add new pipeline

Author: horcham

configs/cls/mobilenetv3/cls_mv3.yaml

@@ -142,3 +142,47 @@ eval:
       drop_remainder: False
       max_rowsize: 12
       num_workers: 8
+
+predict:
+  backend: MindSpore
+  deive_target: Ascend
+  device_id: 0
+  max_device_memory: 8GB
+  amp_level: O0
+  mode: 0
+  ckpt_load_path: /root/.mindspore/models/cls_mobilenetv3-92db9c58.ckpt
+  dataset_sink_mode: False
+  dataset:
+    type: RecDataset
+    dataset_root: dir/to/dataset
+    data_dir: all_images
+    label_file: val_cls_gt.txt
+    sample_ratio: 1.0
+    shuffle: False
+    transform_pipeline:
+      - DecodeImage:
+          img_mode: BGR
+          to_float32: False
+      - Rotate90IfVertical:
+          threshold: 2.0
+          direction: counterclockwise
+      - RecResizeImg:
+          image_shape: [48, 192] # H, W
+          padding: False # aspect ratio will be preserved if true.
+      - NormalizeImage:
+          bgr_to_rgb: True
+          is_hwc: True
+          mean : [127.0, 127.0, 127.0]
+          std : [127.0, 127.0, 127.0]
+      - ToCHWImage:
+    #  the order of the dataloader list, matching the network input and the input labels for the loss function, and optional data for debug/visaulize
+    output_columns: ['image', 'label']  # TODO return text string padding w/ fixed length, and a scaler to indicate the length
+    net_input_column_index: [0] # input indices for network forward func in output_columns
+    label_column_index: [1] # input indices marked as label
+
+  loader:
+      shuffle: False
+      batch_size: 8
+      drop_remainder: False
+      max_rowsize: 12
+      num_workers: 8

configs/det/dbnet/db_r50_icdar15.yaml

@@ -157,7 +157,13 @@ eval:
     num_workers: 2
 
 predict:
-  ckpt_load_path: tmp_det/best.ckpt
+  backend: MindSpore
+  deive_target: Ascend
+  device_id: 0
+  max_device_memory: 8GB
+  amp_level: O0
+  mode: 0
+  ckpt_load_path: /root/.mindspore/models/dbnet_resnet50-c3a4aa24.ckpt
   output_save_dir: ./output
   dataset_sink_mode: False
   dataset:

configs/rec/crnn/crnn_resnet34.yaml

@@ -150,7 +150,13 @@ eval:
       num_workers: 8
 
 predict:
-  ckpt_load_path: ./tmp_rec/best.ckpt
+  backend: MindSpore
+  deive_target: Ascend
+  device_id: 0
+  max_device_memory: 8GB
+  amp_level: O3
+  mode: 0
+  ckpt_load_path: /root/.mindspore/models/crnn_resnet34-83f37f07.ckpt
   vis_font_path: tools/utils/simfang.ttf
   dataset_sink_mode: False
   dataset:

mindocr/infer/classification/__init__.py

@@ -0,0 +1,3 @@
+from .cls_infer_node import ClsInferNode
+from .cls_post_node import ClsPostNode
+from .cls_pre_node import ClsPreNode

