使用主动学习进行评论分类
作者: Darshan Deshpande
创建日期 2021/10/29
最后修改日期 2024/05/08
描述: 通过评论分类展示主动学习的优势。
简介
随着以数据为中心的机器学习的兴起，主动学习在企业和研究人员中越来越受欢迎。主动学习旨在逐步训练机器学习模型，以便生成的模型在达到竞争性分数方面需要更少的训练数据。
主动学习管道的结构包括一个分类器和一个预言机。预言机是一个标注者，负责清理、选择、标注数据，并在需要时将其提供给模型。预言机是一个经过培训的个人或一组个人，他们确保对新数据的标注的一致性。
该过程从标注完整数据集的一个小子集开始，并训练一个初始模型。保存最佳模型检查点，然后在平衡的测试集上进行测试。测试集必须进行仔细采样，因为整个训练过程将依赖于它。一旦我们获得初始评估分数，预言机便负责标注更多样本；要采样的数据点数通常由业务需求决定。之后，将新采样的数据添加到训练集中，并重复训练过程。该循环持续进行，直到达到可接受的分数或满足其他业务指标为止。
本教程通过演示基于比率（最小置信度）的采样策略，展示了主动学习的工作原理，与在整个数据集上训练的模型相比，这种策略可以降低整体假阳性和假阴性率。这种采样属于*不确定性采样*领域，其中新数据集是根据模型对相应标签输出的不确定性进行采样的。在我们的示例中，我们比较了模型的假阳性和假阴性率，并根据它们的比率对新数据进行标注。
其他一些采样技术包括
导入所需的库
import os
os.environ["KERAS_BACKEND"] = "tensorflow" # @param ["tensorflow", "jax", "torch"]
import keras
from keras import ops
from keras import layers
import tensorflow_datasets as tfds
import tensorflow as tf
import matplotlib.pyplot as plt
import re
import string
tfds.disable_progress_bar()
加载和预处理数据
我们将使用 IMDB 评论数据集进行我们的实验。该数据集总共有 50,000 条评论，包括训练和测试拆分。我们将合并这些拆分，并对我们自己的平衡训练、验证和测试集进行采样。
dataset = tfds.load(
"imdb_reviews",
split="train + test",
as_supervised=True,
batch_size=-1,
shuffle_files=False,
)
reviews, labels = tfds.as_numpy(dataset)
print("Total examples:", reviews.shape[0])
Total examples: 50000
主动学习从标注一个数据子集开始。对于我们将使用的比率采样技术，我们需要平衡良好的训练、验证和测试拆分。
val_split = 2500
test_split = 2500
train_split = 7500
# Separating the negative and positive samples for manual stratification
x_positives, y_positives = reviews[labels == 1], labels[labels == 1]
x_negatives, y_negatives = reviews[labels == 0], labels[labels == 0]
# Creating training, validation and testing splits
x_val, y_val = (
tf.concat((x_positives[:val_split], x_negatives[:val_split]), 0),
tf.concat((y_positives[:val_split], y_negatives[:val_split]), 0),
)
x_test, y_test = (
tf.concat(
(
x_positives[val_split : val_split + test_split],
x_negatives[val_split : val_split + test_split],
),
0,
),
tf.concat(
(
y_positives[val_split : val_split + test_split],
y_negatives[val_split : val_split + test_split],
),
0,
),
)
x_train, y_train = (
tf.concat(
(
x_positives[val_split + test_split : val_split + test_split + train_split],
x_negatives[val_split + test_split : val_split + test_split + train_split],
),
0,
),
tf.concat(
(
y_positives[val_split + test_split : val_split + test_split + train_split],
y_negatives[val_split + test_split : val_split + test_split + train_split],
),
0,
),
)
# Remaining pool of samples are stored separately. These are only labeled as and when required
x_pool_positives, y_pool_positives = (
x_positives[val_split + test_split + train_split :],
y_positives[val_split + test_split + train_split :],
)
x_pool_negatives, y_pool_negatives = (
x_negatives[val_split + test_split + train_split :],
y_negatives[val_split + test_split + train_split :],
)
# Creating TF Datasets for faster prefetching and parallelization
train_dataset = tf.data.Dataset.from_tensor_slices((x_train, y_train))
val_dataset = tf.data.Dataset.from_tensor_slices((x_val, y_val))
test_dataset = tf.data.Dataset.from_tensor_slices((x_test, y_test))
pool_negatives = tf.data.Dataset.from_tensor_slices(
(x_pool_negatives, y_pool_negatives)
)
pool_positives = tf.data.Dataset.from_tensor_slices(
(x_pool_positives, y_pool_positives)
)
print(f"Initial training set size: {len(train_dataset)}")
print(f"Validation set size: {len(val_dataset)}")
print(f"Testing set size: {len(test_dataset)}")
print(f"Unlabeled negative pool: {len(pool_negatives)}")
print(f"Unlabeled positive pool: {len(pool_positives)}")
Initial training set size: 15000
Validation set size: 5000
Testing set size: 5000
Unlabeled negative pool: 12500
Unlabeled positive pool: 12500
拟合 TextVectorization 层
由于我们处理的是文本数据，因此我们需要将文本字符串编码为向量，然后将其传递给 Embedding 层。为了使这种标记过程更快，我们使用 map() 函数及其并行化功能。
vectorizer = layers.TextVectorization(
3000, standardize="lower_and_strip_punctuation", output_sequence_length=150
)
# Adapting the dataset
vectorizer.adapt(
train_dataset.map(lambda x, y: x, num_parallel_calls=tf.data.AUTOTUNE).batch(256)
)
def vectorize_text(text, label):
text = vectorizer(text)
return text, label
train_dataset = train_dataset.map(
vectorize_text, num_parallel_calls=tf.data.AUTOTUNE
).prefetch(tf.data.AUTOTUNE)
pool_negatives = pool_negatives.map(vectorize_text, num_parallel_calls=tf.data.AUTOTUNE)
pool_positives = pool_positives.map(vectorize_text, num_parallel_calls=tf.data.AUTOTUNE)
val_dataset = val_dataset.batch(256).map(
vectorize_text, num_parallel_calls=tf.data.AUTOTUNE
)
test_dataset = test_dataset.batch(256).map(
vectorize_text, num_parallel_calls=tf.data.AUTOTUNE
)
创建帮助函数
# Helper function for merging new history objects with older ones
def append_history(losses, val_losses, accuracy, val_accuracy, history):
losses = losses + history.history["loss"]
val_losses = val_losses + history.history["val_loss"]
accuracy = accuracy + history.history["binary_accuracy"]
val_accuracy = val_accuracy + history.history["val_binary_accuracy"]
return losses, val_losses, accuracy, val_accuracy
