Merge branch 'PaddlePaddle:dygraph' into dygraph

ppstructure/vqa/train_ser.py

+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# 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.
+
+import sys
+import os
+import random
+import copy
+import logging
+
+import argparse
+import paddle
+import numpy as np
+from seqeval.metrics import classification_report, f1_score, precision_score, recall_score
+from paddlenlp.transformers import LayoutXLMModel, LayoutXLMTokenizer, LayoutXLMForTokenClassification
+from xfun import XFUNDataset
+from utils import parse_args
+from utils import get_bio_label_maps
+
+logger = logging.getLogger(__name__)
+
+
+def set_seed(args):
+    random.seed(args.seed)
+    np.random.seed(args.seed)
+    paddle.seed(args.seed)
+
+
+def train(args):
+    os.makedirs(args.output_dir, exist_ok=True)
+    logging.basicConfig(
+        filename=os.path.join(args.output_dir, "train.log")
+        if paddle.distributed.get_rank() == 0 else None,
+        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        level=logging.INFO
+        if paddle.distributed.get_rank() == 0 else logging.WARN, )
+
+    ch = logging.StreamHandler()
+    ch.setLevel(logging.DEBUG)
+    logger.addHandler(ch)
+
+    label2id_map, id2label_map = get_bio_label_maps(args.label_map_path)
+    pad_token_label_id = paddle.nn.CrossEntropyLoss().ignore_index
+
+    # dist mode
+    if paddle.distributed.get_world_size() > 1:
+        paddle.distributed.init_parallel_env()
+
+    tokenizer = LayoutXLMTokenizer.from_pretrained(args.model_name_or_path)
+    base_model = LayoutXLMModel.from_pretrained(args.model_name_or_path)
+    model = LayoutXLMForTokenClassification(
+        base_model, num_classes=len(label2id_map), dropout=None)
+
+    # dist mode
+    if paddle.distributed.get_world_size() > 1:
+        model = paddle.DataParallel(model)
+
+    train_dataset = XFUNDataset(
+        tokenizer,
+        data_dir=args.train_data_dir,
+        label_path=args.train_label_path,
+        label2id_map=label2id_map,
+        img_size=(224, 224),
+        pad_token_label_id=pad_token_label_id,
+        contains_re=False,
+        add_special_ids=False,
+        return_attention_mask=True,
+        load_mode='all')
+
+    train_sampler = paddle.io.DistributedBatchSampler(
+        train_dataset, batch_size=args.per_gpu_train_batch_size, shuffle=True)
+
+    args.train_batch_size = args.per_gpu_train_batch_size * max(
+        1, paddle.distributed.get_world_size())
+
+    train_dataloader = paddle.io.DataLoader(
+        train_dataset,
+        batch_sampler=train_sampler,
+        num_workers=0,
+        use_shared_memory=True,
+        collate_fn=None, )
+
+    t_total = len(train_dataloader) * args.num_train_epochs
+
+    # build linear decay with warmup lr sch
+    lr_scheduler = paddle.optimizer.lr.PolynomialDecay(
+        learning_rate=args.learning_rate,
+        decay_steps=t_total,
+        end_lr=0.0,
+        power=1.0)
+    if args.warmup_steps > 0:
+        lr_scheduler = paddle.optimizer.lr.LinearWarmup(
+            lr_scheduler,
+            args.warmup_steps,
+            start_lr=0,
+            end_lr=args.learning_rate, )
+
+    optimizer = paddle.optimizer.AdamW(
+        learning_rate=lr_scheduler,
+        parameters=model.parameters(),
+        epsilon=args.adam_epsilon,
+        weight_decay=args.weight_decay)
+
+    # Train!
+    logger.info("***** Running training *****")
+    logger.info("  Num examples = %d", len(train_dataset))
+    logger.info("  Num Epochs = %d", args.num_train_epochs)
+    logger.info("  Instantaneous batch size per GPU = %d",
+                args.per_gpu_train_batch_size)
+    logger.info(
+        "  Total train batch size (w. parallel, distributed) = %d",
+        args.train_batch_size * paddle.distributed.get_world_size(), )
+    logger.info("  Total optimization steps = %d", t_total)
+
+    global_step = 0
+    tr_loss = 0.0
+    set_seed(args)
+    best_metrics = None
+
+    for epoch_id in range(args.num_train_epochs):
+        for step, batch in enumerate(train_dataloader):
+            model.train()
+            outputs = model(**batch)
+            # model outputs are always tuple in ppnlp (see doc)
+            loss = outputs[0]
+            loss = loss.mean()
+            logger.info(
+                "[epoch {}/{}][iter: {}/{}] lr: {:.5f}, train loss: {:.5f}, ".
+                format(epoch_id, args.num_train_epochs, step,
+                       len(train_dataloader),
+                       lr_scheduler.get_lr(), loss.numpy()[0]))
+
+            loss.backward()
+            tr_loss += loss.item()
+            optimizer.step()
+            lr_scheduler.step()  # Update learning rate schedule
+            optimizer.clear_grad()
+            global_step += 1
+
+            if (paddle.distributed.get_rank() == 0 and args.eval_steps > 0 and
+                    global_step % args.eval_steps == 0):
+                # Log metrics
+                # Only evaluate when single GPU otherwise metrics may not average well
+                if paddle.distributed.get_rank(
+                ) == 0 and args.evaluate_during_training:
+                    results, _ = evaluate(
+                        args,
+                        model,
+                        tokenizer,
+                        label2id_map,
+                        id2label_map,
+                        pad_token_label_id, )
+
+                    if best_metrics is None or results["f1"] >= best_metrics[
+                            "f1"]:
+                        best_metrics = copy.deepcopy(results)
+                        output_dir = os.path.join(args.output_dir, "best_model")
+                        os.makedirs(output_dir, exist_ok=True)
+                        if paddle.distributed.get_rank() == 0:
+                            model.save_pretrained(output_dir)
+                            tokenizer.save_pretrained(output_dir)
+                            paddle.save(
+                                args,
+                                os.path.join(output_dir, "training_args.bin"))
+                            logger.info("Saving model checkpoint to %s",
+                                        output_dir)
+
+                    logger.info("[epoch {}/{}][iter: {}/{}] results: {}".format(
+                        epoch_id, args.num_train_epochs, step,
+                        len(train_dataloader), results))
+                    if best_metrics is not None:
+                        logger.info("best metrics: {}".format(best_metrics))
+
+            if paddle.distributed.get_rank(
+            ) == 0 and args.save_steps > 0 and global_step % args.save_steps == 0:
+                # Save model checkpoint
+                output_dir = os.path.join(args.output_dir,
+                                          "checkpoint-{}".format(global_step))
+                os.makedirs(output_dir, exist_ok=True)
+                if paddle.distributed.get_rank() == 0:
+                    model.save_pretrained(output_dir)
+                    tokenizer.save_pretrained(output_dir)
+                    paddle.save(args,
+                                os.path.join(output_dir, "training_args.bin"))
+                    logger.info("Saving model checkpoint to %s", output_dir)
+
+    return global_step, tr_loss / global_step
+
+
+def evaluate(args,
+             model,
+             tokenizer,
+             label2id_map,
+             id2label_map,
+             pad_token_label_id,
+             prefix=""):
+    eval_dataset = XFUNDataset(
+        tokenizer,
+        data_dir=args.eval_data_dir,
+        label_path=args.eval_label_path,
+        label2id_map=label2id_map,
+        img_size=(224, 224),
+        pad_token_label_id=pad_token_label_id,
+        contains_re=False,
+        add_special_ids=False,
+        return_attention_mask=True,
+        load_mode='all')
+
+    args.eval_batch_size = args.per_gpu_eval_batch_size * max(
+        1, paddle.distributed.get_world_size())
+
+    eval_dataloader = paddle.io.DataLoader(
+        eval_dataset,
+        batch_size=args.eval_batch_size,
+        num_workers=0,
+        use_shared_memory=True,
+        collate_fn=None, )
+
+    # Eval!
+    logger.info("***** Running evaluation %s *****", prefix)
+    logger.info("  Num examples = %d", len(eval_dataset))
+    logger.info("  Batch size = %d", args.eval_batch_size)
+    eval_loss = 0.0
+    nb_eval_steps = 0
+    preds = None
+    out_label_ids = None
+    model.eval()
+    for idx, batch in enumerate(eval_dataloader):
+        with paddle.no_grad():
+            outputs = model(**batch)
+            tmp_eval_loss, logits = outputs[:2]
+
+            tmp_eval_loss = tmp_eval_loss.mean()
+
+            if paddle.distributed.get_rank() == 0:
+                logger.info("[Eval]process: {}/{}, loss: {:.5f}".format(
+                    idx, len(eval_dataloader), tmp_eval_loss.numpy()[0]))
+
+            eval_loss += tmp_eval_loss.item()
+        nb_eval_steps += 1
+        if preds is None:
+            preds = logits.numpy()
+            out_label_ids = batch["labels"].numpy()
+        else:
+            preds = np.append(preds, logits.numpy(), axis=0)
+            out_label_ids = np.append(
+                out_label_ids, batch["labels"].numpy(), axis=0)
+
+    eval_loss = eval_loss / nb_eval_steps
+    preds = np.argmax(preds, axis=2)
+
+    # label_map = {i: label.upper() for i, label in enumerate(labels)}
+
+    out_label_list = [[] for _ in range(out_label_ids.shape[0])]
+    preds_list = [[] for _ in range(out_label_ids.shape[0])]
+
+    for i in range(out_label_ids.shape[0]):
+        for j in range(out_label_ids.shape[1]):
+            if out_label_ids[i, j] != pad_token_label_id:
+                out_label_list[i].append(id2label_map[out_label_ids[i][j]])
+                preds_list[i].append(id2label_map[preds[i][j]])
+
+    results = {
+        "loss": eval_loss,
+        "precision": precision_score(out_label_list, preds_list),
+        "recall": recall_score(out_label_list, preds_list),
+        "f1": f1_score(out_label_list, preds_list),
+    }
+
+    with open(os.path.join(args.output_dir, "test_gt.txt"), "w") as fout:
+        for lbl in out_label_list:
+            for l in lbl:
+                fout.write(l + "\t")
+            fout.write("\n")
+    with open(os.path.join(args.output_dir, "test_pred.txt"), "w") as fout:
+        for lbl in preds_list:
+            for l in lbl:
+                fout.write(l + "\t")
+            fout.write("\n")
+
+    report = classification_report(out_label_list, preds_list)
+    logger.info("\n" + report)
+
+    logger.info("***** Eval results %s *****", prefix)
+    for key in sorted(results.keys()):
+        logger.info("  %s = %s", key, str(results[key]))
+
+    return results, preds_list
+
+
+def print_arguments(args):
+    """print arguments"""
+    print('-----------  Configuration Arguments -----------')
+    for arg, value in sorted(vars(args).items()):
+        print('%s: %s' % (arg, value))
+    print('------------------------------------------------')
+
+
+if __name__ == "__main__":
+    args = parse_args()
+    print_arguments(args)
+    train(args)

