feat: add crf and lstm-crf example; update the config file

research/huawei-gts/CRF/default_config.yaml

@@ -1,27 +1,37 @@
+# --------------------------------------------------
 # Builtin Configurations
-data_path: "../conll2003"
+# 训练数据存储的地方
+data_path: '../conll2003'
+# 输出ckpt文件的名称
 ckpt_save_path: '../ckpt_lstm_crf'
+# 设备类型： Ascend, CPU。默认Ascend
 device_target: 'CPU'
-device_id: 1
-enable_profiling: False
+# 在device_target位Ascend时，指定参与计算的NPU
+device_id: 2
+# 导出文件的前缀
+export_prefix: 'model-crf'
+export_suffix: ''
 
+# --------------------------------------------------
 # LSTM_CRF CONFIG
-num_epochs: 20
-batch_size: 20
-embed_size: 300
-num_hidden: 320
+# 训练的轮数，默认20轮
+num_epochs: 2
+# 单次处理数据量的大小，默认16
+batch_size: 16
+# 组成最大长度
+vocab_max_length: 113
+# embedding_dim数量，默认384
+embedding_dim: 128
+# hidden_dim数量，默认384
+hidden_dim: 128
 num_layers: 2
-bidirectional: True
-
-# train.py
-device_num: 1
-data_CoNLL_path: "../data/conll2003"
+learning_ratelearning_rate: True
+# 学习率过小会导致loss不容易对其，学习率过大会导致预测准确率位0（可以加大epoch解决）
+learning_rate: 0.00001
 
 # export.py
-model_format: "MINDIR"
 model_path: "./"
-ckpt_path: "./lstm-crf.ckpt"
 
 ---
 device_target: ['Ascend', 'CPU']
-file_format: ['AIR', 'MINDIR']
+file_format: ['AIR', 'MINDIR']

research/huawei-gts/CRF/scripts/export_ascend.sh

@@ -1,23 +1,21 @@
 #!/bin/bash
 echo "=============================================================================================================="
 echo "Please run the script as: "
-echo "bash export_ascend.sh DEVICE_ID TRAIN_DATA_FOLDER CKPT_FILE"
-echo "for example: bash export_ascend.sh 0 ../conll2003 lstm_crf.ckpt"
+echo "bash export_ascend.sh DEVICE_ID TRAIN_DATA_FOLDER EXPORT_SUFFIX"
+echo "for example: bash export_ascend.sh 0 ../conll2003 my_suffix"
 echo "=============================================================================================================="
 
 DEVICE_ID=$1
 TRAIN_DATA_FOLDER=$2
-CKPT_FILE=$3
+EXPORT_SUFFIX=$3
 
 BASE_PATH=$(cd ./"`dirname $0`" || exit; pwd)
 CONFIG_FILE="${BASE_PATH}/../default_config.yaml"
 
 
 python "${BASE_PATH}/../src/export.py"  \
     --config_path=$CONFIG_FILE \
-    --device_target="Ascend" \
-    --model_format="ckpt" \
     --device_id=${DEVICE_ID}\
     --data_path=${TRAIN_DATA_FOLDER}\
-    --ckpt_path=${CKPT_FILE}
+    --export_suffix=${EXPORT_SUFFIX}
 #    --ckpt_path=${CKPT_FILE}  > log_export.txt 2>&1 &

research/huawei-gts/CRF/src/example/crf_decode.py

@@ -0,0 +1,26 @@
+import sys
+from pathlib import Path
+
+sys.path.append(f'{Path.cwd()}')
+print(f'======CurrentPath: {Path.cwd()}')
+import mindspore as ms
+
+# 设置mindspore的执行目标，可以使Ascend、CPU、GPU，mode建议位图模式。注意，ms需要放到import的首行，避免context设置不生效
+ms.set_context(device_target="CPU", mode=ms.GRAPH_MODE)
+from model.lstm_crf_model import CRF
+import mindspore.numpy as mnp
+
+if __name__ == '__main__':
+    # 需要使用的实体索引，可以根据需要使用BIO或者BIOES作为标注模式
+    tag_to_idx = {"B": 0, "I": 1, "O": 2}
+
+    # 初始化模型，这里需要传入待使用实体映射的个数
+    model = CRF(len(tag_to_idx))
+
+    # 定义decode需要的emissions
+    seq_length = 3
+    batch_size = 2
+    emissions = mnp.randn(seq_length, batch_size, len(tag_to_idx))
+    score, history = model(emissions)
+    best_tags_list = CRF.post_decode(score, history, mnp.full(batch_size, 1))
+    print(best_tags_list)

