训练空间NER模型，损失会随着迭代和批次的增加而不断增加

Question

# Create a blank 'en' model
nlp = spacy.blank("en")

# Create a new entity recognizer and add it to the pipeline
ner = nlp.create_pipe("ner")
nlp.add_pipe(ner)
# Add a new label
ner.add_label('LABEL')

# Start the training
optimizer = nlp.begin_training()
# Loop for 10 iterations
for itn in range(10):
# Shuffle the training data
    random.shuffle(spacy_train)
    losses = {}
    # Batch the examples and iterate over them
    for batch in spacy.util.minibatch(spacy_train, size=spacy.util.compounding(4.0, 32.0, 1.001)):
        texts = [text for text, entities in batch]
        annotations = [entities for text, entities in batch]
        print("epoch: {} Losses: {}".format(itn, str(losses)))
    # Update the model
        nlp.update(texts, annotations, drop=0.5, losses=losses, sgd=optimizer)

我正在尝试训练spaCy NER模型，并遵循spaCy训练指南。随着时间的流逝，损失应该减少，但是随着每个时期的不断增加。我尝试调整批处理大小和迭代无济于事。

示例：
时代：0损失：{}
时代：0损失：{'ner'：37.49999785423279}
时代：0损失：{'ner'：72.21390223503113}
时代：0损失：{'ner'：93.70724439620972}
时代：0失败：{'ner'：124.94790315628052}
时代：0损失：{'ner'：164.6911883354187}
时代：0损失：{'ner'：182.06093049049377}
时代：0损失：{'ner'：200.32691740989685}
时代：0失败：{'ner'：210.71145126968622}
时代：0失败：{'ner'：222.89578241482377}
时代：0损失：{'ner'：233.59122055233456}
时代：0损失：{'ner'：245.26212133839726}
时代：0损失：{'ner'：258.0684297736734}

对于该时期的最后一批，损失为11,000。感谢帮助。