ppstructure/vqa/utils.py

+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# 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.
+
+import os
+import argparse
+import cv2
+import random
+import numpy as np
+import imghdr
+from copy import deepcopy
+
+import paddle
+
+from PIL import Image, ImageDraw, ImageFont
+
+from paddleocr import PaddleOCR
+
+
+def get_bio_label_maps(label_map_path):
+    with open(label_map_path, "r") as fin:
+        lines = fin.readlines()
+    lines = [line.strip() for line in lines]
+    if "O" not in lines:
+        lines.insert(0, "O")
+    labels = []
+    for line in lines:
+        if line == "O":
+            labels.append("O")
+        else:
+            labels.append("B-" + line)
+            labels.append("I-" + line)
+    label2id_map = {label: idx for idx, label in enumerate(labels)}
+    id2label_map = {idx: label for idx, label in enumerate(labels)}
+    return label2id_map, id2label_map
+
+
+def get_image_file_list(img_file):
+    imgs_lists = []
+    if img_file is None or not os.path.exists(img_file):
+        raise Exception("not found any img file in {}".format(img_file))
+
+    img_end = {'jpg', 'bmp', 'png', 'jpeg', 'rgb', 'tif', 'tiff', 'gif', 'GIF'}
+    if os.path.isfile(img_file) and imghdr.what(img_file) in img_end:
+        imgs_lists.append(img_file)
+    elif os.path.isdir(img_file):
+        for single_file in os.listdir(img_file):
+            file_path = os.path.join(img_file, single_file)
+            if os.path.isfile(file_path) and imghdr.what(file_path) in img_end:
+                imgs_lists.append(file_path)
+    if len(imgs_lists) == 0:
+        raise Exception("not found any img file in {}".format(img_file))
+    imgs_lists = sorted(imgs_lists)
+    return imgs_lists
+
+
+def draw_ser_results(image,
+                     ocr_results,
+                     font_path="../doc/fonts/simfang.ttf",
+                     font_size=18):
+    np.random.seed(0)
+    color = (np.random.permutation(range(255)),
+             np.random.permutation(range(255)),
+             np.random.permutation(range(255)))
+    color_map = {
+        idx: (color[0][idx], color[1][idx], color[2][idx])
+        for idx in range(1, 255)
+    }
+    if isinstance(image, np.ndarray):
+        image = Image.fromarray(image)
+    img_new = image.copy()
+    draw = ImageDraw.Draw(img_new)
+
+    font = ImageFont.truetype(font_path, font_size, encoding="utf-8")
+
+    for ocr_info in ocr_results:
+        if ocr_info["pred_id"] not in color_map:
+            continue
+        color = color_map[ocr_info["pred_id"]]
+
+        # draw ocr results outline
+        bbox = ocr_info["bbox"]
+        bbox = ((bbox[0], bbox[1]), (bbox[2], bbox[3]))
+        draw.rectangle(bbox, fill=color)
+
+        # draw ocr results
+        text = "{}: {}".format(ocr_info["pred"], ocr_info["text"])
+        start_y = max(0, bbox[0][1] - font_size)
+        tw = font.getsize(text)[0]
+        draw.rectangle(
+            [(bbox[0][0] + 1, start_y), (bbox[0][0] + tw + 1,
+                                         start_y + font_size)],
+            fill=(0, 0, 255))
+        draw.text(
+            (bbox[0][0] + 1, start_y), text, fill=(255, 255, 255), font=font)
+
+    img_new = Image.blend(image, img_new, 0.5)
+    return np.array(img_new)
+
+
+def build_ocr_engine(rec_model_dir, det_model_dir):
+    ocr_engine = PaddleOCR(
+        rec_model_dir=rec_model_dir,
+        det_model_dir=det_model_dir,
+        use_angle_cls=False)
+    return ocr_engine
+
+
+# pad sentences
+def pad_sentences(tokenizer,
+                  encoded_inputs,
+                  max_seq_len=512,
+                  pad_to_max_seq_len=True,
+                  return_attention_mask=True,
+                  return_token_type_ids=True,
+                  return_overflowing_tokens=False,
+                  return_special_tokens_mask=False):
+    # Padding with larger size, reshape is carried out
+    max_seq_len = (
+        len(encoded_inputs["input_ids"]) // max_seq_len + 1) * max_seq_len
+
+    needs_to_be_padded = pad_to_max_seq_len and \
+                         max_seq_len and len(encoded_inputs["input_ids"]) < max_seq_len
+
+    if needs_to_be_padded:
+        difference = max_seq_len - len(encoded_inputs["input_ids"])
+        if tokenizer.padding_side == 'right':
+            if return_attention_mask:
+                encoded_inputs["attention_mask"] = [1] * len(encoded_inputs[
+                    "input_ids"]) + [0] * difference
+            if return_token_type_ids:
+                encoded_inputs["token_type_ids"] = (
+                    encoded_inputs["token_type_ids"] +
+                    [tokenizer.pad_token_type_id] * difference)
+            if return_special_tokens_mask:
+                encoded_inputs["special_tokens_mask"] = encoded_inputs[
+                    "special_tokens_mask"] + [1] * difference
+            encoded_inputs["input_ids"] = encoded_inputs[
+                "input_ids"] + [tokenizer.pad_token_id] * difference
+            encoded_inputs["bbox"] = encoded_inputs["bbox"] + [[0, 0, 0, 0]
+                                                               ] * difference
+    else:
+        if return_attention_mask:
+            encoded_inputs["attention_mask"] = [1] * len(encoded_inputs[
+                "input_ids"])
+
+    return encoded_inputs
+
+
+def split_page(encoded_inputs, max_seq_len=512):
+    """
+    truncate is often used in training process
+    """
+    for key in encoded_inputs:
+        encoded_inputs[key] = paddle.to_tensor(encoded_inputs[key])
+        if encoded_inputs[key].ndim <= 1:  # for input_ids, att_mask and so on
+            encoded_inputs[key] = encoded_inputs[key].reshape([-1, max_seq_len])
+        else:  # for bbox
+            encoded_inputs[key] = encoded_inputs[key].reshape(
+                [-1, max_seq_len, 4])
+    return encoded_inputs
+
+
+def preprocess(
+        tokenizer,
+        ori_img,
+        ocr_info,
+        img_size=(224, 224),
+        pad_token_label_id=-100,
+        max_seq_len=512,
+        add_special_ids=False,
+        return_attention_mask=True, ):