research/huawei-gts/CRF/src/example/crf_init.py

@@ -0,0 +1,22 @@
+import sys
+from pathlib import Path
+
+sys.path.append(f'{Path.cwd()}')
+print(f'======CurrentPath: {Path.cwd()}')
+
+import mindspore as ms
+
+# 设置mindspore的执行目标，可以使Ascend、CPU、GPU，mode建议位图模式。注意，ms需要放到import的首行，避免context设置不生效
+ms.set_context(device_target="CPU", mode=ms.GRAPH_MODE)
+
+from model.lstm_crf_model import CRF
+
+if __name__ == '__main__':
+    # 需要使用的实体索引，可以根据需要使用BIO或者BIOES作为标注模式
+    tag_to_idx = {"B": 0, "I": 1, "O": 2}
+
+    # 初始化模型，这里需要传入待使用实体映射的个数
+    model = CRF(len(tag_to_idx))
+
+    # 初始化完成，可以打印模型进行查看
+    print(model)

research/huawei-gts/CRF/src/example/crf_loss.py

@@ -0,0 +1,27 @@
+import sys
+from pathlib import Path
+
+sys.path.append(f'{Path.cwd()}')
+print(f'======CurrentPath: {Path.cwd()}')
+import mindspore as ms
+
+# 设置mindspore的执行目标，可以使Ascend、CPU、GPU，mode建议位图模式。注意，ms需要放到import的首行，避免context设置不生效
+ms.set_context(device_target="CPU", mode=ms.GRAPH_MODE)
+from model.lstm_crf_model import CRF
+
+import mindspore.numpy as mnp
+
+if __name__ == '__main__':
+    # 需要使用的实体索引，可以根据需要使用BIO或者BIOES作为标注模式
+    tag_to_idx = {"B": 0, "I": 1, "O": 2}
+
+    # 初始化模型，这里需要传入待使用实体映射的个数
+    model = CRF(len(tag_to_idx))
+
+    # 定义计算loss相关的输入emissions与tag
+    seq_length = 3
+    batch_size = 2
+    emissions = mnp.randn(seq_length, batch_size, len(tag_to_idx))
+    tags = mnp.array([[0, 1], [1, 2], [2, 1]])
+    loss = model(emissions, tags)
+    print(loss)

research/huawei-gts/CRF/src/example/lstm_crf_ckpt_create.py

@@ -0,0 +1,69 @@
+import sys
+from pathlib import Path
+
+import numpy as np
+from tqdm import tqdm
+
+sys.path.append(f'{Path.cwd()}')
+print(f'======CurrentPath: {Path.cwd()}')
+import mindspore as ms
+import mindspore.dataset as ds
+import mindspore.nn as nn
+
+# 设置mindspore的执行目标，可以使Ascend、CPU、GPU，mode建议位图模式。注意，ms需要放到import的首行，避免context设置不生效
+ms.set_context(device_target="CPU", mode=ms.GRAPH_MODE)
+from model.lstm_crf_model import BiLSTM_CRF
+from utils.dataset import read_data, GetDatasetGenerator, get_dict, COLUMN_NAME
+from utils.config import config
+
+if __name__ == '__main__':
+    # Step1： 定义初始化参数
+    embedding_dim = config.embedding_dim
+    hidden_dim = config.hidden_dim
+    Max_Len = config.vocab_max_length
+    batch_size = config.batch_size
+
+    # BIOES标注模式： 一般一共分为四大类：PER（人名），LOC（位置[地名]），ORG（组织）以及MISC(杂项)，而且B表示开始，I表示中间，O表示不是实体。
+    Entity = ['PER', 'LOC', 'ORG', 'MISC']
+    labels_text_mp = {k: v for k, v in enumerate(Entity)}
+    LABEL_MAP = {'O': 0}  # 非实体
+    for i, e in enumerate(Entity):
+        LABEL_MAP[f'B-{e}'] = 2 * (i + 1) - 1  # 实体首字
+        LABEL_MAP[f'I-{e}'] = 2 * (i + 1)  # 实体非首字
+
+    # Step2: 读取数据集
+    train_dataset = read_data('../../conll2003/train.txt')
+    char_number, id_indexs = get_dict(train_dataset[0])
+
+    train_dataset_generator = GetDatasetGenerator(train_dataset, id_indexs)
+    train_dataset_ds = ds.GeneratorDataset(train_dataset_generator, COLUMN_NAME, shuffle=False)
+    train_dataset_batch = train_dataset_ds.batch(batch_size, drop_remainder=True)
+
+    # Step3: 初始化模型与优化器
+    model = BiLSTM_CRF(vocab_size=len(id_indexs), embedding_dim=embedding_dim, hidden_dim=hidden_dim,
+                       num_tags=len(Entity) * 2 + 1)
+    optimizer = nn.Adam(model.trainable_params(), learning_rate=config.learning_rate)
+    grad_fn = ms.value_and_grad(model, None, optimizer.parameters)
+
+
+    def train_step(token_ids, seq_len, labels):
+        loss, grads = grad_fn(token_ids, seq_len, labels)
+        optimizer(grads)
+        return loss
+
+
+    # Step5: 训练
+    tloss = []
+    for epoch in range(config.num_epochs):
+        model.set_train()
+        with tqdm(total=train_dataset_batch.get_dataset_size()) as t:
+            for batch, (token_ids, seq_len, labels) in enumerate(train_dataset_batch.create_tuple_iterator()):
+                loss = train_step(token_ids, seq_len, labels)
+                tloss.append(loss.asnumpy())
+                t.set_postfix(loss=np.array(tloss).mean())
+                t.update(1)
+
+    # Step6: 导出CKPT
+    file_name = 'lstm_crf.ckpt'
+    ms.save_checkpoint(model, ckpt_file_name=file_name)
+    print(f'======Create CKPT SUCCEEDED, file: {file_name}')

