AI线上部署之ChatGLM 使用趋动云的在线Jupyter Notebook部署模型环境配置很简单，直接使用预置的镜像即可环境：Python 3PyTorch代码 + 注释：# cli_demo.py import os import platform import signal from transformers import AutoTokenizer, AutoModel import readline # 从指定路径加载预训练的tokenizer和model，允许信任远程代码 tokenizer = AutoTokenizer.from_pretrained("/gemini/data-2", trust_remote_code=True) model = AutoModel.from_pretrained("/gemini/data-2", trust_remote_code=True).cuda() # 若要支持多显卡，请使用下面两行代码替换上面一行，并根据实际显卡数量设置num_gpus # from utils import load_model_on_gpus # model = load_model_on_gpus("THUDM/chatglm2-6b", num_gpus=2) model = model.eval() # 获取操作系统信息，以便在不同操作系统上清屏 os_name = platform.system() clear_command = 'cls' if os_name == 'Windows' else 'clear' stop_stream = False # 构建用户提示信息，包括对话历史、清空历史和终止程序的说明 def build_prompt(history): prompt = "欢迎使用 ChatGLM2-6B 模型，输入内容即可进行对话，clear 清空对话历史，stop 终止程序" for query, response in history: prompt += f"\n\n用户：{query}" prompt += f"\n\nChatGLM2-6B：{response}" return prompt # 处理终止信号的函数 def signal_handler(signal, frame): global stop_stream stop_stream = True # 主函数 def main(): past_key_values, history = None, [] global stop_stream print("欢迎使用 ChatGLM2-6B 模型，输入内容即可进行对话，clear 清空对话历史，stop 终止程序") while True: query = input("\n用户：") if query.strip() == "stop": # 如果用户输入"stop"，退出程序 break if query.strip() == "clear": # 如果用户输入"clear"，清空对话历史并清屏 past_key_values, history = None, [] os.system(clear_command) print("欢迎使用 ChatGLM2-6B 模型，输入内容即可进行对话，clear 清空对话历史，stop 终止程序") continue print("\nChatGLM：", end="") current_length = 0 # 使用模型进行对话生成 for response, history, past_key_values in model.stream_chat(tokenizer, query, history=history, past_key_values=past_key_values, return_past_key_values=True): if stop_stream: # 如果接收到终止信号，停止对话生成 stop_stream = False break else: print(response[current_length:], end="", flush=True) current_length = len(response) print("") if __name__ == "__main__": main() 命令行模式命令：python ChatGLM2-6B/cli_demo.py最终结果：web模式：需要安装streamlitpip install streamlit然后在命令行使用streamlit运行web_demo2.pystreamlit run web_demo2.py --server.port=77使用streamlit可以很方便地实现web界面，无需html、css、js的知识都可以用 https://zhuanlan.zhihu.com/p/448853407

AI线上部署之猫狗识别 使用趋动云的在线Jupyter Notebook部署模型环境配置很简单，直接使用预置的镜像即可环境：Python 3TensorFlow 2代码 + 注释：import argparse import tensorflow as tf import os parser = argparse.ArgumentParser(description='Process some integers') # 使用argparse创建一个可以解析命令行参数的工具，description为描述信息 parser.add_argument('--mode', default='train', help='train or test') parser.add_argument("--num_epochs", default=5, type=int) # 设置epochs parser.add_argument("--batch_size", default=32, type=int) # 设置batch_seze parser.add_argument("--learning_rate", default=0.001) # 设置learning_rate parser.add_argument("--data_dir", default="/gemini/data-1") parser.add_argument("--train_dir", default="/gemini/output") args = parser.parse_args() # 将解析的结果存储在args中 # 图片的前置处理 def _decode_and_resize(filename, label): image_string = tf.io.read_file(filename) # 从文件中读取 image_decoded = tf.image.decode_jpeg(image_string, channels=3) # 解码jpeg为rgb三通道 image_resized = tf.image.resize(image_decoded, [150, 150]) / 255.0 # 将图片缩放、归一化 return image_resized, label if __name__ == "__main__": train_dir = args.data_dir + "/train" cats = [] dogs = [] for file in os.listdir(train_dir): # 遍历文件夹 if file.startswith("dog"): dogs.append(train_dir + "/" + file) else: cats.append(train_dir + "/" + file) print("dogSize:%d catSize:%d" % (len(cats), len(dogs))) train_cat_filenames = tf.constant(cats[:10000]) # 将文件名列表转化为张量 train_dog_filenames = tf.constant(dogs[:10000]) train_filenames = tf.concat([train_cat_filenames, train_dog_filenames], axis=-1) # axis=-1表示最后一个维度 train_labels = tf.concat([ tf.zeros(train_cat_filenames.shape, dtype=tf.int32), # 猫标为0 tf.ones(train_dog_filenames.shape, dtype=tf.int32) # 狗标为1 ], axis=-1) train_dataset = tf.data.Dataset.from_tensor_slices((train_filenames, train_labels)) # 创建一个dataset train_dataset = train_dataset.map( map_func=_decode_and_resize, num_parallel_calls=tf.data.experimental.AUTOTUNE # 指定为并行调用 ) train_dataset = train_dataset.shuffle(buffer_size=20000) # 打乱 train_dataset = train_dataset.batch(args.batch_size) # 设置batch_size train_dataset = train_dataset.prefetch(tf.data.experimental.AUTOTUNE) # 自动调整并行化程度，使训练时可异步预取数据，减少训练时间 model = tf.keras.Sequential([ # 定义模型网络 tf.keras.layers.Conv2D(32, 3, activation="relu", input_shape=(150, 150, 3)), tf.keras.layers.MaxPool2D(), tf.keras.layers.Conv2D(64, 3, activation="relu"), tf.keras.layers.MaxPool2D(), tf.keras.layers.Conv2D(128, 3, activation="relu"), tf.keras.layers.MaxPool2D(), tf.keras.layers.Conv2D(128, 3, activation="relu"), tf.keras.layers.MaxPool2D(), tf.keras.layers.Flatten(), tf.keras.layers.Dropout(0.5), tf.keras.layers.Dense(512, activation="relu"), tf.keras.layers.Dense(2, activation="softmax") ]) model.compile( optimizer=tf.keras.optimizers.Adam(learning_rate=args.learning_rate), # 定义优化器为Adam loss=tf.keras.losses.sparse_categorical_crossentropy, # 定义损失函数为交叉熵 metrics=[tf.keras.metrics.sparse_categorical_accuracy] # 定义模型评估指标为稀疏分类准确率 ) model.fit(train_dataset, epochs=args.num_epochs) # 开始训练 model.save(args.train_dir) # 保存模型 # 构建测试数据集 test_cat_filenames = tf.constant(cats[10000:]) test_dog_filenames = tf.constant(dogs[10000:]) test_filenames = tf.concat([test_cat_filenames, test_dog_filenames], axis=-1) test_labels = tf.concat([ tf.zeros(test_cat_filenames.shape, dtype=tf.int32), tf.ones(test_dog_filenames.shape, dtype=tf.int32) ], axis=-1) test_dataset = tf.data.Dataset.from_tensor_slices((test_filenames, test_labels)) test_dataset = test_dataset.map(_decode_and_resize) test_dataset = test_dataset.batch(args.batch_size) sparse_categorical_accuracy = tf.keras.metrics.SparseCategoricalAccuracy() # 稀疏分类指标 for images, label in test_dataset: y_pred = model.predict(images) sparse_categorical_accuracy.update_state(y_true=label, y_pred=y_pred) # 更新准确率指标的状态 print("test accuracy:%f" % sparse_categorical_accuracy.result()) # 输出结果 训练命令：python $GEMINI_RUN/DogsVsCats.py --num_epochs 5 --data_dir $GEMINI_DATA_IN1/DogsVsCats/ --train_dir $GEMINI_DATA_OUT最终结果：学习心得：由于之前接触的都是pytorch，这次第一次接触了TensorFlow，学习到了很多TensorFlow的api和这次demo简洁优美的代码风格

【Datawhale夏令营第二期】AI量化模型预测挑战赛 前言参加了Datawhale的AI夏令营第二期的机器学习赛道~，没错这次还是机器学习baseline：https://aistudio.baidu.com/aistudio/projectdetail/6598302?sUid=2554132&shared=1&ts=1690895519028赛事任务简单地说就是通过模型预测股票价格看这里： https://challenge.xfyun.cn/topic/info?type=quantitative-model数据说明date：日期time：时间戳close：最新价/收盘价amount_delta：成交量变化 从上个tick到当前tick发生的成交金额n_midprice：中间价 标准化后的中间价，以涨跌幅表示n_bidN: 买N价n_bsizeN：买N量n_ask：卖N价n_asize1：卖N量labelN：Ntick价格移动方向 Ntick之后中间价相对于当前tick的移动方向，0为下跌，1为不变，2为上涨评估指标采用macro-F1 score进行评价，取label_5, label_10, label_20, label_40, label_60五项中的最高分作为最终得分。