# Plotter function
def plot_history(losses, val_losses, accuracies, val_accuracies):
plt.plot(losses)
plt.plot(val_losses)
plt.legend(["train_loss", "val_loss"])
plt.xlabel("Epochs")
plt.ylabel("Loss")
plt.show()
plt.plot(accuracies)
plt.plot(val_accuracies)
plt.legend(["train_accuracy", "val_accuracy"])
plt.xlabel("Epochs")
plt.ylabel("Accuracy")
plt.show()
创建模型
我们创建了一个小型双向 LSTM 模型。在使用主动学习时，应确保模型架构能够过度拟合初始数据。过度拟合强烈暗示模型将有足够的容量处理未来的未见数据。
def create_model():
model = keras.models.Sequential(
[
layers.Input(shape=(150,)),
layers.Embedding(input_dim=3000, output_dim=128),
layers.Bidirectional(layers.LSTM(32, return_sequences=True)),
layers.GlobalMaxPool1D(),
layers.Dense(20, activation="relu"),
layers.Dropout(0.5),
layers.Dense(1, activation="sigmoid"),
]
)
model.summary()
return model
在整个数据集上进行训练
为了展示主动学习的有效性，我们首先将在包含 40,000 个标记样本的整个数据集上训练模型。该模型将用于以后进行比较。
def train_full_model(full_train_dataset, val_dataset, test_dataset):
model = create_model()
model.compile(
loss="binary_crossentropy",
optimizer="rmsprop",
metrics=[
keras.metrics.BinaryAccuracy(),
keras.metrics.FalseNegatives(),
keras.metrics.FalsePositives(),
],
)
# We will save the best model at every epoch and load the best one for evaluation on the test set
history = model.fit(
full_train_dataset.batch(256),
epochs=20,
validation_data=val_dataset,
callbacks=[
keras.callbacks.EarlyStopping(patience=4, verbose=1),
keras.callbacks.ModelCheckpoint(
"FullModelCheckpoint.keras", verbose=1, save_best_only=True
),
],
)
# Plot history
plot_history(
history.history["loss"],
history.history["val_loss"],
history.history["binary_accuracy"],
history.history["val_binary_accuracy"],
)
# Loading the best checkpoint
model = keras.models.load_model("FullModelCheckpoint.keras")
print("-" * 100)
print(
"Test set evaluation: ",
model.evaluate(test_dataset, verbose=0, return_dict=True),
)
print("-" * 100)
return model
# Sampling the full train dataset to train on
full_train_dataset = (
train_dataset.concatenate(pool_positives)
.concatenate(pool_negatives)
.cache()
.shuffle(20000)
)
# Training the full model
full_dataset_model = train_full_model(full_train_dataset, val_dataset, test_dataset)
Model: "sequential"
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓ ┃ Layer (type) ┃ Output Shape ┃ Param # ┃ ┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩ │ embedding (Embedding) │ (None, 150, 128) │ 384,000 │ ├─────────────────────────────────┼────────────────────────┼───────────────┤ │ bidirectional (Bidirectional) │ (None, 150, 64) │ 41,216 │ ├─────────────────────────────────┼────────────────────────┼───────────────┤ │ global_max_pooling1d │ (None, 64) │ 0 │ │ (GlobalMaxPooling1D) │ │ │ ├─────────────────────────────────┼────────────────────────┼───────────────┤ │ dense (Dense) │ (None, 20) │ 1,300 │ ├─────────────────────────────────┼────────────────────────┼───────────────┤ │ dropout (Dropout) │ (None, 20) │ 0 │ ├─────────────────────────────────┼────────────────────────┼───────────────┤ │ dense_1 (Dense) │ (None, 1) │ 21 │ └─────────────────────────────────┴────────────────────────┴───────────────┘
Total params: 426,537 (1.63 MB)
Trainable params: 426,537 (1.63 MB)
Non-trainable params: 0 (0.00 B)
Epoch 1/20
156/157 ━━━━━━━━━━━━━━━━━━━ [37m━ 0s 73ms/step - binary_accuracy: 0.6412 - false_negatives: 2084.3333 - false_positives: 5252.1924 - loss: 0.6507
Epoch 1: val_loss improved from inf to 0.57198, saving model to FullModelCheckpoint.keras
157/157 ━━━━━━━━━━━━━━━━━━━━ 15s 79ms/step - binary_accuracy: 0.6411 - false_negatives: 2135.1772 - false_positives: 5292.4053 - loss: 0.6506 - val_binary_accuracy: 0.7356 - val_false_negatives: 898.0000 - val_false_positives: 424.0000 - val_loss: 0.5720
Epoch 2/20
156/157 ━━━━━━━━━━━━━━━━━━━ [37m━ 0s 72ms/step - binary_accuracy: 0.7448 - false_negatives: 1756.2756 - false_positives: 3249.1411 - loss: 0.5416
Epoch 2: val_loss improved from 0.57198 to 0.41756, saving model to FullModelCheckpoint.keras
157/157 ━━━━━━━━━━━━━━━━━━━━ 12s 76ms/step - binary_accuracy: 0.7450 - false_negatives: 1783.8925 - false_positives: 3279.8101 - loss: 0.5412 - val_binary_accuracy: 0.8156 - val_false_negatives: 531.0000 - val_false_positives: 391.0000 - val_loss: 0.4176
Epoch 3/20
156/157 ━━━━━━━━━━━━━━━━━━━ [37m━ 0s 72ms/step - binary_accuracy: 0.8162 - false_negatives: 1539.7693 - false_positives: 2197.1475 - loss: 0.4254
Epoch 3: val_loss improved from 0.41756 to 0.38233, saving model to FullModelCheckpoint.keras
157/157 ━━━━━━━━━━━━━━━━━━━━ 12s 76ms/step - binary_accuracy: 0.8161 - false_negatives: 1562.6962 - false_positives: 2221.5886 - loss: 0.4254 - val_binary_accuracy: 0.8340 - val_false_negatives: 496.0000 - val_false_positives: 334.0000 - val_loss: 0.3823
Epoch 4/20
156/157 ━━━━━━━━━━━━━━━━━━━ [37m━ 0s 72ms/step - binary_accuracy: 0.8413 - false_negatives: 1400.6538 - false_positives: 1818.7372 - loss: 0.3837
Epoch 4: val_loss improved from 0.38233 to 0.36235, saving model to FullModelCheckpoint.keras
