[Enhancement] Add a visualize script for inference result comparation.

docs/en/02-how-to-run/useful_tools.md

@@ -203,6 +203,23 @@ And the output look like this:
 +--------+------------+---------+
 ```
 
+## visualize
+
+This tool can be used to visualize model inference results in different backend.
+
+### Usage
+
+```bash
+python tools/visualize.py \
+    --deploy-cfg {DEPLOY_CFG} \
+    --model-cfg {MODEL_CFG} \
+    --deploy-path {DEPLOY_PATH} \
+    --test-img {TEST_IMGS} \
+    --checkpoint {CHECKPOINTS} \
+    --save-dir {SAVE_DIR} \
+    --device {DEVICE}
+```
+
 ## generate_md_table
 
 This tool can be used to generate supported-backends markdown table.
@@ -244,3 +261,32 @@ And the output look like this:
 | [SAR](https://github.com/open-mmlab/mmocr/tree/main/configs/textrecog/sar)           | TextRecognition |      Y      |    N     |      Y      |   N   |    N     |  N   |
 | [SATRN](https://github.com/open-mmlab/mmocr/tree/main/configs/textrecog/satrn)       | TextRecognition |      Y      |    Y     |      Y      |   N   |    N     |  N   |
 | [ABINet](https://github.com/open-mmlab/mmocr/tree/main/configs/textrecog/abinet)     | TextRecognition |      Y      |    Y     |      Y      |   N   |    N     |  N   |
+
+## visualize
+
+This tool can be used to visualize model inference results in different backend.
+
+### Usage
+
+```bash
+python tools/visualize.py \
+    --deploy-cfg {DEPLOY_CFG} \
+    --model-cfg {MODEL_CFG} \
+    --deploy-path {DEPLOY_PATH} \
+    --test-img {TEST_IMGS} \
+    --checkpoint {CHECKPOINTS} \
+    --batch {BATCH}
+    --save-dir {SAVE_DIR} \
+    --device {DEVICE}
+```
+
+### Description of all arguments
+
+- `deploy-cfg` : The path of the deploy config file in MMDeploy codebase.
+- `model-cfg` : The path of model config file in OpenMMLab codebase.
+- `deploy-path` : The path of the model to be tested, if the backend contains multiple files, you can use it multiple times.
+- `test-img` : The path of the images to be tested, you can use it multiple times.
+- `checkpoint` : The path of the checkpoint to be tested, if it is used, the result will be cancated to right part.
+- `batch` : The batch size of the visual result, if not specified, it will be set to 1.
+- `save-dir` : The path to save the visualization results, if it not specified, it will be set to '.'.
+- `device` : The device type. If not specified, it will be set to `cpu`.

docs/zh_cn/02-how-to-run/useful_tools.md

@@ -244,3 +244,32 @@ python tools/generate_md_table.py tests/regression/mmocr.yml tests/regression/mm
 | [SAR](https://github.com/open-mmlab/mmocr/tree/main/configs/textrecog/sar)           | TextRecognition |      Y      |    N     |      Y      |   N   |    N     |  N   |
 | [SATRN](https://github.com/open-mmlab/mmocr/tree/main/configs/textrecog/satrn)       | TextRecognition |      Y      |    Y     |      Y      |   N   |    N     |  N   |
 | [ABINet](https://github.com/open-mmlab/mmocr/tree/main/configs/textrecog/abinet)     | TextRecognition |      Y      |    Y     |      Y      |   N   |    N     |  N   |
+
+## visualize
+
+这个工具可以用来可视化不同推理引擎的推理结果。
+
+### 用法
+
+```bash
+python tools/visualize.py \
+    --deploy-cfg {DEPLOY_CFG} \
+    --model-cfg {MODEL_CFG} \
+    --deploy-path {DEPLOY_PATH} \
+    --test-img {TEST_IMGS} \
+    --checkpoint {CHECKPOINTS} \
+    --batch {BATCH} \
+    --save-dir {SAVE_DIR} \
+    --device {DEVICE}
+```
+
+### 参数说明
+
+- `deploy-cfg` : 输入的配置文件路径
+- `model-cfg` : 输入的模型配置文件路径
+- `deploy-path` : 测试的模型文件路径，如果部分模型含有多个文件，请多次使用该参数
+- `test-img` : 测试的图片路径，可以多次使用测试测试多张图片
+- `checkpoint` : PyTorch的权重文件，如果使用这个参数，推理的结果会被拼接到图像的右侧
+- `batch` : 可视化的batch大小，默认为1
+- `save-dir` : 保存可视化结果，如果没有指定则会被设为当前目录
+- `device` : 运行的设备类型，如果没有指定则会默认设置为`cpu`

tools/visualize.py

@@ -0,0 +1,125 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+import argparse
+import logging
+import os.path as osp
+
+import mmcv
+import mmengine
+import numpy as np
+
+from mmdeploy.apis.utils import build_task_processor
+from mmdeploy.utils import get_input_shape, get_root_logger, load_config
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(
+        description='Model inference visualization.')