F1 score解释： https://www.9998k.cn/archives/169.htmlF1 scoreF1 score = 2 * (precision * recall) / (precision + recall)precision and recall第一次看的时候还不太懂precision和recall的含义，也总结一下首先定义以下几个概念：TP（True Positive）：将本类归为本类TN (True Negative) : 将其他类归为其他类FP（False Positive）：错将其他类预测为本类FN（False Negative）：本类标签预测为其他类标precision = TP / (TP + FP)recall = TP / (TP + FN)accuracy = (TP + TN) / (TP + TN + FP + FN)分析baseline分析本次baseline使用了catboost，相比于上次的LightGBM，可以方便地调用显卡进行训练： 【Datawhale夏令营第二期】CatBoost如何使用GPU在本次的baseline中，只需在cv_model函数中的line15中的params: dict中添加一个键即可：'task_type' : 'GPU'代码解析path = 'AI量化模型预测挑战赛公开数据/' # 数据目录 train_files = os.listdir(path+'train') # 获取目录下的文件 train_df = pd.DataFrame() # 定义一个空的DataFrame for filename in tqdm.tqdm(train_files): # 遍历文件并使用tqdm显示 if os.path.isdir(path+'train/'+filename): # 先判断是否是缓存文件夹直接跳过，防止bug continue tmp = pd.read_csv(path+'train/'+filename) # 使用tmp变量存储读取转换后的DataFrame tmp['file'] = filename # 记录文件名 train_df = pd.concat([train_df, tmp], axis=0, ignore_index=True) # 合并tmp数据到之前的df test_files = os.listdir(path+'test') test_df = pd.DataFrame() for filename in tqdm.tqdm(test_files): if os.path.isdir(path+'train/'+filename): continue tmp = pd.read_csv(path+'test/'+filename) tmp['file'] = filename test_df = pd.concat([test_df, tmp], axis=0, ignore_index=True)特征工程1.0wap(Weighted Average Price)加权平均价格 = (买价 买量 + 卖价 买量) / (买量 + 卖量)# 计算wap数值 train_df['wap1'] = (train_df['n_bid1']*train_df['n_bsize1'] + train_df['n_ask1']*train_df['n_asize1'])/(train_df['n_bsize1'] + train_df['n_asize1']) test_df['wap1'] = (test_df['n_bid1']*test_df['n_bsize1'] + test_df['n_ask1']*test_df['n_asize1'])/(test_df['n_bsize1'] + test_df['n_asize1']) train_df['wap2'] = (train_df['n_bid2']*train_df['n_bsize2'] + train_df['n_ask2']*train_df['n_asize2'])/(train_df['n_bsize2'] + train_df['n_asize2']) test_df['wap2'] = (test_df['n_bid2']*test_df['n_bsize2'] + test_df['n_ask2']*test_df['n_asize2'])/(test_df['n_bsize2'] + test_df['n_asize2']) train_df['wap3'] = (train_df['n_bid3']*train_df['n_bsize3'] + train_df['n_ask3']*train_df['n_asize3'])/(train_df['n_bsize3'] + train_df['n_asize3']) test_df['wap3'] = (test_df['n_bid3']*test_df['n_bsize3'] + test_df['n_ask3']*test_df['n_asize3'])/(test_df['n_bsize3'] + test_df['n_asize3'])wap1wap2wap3特征工程2.0 # 为了保证时间顺序的一致性，故进行排序 train_df = train_df.sort_values(['file','time']) test_df = test_df.sort_values(['file','time']) # 当前时间特征 # 围绕买卖价格和买卖量进行构建 # 暂时只构建买一卖一和买二卖二相关特征，进行优化时可以加上其余买卖信息 train_df['wap1'] = (train_df['n_bid1']*train_df['n_bsize1'] + train_df['n_ask1']*train_df['n_asize1'])/(train_df['n_bsize1'] + train_df['n_asize1']) test_df['wap1'] = (test_df['n_bid1']*test_df['n_bsize1'] + test_df['n_ask1']*test_df['n_asize1'])/(test_df['n_bsize1'] + test_df['n_asize1']) train_df['wap2'] = (train_df['n_bid2']*train_df['n_bsize2'] + train_df['n_ask2']*train_df['n_asize2'])/(train_df['n_bsize2'] + train_df['n_asize2']) test_df['wap2'] = (test_df['n_bid2']*test_df['n_bsize2'] + test_df['n_ask2']*test_df['n_asize2'])/(test_df['n_bsize2'] + test_df['n_asize2']) train_df['wap_balance'] = abs(train_df['wap1'] - train_df['wap2']) train_df['price_spread'] = (train_df['n_ask1'] - train_df['n_bid1']) / ((train_df['n_ask1'] + train_df['n_bid1'])/2) train_df['bid_spread'] = train_df['n_bid1'] - train_df['n_bid2'] train_df['ask_spread'] = train_df['n_ask1'] - train_df['n_ask2'] train_df['total_volume'] = (train_df['n_asize1'] + train_df['n_asize2']) + (train_df['n_bsize1'] + train_df['n_bsize2']) train_df['volume_imbalance'] = abs((train_df['n_asize1'] + train_df['n_asize2']) - (train_df['n_bsize1'] + train_df['n_bsize2'])) test_df['wap_balance'] = abs(test_df['wap1'] - test_df['wap2']) test_df['price_spread'] = (test_df['n_ask1'] - test_df['n_bid1']) / ((test_df['n_ask1'] + test_df['n_bid1'])/2) test_df['bid_spread'] = test_df['n_bid1'] - test_df['n_bid2'] test_df['ask_spread'] = test_df['n_ask1'] - test_df['n_ask2'] test_df['total_volume'] = (test_df['n_asize1'] + test_df['n_asize2']) + (test_df['n_bsize1'] + test_df['n_bsize2']) test_df['volume_imbalance'] = abs((test_df['n_asize1'] + test_df['n_asize2']) - (test_df['n_bsize1'] + test_df['n_bsize2'])) # 