Skip to content

baselines.md

Latest commit

 

History

History
58 lines (38 loc) · 3.46 KB

baselines.md

File metadata and controls

58 lines (38 loc) · 3.46 KB

Full List of Implemented Baselines

We introduce all supported baselines and backbones.

  • Paper: the model name and its reference paper
  • baseline: the baseline name (argument --baseline)
    • ncl: Naive Continual Learning, meaning no special continual learning is applied by default.
    • one: Trains each task independently (ONE model per task)
    • mtl: Trains all task together via Multi-Task Learning
  • backbone: the supported backbones for the corresponding baseline (argument --backbone)
    • bert: use BERT (fine-tuning) as backbone network
    • bert_frozen: use BERT (frozen) as backbone netowok (a fronzen BERT + a CNN/GRU classification network on top)
    • bert_adapter: use BERT (adapter) as backbone network
    • w2v: use W2V as backbone network (W2V embedding + a CNN classification network on top)
    • w2v_as (specific for Aspect Sentiment Classification): use W2V as backbone network (W2V embedding which includes both review and the aspect + a CNN classification network on top)
    • cnn: use CNN as backbone network (for image dataset only)
    • mlp: use MLP as backbone network (for image dataset only)
  • task: the supported task for the corresponding baseline (argument --task)
    • language dataset includes asc/dsc/ssc/nli/newsgroup'
    • image dataset includes celeba/femnist/vlcs/cifar10/mnist/fashionmnist/cifar100
Paper Baseline Backbone task
NCL ncl bert/bert_frozen/bert_adapter/w2v_as/w2v language, image
ONE one bert/bert_frozen/bert_adapter/w2v_as/w2v language, image
MTL mtl bert/bert_frozen/bert_adapter/w2v_as/w2v language, image
L2 l2 bert/bert_frozen/bert_adapter/w2v_as/w2v language, image
A-GEM a-gem bert/bert_frozen/bert_adapter/w2v_as/w2v/cnn/mlp language, image
DER++ derpp bert/bert_frozen/bert_adapter/w2v_as/w2v/cnn/mlp language
KAN kan bert_frozen/w2v_as/w2v language
SRK srk bert_frozen/w2v_as/w2v language, image
EWC ewc bert/bert_frozen/bert_adapter/w2v_as/w2v/cnn/mlp language, image
HAL hal bert/bert_frozen/bert_adapter/w2v_as/w2v/cnn/mlp language, image
UCL ucl bert/bert_frozen/bert_adapter/w2v_as/w2v/cnn/mlp language, image
OWM owm bert/bert_frozen/bert_adapter/w2v_as/w2v/cnn/mlp language, image
ACL acl cnn/mlp image
HAT hat bert/bert_frozen/bert_adapter/w2v_as/w2v/cnn/mlp language, image
CAT cat bert/bert_frozen/w2v_as/w2v/cnn/mlp language, image
B-CL b-cl bert_adapter language
CLASSIC classic bert_adapter language
CTR ctr bert_adapter language