157/157 ━━━━━━━━━━━━━━━━━━━━ 12s 76ms/step - binary_accuracy: 0.8412 - false_negatives: 1421.5063 - false_positives: 1839.3102 - loss: 0.3838 - val_binary_accuracy: 0.8396 - val_false_negatives: 548.0000 - val_false_positives: 254.0000 - val_loss: 0.3623
Epoch 5/20
156/157 ━━━━━━━━━━━━━━━━━━━ [37m━ 0s 72ms/step - binary_accuracy: 0.8611 - false_negatives: 1264.5256 - false_positives: 1573.5962 - loss: 0.3468
Epoch 5: val_loss did not improve from 0.36235
157/157 ━━━━━━━━━━━━━━━━━━━━ 12s 75ms/step - binary_accuracy: 0.8611 - false_negatives: 1283.0632 - false_positives: 1592.3228 - loss: 0.3468 - val_binary_accuracy: 0.8222 - val_false_negatives: 734.0000 - val_false_positives: 155.0000 - val_loss: 0.4081
Epoch 6/20
156/157 ━━━━━━━━━━━━━━━━━━━ [37m━ 0s 72ms/step - binary_accuracy: 0.8706 - false_negatives: 1186.9166 - false_positives: 1427.9487 - loss: 0.3301
Epoch 6: val_loss improved from 0.36235 to 0.35041, saving model to FullModelCheckpoint.keras
157/157 ━━━━━━━━━━━━━━━━━━━━ 12s 76ms/step - binary_accuracy: 0.8705 - false_negatives: 1204.8038 - false_positives: 1444.9368 - loss: 0.3302 - val_binary_accuracy: 0.8412 - val_false_negatives: 569.0000 - val_false_positives: 225.0000 - val_loss: 0.3504
Epoch 7/20
156/157 ━━━━━━━━━━━━━━━━━━━ [37m━ 0s 72ms/step - binary_accuracy: 0.8768 - false_negatives: 1162.4423 - false_positives: 1342.4807 - loss: 0.3084
Epoch 7: val_loss improved from 0.35041 to 0.32680, saving model to FullModelCheckpoint.keras
157/157 ━━━━━━━━━━━━━━━━━━━━ 12s 76ms/step - binary_accuracy: 0.8768 - false_negatives: 1179.5253 - false_positives: 1358.4114 - loss: 0.3085 - val_binary_accuracy: 0.8590 - val_false_negatives: 364.0000 - val_false_positives: 341.0000 - val_loss: 0.3268
Epoch 8/20
156/157 ━━━━━━━━━━━━━━━━━━━ [37m━ 0s 73ms/step - binary_accuracy: 0.8865 - false_negatives: 1079.3206 - false_positives: 1250.2693 - loss: 0.2924
Epoch 8: val_loss did not improve from 0.32680
157/157 ━━━━━━━━━━━━━━━━━━━━ 12s 76ms/step - binary_accuracy: 0.8864 - false_negatives: 1094.9873 - false_positives: 1265.0632 - loss: 0.2926 - val_binary_accuracy: 0.8460 - val_false_negatives: 548.0000 - val_false_positives: 222.0000 - val_loss: 0.3432
Epoch 9/20
156/157 ━━━━━━━━━━━━━━━━━━━ [37m━ 0s 73ms/step - binary_accuracy: 0.8912 - false_negatives: 1019.1987 - false_positives: 1189.4551 - loss: 0.2807
Epoch 9: val_loss did not improve from 0.32680
157/157 ━━━━━━━━━━━━━━━━━━━━ 12s 77ms/step - binary_accuracy: 0.8912 - false_negatives: 1033.9684 - false_positives: 1203.5632 - loss: 0.2808 - val_binary_accuracy: 0.8588 - val_false_negatives: 330.0000 - val_false_positives: 376.0000 - val_loss: 0.3302
Epoch 10/20
156/157 ━━━━━━━━━━━━━━━━━━━ [37m━ 0s 72ms/step - binary_accuracy: 0.8997 - false_negatives: 968.6346 - false_positives: 1109.9103 - loss: 0.2669
Epoch 10: val_loss did not improve from 0.32680
157/157 ━━━━━━━━━━━━━━━━━━━━ 12s 76ms/step - binary_accuracy: 0.8996 - false_negatives: 983.1202 - false_positives: 1123.3418 - loss: 0.2671 - val_binary_accuracy: 0.8558 - val_false_negatives: 445.0000 - val_false_positives: 276.0000 - val_loss: 0.3413
Epoch 11/20
156/157 ━━━━━━━━━━━━━━━━━━━ [37m━ 0s 72ms/step - binary_accuracy: 0.9055 - false_negatives: 937.0320 - false_positives: 1000.8589 - loss: 0.2520
Epoch 11: val_loss did not improve from 0.32680
157/157 ━━━━━━━━━━━━━━━━━━━━ 12s 76ms/step - binary_accuracy: 0.9055 - false_negatives: 950.3608 - false_positives: 1013.6456 - loss: 0.2521 - val_binary_accuracy: 0.8602 - val_false_negatives: 402.0000 - val_false_positives: 297.0000 - val_loss: 0.3281
Epoch 11: early stopping
----------------------------------------------------------------------------------------------------
Test set evaluation: {'binary_accuracy': 0.8507999777793884, 'false_negatives': 397.0, 'false_positives': 349.0, 'loss': 0.3372706174850464}
----------------------------------------------------------------------------------------------------
通过主动学习进行训练
我们在执行主动学习时遵循的一般过程如下所示
该管道可以概括为五个部分
- 抽取并标注一个小的、平衡的训练数据集
- 在这个小数据集上训练模型
- 在一个平衡的测试集上评估模型
- 如果模型满足业务标准，则将其部署在实时环境中
- 如果模型没有通过标准，则根据误报和漏报的比例抽取一些样本，将其添加到训练集中，并从步骤 2 开始重复，直到模型通过测试或所有可用数据都被用尽。
对于下面的代码，我们将使用以下公式进行采样
主动学习技术广泛使用回调来跟踪进度。在本例中，我们将使用模型检查点和提前停止。提前停止的 patience 参数可以帮助最大程度地减少过度拟合和所需时间。我们现在将其设置为 patience=4，但由于模型很健壮，如果需要，我们可以增加耐心级别。
注意：我们不会在第一次训练迭代后加载检查点。根据我使用主动学习技术的经验，这有助于模型探测新形成的损失情况。即使模型在第二次迭代中没有改进，我们仍然可以洞悉可能的未来误报和漏报率。这将帮助我们在下一次迭代中采样一个更好的集合，在该迭代中，模型将有更大的机会改进。
def train_active_learning_models(
train_dataset,
pool_negatives,
pool_positives,
val_dataset,
test_dataset,
num_iterations=3,
sampling_size=5000,
):
# Creating lists for storing metrics
losses, val_losses, accuracies, val_accuracies = [], [], [], []
model = create_model()
# We will monitor the false positives and false negatives predicted by our model
# These will decide the subsequent sampling ratio for every Active Learning loop
model.compile(
loss="binary_crossentropy",
optimizer="rmsprop",
metrics=[
keras.metrics.BinaryAccuracy(),
keras.metrics.FalseNegatives(),
keras.metrics.FalsePositives(),
],
)