历史平移 # 获取历史信息 for val in ['wap1','wap2','wap_balance','price_spread','bid_spread','ask_spread','total_volume','volume_imbalance']: for loc in [1,5,10,20,40,60]: train_df[f'file_{val}_shift{loc}'] = train_df.groupby(['file'])[val].shift(loc) test_df[f'file_{val}_shift{loc}'] = test_df.groupby(['file'])[val].shift(loc) # 差分特征 # 获取与历史数据的增长关系 for val in ['wap1','wap2','wap_balance','price_spread','bid_spread','ask_spread','total_volume','volume_imbalance']: for loc in [1,5,10,20,40,60]: train_df[f'file_{val}_diff{loc}'] = train_df.groupby(['file'])[val].diff(loc) test_df[f'file_{val}_diff{loc}'] = test_df.groupby(['file'])[val].diff(loc) # 窗口统计 # 获取历史信息分布变化信息 # 可以尝试更多窗口大小已经统计方式，如min、max、median等 for val in ['wap1','wap2','wap_balance','price_spread','bid_spread','ask_spread','total_volume','volume_imbalance']: train_df[f'file_{val}_win7_mean'] = train_df.groupby(['file'])[val].transform(lambda x: x.rolling(window=7, min_periods=3).mean()) train_df[f'file_{val}_win7_std'] = train_df.groupby(['file'])[val].transform(lambda x: x.rolling(window=7, min_periods=3).std()) test_df[f'file_{val}_win7_mean'] = test_df.groupby(['file'])[val].transform(lambda x: x.rolling(window=7, min_periods=3).mean()) test_df[f'file_{val}_win7_std'] = test_df.groupby(['file'])[val].transform(lambda x: x.rolling(window=7, min_periods=3).std()) # 时间相关特征 train_df['hour'] = train_df['time'].apply(lambda x:int(x.split(':')[0])) test_df['hour'] = test_df['time'].apply(lambda x:int(x.split(':')[0])) train_df['minute'] = train_df['time'].apply(lambda x:int(x.split(':')[1])) test_df['minute'] = test_df['time'].apply(lambda x:int(x.split(':')[1])) # 入模特征 cols = [f for f in test_df.columns if f not in ['uuid','time','file']]模型训练def cv_model(clf, train_x, train_y, test_x, clf_name, seed = 2023): folds = 5 kf = KFold(n_splits=folds, shuffle=True, random_state=seed) oof = np.zeros([train_x.shape[0], 3]) test_predict = np.zeros([test_x.shape[0], 3]) cv_scores = [] for i, (train_index, valid_index) in enumerate(kf.split(train_x, train_y)): print('************************************ {} ************************************'.format(str(i+1))) trn_x, trn_y, val_x, val_y = train_x.iloc[train_index], train_y[train_index], train_x.iloc[valid_index], train_y[valid_index] if clf_name == "cat": params = {'learning_rate': 0.12, 'depth': 9, 'bootstrap_type':'Bernoulli','random_seed':2023, 'od_type': 'Iter', 'od_wait': 100, 'random_seed': 11, 'allow_writing_files': False, 'loss_function': 'MultiClass', 'task_type' : 'GPU'} # 使用'task_type' : 'GPU'可利用gpu加速训练 model = clf(iterations=2000, **params) model.fit(trn_x, trn_y, eval_set=(val_x, val_y), metric_period=200, use_best_model=True, cat_features=[], verbose=1) val_pred = model.predict_proba(val_x) test_pred = model.predict_proba(test_x) oof[valid_index] = val_pred test_predict += test_pred / kf.n_splits F1_score = f1_score(val_y, np.argmax(val_pred, axis=1), average='macro') cv_scores.append(F1_score) print(cv_scores) return oof, test_predict for label in ['label_5','label_10','label_20','label_40','label_60']: print(f'=================== {label} ===================') cat_oof, cat_test = cv_model(CatBoostClassifier, train_df[cols], train_df[label], test_df[cols], 'cat') train_df[label] = np.argmax(cat_oof, axis=1) test_df[label] = np.argmax(cat_test, axis=1)疑问为什么没有使用rnn或lstm进行预测？代码