模型训练日志，刷屏使我痛苦，由此呢研发了VLog（附源码）-外链论坛

6hz7vif 发表于 2024-10-8 14:44:07

模型训练日志，刷屏使我痛苦，由此呢研发了VLog（附源码）

<a style="color: black;"><img src="https://mmbiz.qpic.cn/mmbiz_gif/QWAZ7Ok6ecxlg4UicZFDZtejx9tbsaBteIhLotBAZb5WsBvo5jh6KnQe2Cjic2HxIMEib18dibIXoI7hgSQEfTC7FQ/640?wx_fmt=gif&wxfrom=5&wx_lazy=1&tp=webp" style="width: 50%; margin-bottom: 20px;"></a>
以下文案源自于微X公众号：算法美食屋
作者：梁云1991
链接：https://mp.weixin.qq.com/s/KtXsPjmWhOpylO1EDdM7Fg
本文仅用于学术分享，如有侵权，请联系后台作删文处理<img src="https://mmbiz.qpic.cn/mmbiz_gif/Ljib4So7yuWhPmGHfBz7vlEFLwDwiaiaQoCQ1iaUGFFpV3z4Qq9uTTh0gibzDT52BKUwfAM5Kf2jsDpKRmuQuO2asXQ/640?wx_fmt=gif&wxfrom=5&wx_lazy=1&tp=webp" style="width: 50%; margin-bottom: 20px;">导读在模型训练过程中，常常会打印每一个epoch训练的日志，令人眼花缭乱。本文运用torchkeras库中的VLog，分别仔细的演示了在LightGBM、ultralytics、transformers中运用VLog，并供给了源码，期盼对大众有帮忙。训练日志刷屏使我痛苦，我研发了VLog，能够在任意训练代码中容易运用~例如，经过callback嵌入到lightgbm/catboost/transformers/ultralytics，乃至keras库的训练代码流程中~
before:
<img src="https://mmbiz.qpic.cn/mmbiz_png/4WgILHBwVH8XJIyFWnENYfuRE1ianebUlpwZwbtklhw2jJicuLyvkiaz6GTn0dfNR6xuIdbaEsh3OPjgGwVL9KoeQ/640?wx_fmt=png&tp=webp&wxfrom=5&wx_lazy=1&wx_co=1" style="width: 50%; margin-bottom: 20px;">
after：
<img src="https://mmbiz.qpic.cn/mmbiz_gif/4WgILHBwVH8XJIyFWnENYfuRE1ianebUlPicLe3dSXkVMxX9xFDnt03MuYwnibYFYBU60O3gT0Ry85pNOWU8D57cw/640?wx_fmt=gif&tp=webp&wxfrom=5&wx_lazy=1" style="width: 50%; margin-bottom: 20px;">
为何不消tensorboard或wandb?tensorboard需要开端口权限，服务器研发环境有时候无端口权限~wandb需要联网，有时候网速很差或无网，影响体验~综合对比思虑如下表
<img src="https://mmbiz.qpic.cn/mmbiz_png/4WgILHBwVH8XJIyFWnENYfuRE1ianebUlxwtOAaCtVjmPLwHOM1Eh9uSUXKYBhvDLMuegiblOL1oy3sFZRhu7vYg/640?wx_fmt=png&tp=webp&wxfrom=5&wx_lazy=1&wx_co=1" style="width: 50%; margin-bottom: 20px;">
<h2 style="color: black; text-align: left; margin-bottom: 10px;">一，VLog基本原理</h2>VLog类重点有以下5个办法。from torchkeras import
 VLog#1, 初始化办法vlog = VLog(epochs=20, monitor_metric=val_loss, monitor_mode=min
) #2, 表示起始空图表
vlog.log_start()#3, 更新step级别日志vlog.log_step({train_loss:0.003,val_loss:0.002
}) #4, 更新epoch级别日志vlog.log_epoch({train_acc:0.9,val_acc:0.87,train_loss:0.002,val_loss:0.03
})#5, 输出最后稳定状态图表
vlog.log_end()import
 timeimport
 math,randomfrom torchkeras import
 VLogepochs = 10batchs = 30#0, 指定监控北极星指标，以及指标优化方向vlog = VLog(epochs, monitor_metric=val_loss, monitor_mode=min
) #1, log_start 初始化动态图表
vlog.log_start() for epoch in
 range(epochs):    #train    for step in
 range(batchs):        #2, log_step 更新step级别日志信息，打日志，并用小进度条表示进度        vlog.log_step({train_loss:100-2.5*epoch+math.sin(2
*step/batchs)})         time.sleep(0.05
)    #eval        for step in range(20
):        #3, log_step 更新step级别日志信息，指定training=False说明在验证模式，只打日志不更新小进度条        vlog.log_step({val_loss:100-2*epoch+math.sin(2*step/batchs)},training=False
)        time.sleep(0.05
)    #4, log_epoch 更新epoch级别日志信息，每一个epoch刷新一次动态图表和大进度条进度    vlog.log_epoch({val_loss:100 - 2*epoch+2*random.random()-1
,                    train_loss:100-2.5*epoch+2*random.random()-1
})  # 5, log_end 调节坐标轴范围，输出最后指标可视化图表
vlog.log_end()
<img src="https://mmbiz.qpic.cn/mmbiz_png/4WgILHBwVH8XJIyFWnENYfuRE1ianebUlPXFibxtud3EEARtF2H8coZt6iafaOgHYcGKskOHURSUgNoWLayMBSiaug/640?wx_fmt=png&tp=webp&wxfrom=5&wx_lazy=1&wx_co=1" style="width: 50%; margin-bottom: 20px;">
<h2 style="color: black; text-align: left; margin-bottom: 10px;">二，在LightGBM中运用VLog</h2>
设计一个简单的回调，就能够搞定~from torchkeras import
 VLogclass VLogCallback:    def __init__
 (self, num_boost_round,                  monitor_metric=val_loss
 ,                 monitor_mode=min)
 :        self.order = 20
        self.num_boost_round = num_boost_round