+    ocr_info = deepcopy(ocr_info)
+    height = ori_img.shape[0]
+    width = ori_img.shape[1]
+
+    img = cv2.resize(ori_img,
+                     (224, 224)).transpose([2, 0, 1]).astype(np.float32)
+
+    segment_offset_id = []
+    words_list = []
+    bbox_list = []
+    input_ids_list = []
+    token_type_ids_list = []
+
+    for info in ocr_info:
+        # x1, y1, x2, y2
+        bbox = info["bbox"]
+        bbox[0] = int(bbox[0] * 1000.0 / width)
+        bbox[2] = int(bbox[2] * 1000.0 / width)
+        bbox[1] = int(bbox[1] * 1000.0 / height)
+        bbox[3] = int(bbox[3] * 1000.0 / height)
+
+        text = info["text"]
+        encode_res = tokenizer.encode(
+            text, pad_to_max_seq_len=False, return_attention_mask=True)
+
+        if not add_special_ids:
+            # TODO: use tok.all_special_ids to remove
+            encode_res["input_ids"] = encode_res["input_ids"][1:-1]
+            encode_res["token_type_ids"] = encode_res["token_type_ids"][1:-1]
+            encode_res["attention_mask"] = encode_res["attention_mask"][1:-1]
+
+        input_ids_list.extend(encode_res["input_ids"])
+        token_type_ids_list.extend(encode_res["token_type_ids"])
+        bbox_list.extend([bbox] * len(encode_res["input_ids"]))
+        words_list.append(text)
+        segment_offset_id.append(len(input_ids_list))
+
+    encoded_inputs = {
+        "input_ids": input_ids_list,
+        "token_type_ids": token_type_ids_list,
+        "bbox": bbox_list,
+        "attention_mask": [1] * len(input_ids_list),
+    }
+
+    encoded_inputs = pad_sentences(
+        tokenizer,
+        encoded_inputs,
+        max_seq_len=max_seq_len,
+        return_attention_mask=return_attention_mask)
+
+    encoded_inputs = split_page(encoded_inputs)
+
+    fake_bs = encoded_inputs["input_ids"].shape[0]
+
+    encoded_inputs["image"] = paddle.to_tensor(img).unsqueeze(0).expand(
+        [fake_bs] + list(img.shape))
+
+    encoded_inputs["segment_offset_id"] = segment_offset_id
+
+    return encoded_inputs
+
+
+def postprocess(attention_mask, preds, id2label_map):
+    if isinstance(preds, paddle.Tensor):
+        preds = preds.numpy()
+    preds = np.argmax(preds, axis=2)
+
+    preds_list = [[] for _ in range(preds.shape[0])]
+
+    # keep batch info
+    for i in range(preds.shape[0]):
+        for j in range(preds.shape[1]):
+            if attention_mask[i][j] == 1:
+                preds_list[i].append(id2label_map[preds[i][j]])
+
+    return preds_list
+
+
+def merge_preds_list_with_ocr_info(ocr_info, segment_offset_id, preds_list,
+                                   label2id_map_for_draw):
+    # must ensure the preds_list is generated from the same image
+    preds = [p for pred in preds_list for p in pred]
+
+    id2label_map = dict()
+    for key in label2id_map_for_draw:
+        val = label2id_map_for_draw[key]
+        if key == "O":
+            id2label_map[val] = key
+        if key.startswith("B-") or key.startswith("I-"):
+            id2label_map[val] = key[2:]
+        else:
+            id2label_map[val] = key
+
+    for idx in range(len(segment_offset_id)):
+        if idx == 0:
+            start_id = 0
+        else:
+            start_id = segment_offset_id[idx - 1]
+
+        end_id = segment_offset_id[idx]
+
+        curr_pred = preds[start_id:end_id]
+        curr_pred = [label2id_map_for_draw[p] for p in curr_pred]
+
+        if len(curr_pred) <= 0:
+            pred_id = 0
+        else:
+            counts = np.bincount(curr_pred)
+            pred_id = np.argmax(counts)
+        ocr_info[idx]["pred_id"] = int(pred_id)
+        ocr_info[idx]["pred"] = id2label_map[int(pred_id)]
+    return ocr_info
+
+
+def parse_args():
+    parser = argparse.ArgumentParser()
+    # Required parameters
+    # yapf: disable
+    parser.add_argument("--model_name_or_path", default=None, type=str, required=True,)
+    parser.add_argument("--train_data_dir", default=None, type=str, required=False,)
+    parser.add_argument("--train_label_path", default=None, type=str, required=False,)
+    parser.add_argument("--eval_data_dir", default=None, type=str, required=False,)
+    parser.add_argument("--eval_label_path", default=None, type=str, required=False,)
+    parser.add_argument("--output_dir", default=None, type=str, required=True,)
+    parser.add_argument("--max_seq_length", default=512, type=int,)
+    parser.add_argument("--evaluate_during_training", action="store_true",)
+    parser.add_argument("--per_gpu_train_batch_size", default=8, type=int, help="Batch size per GPU/CPU for training.",)
+    parser.add_argument("--per_gpu_eval_batch_size", default=8, type=int, help="Batch size per GPU/CPU for eval.",)
+    parser.add_argument("--learning_rate", default=5e-5, type=float, help="The initial learning rate for Adam.",)
+    parser.add_argument("--weight_decay", default=0.0, type=float, help="Weight decay if we apply some.",)
+    parser.add_argument("--adam_epsilon", default=1e-8, type=float, help="Epsilon for Adam optimizer.",)
+    parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.",)
+    parser.add_argument("--num_train_epochs", default=3, type=int, help="Total number of training epochs to perform.",)
+    parser.add_argument("--warmup_steps", default=0, type=int, help="Linear warmup over warmup_steps.",)
+    parser.add_argument("--eval_steps", type=int, default=10, help="eval every X updates steps.",)
+    parser.add_argument("--save_steps", type=int, default=50, help="Save checkpoint every X updates steps.",)
+    parser.add_argument("--seed", type=int, default=2048, help="random seed for initialization",)
+
+    parser.add_argument("--ocr_rec_model_dir", default=None, type=str, )
+    parser.add_argument("--ocr_det_model_dir", default=None, type=str, )
+    parser.add_argument("--label_map_path", default="./labels/labels_ser.txt", type=str, required=False, )
+    parser.add_argument("--infer_imgs", default=None, type=str, required=False)
+    parser.add_argument("--ocr_json_path", default=None, type=str, required=False, help="ocr prediction results")
+    # yapf: enable
+    args = parser.parse_args()
+    return args