# Defining checkpoints.
# The checkpoint callback is reused throughout the training since it only saves the best overall model.
checkpoint = keras.callbacks.ModelCheckpoint(
"AL_Model.keras", save_best_only=True, verbose=1
)
# Here, patience is set to 4. This can be set higher if desired.
early_stopping = keras.callbacks.EarlyStopping(patience=4, verbose=1)
print(f"Starting to train with {len(train_dataset)} samples")
# Initial fit with a small subset of the training set
history = model.fit(
train_dataset.cache().shuffle(20000).batch(256),
epochs=20,
validation_data=val_dataset,
callbacks=[checkpoint, early_stopping],
)
# Appending history
losses, val_losses, accuracies, val_accuracies = append_history(
losses, val_losses, accuracies, val_accuracies, history
)
for iteration in range(num_iterations):
# Getting predictions from previously trained model
predictions = model.predict(test_dataset)
# Generating labels from the output probabilities
rounded = ops.where(ops.greater(predictions, 0.5), 1, 0)
# Evaluating the number of zeros and ones incorrrectly classified
_, _, false_negatives, false_positives = model.evaluate(test_dataset, verbose=0)
print("-" * 100)
print(
f"Number of zeros incorrectly classified: {false_negatives}, Number of ones incorrectly classified: {false_positives}"
)
# This technique of Active Learning demonstrates ratio based sampling where
# Number of ones/zeros to sample = Number of ones/zeros incorrectly classified / Total incorrectly classified
if false_negatives != 0 and false_positives != 0:
total = false_negatives + false_positives
sample_ratio_ones, sample_ratio_zeros = (
false_positives / total,
false_negatives / total,
)
# In the case where all samples are correctly predicted, we can sample both classes equally
else:
sample_ratio_ones, sample_ratio_zeros = 0.5, 0.5
print(
f"Sample ratio for positives: {sample_ratio_ones}, Sample ratio for negatives:{sample_ratio_zeros}"
)
# Sample the required number of ones and zeros
sampled_dataset = pool_negatives.take(
int(sample_ratio_zeros * sampling_size)
).concatenate(pool_positives.take(int(sample_ratio_ones * sampling_size)))
# Skip the sampled data points to avoid repetition of sample
pool_negatives = pool_negatives.skip(int(sample_ratio_zeros * sampling_size))
pool_positives = pool_positives.skip(int(sample_ratio_ones * sampling_size))
# Concatenating the train_dataset with the sampled_dataset
train_dataset = train_dataset.concatenate(sampled_dataset).prefetch(
tf.data.AUTOTUNE
)
print(f"Starting training with {len(train_dataset)} samples")
print("-" * 100)
# We recompile the model to reset the optimizer states and retrain the model
model.compile(
loss="binary_crossentropy",
optimizer="rmsprop",
metrics=[
keras.metrics.BinaryAccuracy(),
keras.metrics.FalseNegatives(),
keras.metrics.FalsePositives(),
],
)
history = model.fit(
train_dataset.cache().shuffle(20000).batch(256),
validation_data=val_dataset,
epochs=20,
callbacks=[
checkpoint,
keras.callbacks.EarlyStopping(patience=4, verbose=1),
],
)
# Appending the history
losses, val_losses, accuracies, val_accuracies = append_history(
losses, val_losses, accuracies, val_accuracies, history
)
# Loading the best model from this training loop
model = keras.models.load_model("AL_Model.keras")
# Plotting the overall history and evaluating the final model
plot_history(losses, val_losses, accuracies, val_accuracies)
print("-" * 100)
print(
"Test set evaluation: ",
model.evaluate(test_dataset, verbose=0, return_dict=True),
)
print("-" * 100)
return model
active_learning_model = train_active_learning_models(
train_dataset, pool_negatives, pool_positives, val_dataset, test_dataset
)
Model: "sequential_1"
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓ ┃ Layer (type) ┃ Output Shape ┃ Param # ┃ ┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩ │ embedding_1 (Embedding) │ (None, 150, 128) │ 384,000 │ ├─────────────────────────────────┼────────────────────────┼───────────────┤ │ bidirectional_1 (Bidirectional) │ (None, 150, 64) │ 41,216 │ ├─────────────────────────────────┼────────────────────────┼───────────────┤ │ global_max_pooling1d_1 │ (None, 64) │ 0 │ │ (GlobalMaxPooling1D) │ │ │ ├─────────────────────────────────┼────────────────────────┼───────────────┤ │ dense_2 (Dense) │ (None, 20) │ 1,300 │ ├─────────────────────────────────┼────────────────────────┼───────────────┤ │ dropout_1 (Dropout) │ (None, 20) │ 0 │ ├─────────────────────────────────┼────────────────────────┼───────────────┤ │ dense_3 (Dense) │ (None, 1) │ 21 │ └─────────────────────────────────┴────────────────────────┴───────────────┘
