BART

BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension

See https://arxiv.org/abs/1910.13461.

The BART agent can be instantiated as simply -m bart, however it is recommended to specify --init-model zoo:bart/bart_large/model or -mf zoo:bart/bart_large/model to ensure correct dictionaries are saved.

Basic Examples

Train BART on convai2.

parlai train_model -m bart -mf /tmp/model_file -t convai2 -bs 24 --fp16 true -eps 1 -lr 1e-5 --optimizer adam

or

parlai train_model -m bart --init-model zoo:bart/bart_large/model -mf /tmp/model_file -t convai2 -bs 24 --fp16 true -eps 1 -lr 1e-5 --optimizer adam

Interact with a BART Model fine-tuned in fairseq

parlai interactive -m bart --init-fairseq-model /path/to/fairseq/model

BartAgent Options

optional arguments

Argument	Description
--gpu-beam-blocking	Set to use CUDA kernel for beam search ngram blocking Default: False.
--verbose-topk	Return the topk logits in the act message, if verbose mode is set. Default: -1.

Transformer Arguments

Argument	Description
--embedding-size, --esz	Size of all embedding layers. Must be a multiple of –n-heads. Default: 1024.
--n-layers, --nl	Number of transformer layers. Default: 2.
--ffn-size, --hid	Hidden size of the FFN layers Default: 4096.
--dropout	Dropout used around embeddings and before layer layer normalizations. This is used in Vaswani 2017 and works well on large datasets. Default: 0.1.
--attention-dropout	Dropout used after attention softmax. This is not used in Vaswani 2017. Default: 0.1.
--relu-dropout	Dropout used after the ReLU in the FFN. Not used in Vaswani 2017, but used in Tensor2Tensor. Default: 0.0.
--n-heads	Number of multihead attention heads Default: 16.
--learn-positional-embeddings	If off, sinusoidal embeddings are used. If on, position embeddings are learned from scratch. Default: True.
--embeddings-scale	Default: False.
--n-segments	The number of segments that support the model. If zero no segment and no langs_embedding. Default: 0.
--variant	Chooses locations of layer norms, etc. prelayernorm is used to match some fairseq models Choices: xlm, prelayernorm, bart, aiayn. Default: bart. Recommended: xlm.
--activation	Nonlinear activation to use. AIAYN uses relu, but more recent papers prefer gelu. Choices: gelu, relu. Default: gelu. Recommended: gelu.
--output-scaling	Scale the output of every transformer by this quantity. Default: 1.0.
--share-word-embeddings	Share word embeddings table for candidate and contextin the memory network Default: True.
--n-encoder-layers, --nel	This will overidde the n-layers for asymmetrical transformers Default: 12.
--n-decoder-layers, --ndl	This will overidde the n-layers for asymmetrical transformers Default: 12.
--model-parallel	Shard the layers across multiple GPUs. Default: False.
--checkpoint-activations	Recompute activations on backward pass to conserve memory. Default: False.

Torch Generator Agent

Argument	Description
--beam-size	Beam size, if 1 then greedy search Default: 1.
--beam-min-length	Minimum length of prediction to be generated by the beam search Default: 1.
--beam-context-block-ngram	Size n-grams to block in beam search from the context. val <= 0 implies no blocking Default: -1.
--beam-block-ngram	Size n-grams to block in beam search. val <= 0 implies no blocking Default: -1.
--beam-block-full-context	Block n-grams from the full history context. Specify False to block up to m tokens in the past, where m is truncation parameter for agent Default: True.
--beam-length-penalty	Applies a length penalty. Set to 0 for no penalty. Default: 0.65.
--inference	Generation algorithm Choices: beam, nucleus, delayedbeam, greedy, delayednucleusbeam, topk, factual_nucleus. Default: greedy.
--topk	K used in Top K sampling Default: 10.
--topp	P used in nucleus sampling Default: 0.9.
--beam-delay	Used in delayedbeam search Default: 30.
--lambda-decay	Decay factor in factual nucleus sampling Default: 0.9.
--omega-bound	Lower bound in factual nucleus sampling Default: 0.3.
--p-reset	Whether to reset p value in factual nucleus at full stops Default: True.
--beam-block-list-filename	Load a text file of hard blocks for beam search to never say.
--temperature	Temperature to add during decoding Default: 1.0.
--compute-tokenized-bleu	If true, compute tokenized bleu scores Default: False.