ppstructure/vqa/xfun.py

+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# 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.
+
+import json
+import os
+import cv2
+import numpy as np
+import paddle
+import copy
+from paddle.io import Dataset
+
+__all__ = ["XFUNDataset"]
+
+
+class XFUNDataset(Dataset):
+    """
+    Example:
+        print("=====begin to build dataset=====")
+        from paddlenlp.transformers import LayoutXLMTokenizer
+        tokenizer = LayoutXLMTokenizer.from_pretrained("/paddle/models/transformers/layoutxlm-base-paddle/")
+        tok_res = tokenizer.tokenize("Maribyrnong")
+        # res = tokenizer.convert_ids_to_tokens(val_data["input_ids"][0])
+        dataset = XfunDatasetForSer(
+            tokenizer,
+            data_dir="./zh.val/",
+            label_path="zh.val/xfun_normalize_val.json",
+            img_size=(224,224))
+        print(len(dataset))
+
+        data = dataset[0]
+        print(data.keys())
+        print("input_ids: ", data["input_ids"])
+        print("labels: ", data["labels"])
+        print("token_type_ids: ", data["token_type_ids"])
+        print("words_list: ", data["words_list"])
+        print("image shape: ", data["image"].shape)
+    """
+
+    def __init__(self,
+                 tokenizer,
+                 data_dir,
+                 label_path,
+                 contains_re=False,
+                 label2id_map=None,
+                 img_size=(224, 224),
+                 pad_token_label_id=None,
+                 add_special_ids=False,
+                 return_attention_mask=True,
+                 load_mode='all',
+                 max_seq_len=512):
+        super().__init__()
+        self.tokenizer = tokenizer
+        self.data_dir = data_dir
+        self.label_path = label_path
+        self.contains_re = contains_re
+        self.label2id_map = label2id_map
+        self.img_size = img_size
+        self.pad_token_label_id = pad_token_label_id
+        self.add_special_ids = add_special_ids
+        self.return_attention_mask = return_attention_mask
+        self.load_mode = load_mode
+        self.max_seq_len = max_seq_len
+
+        if self.pad_token_label_id is None:
+            self.pad_token_label_id = paddle.nn.CrossEntropyLoss().ignore_index
+
+        self.all_lines = self.read_all_lines()
+
+        self.entities_labels = {'HEADER': 0, 'QUESTION': 1, 'ANSWER': 2}
+        self.return_keys = {
+            'bbox': 'np',
+            'input_ids': 'np',
+            'labels': 'np',
+            'attention_mask': 'np',
+            'image': 'np',
+            'token_type_ids': 'np',
+            'entities': 'dict',
+            'relations': 'dict',
+        }
+
+        if load_mode == "all":
+            self.encoded_inputs_all = self._parse_label_file_all()
+
+    def pad_sentences(self,
+                      encoded_inputs,
+                      max_seq_len=512,
+                      pad_to_max_seq_len=True,
+                      return_attention_mask=True,
+                      return_token_type_ids=True,
+                      truncation_strategy="longest_first",
+                      return_overflowing_tokens=False,
+                      return_special_tokens_mask=False):
+        # Padding
+        needs_to_be_padded = pad_to_max_seq_len and \
+                             max_seq_len and len(encoded_inputs["input_ids"]) < max_seq_len
+
+        if needs_to_be_padded:
+            difference = max_seq_len - len(encoded_inputs["input_ids"])
+            if self.tokenizer.padding_side == 'right':
+                if return_attention_mask:
+                    encoded_inputs["attention_mask"] = [1] * len(encoded_inputs[
+                        "input_ids"]) + [0] * difference
+                if return_token_type_ids:
+                    encoded_inputs["token_type_ids"] = (
+                        encoded_inputs["token_type_ids"] +
+                        [self.tokenizer.pad_token_type_id] * difference)
+                if return_special_tokens_mask:
+                    encoded_inputs["special_tokens_mask"] = encoded_inputs[
+                        "special_tokens_mask"] + [1] * difference
+                encoded_inputs["input_ids"] = encoded_inputs[
+                    "input_ids"] + [self.tokenizer.pad_token_id] * difference
+                encoded_inputs["labels"] = encoded_inputs[
+                    "labels"] + [self.pad_token_label_id] * difference
+                encoded_inputs["bbox"] = encoded_inputs[
+                    "bbox"] + [[0, 0, 0, 0]] * difference
+            elif self.tokenizer.padding_side == 'left':
+                if return_attention_mask:
+                    encoded_inputs["attention_mask"] = [0] * difference + [
+                        1
+                    ] * len(encoded_inputs["input_ids"])
+                if return_token_type_ids:
+                    encoded_inputs["token_type_ids"] = (
+                        [self.tokenizer.pad_token_type_id] * difference +
+                        encoded_inputs["token_type_ids"])
+                if return_special_tokens_mask:
+                    encoded_inputs["special_tokens_mask"] = [
+                        1
+                    ] * difference + encoded_inputs["special_tokens_mask"]
+                encoded_inputs["input_ids"] = [
+                    self.tokenizer.pad_token_id
+                ] * difference + encoded_inputs["input_ids"]
+                encoded_inputs["labels"] = [
+                    self.pad_token_label_id
+                ] * difference + encoded_inputs["labels"]
+                encoded_inputs["bbox"] = [
+                    [0, 0, 0, 0]
+                ] * difference + encoded_inputs["bbox"]
+        else:
+            if return_attention_mask:
+                encoded_inputs["attention_mask"] = [1] * len(encoded_inputs[
+                    "input_ids"])
+
+        return encoded_inputs
+
+    def truncate_inputs(self, encoded_inputs, max_seq_len=512):
+        for key in encoded_inputs:
+            if key == "sample_id":
+                continue
+            length = min(len(encoded_inputs[key]), max_seq_len)
+            encoded_inputs[key] = encoded_inputs[key][:length]
+        return encoded_inputs
+
+    def read_all_lines(self, ):
+        with open(self.label_path, "r") as fin:
+            lines = fin.readlines()
+        return lines
+
+    def _parse_label_file_all(self):
+        """
+        parse all samples
+        """
+        encoded_inputs_all = []
+        for line in self.all_lines:
+            encoded_inputs_all.extend(self._parse_label_file(line))
+        return encoded_inputs_all
+
+    def _parse_label_file(self, line):
+        """
+        parse single sample
+        """
+
+        image_name, info_str = line.split("\t")
+        image_path = os.path.join(self.data_dir, image_name)
+
+        def add_imgge_path(x):
+            x['image_path'] = image_path
+            return x
+
+        encoded_inputs = self._read_encoded_inputs_sample(info_str)
+        if self.contains_re:
+            encoded_inputs = self._chunk_re(encoded_inputs)
+        else:
+            encoded_inputs = self._chunk_ser(encoded_inputs)
+        encoded_inputs = list(map(add_imgge_path, encoded_inputs))
+        return encoded_inputs
+
+    def _read_encoded_inputs_sample(self, info_str):
+        """
+        parse label info
+        """
+        # read text info
+        info_dict = json.loads(info_str)
+        height = info_dict["height"]
+        width = info_dict["width"]
+
+        words_list = []
+        bbox_list = []
+        input_ids_list = []
+        token_type_ids_list = []
+        gt_label_list = []
+
+        if self.contains_re:
+            # for re
+            entities = []
+            relations = []
+            id2label = {}
+            entity_id_to_index_map = {}
+            empty_entity = set()
+        for info in info_dict["ocr_info"]:
+            if self.contains_re:
+                # for re
+                if len(info["text"]) == 0:
+                    empty_entity.add(info["id"])
+                    continue
+                id2label[info["id"]] = info["label"]
+                relations.extend([tuple(sorted(l)) for l in info["linking"]])
+
+            # x1, y1, x2, y2
+            bbox = info["bbox"]
+            label = info["label"]
+            bbox[0] = int(bbox[0] * 1000.0 / width)
+            bbox[2] = int(bbox[2] * 1000.0 / width)
+            bbox[1] = int(bbox[1] * 1000.0 / height)
+            bbox[3] = int(bbox[3] * 1000.0 / height)
+
+            text = info["text"]
+            encode_res = self.tokenizer.encode(
+                text, pad_to_max_seq_len=False, return_attention_mask=True)
+
+            gt_label = []