Total params: 426,537 (1.63 MB)
Trainable params: 426,537 (1.63 MB)
Non-trainable params: 0 (0.00 B)
Starting to train with 15000 samples
Epoch 1/20
59/59 ━━━━━━━━━━━━━━━━━━━━ 0s 72ms/step - binary_accuracy: 0.5197 - false_negatives_1: 1686.7457 - false_positives_1: 1938.3051 - loss: 0.6918
Epoch 1: val_loss improved from inf to 0.67428, saving model to AL_Model.keras
59/59 ━━━━━━━━━━━━━━━━━━━━ 8s 89ms/step - binary_accuracy: 0.5202 - false_negatives_1: 1716.9833 - false_positives_1: 1961.4667 - loss: 0.6917 - val_binary_accuracy: 0.6464 - val_false_negatives_1: 279.0000 - val_false_positives_1: 1489.0000 - val_loss: 0.6743
Epoch 2/20
59/59 ━━━━━━━━━━━━━━━━━━━━ 0s 72ms/step - binary_accuracy: 0.6505 - false_negatives_1: 1216.0170 - false_positives_1: 1434.2373 - loss: 0.6561
Epoch 2: val_loss improved from 0.67428 to 0.59133, saving model to AL_Model.keras
59/59 ━━━━━━━━━━━━━━━━━━━━ 5s 83ms/step - binary_accuracy: 0.6507 - false_negatives_1: 1234.9833 - false_positives_1: 1455.7667 - loss: 0.6558 - val_binary_accuracy: 0.7032 - val_false_negatives_1: 235.0000 - val_false_positives_1: 1249.0000 - val_loss: 0.5913
Epoch 3/20
59/59 ━━━━━━━━━━━━━━━━━━━━ 0s 72ms/step - binary_accuracy: 0.7103 - false_negatives_1: 939.5255 - false_positives_1: 1235.8983 - loss: 0.5829
Epoch 3: val_loss improved from 0.59133 to 0.51602, saving model to AL_Model.keras
59/59 ━━━━━━━━━━━━━━━━━━━━ 5s 83ms/step - binary_accuracy: 0.7106 - false_negatives_1: 953.0500 - false_positives_1: 1255.3167 - loss: 0.5827 - val_binary_accuracy: 0.7686 - val_false_negatives_1: 812.0000 - val_false_positives_1: 345.0000 - val_loss: 0.5160
Epoch 4/20
59/59 ━━━━━━━━━━━━━━━━━━━━ 0s 72ms/step - binary_accuracy: 0.7545 - false_negatives_1: 787.4237 - false_positives_1: 1070.0339 - loss: 0.5214
Epoch 4: val_loss improved from 0.51602 to 0.43948, saving model to AL_Model.keras
59/59 ━━━━━━━━━━━━━━━━━━━━ 5s 83ms/step - binary_accuracy: 0.7547 - false_negatives_1: 799.2667 - false_positives_1: 1085.8833 - loss: 0.5212 - val_binary_accuracy: 0.8028 - val_false_negatives_1: 342.0000 - val_false_positives_1: 644.0000 - val_loss: 0.4395
Epoch 5/20
59/59 ━━━━━━━━━━━━━━━━━━━━ 0s 72ms/step - binary_accuracy: 0.7919 - false_negatives_1: 676.7458 - false_positives_1: 907.4915 - loss: 0.4657
Epoch 5: val_loss improved from 0.43948 to 0.41679, saving model to AL_Model.keras
59/59 ━━━━━━━━━━━━━━━━━━━━ 5s 83ms/step - binary_accuracy: 0.7920 - false_negatives_1: 687.3834 - false_positives_1: 921.1667 - loss: 0.4655 - val_binary_accuracy: 0.8158 - val_false_negatives_1: 598.0000 - val_false_positives_1: 323.0000 - val_loss: 0.4168
Epoch 6/20
59/59 ━━━━━━━━━━━━━━━━━━━━ 0s 72ms/step - binary_accuracy: 0.7994 - false_negatives_1: 661.3560 - false_positives_1: 828.0847 - loss: 0.4498
Epoch 6: val_loss improved from 0.41679 to 0.39680, saving model to AL_Model.keras
59/59 ━━━━━━━━━━━━━━━━━━━━ 5s 83ms/step - binary_accuracy: 0.7997 - false_negatives_1: 671.3666 - false_positives_1: 840.2500 - loss: 0.4495 - val_binary_accuracy: 0.8260 - val_false_negatives_1: 382.0000 - val_false_positives_1: 488.0000 - val_loss: 0.3968
Epoch 7/20
59/59 ━━━━━━━━━━━━━━━━━━━━ 0s 72ms/step - binary_accuracy: 0.8311 - false_negatives_1: 589.1187 - false_positives_1: 707.0170 - loss: 0.4017
Epoch 7: val_loss did not improve from 0.39680
59/59 ━━━━━━━━━━━━━━━━━━━━ 5s 83ms/step - binary_accuracy: 0.8312 - false_negatives_1: 598.3500 - false_positives_1: 717.8167 - loss: 0.4016 - val_binary_accuracy: 0.7706 - val_false_negatives_1: 1004.0000 - val_false_positives_1: 143.0000 - val_loss: 0.4884
Epoch 8/20
59/59 ━━━━━━━━━━━━━━━━━━━━ 0s 72ms/step - binary_accuracy: 0.8365 - false_negatives_1: 566.7288 - false_positives_1: 649.9322 - loss: 0.3896
Epoch 8: val_loss did not improve from 0.39680
59/59 ━━━━━━━━━━━━━━━━━━━━ 5s 83ms/step - binary_accuracy: 0.8366 - false_negatives_1: 575.2833 - false_positives_1: 660.2167 - loss: 0.3895 - val_binary_accuracy: 0.8216 - val_false_negatives_1: 623.0000 - val_false_positives_1: 269.0000 - val_loss: 0.4043
Epoch 9/20
59/59 ━━━━━━━━━━━━━━━━━━━━ 0s 72ms/step - binary_accuracy: 0.8531 - false_negatives_1: 519.0170 - false_positives_1: 591.6440 - loss: 0.3631
Epoch 9: val_loss improved from 0.39680 to 0.37727, saving model to AL_Model.keras