mindocr/infer/classification/classification.py

@@ -0,0 +1,63 @@
+import logging
+import os
+import time
+import sys
+import numpy as np
+import yaml
+from addict import Dict
+from typing import List
+
+__dir__ = os.path.dirname(os.path.abspath(__file__))
+sys.path.insert(0, os.path.abspath(os.path.join(__dir__, "../../")))
+
+from tools.infer.text.utils import get_ckpt_file
+from mindocr.data.transforms import create_transforms, run_transforms
+from mindocr.postprocess import build_postprocess
+from mindocr.infer.utils.model import MSModel, LiteModel
+
+
+algo_to_model_name = {
+    "MV3": "cls_mobilenet_v3_small_100_model",
+}
+logger = logging.getLogger("mindocr")
+
+class ClsPreprocess(object):
+    def __init__(self, args):
+        self.args = args
+        with open(args.cls_model_name_or_config, "r") as f:
+            self.yaml_cfg = Dict(yaml.safe_load(f))
+        self.transforms = create_transforms(self.yaml_cfg.predict.dataset.transform_pipeline)
+    
+    def __call__(self, img):
+        print(img)
+        data = {"image": img}
+        data = run_transforms(data, self.transforms[1:])
+        return data
+
+
+class ClsModelMS(MSModel):
+    def __init__(self, args):
+        self.args = args
+        self.model_name = algo_to_model_name[args.cls_algorithm]
+        self.config_path = args.cls_config_path
+        self._init_model(self.model_name, self.config_path)
+
+
+class ClsModelLite(LiteModel):
+    def __init__(self, args):
+        self.args = args
+        self.model_name = algo_to_model_name[args.cls_algorithm]
+        self.config_path = args.cls_config_path
+        self._init_model(self.model_name, self.config_path)
+
+INFER_CLS_MAP = {"MindSporeLite": ClsModelLite, "MindSpore": ClsModelMS}
+
+class ClsPostprocess(object):
+    def __init__(self, args):
+        self.args = args
+        with open(args.cls_model_name_or_config, "r") as f:
+            self.yaml_cfg = Dict(yaml.safe_load(f))        
+        self.postprocessor = build_postprocess(self.yaml_cfg.postprocess)
+    
+    def __call__(self, pred):
+        return self.postprocessor(pred)

mindocr/infer/classification/cls_infer_node.py

@@ -0,0 +1,57 @@
+import os
+import time
+import sys
+import numpy as np
+import yaml
+from addict import Dict
+
+__dir__ = os.path.dirname(os.path.abspath(__file__))
+sys.path.insert(0, os.path.abspath(os.path.join(__dir__, "../../")))
+
+from pipeline.framework.module_base import ModuleBase
+from pipeline.tasks import TaskType
+from .classification import INFER_CLS_MAP
+
+
+class ClsInferNode(ModuleBase):
+    def __init__(self, args, msg_queue, tqdm_info):
+        super(ClsInferNode, self).__init__(args, msg_queue, tqdm_info)
+        self.args = args
+        self.cls_model = None
+        self.task_type = self.args.task_type
+
+    def init_self_args(self):
+        with open(self.args.cls_model_name_or_config, "r") as f:
+            self.yaml_cfg = Dict(yaml.safe_load(f))
+        self.batch_size = self.yaml_cfg.predict.loader.batch_size
+        ClsModel = INFER_CLS_MAP[self.yaml_cfg.predict.backend]
+        self.cls_model = ClsModel(self.args)
+        super().init_self_args()
+
+    def process(self, input_data):
+        """
+        Input:
+          - input_data.data: [np.ndarray], shape:[3,w,h], e.g. [3,48,192]
+        Output:
+          - input_data.data: [np.ndarray], shape:[?,2]
+        """
+        if input_data.skip:
+            self.send_to_next_module(input_data)
+            return
+
+        data = input_data.data
+        data = [np.expand_dims(d, 0) for d in data if len(d.shape) == 3]
+        data = np.concatenate(data, axis=0)
+
+        preds = []
+        for batch_i in range(data.shape[0] // self.batch_size + 1):
+            start_i = batch_i * self.batch_size
+            end_i = (batch_i + 1) * self.batch_size
+            d = data[start_i:end_i]
+            if d.shape[0] == 0:
+                continue
+            pred = self.cls_model([d])
+            preds.append(pred[0])
+        preds = np.concatenate(preds, axis=0)
+        input_data.data = {"pred": preds}
+        self.send_to_next_module(input_data)

