[P2] Week 2 Day 2

今日の私の勉強目標は何ですか.

xlm-roberta-large

提出機会5回全部使用

今日私は私の学習目標を達成するために何をしましたか?

accuracy : 76.0000%

training_args = TrainingArguments(
  output_dir='./results',
  save_total_limit=5,
  save_steps=100,
  num_train_epochs=10,
  learning_rate=5e-5,
  per_device_train_batch_size=64,
  per_device_eval_batch_size=64,
  warmup_steps=300,
  weight_decay=0.01,
  logging_dir='./logs',
  logging_steps=100,
  evaluation_strategy='steps',
  eval_steps = 100,
  fp16=True,
  dataloader_num_workers=4,
  label_smoothing_factor=0.5
)

accuracy : 76.7000%

  training_args = TrainingArguments(
    output_dir='./results',
    save_total_limit=5,
    save_steps=100,
    num_train_epochs=10,
    learning_rate=5e-5,
    per_device_train_batch_size=32,
    per_device_eval_batch_size=32,
    warmup_steps=300,
    weight_decay=0.01,
    logging_dir='./logs',
    logging_steps=100,
    evaluation_strategy='steps',
    eval_steps = 100,
    fp16=True,
    dataloader_num_workers=4,
    label_smoothing_factor=0.5
  )

accuracy : 77.7000%

ensemble 3

accuracy : 79.0000%

  training_args = TrainingArguments(
    output_dir='./results',
    save_total_limit=5,
    save_steps=100,
    num_train_epochs=15,
    learning_rate=1e-5,
    per_device_train_batch_size=32,
    per_device_eval_batch_size=32,
    warmup_steps=300,
    weight_decay=0.01,
    logging_dir='./logs',
    logging_steps=100,
    evaluation_strategy='steps',
    eval_steps = 100,
    dataloader_num_workers=4,
    label_smoothing_factor=0.5
  )

accuracy : 78.7000%

  training_args = TrainingArguments(
    output_dir='./results',
    save_total_limit=5,
    save_steps=100,
    num_train_epochs=15,
    learning_rate=1e-5,
    per_device_train_batch_size=32,
    per_device_eval_batch_size=32,
    warmup_steps=300,
    weight_decay=0.01,
    logging_dir='./logs',
    logging_steps=100,
    evaluation_strategy='steps',
    eval_steps = 100,
    dataloader_num_workers=4,
    label_smoothing_factor=0.5
  )

今日はどんな方法でモデルを改善しましたか?

xlm−Roberta−LazeおよびBaseline CodeのHyperパラメータを変更することによってモデルを改良した.

明日は何か違う試みがありますか?

は明日、train dataset 100%使用ポリシーとtest datasetパーセントの統合ポリシーを試みます.

の最後の部分
今日よりもっと成长する明日を楽しみにしている私にまた明日会いましょう
読んでくれてありがとう!