Open learner-crapy opened 1 year ago
我使用了一块RTX4090分别训练了中文和英文的OntoNote数据，得到下面的结果 中文：RoBERTa_zh_L12_PyTorch
英文：spanbert_base
使用参数如下： `# Computation limits. max_top_antecedents = 50 max_training_sentences = 11 top_span_ratio = 0.4 max_num_speakers = 20 max_segment_len = 128
bert_learning_rate = 1e-05 task_learning_rate = 0.0002 adam_eps = 1e-6 dropout_rate = 0.3
num_docs = 2802 num_epochs = 30 do_train = true do_eval = true do_test = true do_one_example_test = true eval_frequency = 100 report_frequency = 10
genres = ["bc", "bn", "mz", "nw", "tc", "wb"] coref_depth = 2 ffnn_size = 3000 feature_size = 20 max_span_width = 30 use_metadata = true use_features = true use_segment_distance = true model_heads = true fine_grained = true use_prior = true single_example = true`
但是我看coref官方使用base版本在英文预料下的F1指标达到77，请问可能的问题在哪里？
我使用了一块RTX4090分别训练了中文和英文的OntoNote数据，得到下面的结果 中文：RoBERTa_zh_L12_PyTorch
英文：spanbert_base
使用参数如下： `# Computation limits. max_top_antecedents = 50 max_training_sentences = 11 top_span_ratio = 0.4 max_num_speakers = 20 max_segment_len = 128
Learning
bert_learning_rate = 1e-05 task_learning_rate = 0.0002 adam_eps = 1e-6 dropout_rate = 0.3
Task choice
num_docs = 2802 num_epochs = 30 do_train = true do_eval = true do_test = true do_one_example_test = true eval_frequency = 100 report_frequency = 10
Model hyperparameters.
genres = ["bc", "bn", "mz", "nw", "tc", "wb"] coref_depth = 2 ffnn_size = 3000 feature_size = 20 max_span_width = 30 use_metadata = true use_features = true use_segment_distance = true model_heads = true fine_grained = true use_prior = true single_example = true`
但是我看coref官方使用base版本在英文预料下的F1指标达到77，请问可能的问题在哪里？