research/huawei-gts/CRF/src/example/lstm_crf_ckpt_predict.py

@@ -0,0 +1,62 @@
+import sys
+from pathlib import Path
+
+from tqdm import tqdm
+
+sys.path.append(f'{Path.cwd()}')
+print(f'======CurrentPath: {Path.cwd()}')
+import mindspore as ms
+import mindspore.dataset as ds
+from mindspore.train.serialization import load_checkpoint, load_param_into_net
+
+# 设置mindspore的执行目标，可以使Ascend、CPU、GPU，mode建议位图模式。注意，ms需要放到import的首行，避免context设置不生效
+ms.set_context(device_target="CPU", mode=ms.GRAPH_MODE)
+from model.lstm_crf_model import BiLSTM_CRF, CRF
+from utils.dataset import read_data, GetDatasetGenerator, get_dict, COLUMN_NAME, get_entity
+from utils.metrics import get_metric
+from utils.config import config
+
+if __name__ == '__main__':
+    # Step1： 定义初始化参数
+    batch_size = config.batch_size
+
+    # BIOES标注模式： 一般一共分为四大类：PER（人名），LOC（位置[地名]），ORG（组织）以及MISC(杂项)，而且B表示开始，I表示中间，O表示不是实体。
+    Entity = ['PER', 'LOC', 'ORG', 'MISC']
+
+    # Step2: 加载ckpt，传入文件路径与名称
+    file_name = 'lstm_crf.ckpt'
+    param_dict = load_checkpoint(file_name)
+
+    # Step3: 获取模型初始化参数
+    embedding_shape = param_dict.get('embedding.embedding_table').shape
+
+    # Step4: 初始化模型
+    model = BiLSTM_CRF(vocab_size=embedding_shape[0], embedding_dim=embedding_shape[1], hidden_dim=embedding_shape[1],
+                       num_tags=len(Entity) * 2 + 1)
+
+    # Step5: 将ckpt导入model
+    load_param_into_net(model, param_dict)
+    print(model)
+
+    # Step6: 读取数据集
+    train_dataset = read_data('../../conll2003/train.txt')
+    test_dataset = read_data('../../conll2003/test.txt')
+    char_number, id_indexs = get_dict(train_dataset[0])
+
+    test_dataset_generator = GetDatasetGenerator(test_dataset, id_indexs)
+    test_dataset_ds = ds.GeneratorDataset(test_dataset_generator, COLUMN_NAME, shuffle=False)
+    test_dataset_batch = test_dataset_ds.batch(batch_size, drop_remainder=True)
+
+    # Step7: 进行预测
+    decodes = []
+    model.set_train(False)
+    with tqdm(total=test_dataset_batch.get_dataset_size()) as t:
+        for batch, (token_ids, seq_len, labels) in enumerate(test_dataset_batch.create_tuple_iterator()):
+            score, history = model(token_ids, seq_len)
+            best_tag = CRF.post_decode(score, history, seq_len)
+            decode = [[y for y in x] for x in best_tag]
+            decodes.extend(list(decode))
+            t.update(1)
+
+    pred = [get_entity(x) for x in decodes]
+    get_metric(pred, test_dataset_generator)

