python - 新任务的 RoBERTa 微调中的 from_logits=True 与 from_logits=False

Question

我正在训练/微调一个西班牙 RoBERTa 模型，该模型最近已针对除文本分类之外的各种 NLP 任务进行了预训练。

由于基线模型似乎很有希望，我想针对不同的任务对其进行微调：文本分类，更准确地说，是西班牙推文的情感分析。
我有一个很好的选择西班牙，我可以用来微调的标签推文。

预处理没有任何问题。但是，当我训练模型时，它不会提高，即准确率不会上升。

代码：
我将省略预处理部分，因为我认为似乎没有问题。

# Training with native TensorFlow 
from transformers import TFAutoModelForSequenceClassification

## Model Definition
model = TFAutoModelForSequenceClassification.from_pretrained("BSC-TeMU/roberta-base-bne", from_pt=True, num_labels=3)

## Model Compilation
optimizer = tf.keras.optimizers.Adam(learning_rate=5e-5)
loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=False)
metric = tf.metrics.SparseCategoricalAccuracy()
model.compile(optimizer=optimizer, 
              loss=loss,
              metrics=metric) 

## Fitting the data 
history = model.fit(train_dataset.shuffle(1000).batch(64), epochs=3, batch_size=64)

输出：
为此，loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=False)我得到：

/usr/local/lib/python3.7/dist-packages/transformers/configuration_utils.py:337: UserWarning: Passing `gradient_checkpointing` to a config initialization is deprecated and will be removed in v5 Transformers. Using `model.gradient_checkpointing_enable()` instead, or if you are using the `Trainer` API, pass `gradient_checkpointing=True` in your `TrainingArguments`.
  "Passing `gradient_checkpointing` to a config initialization is deprecated and will be removed in v5 "
Some weights of the PyTorch model were not used when initializing the TF 2.0 model TFRobertaForSequenceClassification: ['roberta.embeddings.position_ids']
- This IS expected if you are initializing TFRobertaForSequenceClassification from a PyTorch model trained on another task or with another architecture (e.g. initializing a TFBertForSequenceClassification model from a BertForPreTraining model).
- This IS NOT expected if you are initializing TFRobertaForSequenceClassification from a PyTorch model that you expect to be exactly identical (e.g. initializing a TFBertForSequenceClassification model from a BertForSequenceClassification model).
Some weights or buffers of the TF 2.0 model TFRobertaForSequenceClassification were not initialized from the PyTorch model and are newly initialized: ['classifier.dense.weight', 'classifier.dense.bias', 'classifier.out_proj.weight', 'classifier.out_proj.bias']
You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
Epoch 1/5
16/16 [==============================] - 34s 1s/step - loss: 1.0986 - sparse_categorical_accuracy: 0.2734
Epoch 2/5
16/16 [==============================] - 17s 1s/step - loss: 1.0986 - sparse_categorical_accuracy: 0.2734
Epoch 3/5
16/16 [==============================] - 17s 1s/step - loss: 1.0986 - sparse_categorical_accuracy: 0.2734
Epoch 4/5
16/16 [==============================] - 17s 1s/step - loss: 1.0986 - sparse_categorical_accuracy: 0.2734
Epoch 5/5
16/16 [==============================] - 17s 1s/step - loss: 1.0986 - sparse_categorical_accuracy: 0.2734

但是，当我使用loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=False)时，我得到：

/usr/local/lib/python3.7/dist-packages/transformers/configuration_utils.py:337: UserWarning: Passing `gradient_checkpointing` to a config initialization is deprecated and will be removed in v5 Transformers. Using `model.gradient_checkpointing_enable()` instead, or if you are using the `Trainer` API, pass `gradient_checkpointing=True` in your `TrainingArguments`.
  "Passing `gradient_checkpointing` to a config initialization is deprecated and will be removed in v5 "
Some weights of the PyTorch model were not used when initializing the TF 2.0 model TFRobertaForSequenceClassification: ['roberta.embeddings.position_ids']
- This IS expected if you are initializing TFRobertaForSequenceClassification from a PyTorch model trained on another task or with another architecture (e.g. initializing a TFBertForSequenceClassification model from a BertForPreTraining model).
- This IS NOT expected if you are initializing TFRobertaForSequenceClassification from a PyTorch model that you expect to be exactly identical (e.g. initializing a TFBertForSequenceClassification model from a BertForSequenceClassification model).
Some weights or buffers of the TF 2.0 model TFRobertaForSequenceClassification were not initialized from the PyTorch model and are newly initialized: ['classifier.dense.weight', 'classifier.dense.bias', 'classifier.out_proj.weight', 'classifier.out_proj.bias']
You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
Epoch 1/5
16/16 [==============================] - 35s 1s/step - loss: 1.0184 - sparse_categorical_accuracy: 0.4778
Epoch 2/5
16/16 [==============================] - 18s 1s/step - loss: 0.6688 - sparse_categorical_accuracy: 0.7394
Epoch 3/5
16/16 [==============================] - 18s 1s/step - loss: 0.3270 - sparse_categorical_accuracy: 0.8845
Epoch 4/5
16/16 [==============================] - 18s 1s/step - loss: 0.1200 - sparse_categorical_accuracy: 0.9654
Epoch 5/5
16/16 [==============================] - 18s 1s/step - loss: 0.0500 - sparse_categorical_accuracy: 0.9872

问题：
我应该使用哪一个，即哪一个是正确的，为什么？
另外，我想知道我是否使用了正确的指标和损失。

标签： pythontensorflowkerastext-classificationhuggingface-transformers