59/59 ━━━━━━━━━━━━━━━━━━━━ 5s 83ms/step - binary_accuracy: 0.8531 - false_negatives_1: 527.2667 - false_positives_1: 601.2500 - loss: 0.3631 - val_binary_accuracy: 0.8348 - val_false_negatives_1: 296.0000 - val_false_positives_1: 530.0000 - val_loss: 0.3773
Epoch 10/20
59/59 ━━━━━━━━━━━━━━━━━━━━ 0s 72ms/step - binary_accuracy: 0.8686 - false_negatives_1: 475.7966 - false_positives_1: 569.0508 - loss: 0.3387
Epoch 10: val_loss improved from 0.37727 to 0.37354, saving model to AL_Model.keras
59/59 ━━━━━━━━━━━━━━━━━━━━ 5s 83ms/step - binary_accuracy: 0.8685 - false_negatives_1: 483.5000 - false_positives_1: 577.9667 - loss: 0.3387 - val_binary_accuracy: 0.8400 - val_false_negatives_1: 327.0000 - val_false_positives_1: 473.0000 - val_loss: 0.3735
Epoch 11/20
59/59 ━━━━━━━━━━━━━━━━━━━━ 0s 72ms/step - binary_accuracy: 0.8716 - false_negatives_1: 452.1356 - false_positives_1: 522.1187 - loss: 0.3303
Epoch 11: val_loss improved from 0.37354 to 0.37074, saving model to AL_Model.keras
59/59 ━━━━━━━━━━━━━━━━━━━━ 5s 83ms/step - binary_accuracy: 0.8716 - false_negatives_1: 459.3833 - false_positives_1: 530.6667 - loss: 0.3303 - val_binary_accuracy: 0.8390 - val_false_negatives_1: 362.0000 - val_false_positives_1: 443.0000 - val_loss: 0.3707
Epoch 12/20
59/59 ━━━━━━━━━━━━━━━━━━━━ 0s 72ms/step - binary_accuracy: 0.8833 - false_negatives_1: 433.0678 - false_positives_1: 481.1864 - loss: 0.3065
Epoch 12: val_loss did not improve from 0.37074
59/59 ━━━━━━━━━━━━━━━━━━━━ 5s 83ms/step - binary_accuracy: 0.8833 - false_negatives_1: 439.8333 - false_positives_1: 488.9667 - loss: 0.3066 - val_binary_accuracy: 0.8236 - val_false_negatives_1: 208.0000 - val_false_positives_1: 674.0000 - val_loss: 0.4046
Epoch 13/20
59/59 ━━━━━━━━━━━━━━━━━━━━ 0s 72ms/step - binary_accuracy: 0.8876 - false_negatives_1: 384.8305 - false_positives_1: 476.5254 - loss: 0.2978
Epoch 13: val_loss did not improve from 0.37074
59/59 ━━━━━━━━━━━━━━━━━━━━ 5s 82ms/step - binary_accuracy: 0.8876 - false_negatives_1: 391.2667 - false_positives_1: 484.2500 - loss: 0.2978 - val_binary_accuracy: 0.8380 - val_false_negatives_1: 364.0000 - val_false_positives_1: 446.0000 - val_loss: 0.3783
Epoch 14/20
59/59 ━━━━━━━━━━━━━━━━━━━━ 0s 72ms/step - binary_accuracy: 0.8976 - false_negatives_1: 378.0169 - false_positives_1: 433.9831 - loss: 0.2754
Epoch 14: val_loss did not improve from 0.37074
59/59 ━━━━━━━━━━━━━━━━━━━━ 5s 83ms/step - binary_accuracy: 0.8975 - false_negatives_1: 384.2333 - false_positives_1: 441.3833 - loss: 0.2757 - val_binary_accuracy: 0.8310 - val_false_negatives_1: 525.0000 - val_false_positives_1: 320.0000 - val_loss: 0.3957
Epoch 15/20
59/59 ━━━━━━━━━━━━━━━━━━━━ 0s 72ms/step - binary_accuracy: 0.9013 - false_negatives_1: 354.9322 - false_positives_1: 403.1695 - loss: 0.2709
Epoch 15: val_loss did not improve from 0.37074
59/59 ━━━━━━━━━━━━━━━━━━━━ 5s 83ms/step - binary_accuracy: 0.9013 - false_negatives_1: 360.4000 - false_positives_1: 409.5833 - loss: 0.2709 - val_binary_accuracy: 0.8298 - val_false_negatives_1: 302.0000 - val_false_positives_1: 549.0000 - val_loss: 0.4015
Epoch 15: early stopping
20/20 ━━━━━━━━━━━━━━━━━━━━ 1s 39ms/step
----------------------------------------------------------------------------------------------------
Number of zeros incorrectly classified: 290.0, Number of ones incorrectly classified: 538.0
Sample ratio for positives: 0.6497584541062802, Sample ratio for negatives:0.3502415458937198
Starting training with 19999 samples
----------------------------------------------------------------------------------------------------
Epoch 1/20
78/79 ━━━━━━━━━━━━━━━━━━━ [37m━ 0s 72ms/step - binary_accuracy: 0.8735 - false_negatives_2: 547.2436 - false_positives_2: 650.2436 - loss: 0.3527
Epoch 1: val_loss did not improve from 0.37074
79/79 ━━━━━━━━━━━━━━━━━━━━ 9s 84ms/step - binary_accuracy: 0.8738 - false_negatives_2: 559.2125 - false_positives_2: 665.3375 - loss: 0.3518 - val_binary_accuracy: 0.7932 - val_false_negatives_2: 119.0000 - val_false_positives_2: 915.0000 - val_loss: 0.4949
Epoch 2/20
78/79 ━━━━━━━━━━━━━━━━━━━ [37m━ 0s 72ms/step - binary_accuracy: 0.8961 - false_negatives_2: 470.2436 - false_positives_2: 576.1539 - loss: 0.2824
Epoch 2: val_loss did not improve from 0.37074
79/79 ━━━━━━━━━━━━━━━━━━━━ 6s 80ms/step - binary_accuracy: 0.8962 - false_negatives_2: 481.4125 - false_positives_2: 589.6750 - loss: 0.2823 - val_binary_accuracy: 0.8014 - val_false_negatives_2: 809.0000 - val_false_positives_2: 184.0000 - val_loss: 0.4580
Epoch 3/20
78/79 ━━━━━━━━━━━━━━━━━━━ [37m━ 0s 72ms/step - binary_accuracy: 0.9059 - false_negatives_2: 442.2051 - false_positives_2: 500.5385 - loss: 0.2628
