RoBERTa-large-TDA-frozen
This model is a pre-trained version of roberta-large with frozen weights and a linear layer trained over [CLS]-pooled text representations on CoLA. It achieves the following results on the evaluation set:
- Loss: 0.6793
- Accuracy: 0.7400
- Mcc: 0.3172
Features extracted from Transformer
The features extracted from attention maps include the following:
Topological features are properties of attention graphs. Features of directed attention graphs include the number of strongly connected components, edges, simple cycles and average vertex degree. The properties of undirected graphs include the first two Betti numbers: the number of connected components and the number of simple cycles, the matching number and the chordality.
Features derived from barcodes include descriptive characteristics of 0/1-dimensional barcodes and reflect the survival (death and birth) of connected components and edges throughout the filtration.
Distance-to-pattern features measure the distance between attention matrices and identity matrices of pre-defined attention patterns, such as attention to the first token [CLS] and to the last [SEP] of the sequence, attention to previous and next token and to punctuation marks.
The computed features and barcodes can be found in the subdirectories of the repository. test_sub features and barcodes were computed on the out of domain test CoLA dataset. Refer to notebooks 4* and 5* from the repository to construct the classification pipeline with TDA features.
Training procedure
Training hyperparameters
Only a linear layer was trained over [CLS]-pooled text representations during training. The following hyperparameters were used during training:
- learning_rate: 2e-05
- train_batch_size: 32
- eval_batch_size: 8
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- num_epochs: 5.0
Framework versions
- Transformers 4.27.0.dev0
- Pytorch 1.13.1+cu116
- Datasets 2.9.0
- Tokenizers 0.13.2
- Downloads last month
- 4