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

循环神经网路 RNN 普通的RNN：小数据集 低算力S=f(winXt+b) St=f(WinXt+WSSt−1+b)变种输入序列，单输出单输入，输出序列输入不随序列变化原始的N to N的RNN要求序列等长，然而我们遇到的大部分问题序列都是不等长的，如机器翻译中，源语言和目标语言的句子往往并没有相同的长度。下面介绍RNN最重要的一个变种：N to M。这种结构又叫Encoder-Decoder模型，也可以称之为Seq2Seq模型。从名字就能看出，这个结构的原理是先编码后解码。左侧的RNN用来编码得到c，拿到c后再用右侧的RNN进行解码。得到c有多种方式：长短期记忆网络 LSTMforget gate 遗忘门f1=sigmoid(w1[St−1xt]+b1)input gate 输入门f2=sigmoid(w2[St−1xt]+b2)∗tanh(w2′[St−1xt]+b2′) ct=f1∗ct−1+f2

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

Jupyter Notebook添加删除查看kernel 查看 Jupyter notebook kerneljupyter kernelspec list添加kernel# 首先进入道相应环境中 python -m pip install ipykernel python -m ipykernel install --user --name=kernelname --display-name showname # name为创建的文件夹名，showname为jupyter notebook展示的内核名删除jupyter kerneljupyter kernelspec remove kernelname