Epoch 3: val_loss did not improve from 0.37074
79/79 ━━━━━━━━━━━━━━━━━━━━ 6s 80ms/step - binary_accuracy: 0.9059 - false_negatives_2: 452.6750 - false_positives_2: 513.5250 - loss: 0.2629 - val_binary_accuracy: 0.8294 - val_false_negatives_2: 302.0000 - val_false_positives_2: 551.0000 - val_loss: 0.3868
Epoch 4/20
78/79 ━━━━━━━━━━━━━━━━━━━ [37m━ 0s 72ms/step - binary_accuracy: 0.9188 - false_negatives_2: 394.5513 - false_positives_2: 462.4359 - loss: 0.2391
Epoch 4: val_loss did not improve from 0.37074
79/79 ━━━━━━━━━━━━━━━━━━━━ 6s 80ms/step - binary_accuracy: 0.9187 - false_negatives_2: 405.0625 - false_positives_2: 474.1250 - loss: 0.2393 - val_binary_accuracy: 0.8268 - val_false_negatives_2: 225.0000 - val_false_positives_2: 641.0000 - val_loss: 0.4197
Epoch 5/20
78/79 ━━━━━━━━━━━━━━━━━━━ [37m━ 0s 72ms/step - binary_accuracy: 0.9255 - false_negatives_2: 349.8718 - false_positives_2: 413.0898 - loss: 0.2270
Epoch 5: val_loss did not improve from 0.37074
79/79 ━━━━━━━━━━━━━━━━━━━━ 6s 79ms/step - binary_accuracy: 0.9254 - false_negatives_2: 358.6500 - false_positives_2: 423.5625 - loss: 0.2270 - val_binary_accuracy: 0.8228 - val_false_negatives_2: 611.0000 - val_false_positives_2: 275.0000 - val_loss: 0.4233
Epoch 6/20
78/79 ━━━━━━━━━━━━━━━━━━━ [37m━ 0s 73ms/step - binary_accuracy: 0.9265 - false_negatives_2: 349.8590 - false_positives_2: 389.9359 - loss: 0.2147
Epoch 6: val_loss did not improve from 0.37074
79/79 ━━━━━━━━━━━━━━━━━━━━ 6s 80ms/step - binary_accuracy: 0.9265 - false_negatives_2: 358.8375 - false_positives_2: 399.9875 - loss: 0.2148 - val_binary_accuracy: 0.8272 - val_false_negatives_2: 581.0000 - val_false_positives_2: 283.0000 - val_loss: 0.4415
Epoch 7/20
78/79 ━━━━━━━━━━━━━━━━━━━ [37m━ 0s 72ms/step - binary_accuracy: 0.9409 - false_negatives_2: 286.7820 - false_positives_2: 322.7949 - loss: 0.1877
Epoch 7: val_loss did not improve from 0.37074
79/79 ━━━━━━━━━━━━━━━━━━━━ 6s 79ms/step - binary_accuracy: 0.9408 - false_negatives_2: 294.4375 - false_positives_2: 331.4000 - loss: 0.1880 - val_binary_accuracy: 0.8266 - val_false_negatives_2: 528.0000 - val_false_positives_2: 339.0000 - val_loss: 0.4419
Epoch 7: early stopping
20/20 ━━━━━━━━━━━━━━━━━━━━ 1s 39ms/step
----------------------------------------------------------------------------------------------------
Number of zeros incorrectly classified: 376.0, Number of ones incorrectly classified: 442.0
Sample ratio for positives: 0.5403422982885085, Sample ratio for negatives:0.45965770171149145
Starting training with 24998 samples
----------------------------------------------------------------------------------------------------
Epoch 1/20
98/98 ━━━━━━━━━━━━━━━━━━━━ 0s 73ms/step - binary_accuracy: 0.8509 - false_negatives_3: 809.9184 - false_positives_3: 1018.9286 - loss: 0.3732
Epoch 1: val_loss improved from 0.37074 to 0.36196, saving model to AL_Model.keras
98/98 ━━━━━━━━━━━━━━━━━━━━ 11s 83ms/step - binary_accuracy: 0.8509 - false_negatives_3: 817.5757 - false_positives_3: 1028.7980 - loss: 0.3731 - val_binary_accuracy: 0.8424 - val_false_negatives_3: 368.0000 - val_false_positives_3: 420.0000 - val_loss: 0.3620
Epoch 2/20
98/98 ━━━━━━━━━━━━━━━━━━━━ 0s 72ms/step - binary_accuracy: 0.8744 - false_negatives_3: 734.7449 - false_positives_3: 884.7755 - loss: 0.3185
Epoch 2: val_loss did not improve from 0.36196
98/98 ━━━━━━━━━━━━━━━━━━━━ 8s 79ms/step - binary_accuracy: 0.8744 - false_negatives_3: 741.9697 - false_positives_3: 893.7172 - loss: 0.3186 - val_binary_accuracy: 0.8316 - val_false_negatives_3: 202.0000 - val_false_positives_3: 640.0000 - val_loss: 0.3792
Epoch 3/20
98/98 ━━━━━━━━━━━━━━━━━━━━ 0s 72ms/step - binary_accuracy: 0.8830 - false_negatives_3: 684.1326 - false_positives_3: 807.8878 - loss: 0.3090
Epoch 3: val_loss did not improve from 0.36196
98/98 ━━━━━━━━━━━━━━━━━━━━ 8s 79ms/step - binary_accuracy: 0.8830 - false_negatives_3: 691.0707 - false_positives_3: 816.2222 - loss: 0.3090 - val_binary_accuracy: 0.8118 - val_false_negatives_3: 738.0000 - val_false_positives_3: 203.0000 - val_loss: 0.4112
Epoch 4/20
98/98 ━━━━━━━━━━━━━━━━━━━━ 0s 72ms/step - binary_accuracy: 0.8892 - false_negatives_3: 651.9898 - false_positives_3: 776.4388 - loss: 0.2928
Epoch 4: val_loss did not improve from 0.36196
98/98 ━━━━━━━━━━━━━━━━━━━━ 8s 79ms/step - binary_accuracy: 0.8892 - false_negatives_3: 658.4041 - false_positives_3: 784.3839 - loss: 0.2928 - val_binary_accuracy: 0.8344 - val_false_negatives_3: 557.0000 - val_false_positives_3: 271.0000 - val_loss: 0.3734
