Deployment¶

We provide deployment tools under tools/deployment directory.

Convert to ONNX (experimental)¶

We provide a script to convert the model to ONNX format. The converted model could be visualized by tools like Netron. Besides, we also support comparing the output results between Pytorch and ONNX model.

python tools/deployment/pytorch2onnx.py
    ${MODEL_CONFIG_PATH} \
    ${MODEL_CKPT_PATH} \
    ${MODEL_TYPE} \
    ${IMAGE_PATH} \
    --output-file ${OUTPUT_FILE} \
    --device-id ${DEVICE_ID} \
    --opset-version ${OPSET_VERSION} \
    --verify \
    --verbose \
    --show \
    --dynamic-export

Description of arguments:

ARGS	Type	Description
`model_config`	str	The path to a model config file.
`model_ckpt`	str	The path to a model checkpoint file.
`model_type`	'recog', 'det'	The model type of the config file.
`image_path`	str	The path to input image file.
`--output-file`	str	The path to output ONNX model. Defaults to `tmp.onnx`.
`--device-id`	int	Which GPU to use. Defaults to 0.
`--opset-version`	int	ONNX opset version. Defaults to 11.
`--verify`	bool	Determines whether to verify the correctness of an exported model. Defaults to `False`.
`--verbose`	bool	Determines whether to print the architecture of the exported model. Defaults to `False`.
`--show`	bool	Determines whether to visualize outputs of ONNXRuntime and PyTorch. Defaults to `False`.
`--dynamic-export`	bool	Determines whether to export ONNX model with dynamic input and output shapes. Defaults to `False`.

Note: This tool is still experimental. For now, some customized operators are not supported, and we only support a subset of detection and recognition algorithms.

List of supported models exportable to ONNX¶

The table below lists the models that are guaranteed to be exportable to ONNX and runnable in ONNX Runtime.

Model	Config	Dynamic Shape	Batch Inference	Note
DBNet	dbnet_r18_fpnc_1200e_icdar2015.py	Y	N
PSENet	psenet_r50_fpnf_600e_ctw1500.py	Y	Y
PSENet	psenet_r50_fpnf_600e_icdar2015.py	Y	Y
PANet	panet_r18_fpem_ffm_600e_ctw1500.py	Y	Y
PANet	panet_r18_fpem_ffm_600e_icdar2015.py	Y	Y
CRNN	crnn_academic_dataset.py	Y	Y	CRNN only accepts input with height 32

Notes:

All models above are tested with Pytorch==1.8.1 and onnxruntime==1.7.0
If you meet any problem with the listed models above, please create an issue and it would be taken care of soon.
Because this feature is experimental and may change fast, please always try with the latest mmcv and mmocr.

Convert ONNX to TensorRT (experimental)¶

We also provide a script to convert ONNX model to TensorRT format. Besides, we support comparing the output results between ONNX and TensorRT model.

python tools/deployment/onnx2tensorrt.py
    ${MODEL_CONFIG_PATH} \
    ${MODEL_TYPE} \
    ${IMAGE_PATH} \
    ${ONNX_FILE} \
    --trt-file ${OUT_TENSORRT} \
    --max-shape INT INT INT INT \
    --min-shape INT INT INT INT \
    --workspace-size INT \
    --fp16 \
    --verify \
    --show \
    --verbose

Description of arguments:

ARGS	Type	Description
`model_config`	str	The path to a model config file.
`model_type`	'recog', 'det'	The model type of the config file.
`image_path`	str	The path to input image file.
`onnx_file`	str	The path to input ONNX file.
`--trt-file`	str	The path of output TensorRT model. Defaults to `tmp.trt`.
`--max-shape`	int * 4	Maximum shape of model input.
`--min-shape`	int * 4	Minimum shape of model input.
`--workspace-size`	int	Max workspace size in GiB. Defaults to 1.
`--fp16`	bool	Determines whether to export TensorRT with fp16 mode. Defaults to `False`.
`--verify`	bool	Determines whether to verify the correctness of an exported model. Defaults to `False`.
`--show`	bool	Determines whether to show the output of ONNX and TensorRT. Defaults to `False`.
`--verbose`	bool	Determines whether to verbose logging messages while creating TensorRT engine. Defaults to `False`.

Note: This tool is still experimental. For now, some customized operators are not supported, and we only support a subset of detection and recognition algorithms.