research/huawei-gts/CRF/src/example/lstm_crf_init.py

@@ -0,0 +1,22 @@
+import sys
+from pathlib import Path
+
+sys.path.append(f'{Path.cwd()}')
+print(f'======CurrentPath: {Path.cwd()}')
+import mindspore as ms
+
+# 设置mindspore的执行目标，可以使Ascend、CPU、GPU，mode建议位图模式。注意，ms需要放到import的首行，避免context设置不生效
+ms.set_context(device_target="CPU", mode=ms.GRAPH_MODE)
+from model.lstm_crf_model import BiLSTM_CRF
+from utils.config import config
+
+if __name__ == '__main__':
+    # 需要使用的实体索引，可以根据需要使用BIO或者BIOES作为标注模式
+    tag_to_idx = {"B": 0, "I": 1, "O": 2}
+
+    len_id_index = 1024
+
+    # 初始化模型
+    model = BiLSTM_CRF(vocab_size=len_id_index, embedding_dim=config.embedding_dim, hidden_dim=config.hidden_dim,
+                       num_tags=len(tag_to_idx))
+    print(model)

research/huawei-gts/CRF/src/example/lstm_crf_mindir_create.py

@@ -0,0 +1,71 @@
+import sys
+from pathlib import Path
+
+import numpy as np
+from tqdm import tqdm
+
+sys.path.append(f'{Path.cwd()}')
+print(f'======CurrentPath: {Path.cwd()}')
+import mindspore as ms
+import mindspore.ops as ops
+import mindspore.dataset as ds
+import mindspore.nn as nn
+
+# 设置mindspore的执行目标，可以使Ascend、CPU、GPU，mode建议位图模式。注意，ms需要放到import的首行，避免context设置不生效
+ms.set_context(device_target="CPU", mode=ms.GRAPH_MODE)
+from model.lstm_crf_model import BiLSTM_CRF
+from utils.dataset import read_data, GetDatasetGenerator, get_dict, COLUMN_NAME
+from utils.config import config
+
+if __name__ == '__main__':
+    # Step1： 定义初始化参数
+    embedding_dim = config.embedding_dim
+    hidden_dim = config.hidden_dim
+    Max_Len = config.vocab_max_length
+    batch_size = config.batch_size
+
+    # BIOES标注模式： 一般一共分为四大类：PER（人名），LOC（位置[地名]），ORG（组织）以及MISC(杂项)，而且B表示开始，I表示中间，O表示不是实体。
+    Entity = ['PER', 'LOC', 'ORG', 'MISC']
+    labels_text_mp = {k: v for k, v in enumerate(Entity)}
+    LABEL_MAP = {'O': 0}  # 非实体
+    for i, e in enumerate(Entity):
+        LABEL_MAP[f'B-{e}'] = 2 * (i + 1) - 1  # 实体首字
+        LABEL_MAP[f'I-{e}'] = 2 * (i + 1)  # 实体非首字
+
+    # Step2: 读取数据集
+    train_dataset = read_data('../../conll2003/train.txt')
+    char_number, id_indexs = get_dict(train_dataset[0])
+
+    train_dataset_generator = GetDatasetGenerator(train_dataset, id_indexs)
+    train_dataset_ds = ds.GeneratorDataset(train_dataset_generator, COLUMN_NAME, shuffle=False)
+    train_dataset_batch = train_dataset_ds.batch(batch_size, drop_remainder=True)
+
+    # Step3: 初始化模型与优化器
+    model = BiLSTM_CRF(vocab_size=len(id_indexs), embedding_dim=embedding_dim, hidden_dim=hidden_dim,
+                       num_tags=len(Entity) * 2 + 1)
+    optimizer = nn.Adam(model.trainable_params(), learning_rate=config.learning_rate)
+    grad_fn = ms.value_and_grad(model, None, optimizer.parameters)
+
+
+    def train_step(token_ids, seq_len, labels):
+        loss, grads = grad_fn(token_ids, seq_len, labels)
+        optimizer(grads)
+        return loss
+
+
+    # Step5: 训练
+    tloss = []
+    for epoch in range(config.num_epochs):
+        model.set_train()
+        with tqdm(total=train_dataset_batch.get_dataset_size()) as t:
+            for batch, (token_ids, seq_len, labels) in enumerate(train_dataset_batch.create_tuple_iterator()):
+                loss = train_step(token_ids, seq_len, labels)
+                tloss.append(loss.asnumpy())
+                t.set_postfix(loss=np.array(tloss).mean())
+                t.update(1)
+
+    # Step6: 导出MindIR
+    file_name = 'lstm_crf.mindir'
+    ms.export(model, ops.ones((batch_size, Max_Len), ms.int64), ops.ones(batch_size, ms.int64), file_name=file_name,
+              file_format='MINDIR')
+    print(f'======Create MINDIR SUCCEEDED, file: {file_name}')