【Datawhale夏令营第三期】用户新增预测挑战赛 前言又又又参加了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（外加运营助教）baseline：https://aistudio.baidu.com/aistudio/projectdetail/6618108赛事任务基于提供的样本构建模型，预测用户的新增情况数据说明udmap: 以dict形式给出，需要自定义函数解析common_ts: 事件发生时间，可使用df.dt进行解析x1-x8: 某种特征，但不清楚含义，可通过后续画图分析进行处理target: 预测目标0或1二分类评估指标还是f1，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）：真阳性TN (True Negative) : 真阴性FP（False Positive）：假阳性FN（False Negative）：假阴性precision = TP / (TP + FP)recall = TP / (TP + FN)accuracy = (TP + TN) / (TP + TN + FP + FN)分析baseline分析跑出来是0.62+使用了决策树模型进行训练对udmap特征进行了提取计算了eid的freq和mean通过df.dt提取了hour信息特征提取除去baseline的特征外，又借鉴了锂电池那次的baseline，提取了如下特征：dayofweekweekofyeardayofyearis_weekend调整后的分数为：0.75train_data['common_ts_hour'] = train_data['common_ts'].dt.hour test_data['common_ts_hour'] = test_data['common_ts'].dt.hour train_data['common_ts_minute'] = train_data['common_ts'].dt.minute + train_data['common_ts_hour'] * 60 test_data['common_ts_minute'] = test_data['common_ts'].dt.minute + test_data['common_ts_hour'] * 60 train_data['dayofweek'] = train_data['common_ts'].dt.dayofweek test_data['dayofweek'] = test_data['common_ts'].dt.dayofweek train_data["weekofyear"] = train_data["common_ts"].dt.isocalendar().week.astype(int) test_data["weekofyear"] = test_data["common_ts"].dt.isocalendar().week.astype(int) train_data["dayofyear"] = train_data["common_ts"].dt.dayofyear test_data["dayofyear"] = test_data["common_ts"].dt.dayofyear train_data["day"] = train_data["common_ts"].dt.day test_data["day"] = test_data["common_ts"].dt.day train_data['is_weekend'] = train_data['dayofweek'] // 6 test_data['is_weekend'] = test_data['dayofweek'] // 6x1-x8特征提取train_data['x1_freq'] = train_data['x1'].map(train_data['x1'].value_counts()) test_data['x1_freq'] = test_data['x1'].map(train_data['x1'].value_counts()) test_data['x1_freq'].fillna(test_data['x1_freq'].mode()[0], inplace=True) train_data['x1_mean'] = train_data['x1'].map(train_data.groupby('x1')['target'].mean()) test_data['x1_mean'] = test_data['x1'].map(train_data.groupby('x1')['target'].mean()) test_data['x1_mean'].fillna(test_data['x1_mean'].mode()[0], inplace=True) train_data['x2_freq'] = train_data['x2'].map(train_data['x2'].value_counts()) test_data['x2_freq'] = test_data['x2'].map(train_data['x2'].value_counts()) test_data['x2_freq'].fillna(test_data['x2_freq'].mode()[0], inplace=True) train_data['x2_mean'] = train_data['x2'].map(train_data.groupby('x2')['target'].mean()) test_data['x2_mean'] = test_data['x2'].map(train_data.groupby('x2')['target'].mean()) test_data['x2_mean'].fillna(test_data['x2_mean'].mode()[0], inplace=True) train_data['x3_freq'] = train_data['x3'].map(train_data['x3'].value_counts()) test_data['x3_freq'] = test_data['x3'].map(train_data['x3'].value_counts()) test_data['x3_freq'].fillna(test_data['x3_freq'].mode()[0], inplace=True) train_data['x4_freq'] = train_data['x4'].map(train_data['x4'].value_counts()) test_data['x4_freq'] = test_data['x4'].map(train_data['x4'].value_counts()) test_data['x4_freq'].fillna(test_data['x4_freq'].mode()[0], inplace=True) train_data['x6_freq'] = train_data['x6'].map(train_data['x6'].value_counts()) test_data['x6_freq'] = test_data['x6'].map(train_data['x6'].value_counts()) test_data['x6_freq'].fillna(test_data['x6_freq'].mode()[0], inplace=True) train_data['x6_mean'] = train_data['x6'].map(train_data.groupby('x6')['target'].mean()) test_data['x6_mean'] = test_data['x6'].map(train_data.groupby('x6')['target'].mean()) test_data['x6_mean'].fillna(test_data['x6_mean'].mode()[0], inplace=True) train_data['x7_freq'] = train_data['x7'].map(train_data['x7'].value_counts()) test_data['x7_freq'] = test_data['x7'].map(train_data['x7'].value_counts()) test_data['x7_freq'].fillna(test_data['x7_freq'].mode()[0], inplace=True) train_data['x7_mean'] = train_data['x7'].map(train_data.groupby('x7')['target'].mean()) test_data['x7_mean'] = test_data['x7'].map(train_data.groupby('x7')['target'].mean()) test_data['x7_mean'].fillna(test_data['x7_mean'].mode()[0], inplace=True) train_data['x8_freq'] = train_data['x8'].map(train_data['x8'].value_counts()) test_data['x8_freq'] = test_data['x8'].map(train_data['x8'].value_counts()) test_data['x8_freq'].fillna(test_data['x8_freq'].mode()[0], inplace=True) train_data['x8_mean'] = train_data['x8'].map(train_data.groupby('x8')['target'].mean()) test_data['x8_mean'] = test_data['x8'].map(train_data.groupby('x8')['target'].mean()) test_data['x8_mean'].fillna(test_data['x8_mean'].mode()[0], inplace=True)实测使用众数填充会比0填充好一点实测分数 0.76398无脑大招：AutoGluon直接上代码：import pandas as pd import numpy as np train_data = pd.read_csv('用户新增预测挑战赛公开数据/train.csv') test_data = pd.read_csv('用户新增预测挑战赛公开数据/test.csv') #autogluon from autogluon.tabular import TabularDataset, TabularPredictor clf = TabularPredictor(label='target') clf.fit( TabularDataset(train_data.drop(['uuid'], axis=1)), ) print("预测的正确率为：",clf.evaluate( TabularDataset(train_data.drop(['uuid'], axis=1)), ) ) pd.DataFrame({ 'uuid': test_data['uuid'], 'target': clf.predict(test_data.drop(['uuid'], axis=1)) }).to_csv('submit.csv', index=None)AutoGluon分数：0.79868使用x1-x8识别用户特征参考自Ivan大佬import pandas as pd import numpy as np train_data = pd.read_csv('用户新增预测挑战赛公开数据/train.csv') test_data = pd.read_csv('用户新增预测挑战赛公开数据/test.csv') user_df = train_data.groupby(by=['x1', 'x2', 'x3', 'x4', 'x5', 'x6', 'x7', 'x8'])['target'].mean().reset_index( name='user_prob') from sklearn.tree import DecisionTreeClassifier for i in range(user_df.shape[0]): x1 = user_df.iloc[i, 0] x2 = user_df.iloc[i, 1] x3 = user_df.iloc[i, 2] x4 = user_df.iloc[i, 3] x5 = user_df.iloc[i, 4] x6 = user_df.iloc[i, 5] x7 = user_df.iloc[i, 6] x8 = user_df.iloc[i, 7] sub_train = train_data.loc[ (train_data['x1'] == x1) & (train_data['x2'] == x2) & (train_data['x3'] == x3) & (train_data['x4'] == x4) & (train_data['x5'] == x5) & (train_data['x6'] == x6) & (train_data['x7'] == x7) & (train_data['x8'] == x8) ] sub_test = test_data.loc[ (test_data['x1'] == x1) & (test_data['x2'] == x2) & (test_data['x3'] == x3) & (test_data['x4'] == x4) & (test_data['x5'] == x5) & (test_data['x6'] == x6) & (test_data['x7'] == x7) & (test_data['x8'] == x8) ] # print(sub_train.columns) clf = DecisionTreeClassifier() clf.fit( sub_train.loc[:, ['eid', 'common_ts']], sub_train['target'] ) try: test_data.loc[ (test_data['x1'] == x1) & (test_data['x2'] == x2) & (test_data['x3'] == x3) & (test_data['x4'] == x4) & (test_data['x5'] == x5) & (test_data['x6'] == x6) & (test_data['x7'] == x7) & (test_data['x8'] == x8), ['target'] ] = clf.predict( test_data.loc[ (test_data['x1'] == x1) & (test_data['x2'] == x2) & (test_data['x3'] == x3) & (test_data['x4'] == x4) & (test_data['x5'] == x5) & (test_data['x6'] == x6) & (test_data['x7'] == x7) & (test_data['x8'] == x8), ['eid', 'common_ts']] ) except: pass test_data.fillna(0, inplace=True) test_data['target'] = test_data.target.astype(int) test_data[['uuid','target']].to_csv('submit_2.csv', index=None)实测分数：0.831最后 修改代码，把fillna替换为如下代码from sklearn.tree import DecisionTreeClassifier clf = DecisionTreeClassifier() clf.fit( train_data.drop(['udmap', 'common_ts', 'uuid', 'target', 'common_ts_hour'], axis=1), train_data['target'] ) test_data.loc[pd.isna(test_data['target']),'target'] = \ clf.predict( test_data.loc[ pd.isna(test_data['target']), test_data.drop(['udmap', 'common_ts', 'uuid', 'target', 'common_ts_hour'], axis=1).columns] )最终分数： 0.8321