Epoch 5/20
98/98 ━━━━━━━━━━━━━━━━━━━━ 0s 72ms/step - binary_accuracy: 0.8975 - false_negatives_3: 612.0714 - false_positives_3: 688.9184 - loss: 0.2806
Epoch 5: val_loss did not improve from 0.36196
98/98 ━━━━━━━━━━━━━━━━━━━━ 8s 79ms/step - binary_accuracy: 0.8974 - false_negatives_3: 618.4343 - false_positives_3: 696.1313 - loss: 0.2807 - val_binary_accuracy: 0.8456 - val_false_negatives_3: 446.0000 - val_false_positives_3: 326.0000 - val_loss: 0.3658
Epoch 5: early stopping
20/20 ━━━━━━━━━━━━━━━━━━━━ 1s 40ms/step
----------------------------------------------------------------------------------------------------
Number of zeros incorrectly classified: 407.0, Number of ones incorrectly classified: 410.0
Sample ratio for positives: 0.5018359853121175, Sample ratio for negatives:0.4981640146878825
Starting training with 29997 samples
----------------------------------------------------------------------------------------------------
Epoch 1/20
117/118 ━━━━━━━━━━━━━━━━━━━ [37m━ 0s 76ms/step - binary_accuracy: 0.8621 - false_negatives_4: 916.2393 - false_positives_4: 1130.9744 - loss: 0.3527
Epoch 1: val_loss did not improve from 0.36196
118/118 ━━━━━━━━━━━━━━━━━━━━ 13s 85ms/step - binary_accuracy: 0.8621 - false_negatives_4: 931.0924 - false_positives_4: 1149.7479 - loss: 0.3525 - val_binary_accuracy: 0.8266 - val_false_negatives_4: 627.0000 - val_false_positives_4: 240.0000 - val_loss: 0.3802
Epoch 2/20
117/118 ━━━━━━━━━━━━━━━━━━━ [37m━ 0s 76ms/step - binary_accuracy: 0.8761 - false_negatives_4: 876.4872 - false_positives_4: 1005.5726 - loss: 0.3195
Epoch 2: val_loss improved from 0.36196 to 0.35707, saving model to AL_Model.keras
118/118 ━━━━━━━━━━━━━━━━━━━━ 10s 82ms/step - binary_accuracy: 0.8760 - false_negatives_4: 891.0504 - false_positives_4: 1022.9412 - loss: 0.3196 - val_binary_accuracy: 0.8404 - val_false_negatives_4: 479.0000 - val_false_positives_4: 319.0000 - val_loss: 0.3571
Epoch 3/20
117/118 ━━━━━━━━━━━━━━━━━━━ [37m━ 0s 74ms/step - binary_accuracy: 0.8874 - false_negatives_4: 801.1710 - false_positives_4: 941.4786 - loss: 0.2965
Epoch 3: val_loss did not improve from 0.35707
118/118 ━━━━━━━━━━━━━━━━━━━━ 9s 79ms/step - binary_accuracy: 0.8873 - false_negatives_4: 814.8319 - false_positives_4: 957.8571 - loss: 0.2966 - val_binary_accuracy: 0.8226 - val_false_negatives_4: 677.0000 - val_false_positives_4: 210.0000 - val_loss: 0.3948
Epoch 4/20
117/118 ━━━━━━━━━━━━━━━━━━━ [37m━ 0s 76ms/step - binary_accuracy: 0.8977 - false_negatives_4: 740.5385 - false_positives_4: 837.1710 - loss: 0.2768
Epoch 4: val_loss did not improve from 0.35707
118/118 ━━━━━━━━━━━━━━━━━━━━ 10s 81ms/step - binary_accuracy: 0.8976 - false_negatives_4: 753.5378 - false_positives_4: 852.2437 - loss: 0.2770 - val_binary_accuracy: 0.8406 - val_false_negatives_4: 530.0000 - val_false_positives_4: 267.0000 - val_loss: 0.3630
Epoch 5/20
117/118 ━━━━━━━━━━━━━━━━━━━ [37m━ 0s 76ms/step - binary_accuracy: 0.9020 - false_negatives_4: 722.5214 - false_positives_4: 808.2308 - loss: 0.2674
Epoch 5: val_loss did not improve from 0.35707
118/118 ━━━━━━━━━━━━━━━━━━━━ 10s 82ms/step - binary_accuracy: 0.9019 - false_negatives_4: 734.8655 - false_positives_4: 822.4117 - loss: 0.2676 - val_binary_accuracy: 0.8330 - val_false_negatives_4: 592.0000 - val_false_positives_4: 243.0000 - val_loss: 0.3805
Epoch 6/20
117/118 ━━━━━━━━━━━━━━━━━━━ [37m━ 0s 76ms/step - binary_accuracy: 0.9059 - false_negatives_4: 682.1453 - false_positives_4: 737.0513 - loss: 0.2525
Epoch 6: val_loss did not improve from 0.35707
118/118 ━━━━━━━━━━━━━━━━━━━━ 10s 82ms/step - binary_accuracy: 0.9059 - false_negatives_4: 693.6387 - false_positives_4: 749.9412 - loss: 0.2526 - val_binary_accuracy: 0.8454 - val_false_negatives_4: 391.0000 - val_false_positives_4: 382.0000 - val_loss: 0.3620
Epoch 6: early stopping
----------------------------------------------------------------------------------------------------
Test set evaluation: {'binary_accuracy': 0.8424000144004822, 'false_negatives_4': 491.0, 'false_positives_4': 297.0, 'loss': 0.3661557137966156}
----------------------------------------------------------------------------------------------------
结论
主动学习是一个不断发展的研究领域。本示例展示了使用主动学习的成本效益，因为它消除了对大量数据进行标注的需要，从而节省了资源。
以下是一些值得注意的观察结果
- 我们只需要 30,000 个样本就能达到与在完整数据集上训练的模型相同的（甚至更好的）分数。这意味着在现实环境中，我们节省了对 10,000 张图像进行标注的工作量！
- 与完整训练获得的偏斜比例相比，在训练结束时，误报和漏报的数量得到了很好的平衡。这使得该模型在现实场景中稍微更有用，因为这两个标签都同等重要。
有关采样比率类型、训练技术或可用开源库/实现的更多信息，您可以参考以下资源