List of supported models exportable to TensorRT¶

The table below lists the models that are guaranteed to be exportable to TensorRT engine and runnable in TensorRT.

Model	Config	Dynamic Shape	Batch Inference	Note
DBNet	dbnet_r18_fpnc_1200e_icdar2015.py	Y	N
PSENet	psenet_r50_fpnf_600e_ctw1500.py	Y	Y
PSENet	psenet_r50_fpnf_600e_icdar2015.py	Y	Y
PANet	panet_r18_fpem_ffm_600e_ctw1500.py	Y	Y
PANet	panet_r18_fpem_ffm_600e_icdar2015.py	Y	Y
CRNN	crnn_academic_dataset.py	Y	Y	CRNN only accepts input with height 32

Notes:

All models above are tested with Pytorch==1.8.1, onnxruntime==1.7.0 and tensorrt==7.2.1.6
If you meet any problem with the listed models above, please create an issue and it would be taken care of soon.
Because this feature is experimental and may change fast, please always try with the latest mmcv and mmocr.

Evaluate ONNX and TensorRT Models (experimental)¶

We provide methods to evaluate TensorRT and ONNX models in tools/deployment/deploy_test.py.

Prerequisite¶

To evaluate ONNX and TensorRT models, ONNX, ONNXRuntime and TensorRT should be installed first. Install mmcv-full with ONNXRuntime custom ops and TensorRT plugins follow ONNXRuntime in mmcv and TensorRT plugin in mmcv.

Usage¶

python tools/deploy_test.py \
    ${CONFIG_FILE} \
    ${MODEL_PATH} \
    ${MODEL_TYPE} \
    ${BACKEND} \
    --eval ${METRICS} \
    --device ${DEVICE}

Description of all arguments¶

ARGS	Type	Description
`model_config`	str	The path to a model config file.
`model_file`	str	The path to a TensorRT or an ONNX model file.
`model_type`	'recog', 'det'	Detection or recognition model to deploy.
`backend`	'TensorRT', 'ONNXRuntime'	The backend for testing.
`--eval`	'acc', 'hmean-iou'	The evaluation metrics. 'acc' for recognition models, 'hmean-iou' for detection models.
`--device`	str	Device for evaluation. Defaults to `cuda:0`.

Results and Models¶

Model	Config	Dataset	Metric	PyTorch	ONNX Runtime	TensorRT FP32	TensorRT FP16
DBNet	dbnet_r18_fpnc_1200e_icdar2015.py	icdar2015	Recall	0.731	0.731	0.678	0.679
			Precision	0.871	0.871	0.844	0.842
			Hmean	0.795	0.795	0.752	0.752
DBNet*	dbnet_r18_fpnc_1200e_icdar2015.py	icdar2015	Recall	0.720	0.720	0.720	0.718
			Precision	0.868	0.868	0.868	0.868
			Hmean	0.787	0.787	0.787	0.786
PSENet	psenet_r50_fpnf_600e_icdar2015.py	icdar2015	Recall	0.753	0.753	0.753	0.752
			Precision	0.867	0.867	0.867	0.867
			Hmean	0.806	0.806	0.806	0.805
PANet	panet_r18_fpem_ffm_600e_icdar2015.py	icdar2015	Recall	0.740	0.740	0.687	N/A
			Precision	0.860	0.860	0.815	N/A
			Hmean	0.796	0.796	0.746	N/A
PANet*	panet_r18_fpem_ffm_600e_icdar2015.py	icdar2015	Recall	0.736	0.736	0.736	N/A
			Precision	0.857	0.857	0.857	N/A
			Hmean	0.792	0.792	0.792	N/A
CRNN	crnn_academic_dataset.py	IIIT5K	Acc	0.806	0.806	0.806	0.806

Notes:

TensorRT upsampling operation is a little different from PyTorch. For DBNet and PANet, we suggest replacing upsampling operations with the nearest mode to operations with bilinear mode. Here for PANet, here and here for DBNet. As is shown in the above table, networks with tag * mean the upsampling mode is changed.
Note that changing upsampling mode reduces less performance compared with using the nearest mode. However, the weights of networks are trained through the nearest mode. To pursue the best performance, using bilinear mode for both training and TensorRT deployment is recommended.
All ONNX and TensorRT models are evaluated with dynamic shapes on the datasets, and images are preprocessed according to the original config file.
This tool is still experimental, and we only support a subset of detection and recognition algorithms for now.