self.vlog = VLog(epochs = num_boost_round, monitor_metric = monitor_metric,
                         monitor_mode = monitor_mode)    def __call__(self, env) -> None:
        metrics = {}        for item in
 env.evaluation_result_list:            if len(item) == 4
:data_name, eval_name, result = item[:3
]                metrics[data_name+_
+eval_name] = result            else
:                data_name, eval_name = item[1
].split()                res_mean = item[2
]                res_stdv = item[4
]                metrics[data_name+_
+eval_name] = res_mean
        self.vlog.log_epoch(metrics)import
 datetimeimport numpy as
 npimport pandas as
 pdimport lightgbm as
 lgbfrom sklearn import
 datasetsfromsklearn.model_selectionimport
 train_test_splitfrom sklearn.metrics import
 accuracy_scoredef printlog(info):    nowtime = datetime.datetime.now().strftime(%Y-%m-%d %H:%M:%S
)    print("\n"+"=========="*8 + "%s"
%nowtime)    print(info+...\n\n
)#================================================================================# 一，读取数据#================================================================================printlog("step1: reading data..."
)# 读取dftrain,dftest
breast = datasets.load_breast_cancer()df = pd.DataFrame(breast.data,columns = [x.replace(,_) for x in
 breast.feature_names])df[label
] = breast.targetdf[mean_radius] = df[mean_radius].apply(lambda
 x:int(x))df[mean_texture] = df[mean_texture].apply(lambda
 x:int(x))
dftrain,dftest = train_test_split(df)categorical_features = [mean_radius,mean_texture
]lgb_train = lgb.Dataset(dftrain.drop([label],axis = 1),label=dftrain[label
],
                        categorical_feature = categorical_features)lgb_valid = lgb.Dataset(dftest.drop([label],axis = 1),label=dftest[label
],
                        categorical_feature = categorical_features,
                        reference=lgb_train)#================================================================================# 二，设置参数#================================================================================printlog("step2: setting parameters..."
)boost_round = 50early_stop_rounds = 10
params = {    boosting_type: gbdt
,    objective:binary
,    metric: [auc], #l2    num_leaves:15
,       learning_rate: 0.05
,    feature_fraction: 0.9
,    bagging_fraction: 0.8
,    bagging_freq: 5
,    verbose: 0
,    early_stopping_round:5
}#================================================================================# 三，训练模型#================================================================================printlog("step3: training model..."
)
result = {}vlog_cb = VLogCallback(boost_round, monitor_metric = val_auc, monitor_mode = max
)
vlog_cb.vlog.log_start()
gbm = lgb.train(params,
                lgb_train,
                num_boost_round= boost_round,
                valid_sets=(lgb_valid, lgb_train),                valid_names=(val,train
),
                callbacks = [lgb.record_evaluation(result),
vlog_cb]
               )
vlog_cb.vlog.log_end()#================================================================================# 四，评定模型#================================================================================printlog("step4: evaluating model ..."
)y_pred_train = gbm.predict(dftrain.drop(label,axis = 1
), num_iteration=gbm.best_iteration)y_pred_test = gbm.predict(dftest.drop(label,axis = 1
), num_iteration=gbm.best_iteration)print(train accuracy: {:.5} .format(accuracy_score(dftrain[label],y_pred_train>0.5
)))print(valid accuracy: {:.5} \n.format(accuracy_score(dftest[label],y_pred_test>0.5