+            if not self.add_special_ids:
+                # TODO: use tok.all_special_ids to remove
+                encode_res["input_ids"] = encode_res["input_ids"][1:-1]
+                encode_res["token_type_ids"] = encode_res["token_type_ids"][1:
+                                                                            -1]
+                encode_res["attention_mask"] = encode_res["attention_mask"][1:
+                                                                            -1]
+            if label.lower() == "other":
+                gt_label.extend([0] * len(encode_res["input_ids"]))
+            else:
+                gt_label.append(self.label2id_map[("b-" + label).upper()])
+                gt_label.extend([self.label2id_map[("i-" + label).upper()]] *
+                                (len(encode_res["input_ids"]) - 1))
+            if self.contains_re:
+                if gt_label[0] != self.label2id_map["O"]:
+                    entity_id_to_index_map[info["id"]] = len(entities)
+                    entities.append({
+                        "start": len(input_ids_list),
+                        "end":
+                        len(input_ids_list) + len(encode_res["input_ids"]),
+                        "label": label.upper(),
+                    })
+            input_ids_list.extend(encode_res["input_ids"])
+            token_type_ids_list.extend(encode_res["token_type_ids"])
+            bbox_list.extend([bbox] * len(encode_res["input_ids"]))
+            gt_label_list.extend(gt_label)
+            words_list.append(text)
+
+        encoded_inputs = {
+            "input_ids": input_ids_list,
+            "labels": gt_label_list,
+            "token_type_ids": token_type_ids_list,
+            "bbox": bbox_list,
+            "attention_mask": [1] * len(input_ids_list),
+            # "words_list": words_list,
+        }
+        encoded_inputs = self.pad_sentences(
+            encoded_inputs,
+            max_seq_len=self.max_seq_len,
+            return_attention_mask=self.return_attention_mask)
+        encoded_inputs = self.truncate_inputs(encoded_inputs)
+
+        if self.contains_re:
+            relations = self._relations(entities, relations, id2label,
+                                        empty_entity, entity_id_to_index_map)
+            encoded_inputs['relations'] = relations
+            encoded_inputs['entities'] = entities
+        return encoded_inputs
+
+    def _chunk_ser(self, encoded_inputs):
+        encoded_inputs_all = []
+        seq_len = len(encoded_inputs['input_ids'])
+        chunk_size = 512
+        for chunk_id, index in enumerate(range(0, seq_len, chunk_size)):
+            chunk_beg = index
+            chunk_end = min(index + chunk_size, seq_len)
+            encoded_inputs_example = {}
+            for key in encoded_inputs:
+                encoded_inputs_example[key] = encoded_inputs[key][chunk_beg:
+                                                                  chunk_end]
+
+            encoded_inputs_all.append(encoded_inputs_example)
+        return encoded_inputs_all
+
+    def _chunk_re(self, encoded_inputs):
+        # prepare data
+        entities = encoded_inputs.pop('entities')
+        relations = encoded_inputs.pop('relations')
+        encoded_inputs_all = []
+        chunk_size = 512
+        for chunk_id, index in enumerate(
+                range(0, len(encoded_inputs["input_ids"]), chunk_size)):
+            item = {}
+            for k in encoded_inputs:
+                item[k] = encoded_inputs[k][index:index + chunk_size]
+
+            # select entity in current chunk
+            entities_in_this_span = []
+            global_to_local_map = {}  #
+            for entity_id, entity in enumerate(entities):
+                if (index <= entity["start"] < index + chunk_size and
+                        index <= entity["end"] < index + chunk_size):
+                    entity["start"] = entity["start"] - index
+                    entity["end"] = entity["end"] - index
+                    global_to_local_map[entity_id] = len(entities_in_this_span)
+                    entities_in_this_span.append(entity)
+
+            # select relations in current chunk
+            relations_in_this_span = []
+            for relation in relations:
+                if (index <= relation["start_index"] < index + chunk_size and
+                        index <= relation["end_index"] < index + chunk_size):
+                    relations_in_this_span.append({
+                        "head": global_to_local_map[relation["head"]],
+                        "tail": global_to_local_map[relation["tail"]],
+                        "start_index": relation["start_index"] - index,
+                        "end_index": relation["end_index"] - index,
+                    })
+            item.update({
+                "entities": reformat(entities_in_this_span),
+                "relations": reformat(relations_in_this_span),
+            })
+            item['entities']['label'] = [
+                self.entities_labels[x] for x in item['entities']['label']
+            ]
+            encoded_inputs_all.append(item)
+        return encoded_inputs_all
+
+    def _relations(self, entities, relations, id2label, empty_entity,
+                   entity_id_to_index_map):
+        """
+        build relations
+        """
+        relations = list(set(relations))
+        relations = [
+            rel for rel in relations
+            if rel[0] not in empty_entity and rel[1] not in empty_entity
+        ]
+        kv_relations = []
+        for rel in relations:
+            pair = [id2label[rel[0]], id2label[rel[1]]]
+            if pair == ["question", "answer"]:
+                kv_relations.append({
+                    "head": entity_id_to_index_map[rel[0]],
+                    "tail": entity_id_to_index_map[rel[1]]
+                })
+            elif pair == ["answer", "question"]:
+                kv_relations.append({
+                    "head": entity_id_to_index_map[rel[1]],
+                    "tail": entity_id_to_index_map[rel[0]]
+                })
+            else:
+                continue
+        relations = sorted(
+            [{
+                "head": rel["head"],
+                "tail": rel["tail"],
+                "start_index": get_relation_span(rel, entities)[0],
+                "end_index": get_relation_span(rel, entities)[1],
+            } for rel in kv_relations],
+            key=lambda x: x["head"], )
+        return relations
+
+    def load_img(self, image_path):
+        # read img
+        img = cv2.imread(image_path)
+        img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
+        resize_h, resize_w = self.img_size
+        im_shape = img.shape[0:2]
+        im_scale_y = resize_h / im_shape[0]
+        im_scale_x = resize_w / im_shape[1]
+        img_new = cv2.resize(
+            img, None, None, fx=im_scale_x, fy=im_scale_y, interpolation=2)
+        mean = np.array([0.485, 0.456, 0.406])[np.newaxis, np.newaxis, :]
+        std = np.array([0.229, 0.224, 0.225])[np.newaxis, np.newaxis, :]
+        img_new = img_new / 255.0
+        img_new -= mean
+        img_new /= std
+        img = img_new.transpose((2, 0, 1))
+        return img
+
+    def __getitem__(self, idx):
+        if self.load_mode == "all":
+            data = copy.deepcopy(self.encoded_inputs_all[idx])
+        else:
+            data = self._parse_label_file(self.all_lines[idx])[0]
+
+        image_path = data.pop('image_path')
+        data["image"] = self.load_img(image_path)
+
+        return_data = {}
+        for k, v in data.items():
+            if k in self.return_keys:
+                if self.return_keys[k] == 'np':
+                    v = np.array(v)
+                return_data[k] = v
+        return return_data
+
+    def __len__(self, ):
+        if self.load_mode == "all":
+            return len(self.encoded_inputs_all)
+        else:
+            return len(self.all_lines)
+
+
+def get_relation_span(rel, entities):
+    bound = []
+    for entity_index in [rel["head"], rel["tail"]]:
+        bound.append(entities[entity_index]["start"])
+        bound.append(entities[entity_index]["end"])
+    return min(bound), max(bound)
+
+
+def reformat(data):
+    new_data = {}
+    for item in data:
+        for k, v in item.items():
+            if k not in new_data:
+                new_data[k] = []
+            new_data[k].append(v)
+    return new_data

test_tipc/common_func.sh

@@ -30,6 +30,7 @@ function func_set_params(){
 
 function func_parser_params(){
     strs=$1
+    MODE=$2
     IFS=":"
     array=(${strs})
     key=${array[0]}

test_tipc/configs/ch_PP-OCRv2/model_linux_gpu_normal_normal_infer_python_linux_gpu_cpu.txt