TorchAgent Arguments

Argument	Description
--interactive-mode, --i	Whether in full interactive mode or not, which means generating text or retrieving from a full set of candidates, which is necessary to actually do full dialogue. However, during training or quick validation (e.g. PPL for generation or ranking a few candidates for ranking models) you might want these set to off. Typically, scripts can set their preferred default behavior at the start, e.g. eval scripts. Default: False.
--embedding-type, --emb	Choose between different strategies for initializing word embeddings. Default is random, but can also preinitialize from Glove or Fasttext. Preinitialized embeddings can also be fixed so they are not updated during training. Choices: random, glove, glove-fixed, fasttext, fasttext-fixed, fasttext_cc, fasttext_cc-fixed. Default: random.
--embedding-projection, --embp	If pretrained embeddings have a different dimensionality than your embedding size, strategy for projecting to the correct size. If the dimensions are the same, this is ignored unless you append “-force” to your choice. Default: random.
--fp16	Use fp16 computations. Default: True.
--fp16-impl	Implementation of FP16 to use Choices: safe, mem_efficient. Default: safe.
--rank-candidates, --rc	Whether the model should parse candidates for ranking. Default: False.
--truncate, --tr	Truncate input lengths to increase speed / use less memory. Default: -1.
--text-truncate	Text input truncation length: if not specified, this will default to truncate
--label-truncate	Label truncation length: if not specified, this will default to truncate
--history-reversed	Reverse the history Default: False.
--history-size, --histsz	Number of past dialog utterances to remember. Default: -1.
--person-tokens, --pt	Add person tokens to history. adds p1 in front of input text and p2 in front of past labels when available or past utterances generated by the model. these are added to the dictionary during initialization. Default: False.
--split-lines	Split the dialogue history on newlines and save in separate vectors Default: False.
--delimiter	Join history lines with this token, defaults to newline Default: \n.
--special-tok-lst	Comma separated list of special tokens. In case of ambiguous parses from special tokens, the ordering provided in this arg sets precedence.
-gpu, --gpu	Which GPU to use Default: -1.
--no-cuda	Disable GPUs even if available. otherwise, will use GPUs if available on the device. Default: False.

Optimizer Arguments

Argument	Description
--optimizer, --opt	Optimizer choice. Possible values: adadelta, adagrad, adam, adamw, sparseadam, adamax, asgd, sgd, radam, rprop, rmsprop, optimizer, nadam, lbfgs, mem_eff_adam, adafactor. Choices: adadelta, adagrad, adam, adamw, sparseadam, adamax, asgd, sgd, radam, rprop, rmsprop, optimizer, nadam, lbfgs, mem_eff_adam, adafactor. Default: sgd.
--learningrate, --lr	Learning rate Default: 1.
--gradient-clip, --clip	Gradient clipping using l2 norm Default: 0.1.
--adafactor-eps	Epsilon values for adafactor optimizer: regularization constants for square gradient and parameter scale respectively Default: 1e-30,1e-3. Recommended: 1e-30,1e-3.
--momentum, --mom	If applicable, momentum value for optimizer. Default: 0.
--nesterov	If applicable, whether to use nesterov momentum. Default: True.
--nus, --nu	If applicable, nu value(s) for optimizer. can use a single value like 0.7 or a comma-separated tuple like 0.7,1.0 Default: 0.7.
--betas, --beta	If applicable, beta value(s) for optimizer. can use a single value like 0.9 or a comma-separated tuple like 0.9,0.999 Default: 0.9,0.999.
--weight-decay, --wdecay	Weight decay on the weights.

BPEHelper Arguments

Argument	Description
--bpe-vocab	Path to pre-trained tokenizer vocab
--bpe-merge	Path to pre-trained tokenizer merge
--bpe-dropout	Use BPE dropout during training.

Learning Rate Scheduler

Argument	Description
--lr-scheduler	Learning rate scheduler. Choices: reduceonplateau, none, fixed, invsqrt, cosine, linear. Default: reduceonplateau.
--lr-scheduler-patience	LR scheduler patience. In number of validation runs. If using fixed scheduler, LR is decayed every validations. Default: 3.
--lr-scheduler-decay	Decay factor for LR scheduler, or how much LR is multiplied by when it is lowered. Default: 0.5.
--invsqrt-lr-decay-gamma	Constant used only to find the lr multiplier for the invsqrt scheduler. Must be set for –lr-scheduler invsqrt Default: -1.