+    parser.add_argument('deploy_cfg', help='deploy config path')
+    parser.add_argument('model_cfg', help='model config path')
+    parser.add_argument(
+        '--model',
+        type=str,
+        nargs='+',
+        required=True,
+        help='deploy model path')
+    parser.add_argument(
+        '--checkpoint', default=None, help='model checkpoint path')
+    parser.add_argument(
+        '--device', help='device type for inference', default='cpu')
+    parser.add_argument(
+        '--test-img',
+        type=str,
+        nargs='+',
+        required=True,
+        help='image used to test model')
+    parser.add_argument(
+        '--batch',
+        type=int,
+        choices=[1, 2],
+        help='batch size for inference, accepts only 1 or 2')
+    parser.add_argument(
+        '--save-dir',
+        default='workdir',
+        help='the dir to save inference results')
+    parser.add_argument(
+        '--show', action='store_true', help='Show detection outputs')
+    parser.add_argument(
+        '--log-level',
+        help='set log level',
+        default='INFO',
+        choices=list(logging._nameToLevel.keys()))
+    args = parser.parse_args()
+    return args
+
+
+def main():
+    args = parse_args()
+    logger = get_root_logger()
+    log_level = logging.getLevelName(args.log_level)
+    logger.setLevel(log_level)
+
+    # load cfgs
+    deploy_cfg, model_cfg = load_config(args.deploy_cfg, args.model_cfg)
+    task_processor = build_task_processor(model_cfg, deploy_cfg, args.device)
+    input_shape = get_input_shape(deploy_cfg)
+    backend_model = task_processor.build_backend_model(
+        args.model,
+        data_preprocessor_updater=task_processor.update_data_preprocessor)
+    torch_model = None
+    if args.checkpoint is not None:
+        torch_model = task_processor.build_pytorch_model(args.checkpoint)
+
+    mmengine.mkdir_or_exist(args.save_dir)
+    # get visualizer
+    visualizer = task_processor.get_visualizer('mmdeploy', args.save_dir)
+
+    for i in range(0, len(args.test_img), args.batch):
+        imgs = args.test_img[i:(i + args.batch)]
+        model_inputs, _ = task_processor.create_input(
+            imgs,
+            input_shape,
+            data_preprocessor=getattr(backend_model, 'data_preprocessor',
+                                      None))
+        backend_results = backend_model.test_step(model_inputs)
+        torch_results = [None] * len(imgs)
+        if torch_model is not None:
+            torch_results = torch_model.test_step(model_inputs)
+
+        # get visualized results
+        for img_path, torch_res, backend_res in zip(imgs, torch_results,
+                                                    backend_results):
+            _, filename = osp.split(img_path)
+            output_file = osp.join(args.save_dir, filename)
+            image = mmcv.imread(img_path, channel_order='rgb')
+            visualizer.add_datasample(
+                filename,
+                image,
+                data_sample=backend_res,
+                draw_gt=False,
+                show=False,
+                out_file=None)
+            drawn_img = visualizer.get_image()
+            if torch_res:
+                visualizer.add_datasample(
+                    filename,
+                    image,
+                    data_sample=torch_res,
+                    draw_gt=False,
+                    show=False,
+                    out_file=None)
+                drawn_img_torch = visualizer.get_image()
+                shape = drawn_img.shape
+                dummy_img = np.full((shape[0], 20, shape[2]),
+                                    255,
+                                    dtype=np.uint8)
+                drawn_img = np.concatenate(
+                    (drawn_img, dummy_img, drawn_img_torch), axis=1)
+            if args.show:
+                visualizer.show(drawn_img, win_name=filename, wait_time=0)
+            drawn_img = mmcv.image.rgb2bgr(drawn_img)
+            mmcv.imwrite(drawn_img, output_file)
+            logger.info(f'Saved to {output_file}')
+
+
+if __name__ == '__main__':
+    main()