mindocr/infer/classification/cls_post_node.py

@@ -0,0 +1,63 @@
+import os
+import time
+import sys
+import numpy as np
+import yaml
+from addict import Dict
+
+__dir__ = os.path.dirname(os.path.abspath(__file__))
+sys.path.insert(0, os.path.abspath(os.path.join(__dir__, "../../")))
+
+import cv2
+import numpy as np
+
+from pipeline.framework.module_base import ModuleBase
+from pipeline.tasks import TaskType
+from .classification import ClsPostprocess
+from tools.infer.text.utils import crop_text_region
+from pipeline.data_process.utils.cv_utils import crop_box_from_image
+
+
+class ClsPostNode(ModuleBase):
+    def __init__(self, args, msg_queue, tqdm_info):
+        super(ClsPostNode, self).__init__(args, msg_queue, tqdm_info)
+        self.cls_postprocess = ClsPostprocess(args)
+        self.task_type = self.args.task_type
+        self.cls_thresh = 0.9
+
+    def init_self_args(self):
+        super().init_self_args()
+
+    def process(self, input_data):
+        """
+        Input:
+          - input_data.data: [np.ndarray], shape:[?,2]
+        Output:
+          - input_data.sub_image_list: [np.ndarray], shape:[1,3,-1,-1], e.g. [1,3,48,192]
+          - input_data.data = None
+          or
+          - input_data.infer_result = [(str, float)]
+        """
+        if input_data.skip:
+            self.send_to_next_module(input_data)
+            return
+
+        data = input_data.data
+        pred = data["pred"]
+        output = self.cls_postprocess(pred)
+        angles = output["angles"]
+        scores = np.array(output["scores"]).tolist()
+
+        batch = input_data.sub_image_size
+        if self.task_type.value == TaskType.DET_CLS_REC.value:
+            sub_images = input_data.sub_image_list
+            for i in range(batch):
+                angle, score = angles[i], scores[i]
+                if "180" == angle and score > self.cls_thresh:
+                    sub_images[i] = cv2.rotate(sub_images[i], cv2.ROTATE_180)
+            input_data.sub_image_list = sub_images
+        else:
+            input_data.infer_result = [(angle, score) for angle, score in zip(angles, scores)]
+
+        input_data.data = None
+        self.send_to_next_module(input_data)

mindocr/infer/classification/cls_pre_node.py

@@ -0,0 +1,40 @@
+import argparse
+import os
+import time
+import sys
+import numpy as np
+
+__dir__ = os.path.dirname(os.path.abspath(__file__))
+sys.path.insert(0, os.path.abspath(os.path.join(__dir__, "../../")))
+
+from pipeline.framework.module_base import ModuleBase
+from pipeline.tasks import TaskType
+from .classification import ClsPreprocess
+
+
+class ClsPreNode(ModuleBase):
+    def __init__(self, args, msg_queue, tqdm_info):
+        super(ClsPreNode, self).__init__(args, msg_queue, tqdm_info)
+        self.cls_preprocesser = ClsPreprocess(args)
+        self.task_type = self.args.task_type
+
+    def init_self_args(self):
+        super().init_self_args()
+        return {"batch_size": 1}
+
+    def process(self, input_data):
+        if input_data.skip:
+            self.send_to_next_module(input_data)
+            return
+
+        if self.task_type.value == TaskType.REC.value:
+            image = input_data.frame[0]
+            data = [self.cls_preprocesser(image)["image"]]
+            input_data.sub_image_size = 1
+            input_data.data = data
+            self.send_to_next_module(input_data)
+        else:
+            sub_image_list = input_data.sub_image_list
+            data = [self.cls_preprocesser(split_image)["image"] for split_image in sub_image_list]
+            input_data.data = data
+            self.send_to_next_module(input_data)

mindocr/infer/common/__init__.py

@@ -0,0 +1,3 @@
+from .collect_node import CollectNode
+from .decode_node import DecodeNode
+from .handout_node import HandoutNode