Bart Args

Argument	Description
--init-fairseq-model	Fairseq checkpoint for bart
--output-conversion-path	Where to save fairseq conversion

TransformerGeneratorAgent Options

optional arguments

Argument	Description
--gpu-beam-blocking	Set to use CUDA kernel for beam search ngram blocking Default: False.
--verbose-topk	Return the topk logits in the act message, if verbose mode is set. Default: -1.

Transformer Arguments

Argument	Description
--embedding-size, --esz	Size of all embedding layers. Must be a multiple of –n-heads. Default: 300.
--n-layers, --nl	Number of transformer layers. Default: 2.
--ffn-size, --hid	Hidden size of the FFN layers Default: 300.
--dropout	Dropout used around embeddings and before layer layer normalizations. This is used in Vaswani 2017 and works well on large datasets. Default: 0.0.
--attention-dropout	Dropout used after attention softmax. This is not used in Vaswani 2017. Default: 0.0.
--relu-dropout	Dropout used after the ReLU in the FFN. Not used in Vaswani 2017, but used in Tensor2Tensor. Default: 0.0.
--n-heads	Number of multihead attention heads Default: 2.
--learn-positional-embeddings	If off, sinusoidal embeddings are used. If on, position embeddings are learned from scratch. Default: False.
--embeddings-scale	Default: True.
--n-segments	The number of segments that support the model. If zero no segment and no langs_embedding. Default: 0.
--variant	Chooses locations of layer norms, etc. prelayernorm is used to match some fairseq models Choices: xlm, prelayernorm, bart, aiayn. Default: aiayn. Recommended: xlm.
--activation	Nonlinear activation to use. AIAYN uses relu, but more recent papers prefer gelu. Choices: gelu, relu. Default: relu. Recommended: gelu.
--output-scaling	Scale the output of every transformer by this quantity. Default: 1.0.
--share-word-embeddings	Share word embeddings table for candidate and contextin the memory network Default: True.
--n-encoder-layers, --nel	This will overidde the n-layers for asymmetrical transformers Default: -1.
--n-decoder-layers, --ndl	This will overidde the n-layers for asymmetrical transformers Default: -1.
--model-parallel	Shard the layers across multiple GPUs. Default: False.
--checkpoint-activations	Recompute activations on backward pass to conserve memory. Default: False.

Torch Generator Agent

Argument	Description
--beam-size	Beam size, if 1 then greedy search Default: 1.
--beam-min-length	Minimum length of prediction to be generated by the beam search Default: 1.
--beam-context-block-ngram	Size n-grams to block in beam search from the context. val <= 0 implies no blocking Default: -1.
--beam-block-ngram	Size n-grams to block in beam search. val <= 0 implies no blocking Default: -1.
--beam-block-full-context	Block n-grams from the full history context. Specify False to block up to m tokens in the past, where m is truncation parameter for agent Default: True.
--beam-length-penalty	Applies a length penalty. Set to 0 for no penalty. Default: 0.65.
--inference	Generation algorithm Choices: beam, nucleus, delayedbeam, greedy, delayednucleusbeam, topk, factual_nucleus. Default: greedy.
--topk	K used in Top K sampling Default: 10.
--topp	P used in nucleus sampling Default: 0.9.
--beam-delay	Used in delayedbeam search Default: 30.
--lambda-decay	Decay factor in factual nucleus sampling Default: 0.9.
--omega-bound	Lower bound in factual nucleus sampling Default: 0.3.
--p-reset	Whether to reset p value in factual nucleus at full stops Default: True.
--beam-block-list-filename	Load a text file of hard blocks for beam search to never say.
--temperature	Temperature to add during decoding Default: 1.0.
--compute-tokenized-bleu	If true, compute tokenized bleu scores Default: False.