+===========================ch_PP-OCRv2===========================
+model_name:ch_PP-OCRv2
+python:python3.7
+infer_model:./inference/ch_PP-OCRv2_det_infer/
+infer_export:null
+infer_quant:False
+inference:tools/infer/predict_system.py
+--use_gpu:False|True
+--enable_mkldnn:False|True
+--cpu_threads:1|6
+--rec_batch_num:1
+--use_tensorrt:False|True
+--precision:fp32
+--det_model_dir:
+--image_dir:./inference/ch_det_data_50/all-sum-510/
+--rec_model_dir:./inference/ch_PP-OCRv2_rec_infer/
+--benchmark:True
+null:null
+null:null

test_tipc/configs/ch_PP-OCRv2/model_linux_gpu_normal_normal_lite_cpp_arm_cpu.txt

+===========================lite_params===========================
+inference:./ocr_db_crnn system
+runtime_device:ARM_CPU
+det_infer_model:ch_PP-OCRv2_det_infer|ch_PP-OCRv2_det_slim_quant_infer
+rec_infer_model:ch_PP-OCRv2_rec_infer|ch_PP-OCRv2_rec_slim_quant_infer
+cls_infer_model:ch_ppocr_mobile_v2.0_cls_infer|ch_ppocr_mobile_v2.0_cls_slim_infer
+--cpu_threads:1|4
+--det_batch_size:1
+--rec_batch_size:1
+--image_dir:./test_data/icdar2015_lite/text_localization/ch4_test_images/
+--config_dir:./config.txt
+--rec_dict_dir:./ppocr_keys_v1.txt
+--benchmark:True

test_tipc/configs/ch_PP-OCRv2/model_linux_gpu_normal_normal_lite_cpp_arm_gpu_opencl.txt

+===========================lite_params===========================
+inference:./ocr_db_crnn system
+runtime_device:ARM_GPU_OPENCL
+det_infer_model:ch_PP-OCRv2_det_infer|ch_PP-OCRv2_det_slim_quant_infer
+rec_infer_model:ch_PP-OCRv2_rec_infer|ch_PP-OCRv2_rec_slim_quant_infer
+cls_infer_model:ch_ppocr_mobile_v2.0_cls_infer|ch_ppocr_mobile_v2.0_cls_slim_infer
+--cpu_threads:1|4
+--det_batch_size:1
+--rec_batch_size:1
+--image_dir:./test_data/icdar2015_lite/text_localization/ch4_test_images/
+--config_dir:./config.txt
+--rec_dict_dir:./ppocr_keys_v1.txt
+--benchmark:True

test_tipc/configs/ch_PP-OCRv2_det/model_linux_gpu_normal_normal_lite_cpp_arm_cpu.txt

+===========================lite_params===========================
+inference:./ocr_db_crnn det
+runtime_device:ARM_CPU
+det_infer_model:ch_PP-OCRv2_det_infer|ch_PP-OCRv2_det_slim_quant_infer
+null:null
+null:null
+--cpu_threads:1|4
+--det_batch_size:1
+null:null
+--image_dir:./test_data/icdar2015_lite/text_localization/ch4_test_images/
+--config_dir:./config.txt
+null:null
+--benchmark:True
\ No newline at end of file

test_tipc/configs/ch_PP-OCRv2_det/model_linux_gpu_normal_normal_lite_cpp_arm_gpu_opencl.txt

+===========================lite_params===========================
+inference:./ocr_db_crnn det
+runtime_device:ARM_GPU_OPENCL
+det_infer_model:ch_PP-OCRv2_det_infer|ch_PP-OCRv2_det_slim_quant_infer
+null:null
+null:null
+--cpu_threads:1|4
+--det_batch_size:1
+null:null
+--image_dir:./test_data/icdar2015_lite/text_localization/ch4_test_images/
+--config_dir:./config.txt
+null:null
+--benchmark:True

test_tipc/configs/ppocrv2_det_mobile/train_infer_python.txt → test_tipc/configs/ch_PP-OCRv2_det/train_infer_python.txt

 ===========================train_params===========================
-model_name:PPOCRv2_ocr_det
+model_name:ch_PPOCRv2_det
 python:python3.7
 gpu_list:0|0,1
 Global.use_gpu:True|True
 Global.auto_cast:fp32
-Global.epoch_num:lite_train_infer=1|whole_train_infer=500
+Global.epoch_num:lite_train_lite_infer=1|whole_train_whole_infer=500
 Global.save_model_dir:./output/
-Train.loader.batch_size_per_card:lite_train_infer=2|whole_train_infer=4
+Train.loader.batch_size_per_card:lite_train_lite_infer=2|whole_train_whole_infer=4
 Global.pretrained_model:null
 train_model_name:latest
 train_infer_img_dir:./train_data/icdar2015/text_localization/ch4_test_images/
 null:null
 ##
 trainer:norm_train|pact_train
-norm_train:tools/train.py -c configs/det/ch_PP-OCRv2/ch_PP-OCR_det_cml.yml -o 
-pact_train:deploy/slim/quantization/quant.py -c configs/det/ch_PP-OCRv2/ch_PP-OCR_det_cml.yml -o
+norm_train:tools/train.py -c configs/det/ch_PP-OCRv2/ch_PP-OCRv2_det_cml.yml -o 
+pact_train:deploy/slim/quantization/quant.py -c configs/det/ch_PP-OCRv2/ch_PP-OCRv2_det_cml.yml -o
 fpgm_train:null
 distill_train:null
 null:null
@@ -27,8 +27,8 @@ null:null
 ===========================infer_params===========================
 Global.save_inference_dir:./output/
 Global.pretrained_model:
-norm_export:tools/export_model.py -c configs/det/ch_PP-OCRv2/ch_PP-OCR_det_cml.yml -o 
-quant_export:deploy/slim/quantization/export_model.py -c configs/det/ch_PP-OCRv2/ch_PP-OCR_det_cml.yml -o 
+norm_export:tools/export_model.py -c configs/det/ch_PP-OCRv2/ch_PP-OCRv2_det_cml.yml -o 
+quant_export:deploy/slim/quantization/export_model.py -c configs/det/ch_PP-OCRv2/ch_PP-OCRv2_det_cml.yml -o 
 fpgm_export: 
 distill_export:null
 export1:null

test_tipc/configs/ch_PP-OCRv2_det_KL/model_linux_gpu_normal_normal_infer_python_linux_gpu_cpu.txt

