training. In this bundle, the encoder is trained with pre-trained weights from some internal data. We provide two options to enable users to load pre-trained weights:
use_imagenet_pretrain
parameter in the config file to True
, ImageNet pre-trained weights from the EfficientNet-PyTorch repo can be loaded. Please note that these weights are for non-commercial use. Each user is responsible for checking the content of the models/datasets and the applicable licenses and determining if suitable for the intended use.CheckpointLoader
as the first handler to the handlers
section of the train.json
config file, weights from a local path can be loaded. Here is an example CheckpointLoader
:{
"_target_": "CheckpointLoader",
"load_path": "/path/to/local/weight/model.pt",
"load_dict": {
"model": "@network"
},
"strict": false,
"map_location": "@device"
}
When executing the training command, if neither adding the CheckpointLoader
to the train.json
nor setting the use_imagenet_pretrain
parameter to True
, a training process would start from scratch.
Datasets used in this work were provided by Activ Surgical.
Since datasets are private, existing public datasets like EndoVis 2017 can be used to train a similar model.
When using EndoVis or any other dataset, it should be divided into "train", "valid" and "test" folders. Samples in each folder would better be images and converted to jpg format. Otherwise, "images", "labels", "val_images" and "val_labels" parameters in configs/train.json
and "datalist" in configs/inference.json
should be modified to fit given dataset. After that, "dataset_dir" parameter in configs/train.json
and configs/inference.json
should be changed to root folder which contains "train", "valid" and "test" folders.
Please notice that loading data operation in this bundle is adaptive. If images and labels are not in the same format, it may lead to a mismatching problem. For example, if images are in jpg format and labels are in npy format, PIL and Numpy readers will be used separately to load images and labels. Since these two readers have their own way to parse file's shape, loaded labels will be transpose of the correct ones and incur a missmatching problem.
The training as performed with the following:
If you face memory issues with CacheDataset, you can either switch to a regular Dataset class or lower the caching rate cache_rate
in the configurations within range [0, 1] to minimize the System RAM requirements.
A three channel video frame
Two channels:
IoU was used for evaluating the performance of the model. This model achieves a mean IoU score of 0.86.
The endoscopic_tool_segmentation
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 | 12.00 | 14.06 | 6.59 | 5.20 | 0.85 | 1.82 | 2.31 | 2.70 |
end2end | 170.04 | 172.20 | 155.26 | 155.57 | 0.99 | 1.10 | 1.09 | 1.11 |
Where:
model computation
means the speedup ratio of model's inference with a random input without preprocessing and postprocessingend2end
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 modelamp vs fp16
is the speedup ratio between the PyTorch amp model and the TensorRT float16 based model.This result is benchmarked under:
In addition to the Pythonic APIs, a few command line interfaces (CLI) are provided to interact with the bundle. The CLI supports flexible use cases, such as overriding configs at runtime and predefining arguments in a file.
For more details usage instructions, visit the MONAI Bundle Configuration Page.
python -m monai.bundle run --config_file configs/train.json
Please note that if the default dataset path is not modified with the actual path in the bundle config files, you can also override it by using --dataset_dir
:
python -m monai.bundle run --config_file configs/train.json --dataset_dir <actual dataset path>
train
config to execute multi-GPU training:torchrun --standalone --nnodes=1 --nproc_per_node=2 -m monai.bundle run --config_file "['configs/train.json','configs/multi_gpu_train.json']"
Please note that the distributed training-related options depend on the actual running environment; thus, users may need to remove --standalone
, modify --nnodes
, or do some other necessary changes according to the machine used. For more details, please refer to pytorch's official tutorial.
train
config to execute evaluation with the trained model:python -m monai.bundle run --config_file "['configs/train.json','configs/evaluate.json']"
train
config and evaluate
config to execute multi-GPU evaluation:torchrun --standalone --nnodes=1 --nproc_per_node=2 -m monai.bundle run --config_file "['configs/train.json','configs/evaluate.json','configs/multi_gpu_evaluate.json']"
python -m monai.bundle run --config_file configs/inference.json
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.json
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.json --precision <fp32/fp16>
python -m monai.bundle run --config_file "['configs/inference.json', 'configs/inference_trt.json']"
[1] Tan, M. and Le, Q. V. Efficientnet: Rethinking model scaling for convolutional neural networks. ICML, 2019a. https://arxiv.org/pdf/1905.11946.pdf
[2] O. Ronneberger, P. Fischer, and T. Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234–241. Springer, 2015. https://arxiv.org/pdf/1505.04597.pdf
Copyright (c) MONAI Consortium
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.