TorchAgent Arguments

Argument	Description
--interactive-mode, --i	Whether in full interactive mode or not, which means generating text or retrieving from a full set of candidates, which is necessary to actually do full dialogue. However, during training or quick validation (e.g. PPL for generation or ranking a few candidates for ranking models) you might want these set to off. Typically, scripts can set their preferred default behavior at the start, e.g. eval scripts. Default: False.
--embedding-type, --emb	Choose between different strategies for initializing word embeddings. Default is random, but can also preinitialize from Glove or Fasttext. Preinitialized embeddings can also be fixed so they are not updated during training. Choices: random, glove, glove-fixed, fasttext, fasttext-fixed, fasttext_cc, fasttext_cc-fixed. Default: random.
--embedding-projection, --embp	If pretrained embeddings have a different dimensionality than your embedding size, strategy for projecting to the correct size. If the dimensions are the same, this is ignored unless you append “-force” to your choice. Default: random.
--fp16	Use fp16 computations. Default: False.
--fp16-impl	Implementation of FP16 to use Choices: safe, mem_efficient. Default: safe.
--rank-candidates, --rc	Whether the model should parse candidates for ranking. Default: False.
--truncate, --tr	Truncate input lengths to increase speed / use less memory. Default: -1.
--text-truncate	Text input truncation length: if not specified, this will default to truncate
--label-truncate	Label truncation length: if not specified, this will default to truncate
--history-reversed	Reverse the history Default: False.
--history-size, --histsz	Number of past dialog utterances to remember. Default: -1.
--person-tokens, --pt	Add person tokens to history. adds p1 in front of input text and p2 in front of past labels when available or past utterances generated by the model. these are added to the dictionary during initialization. Default: False.
--split-lines	Split the dialogue history on newlines and save in separate vectors Default: False.
--delimiter	Join history lines with this token, defaults to newline Default: \n.
--special-tok-lst	Comma separated list of special tokens. In case of ambiguous parses from special tokens, the ordering provided in this arg sets precedence.
-gpu, --gpu	Which GPU to use Default: -1.
--no-cuda	Disable GPUs even if available. otherwise, will use GPUs if available on the device. Default: False.

Optimizer Arguments

Argument	Description
--optimizer, --opt	Optimizer choice. Possible values: adadelta, adagrad, adam, adamw, sparseadam, adamax, asgd, sgd, radam, rprop, rmsprop, optimizer, nadam, lbfgs, mem_eff_adam, adafactor. Choices: adadelta, adagrad, adam, adamw, sparseadam, adamax, asgd, sgd, radam, rprop, rmsprop, optimizer, nadam, lbfgs, mem_eff_adam, adafactor. Default: sgd.
--learningrate, --lr	Learning rate Default: 1.
--gradient-clip, --clip	Gradient clipping using l2 norm Default: 0.1.
--adafactor-eps	Epsilon values for adafactor optimizer: regularization constants for square gradient and parameter scale respectively Default: 1e-30,1e-3. Recommended: 1e-30,1e-3.
--momentum, --mom	If applicable, momentum value for optimizer. Default: 0.
--nesterov	If applicable, whether to use nesterov momentum. Default: True.
--nus, --nu	If applicable, nu value(s) for optimizer. can use a single value like 0.7 or a comma-separated tuple like 0.7,1.0 Default: 0.7.
--betas, --beta	If applicable, beta value(s) for optimizer. can use a single value like 0.9 or a comma-separated tuple like 0.9,0.999 Default: 0.9,0.999.
--weight-decay, --wdecay	Weight decay on the weights.

BPEHelper Arguments

Argument	Description
--bpe-vocab	Path to pre-trained tokenizer vocab
--bpe-merge	Path to pre-trained tokenizer merge
--bpe-dropout	Use BPE dropout during training.

Learning Rate Scheduler

Argument	Description
--lr-scheduler	Learning rate scheduler. Choices: reduceonplateau, none, fixed, invsqrt, cosine, linear. Default: reduceonplateau.
--lr-scheduler-patience	LR scheduler patience. In number of validation runs. If using fixed scheduler, LR is decayed every validations. Default: 3.
--lr-scheduler-decay	Decay factor for LR scheduler, or how much LR is multiplied by when it is lowered. Default: 0.5.
--invsqrt-lr-decay-gamma	Constant used only to find the lr multiplier for the invsqrt scheduler. Must be set for –lr-scheduler invsqrt Default: -1.