+===========================kl_quant_params===========================
+model_name:PPOCRv2_ocr_det_kl
+python:python3.7
+Global.pretrained_model:null
+Global.save_inference_dir:null
+infer_model:./inference/ch_PP-OCRv2_det_infer/
+infer_export:deploy/slim/quantization/quant_kl.py -c configs/det/ch_PP-OCRv2/ch_PP-OCRv2_det_cml.yml -o
+infer_quant:True
+inference:tools/infer/predict_det.py
+--use_gpu:False|True
+--enable_mkldnn:True
+--cpu_threads:1|6
+--rec_batch_num:1
+--use_tensorrt:False|True
+--precision:int8
+--det_model_dir:
+--image_dir:./inference/ch_det_data_50/all-sum-510/
+null:null
+--benchmark:True
+null:null
+null:null

test_tipc/configs/ch_PP-OCRv2_det_PACT/train_infer_python.txt

+===========================train_params===========================
+model_name:PPOCRv2_ocr_det
+python:python3.7
+gpu_list:0|0,1
+Global.use_gpu:True|True
+Global.auto_cast:fp32
+Global.epoch_num:lite_train_lite_infer=1|whole_train_whole_infer=500
+Global.save_model_dir:./output/
+Train.loader.batch_size_per_card:lite_train_lite_infer=2|whole_train_whole_infer=4
+Global.pretrained_model:null
+train_model_name:latest
+train_infer_img_dir:./train_data/icdar2015/text_localization/ch4_test_images/
+null:null
+##
+trainer:pact_train
+norm_train:null
+pact_train:deploy/slim/quantization/quant.py -c configs/det/ch_PP-OCRv2/ch_PP-OCRv2_det_cml.yml -o
+fpgm_train:null
+distill_train:null
+null:null
+null:null
+##
+===========================eval_params=========================== 
+eval:null
+null:null
+##
+===========================infer_params===========================
+Global.save_inference_dir:./output/
+Global.pretrained_model:
+norm_export:null
+quant_export:deploy/slim/quantization/export_model.py -c configs/det/ch_PP-OCRv2/ch_PP-OCRv2_det_cml.yml -o 
+fpgm_export: 
+distill_export:null
+export1:null
+export2:null
+inference_dir:Student
+infer_model:./inference/ch_PP-OCRv2_det_infer/
+infer_export:null
+infer_quant:False
+inference:tools/infer/predict_det.py
+--use_gpu:True|False
+--enable_mkldnn:True|False
+--cpu_threads:1|6
+--rec_batch_num:1
+--use_tensorrt:False|True
+--precision:fp32|fp16|int8
+--det_model_dir:
+--image_dir:./inference/ch_det_data_50/all-sum-510/
+null:null
+--benchmark:True
+null:null

test_tipc/configs/ch_PP-OCRv2_rec/ch_PP-OCRv2_rec_distillation.yml

+Global:
+  debug: false
+  use_gpu: true
+  epoch_num: 800
+  log_smooth_window: 20
+  print_batch_step: 10
+  save_model_dir: ./output/rec_pp-OCRv2_distillation
+  save_epoch_step: 3
+  eval_batch_step: [0, 2000]
+  cal_metric_during_train: true
+  pretrained_model:
+  checkpoints:
+  save_inference_dir:
+  use_visualdl: false
+  infer_img: doc/imgs_words/ch/word_1.jpg
+  character_dict_path: ppocr/utils/ppocr_keys_v1.txt
+  max_text_length: 25
+  infer_mode: false
+  use_space_char: true
+  distributed: true
+  save_res_path: ./output/rec/predicts_pp-OCRv2_distillation.txt
+
+
+Optimizer:
+  name: Adam
+  beta1: 0.9
+  beta2: 0.999
+  lr:
+    name: Piecewise
+    decay_epochs : [700, 800]
+    values : [0.001, 0.0001]
+    warmup_epoch: 5
+  regularizer:
+    name: L2
+    factor: 2.0e-05
+
+Architecture:
+  model_type: &model_type "rec"
+  name: DistillationModel
+  algorithm: Distillation
+  Models:
+    Teacher:
+      pretrained:
+      freeze_params: false
+      return_all_feats: true
+      model_type: *model_type
+      algorithm: CRNN
+      Transform:
+      Backbone:
+        name: MobileNetV1Enhance
+        scale: 0.5
+      Neck:
+        name: SequenceEncoder
+        encoder_type: rnn
+        hidden_size: 64
+      Head:
+        name: CTCHead
+        mid_channels: 96
+        fc_decay: 0.00002
+    Student:
+      pretrained:
+      freeze_params: false
+      return_all_feats: true
+      model_type: *model_type
+      algorithm: CRNN
+      Transform:
+      Backbone:
+        name: MobileNetV1Enhance
+        scale: 0.5
+      Neck:
+        name: SequenceEncoder
+        encoder_type: rnn
+        hidden_size: 64
+      Head:
+        name: CTCHead
+        mid_channels: 96
+        fc_decay: 0.00002
+  
+
+Loss:
+  name: CombinedLoss
+  loss_config_list:
+  - DistillationCTCLoss:
+      weight: 1.0
+      model_name_list: ["Student", "Teacher"]
+      key: head_out
+  - DistillationDMLLoss:
+      weight: 1.0
+      act: "softmax"
+      use_log: true
+      model_name_pairs:
+      - ["Student", "Teacher"]
+      key: head_out
+  - DistillationDistanceLoss:
+      weight: 1.0
+      mode: "l2"
+      model_name_pairs:
+      - ["Student", "Teacher"]
+      key: backbone_out
+
+PostProcess:
+  name: DistillationCTCLabelDecode
+  model_name: ["Student", "Teacher"]
+  key: head_out
+
+Metric:
+  name: DistillationMetric
+  base_metric_name: RecMetric
+  main_indicator: acc
+  key: "Student"
+
+Train:
+  dataset:
+    name: SimpleDataSet
+    data_dir: ./train_data/ic15_data/
+    label_file_list:
+    - ./train_data/ic15_data/rec_gt_train.txt
+    transforms:
+    - DecodeImage:
+        img_mode: BGR
+        channel_first: false
+    - RecAug:
+    - CTCLabelEncode:
+    - RecResizeImg:
+        image_shape: [3, 32, 320]
+    - KeepKeys:
+        keep_keys:
+        - image
+        - label
+        - length
+  loader:
+    shuffle: true
+    batch_size_per_card: 128
+    drop_last: true
+    num_sections: 1
+    num_workers: 8
+Eval:
+  dataset:
+    name: SimpleDataSet
+    data_dir: ./train_data/ic15_data
+    label_file_list:
+    - ./train_data/ic15_data/rec_gt_test.txt
+    transforms:
+    - DecodeImage:
+        img_mode: BGR
+        channel_first: false
+    - CTCLabelEncode:
+    - RecResizeImg:
+        image_shape: [3, 32, 320]
+    - KeepKeys:
+        keep_keys:
+        - image
+        - label
+        - length
+  loader:
+    shuffle: false
+    drop_last: false
+    batch_size_per_card: 128
+    num_workers: 8

test_tipc/configs/ch_PP-OCRv2_rec/train_infer_python.txt

+===========================train_params===========================
+model_name:PPOCRv2_ocr_rec
+python:python3.7
+gpu_list:0|0,1
+Global.use_gpu:True|True
+Global.auto_cast:fp32
+Global.epoch_num:lite_train_lite_infer=3|whole_train_whole_infer=300
+Global.save_model_dir:./output/
+Train.loader.batch_size_per_card:lite_train_lite_infer=128|whole_train_whole_infer=128
+Global.pretrained_model:null
+train_model_name:latest
+train_infer_img_dir:./inference/rec_inference
+null:null
+##
+trainer:norm_train
+norm_train:tools/train.py -c test_tipc/configs/ch_PP-OCRv2_rec/ch_PP-OCRv2_rec_distillation.yml -o 
+pact_train:null
+fpgm_train:null
+distill_train:null
+null:null
+null:null
+##
+===========================eval_params=========================== 
+eval:null
+null:null
+##
+===========================infer_params===========================
+Global.save_inference_dir:./output/
+Global.pretrained_model:
+norm_export:tools/export_model.py -c test_tipc/configs/ch_PP-OCRv2_rec/ch_PP-OCRv2_rec_distillation.yml -o 
+quant_export:
+fpgm_export: 
+distill_export:null
+export1:null
+export2:null
+inference_dir:Student
+infer_model:./inference/ch_PP-OCRv2_rec_infer/
+infer_export:null
+infer_quant:False
+inference:tools/infer/predict_rec.py
+--use_gpu:True|False
+--enable_mkldnn:True|False
+--cpu_threads:1|6
+--rec_batch_num:1|6
+--use_tensorrt:False|True
+--precision:fp32|fp16|int8
+--rec_model_dir:
+--image_dir:/inference/rec_inference
+null:null
+--benchmark:True
+null:null
+
+

test_tipc/configs/ch_PP-OCRv2_rec_KL/model_linux_gpu_normal_normal_infer_python_linux_gpu_cpu.txt