)))#================================================================================# 五，保留模型#================================================================================printlog("step5: saving model ..."
)model_dir = "gbm.model"print("model_dir: %s"
%model_dir)gbm.save_model("gbm.model"
)printlog("task end..."
)######
================================================================================2023-11-10 15:39:38
step1: reading data......
================================================================================2023-11-10 15:39:38
step2: setting parameters......
================================================================================2023-11-10 15:39:38
step3: training model......
<img src="https://mmbiz.qpic.cn/mmbiz_png/4WgILHBwVH8XJIyFWnENYfuRE1ianebUlfYIzS6Qib3lNpbficqsIo4bDQmGOaUI6NjDS64bmNPvfcGOZFLBcW9zQ/640?wx_fmt=png&tp=webp&wxfrom=5&wx_lazy=1&wx_co=1" style="width: 50%; margin-bottom: 20px;">
================================================================================2023-11-10 15:39:44
step4: evaluating model ......
train accuracy: 0.95775 
valid accuracy: 0.94406 
================================================================================2023-11-10 15:39:44
step5: saving model ......
model_dir: gbm.model
================================================================================2023-11-10 15:39:44
task end......
<h2 style="color: black; text-align: left; margin-bottom: 10px;">三，在ultralytics中运用VLog</h2>写个适配的回调~ultralytics能够做 归类，检测，分割等等。这个回调函数是通用的，此处以归类问题为例，改个monitor_metric就可~cats_vs_dogs数据集能够在公众号江大白后台回复:torchkeras 获取~from torchkeras import
 VLogclass VLogCallback:    def __init__(self,epochs,monitor_metric,monitor_mode):
self.vlog = VLog(epochs,monitor_metric,monitor_mode)    def on_train_batch_end(self,trainer):        self.vlog.log_step(trainer.label_loss_items(trainer.tloss, prefix=train
))    def on_fit_epoch_end(self,trainer):        metrics = {k.split(/)[-1]:v for k,v in trainer.metrics.items() if loss not in
 k}
        self.vlog.log_epoch(metrics)    def on_train_epoch_end(self,trainer):        passfrom ultralytics import
 YOLO epochs = 10
vlog_cb = VLogCallback(epochs = epochs,                       monitor_metric=accuracy_top1
,                       monitor_mode=max
)
callbacks = {    "on_train_batch_end"
: vlog_cb.on_train_batch_end,    "on_fit_epoch_end"
: vlog_cb.on_fit_epoch_end
}model = YOLO(model = yolov8n-cls.pt
)forevent,funcin
 callbacks.items():
    model.add_callback(event,func)
vlog_cb.vlog.log_start()results = model.train(data=cats_vs_dogs
,                       epochs=epochs, workers=4)     # train the model
vlog_cb.vlog.log_end()
<img src="https://mmbiz.qpic.cn/mmbiz_png/4WgILHBwVH8XJIyFWnENYfuRE1ianebUlYUGOsRX6lN7rcCIpM875JrClvMX0Mic6zt2nia7f6eIUuFtIIIBjE1Fw/640?wx_fmt=png&tp=webp&wxfrom=5&wx_lazy=1&wx_co=1" style="width: 50%; margin-bottom: 20px;">
<h2 style="color: black; text-align: left; margin-bottom: 10px;">四，在transformers中运用VLog</h2>waimai评论数据集能够在公众号江大白后台回复:torchkeras 获取~#回调给你写好了~from torchkeras.tools.transformers import
VLogCallbackimport numpy as
 np import pandas as
 pd import
 torch import
 datasets from transformers import
 AutoTokenizer,DataCollatorWithPaddingfrom transformers import
AutoModelForSequenceClassificationfrom transformers import
 TrainingArguments,Trainer from transformers import
 EarlyStoppingCallbackfrom tqdm import
 tqdm from transformers import
 AdamW, get_scheduler#一，准备数据df = pd.read_csv("waimai_10k.csv"
)
ds = datasets.Dataset.from_pandas(df)ds = ds.shuffle(42
) ds = ds.rename_columns({"review":"text","label":"labels"
})tokenizer = AutoTokenizer.from_pretrained(bert-base-chinese
