EntityLinker: Cannot get dimension 'nO' for model 'transformer-listener>>trfs2arrays

[adamkgoldfarb]

Hi @svlandeg -- I'm attempting to train an entity linker with a transformer component as a drop-in replacement for the tok2vec layer. I am getting the error:

ValueError: Cannot get dimension 'nO' for model 'transformer-listener>>trfs2arrays': value unset

I checked through this issue but I don't think it's directly related-- happy to be wrong!

I'm likely missing something in the config, but I've tried setting nO to 768 rather than null where it's noted in the config and am still getting that error. Here's the config-- any insight would be super appreciated, and thanks in advance!

---

[paths]
train = "training/lea_nel_train/lea_nel_train.spacy"
dev = "training/lea_nel_train/lea_nel_dev.spacy"
vectors = null
init_tok2vec = null
raw_text = "corpus/rfp_free_responses.jsonl"
kb = "knowledge_bases/lea_entities_roberta_trf"
base_nlp = "models/lea_roberta_base_trf/model-best"

[system]
gpu_allocator = "pytorch"
seed = 30

[nlp]
lang = "en"
pipeline = ["transformer","ner","entity_linker"]
batch_size = 128
disabled = []
before_creation = null
after_creation = null
after_pipeline_creation = null
tokenizer = {"@Tokenizers":"spacy.Tokenizer.v1"}

[components]

[components.entity_linker]
factory = "entity_linker"
entity_vector_length = 768
get_candidates = {"@misc":"spacy.CandidateGenerator.v1"}
incl_context = true
incl_prior = true
n_sents = 1
labels_discard = []

[components.entity_linker.model]
@architectures = "spacy.EntityLinker.v1"
nO = null

[components.entity_linker.model.tok2vec]
@architectures = "spacy-transformers.TransformerListener.v1"
grad_factor = 1.0
pooling = {"@layers":"reduce_mean.v1"}
upstream = "*"

[components.ner]
source = ${paths.base_nlp}

[components.transformer]
source = ${paths.base_nlp}

[corpora]

[corpora.train]
@readers = "MyCorpus.v1"
file = ${paths.train}

[corpora.dev]
@readers = "MyCorpus.v1"
file = ${paths.dev}

[training]
train_corpus = "corpora.train"
dev_corpus = "corpora.dev"
seed = ${system.seed}
gpu_allocator = ${system.gpu_allocator}
dropout = 0.2
patience = 10000
eval_frequency = 200
accumulate_gradient = 2
max_epochs = 0
max_steps = 600
frozen_components = ["transformer","ner"]
before_to_disk = null

[training.logger]
@Loggers = "spacy.ConsoleLogger.v1"
progress_bar = true

[training.batcher]
@batchers = "spacy.batch_by_words.v1"
discard_oversize = false
tolerance = 0.2
get_length = null

[training.batcher.size]
@schedules = "compounding.v1"
start = 100
stop = 1000
compound = 1.001
t = 0.0

[training.optimizer]
@optimizers = "Adam.v1"
beta1 = 0.9
beta2 = 0.999
L2_is_weight_decay = true
L2 = 0.01
grad_clip = 1.0
use_averages = false
eps = 0.00000001
learn_rate = 0.001

[training.score_weights]
nel_micro_p = 0.0
nel_micro_r = 0.0
nel_micro_f = 1.0

[pretraining]

[optimizer]
@optimizers = "Adam.v1"
learn_rate = 0.001
beta1 = 0.9
beta2 = 0.999
L2 = 0.0
eps = 0.00000001
grad_clip = 1.0
L2_is_weight_decay = true
use_averages = true

[initialize]
vectors = null
init_tok2vec = ${paths.init_tok2vec}
vocab_data = null
lookups = null
before_init = null
after_init = null

[initialize.components]

[initialize.components.entity_linker]

[initialize.components.entity_linker.kb_loader]
@misc = "spacy.KBFromFile.v1"
kb_path = ${paths.kb}

[initialize.tokenizer]

[svlandeg]

Hi!

Could you check whether this still happens if you don't put the transformer in the frozen list? If it does still happen - can you paste the full stack trace?

A related discussion & explanation by Adriane is here: #7811 (reply in thread). In short, 3.0 has some "quircks" when freezing some layers of the listener pattern and training others, like in this case, so until 3.1 we'll have to find a way around that.

[adamkgoldfarb]

This article made it all clear for me and really spells it out toward the end! You'll need to add the tensor2attr pipe to your nlp and then you can access the tensor encodings rather than the static vector encodings, because you will have overridden the .vector attr with trf tensors on the doc instance.

[moyid]

thank you @adamkgoldfarb ! That was very helpful!

[moyid]

sadly this did not solve my issue. I think I got the vector right, but I am still getting the same error. @adamkgoldfarb @svlandeg If you don't mind taking a look at my issue - #9310 , it is very similar to this setup. I'd really appreciate it!

[adamkgoldfarb]

Going to keep my reply here because I'm afraid I'm out of my depth-- one of the differences I see between our setups is the sentencizer. Are you sure you need that component? It might throw a wrench in the trf processing. I might try removing it. But I need to emphasize that's a guess!

[moyid]

Thanks, I tried it without sentencizer as well, and still got the same error :/

[svlandeg]

(reposting here to be able to select as answer): Assuming your vectors in the KB are of length 768, this fix should hopefully resolve your problem: #7930

[adamkgoldfarb]

Thank you so much for your attention to this! Looking forward to trying it out! I'll monitor the pull request and test when merged.

Much, much appreciated!

[svlandeg]

Merged :-)