+===========================kl_quant_params===========================
+model_name:PPOCRv2_ocr_rec_kl
+python:python3.7
+Global.pretrained_model:null
+Global.save_inference_dir:null
+infer_model:./inference/ch_PP-OCRv2_rec_infer/
+infer_export:deploy/slim/quantization/quant_kl.py -c test_tipc/configs/ch_PP-OCRv2_rec/ch_PP-OCRv2_rec_distillation.yml -o
+infer_quant:True
+inference:tools/infer/predict_rec.py
+--use_gpu:False|True
+--enable_mkldnn:False|True
+--cpu_threads:1|6
+--rec_batch_num:1|6
+--use_tensorrt:True
+--precision:int8
+--rec_model_dir:
+--image_dir:./inference/rec_inference
+null:null
+--benchmark:True
+null:null
+null:null

test_tipc/configs/ch_PP-OCRv2_rec_PACT/train_infer_python.txt

+===========================train_params===========================
+model_name:PPOCRv2_ocr_rec_pact
+python:python3.7
+gpu_list:0|0,1
+Global.use_gpu:True|True
+Global.auto_cast:fp32
+Global.epoch_num:lite_train_lite_infer=3|whole_train_whole_infer=300
+Global.save_model_dir:./output/
+Train.loader.batch_size_per_card:lite_train_lite_infer=128|whole_train_whole_infer=128
+Global.pretrained_model:null
+train_model_name:latest
+train_infer_img_dir:./inference/rec_inference
+null:null
+##
+trainer:pact_train
+norm_train:deploy/slim/quantization/quant.py -c test_tipc/configs/ch_PP-OCRv2_rec/ch_PP-OCRv2_rec_distillation.yml -o 
+pact_train:null
+fpgm_train:null
+distill_train:null
+null:null
+null:null
+##
+===========================eval_params=========================== 
+eval:null
+null:null
+##
+===========================infer_params===========================
+Global.save_inference_dir:./output/
+Global.pretrained_model:
+norm_export:deploy/slim/quantization/export_model.py -c test_tipc/configs/ch_PP-OCRv2_rec/ch_PP-OCRv2_rec_distillation.yml -o 
+quant_export:
+fpgm_export: 
+distill_export:null
+export1:null
+export2:null
+inference_dir:Student
+infer_model:./inference/ch_PP-OCRv2_rec_infer/
+infer_export:null
+infer_quant:True
+inference:tools/infer/predict_rec.py
+--use_gpu:True|False
+--enable_mkldnn:True|False
+--cpu_threads:1|6
+--rec_batch_num:1|6
+--use_tensorrt:False|True
+--precision:fp32|fp16|int8
+--rec_model_dir:
+--image_dir:/inference/rec_inference
+null:null
+--benchmark:True
+null:null
+
+

test_tipc/configs/ch_ppocr_mobile_V2.0_det_FPGM/train_infer_python.txt

+===========================train_params===========================
+model_name:ocr_det
+python:python3.7
+gpu_list:0|0,1
+Global.use_gpu:True|True
+Global.auto_cast:null
+Global.epoch_num:lite_train_lite_infer=5|whole_train_whole_infer=300
+Global.save_model_dir:./output/
+Train.loader.batch_size_per_card:lite_train_lite_infer=2|whole_train_whole_infer=4
+Global.pretrained_model:null
+train_model_name:latest
+train_infer_img_dir:./train_data/icdar2015/text_localization/ch4_test_images/
+null:null
+##
+trainer:fpgm_train
+norm_train:null
+pact_train:null
+fpgm_train:deploy/slim/prune/sensitivity_anal.py -c configs/det/ch_ppocr_v2.0/ch_det_mv3_db_v2.0.yml -o Global.pretrained_model=./pretrain_models/det_mv3_db_v2.0_train/best_accuracy
+distill_train:null
+null:null
+null:null
+##
+===========================eval_params=========================== 
+eval:null
+null:null
+##
+===========================infer_params===========================
+Global.save_inference_dir:./output/
+Global.pretrained_model:
+norm_export:null
+quant_export:null
+fpgm_export:deploy/slim/prune/export_prune_model.py -c configs/det/ch_ppocr_v2.0/ch_det_mv3_db_v2.0.yml -o 
+distill_export:null
+export1:null
+export2:null
+inference_dir:null
+train_model:null
+infer_export:null
+infer_quant:False
+inference:tools/infer/predict_det.py
+--use_gpu:True|False
+--enable_mkldnn:True|False
+--cpu_threads:1|6
+--rec_batch_num:1
+--use_tensorrt:False|True
+--precision:fp32|fp16|int8
+--det_model_dir:
+--image_dir:./inference/ch_det_data_50/all-sum-510/
+null:null
+--benchmark:True
+null:null
\ No newline at end of file

test_tipc/configs/ch_ppocr_mobile_v2.0/model_linux_gpu_normal_normal_infer_python_linux_gpu_cpu.txt

+===========================ch_ppocr_mobile_v2.0===========================
+model_name:ch_ppocr_mobile_v2.0
+python:python3.7
+infer_model:./inference/ch_ppocr_mobile_v2.0_det_infer/
+infer_export:null
+infer_quant:False
+inference:tools/infer/predict_system.py
+--use_gpu:False|True
+--enable_mkldnn:False|True
+--cpu_threads:1|6
+--rec_batch_num:1
+--use_tensorrt:False|True
+--precision:fp32
+--det_model_dir:
+--image_dir:./inference/ch_det_data_50/all-sum-510/
+--rec_model_dir:./inference/ch_ppocr_mobile_v2.0_rec_infer/
+--benchmark:True
+null:null
+null:null

test_tipc/configs/ch_ppocr_mobile_v2.0/model_linux_gpu_normal_normal_lite_cpp_arm_cpu.txt

+===========================lite_params===========================
+inference:./ocr_db_crnn system
+runtime_device:ARM_CPU
+det_infer_model:ch_ppocr_mobile_v2.0_det_infer|ch_ppocr_db_mobile_v2.0_det_quant_infer
+rec_infer_model:ch_ppocr_mobile_v2.0_rec_infer|ch_ppocr_mobile_v2.0_rec_slim_infer
+cls_infer_model:ch_ppocr_mobile_v2.0_cls_infer|ch_ppocr_mobile_v2.0_cls_slim_infer
+--cpu_threads:1|4
+--det_batch_size:1
+--rec_batch_size:1
+--image_dir:./test_data/icdar2015_lite/text_localization/ch4_test_images/
+--config_dir:./config.txt
+--rec_dict_dir:./ppocr_keys_v1.txt
+--benchmark:True

test_tipc/configs/ch_ppocr_mobile_v2.0/model_linux_gpu_normal_normal_lite_cpp_arm_gpu_opencl.txt

+===========================lite_params===========================
+inference:./ocr_db_crnn system
+runtime_device:ARM_GPU_OPENCL
+det_infer_model:ch_ppocr_mobile_v2.0_det_infer|ch_ppocr_db_mobile_v2.0_det_quant_infer
+rec_infer_model:ch_ppocr_mobile_v2.0_rec_infer|ch_ppocr_mobile_v2.0_rec_slim_infer
+cls_infer_model:ch_ppocr_mobile_v2.0_cls_infer|ch_ppocr_mobile_v2.0_cls_slim_infer
+--cpu_threads:1|4
+--det_batch_size:1
+--rec_batch_size:1
+--image_dir:./test_data/icdar2015_lite/text_localization/ch4_test_images/
+--config_dir:./config.txt
+--rec_dict_dir:./ppocr_keys_v1.txt
+--benchmark:True