) ds_encoded = ds.map(lambda example:tokenizer(example["text"
]),remove_columns = ["text"
],                    batched=True
)#train,val,test splitds_train_val,ds_test = ds_encoded.train_test_split(test_size=0.2
).values()ds_train,ds_val = ds_train_val.train_test_split(test_size=0.2
).values() 
data_collator = DataCollatorWithPadding(tokenizer=tokenizer)dl_train = torch.utils.data.DataLoader(ds_train, batch_size=16
, collate_fn = data_collator)dl_val = torch.utils.data.DataLoader(ds_val, batch_size=16
,  collate_fn = data_collator)dl_test = torch.utils.data.DataLoader(ds_test, batch_size=16
,  collate_fn = data_collator)for batch in
 dl_train:    breakprint({k: v.shape for k, v in
 batch.items()})#二，定义模型
model = AutoModelForSequenceClassification.from_pretrained(    bert-base-chinese,num_labels=2
)#三，训练模型def compute_metrics(eval_preds):
logits, labels = eval_preds    preds = np.argmax(logits, axis=-1
)
    accuracy = np.sum(preds==labels)/len(labels)    precision = np.sum((preds==1)&(labels==1))/np.sum(preds==1
)recall = np.sum((preds==1)&(labels==1))/np.sum(labels==1
)    f1  = 2
*recall*precision/(recall+precision)    return {"accuracy":accuracy,"precision":precision,"recall":recall,f1
:f1}
training_args = TrainingArguments(    output_dir = "bert_waimai"
,    num_train_epochs = 3
,    logging_steps = 20
,    gradient_accumulation_steps = 10
,evaluation_strategy="steps", #epoch    metric_for_best_model=eval_f1
,    greater_is_better=True
,    report_to=none
,    load_best_model_at_end=True
)callbacks = [EarlyStoppingCallback(early_stopping_patience=10
),             VLogCallback()] #监控指标同 metric_for_best_model
trainer = Trainer(
    model,
    training_args,
    train_dataset=ds_train,
eval_dataset=ds_val,
    compute_metrics=compute_metrics,
    callbacks = callbacks,
    data_collator=data_collator,
    tokenizer=tokenizer,
)
trainer.train() #四，评定模型
trainer.evaluate(ds_val)#五，运用模型from transformers import
 pipelinemodel.config.id2label = {0:"差评",1:"好评"
}classifier = pipeline(task="text-classification"
,tokenizer = tokenizer,model=model.cpu())classifier("挺好吃的哦"
)#六，保留模型model.save_pretrained("waimai_10k_bert"
)tokenizer.save_pretrained("waimai_10k_bert"
)classifier = pipeline("text-classification",model="waimai_10k_bert"
)classifier(["味道还不错，下次再来","我去，吃了我吐了三天"
])
<img src="data:image/svg+xml,%3C%3Fxml version=1.0 encoding=UTF-8%3F%3E%3Csvg width=1px height=1px viewBox=0 0 1 1 version=1.1 xmlns=http://www.w3.org/2000/svg xmlns:xlink=http://www.w3.org/1999/xlink%3E%3Ctitle%3E%3C/title%3E%3Cg stroke=none stroke-width=1 fill=none fill-rule=evenodd fill-opacity=0%3E%3Cg transform=translate(-249.000000, -126.000000) fill=%23FFFFFF%3E%3Crect x=249 y=126 width=1 height=1%3E%3C/rect%3E%3C/g%3E%3C/g%3E%3C/svg%3E" style="width: 50%; margin-bottom: 20px;">
<img src="data:image/svg+xml,%3C%3Fxml version=1.0 encoding=UTF-8%3F%3E%3Csvg width=1px height=1px viewBox=0 0 1 1 version=1.1 xmlns=http://www.w3.org/2000/svg xmlns:xlink=http://www.w3.org/1999/xlink%3E%3Ctitle%3E%3C/title%3E%3Cg stroke=none stroke-width=1 fill=none fill-rule=evenodd fill-opacity=0%3E%3Cg transform=translate(-249.000000, -126.000000) fill=%23FFFFFF%3E%3Crect x=249 y=126 width=1 height=1%3E%3C/rect%3E%3C/g%3E%3C/g%3E%3C/svg%3E" style="width: 50%; margin-bottom: 20px;">万水千山总是情，点个赞赞行不行~

qzmjef 发表于 2024-10-23 20:12:27

“板凳”（第三个回帖的人）‌

m5k1umn 发表于 2024-10-28 12:29:45

期待你更多的精彩评论，一起交流学习。

1fy07h 发表于 2024-11-5 08:39:01

期待与你深入交流，共探知识的无穷魅力。

nqkk58 发表于 5 天前

对于这个问题，我有不同的看法...

m5k1umn 发表于 4 天前

楼主的文章深得我心，表示由衷的感谢！

页: [1]

外链论坛's Archiver

模型训练日志，刷屏使我痛苦，由此呢研发了VLog（附源码）