SpanCategorizer and Transformers #10059
-
|
Hello! I have started from the Indonesian spancat example project, and want to replace the default Thanks and regards, |
Beta Was this translation helpful? Give feedback.
Replies: 2 comments 4 replies
-
|
I was going to suggest using [paths]
train = null
dev = null
vectors = null
init_tok2vec = null
[system]
gpu_allocator = "pytorch"
seed = 0
[nlp]
lang = "id"
pipeline = ["transformer","spancat"]
batch_size = 128
disabled = []
before_creation = null
after_creation = null
after_pipeline_creation = null
tokenizer = {"@tokenizers":"spacy.Tokenizer.v1"}
[components]
[components.spancat]
factory = "spancat"
max_positive = null
scorer = {"@scorers":"spacy.spancat_scorer.v1"}
spans_key = "grit"
threshold = 0.5
[components.spancat.model]
@architectures = "spacy.SpanCategorizer.v1"
[components.spancat.model.reducer]
@layers = "spacy.mean_max_reducer.v1"
hidden_size = 128
[components.spancat.model.scorer]
@layers = "spacy.LinearLogistic.v1"
nO = null
nI = null
[components.spancat.model.tok2vec]
@architectures = "spacy-transformers.TransformerListener.v1"
grad_factor = 1.0
pooling = {"@layers":"reduce_mean.v1"}
upstream = "transformer"
[components.spancat.suggester]
@misc = "spacy.ngram_suggester.v1"
sizes = [1,2,3]
[components.transformer]
factory = "transformer"
max_batch_items = 4096
set_extra_annotations = {"@annotation_setters":"spacy-transformers.null_annotation_setter.v1"}
[components.transformer.model]
@architectures = "spacy-transformers.TransformerModel.v3"
name = "indolem/indobert-base-uncased"
mixed_precision = false
[components.transformer.model.get_spans]
@span_getters = "spacy-transformers.strided_spans.v1"
window = 128
stride = 96
[components.transformer.model.grad_scaler_config]
[components.transformer.model.tokenizer_config]
use_fast = true
[components.transformer.model.transformer_config]
[corpora]
[corpora.dev]
@readers = "spacy.Corpus.v1"
path = ${paths.dev}
max_length = 0
gold_preproc = false
limit = 0
augmenter = null
[corpora.train]
@readers = "spacy.Corpus.v1"
path = ${paths.train}
max_length = 0
gold_preproc = false
limit = 0
augmenter = null
[training]
accumulate_gradient = 3
dev_corpus = "corpora.dev"
train_corpus = "corpora.train"
seed = ${system.seed}
gpu_allocator = ${system.gpu_allocator}
dropout = 0.1
patience = 1600
max_epochs = 0
max_steps = 20000
eval_frequency = 200
frozen_components = []
annotating_components = []
before_to_disk = null
[training.batcher]
@batchers = "spacy.batch_by_padded.v1"
discard_oversize = true
size = 2000
buffer = 256
get_length = null
[training.logger]
@loggers = "spacy.ConsoleLogger.v1"
progress_bar = false
[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
[training.optimizer.learn_rate]
@schedules = "warmup_linear.v1"
warmup_steps = 250
total_steps = 20000
initial_rate = 0.00005
[training.score_weights]
spans_sc_f = null
spans_sc_p = null
spans_sc_r = null
spans_grit_f = 1.0
spans_grit_p = 0.0
spans_grit_r = 0.0
[pretraining]
[initialize]
vectors = ${paths.vectors}
init_tok2vec = ${paths.init_tok2vec}
vocab_data = null
lookups = null
before_init = null
after_init = null
[initialize.components]
[initialize.tokenizer] |
Beta Was this translation helpful? Give feedback.
-
|
Thanks for the prompt reply, it works like a charm! What is the motivation of reducing the batch_size from the original 1000 in the spancat example project to 128? Is there any downside of training with larger batch_sizes, other than GPU memory usage? |
Beta Was this translation helpful? Give feedback.
-
|
GPU memory usage is the main reason. In practice it can depend a lot on text length, so you'll have to experiment a bit depending on your corpus and GPU. Note that that is only the eval batch size. The training batch size in is |
Beta Was this translation helpful? Give feedback.
-
|
I see, thanks for the clarification :) |
Beta Was this translation helpful? Give feedback.
-
|
Hi @adrianeboyd I tried the above configuration for spancat on transformer as the base but while training i get all 0's for all precision, recall and F1. Below is my spancat tok2vec configuration which i want to switch to transformers as we have other component in assembled pipeline that depends on trf. [system] [nlp] [components] [components.spancat] [components.spancat.model] [components.spancat.model.reducer] [components.spancat.model.scorer] [components.spancat.model.tok2vec] [components.spancat.model.tok2vec.embed] [components.spancat.model.tok2vec.encode] [components.spancat.suggester] [corpora] [corpora.dev] [corpora.train] [training] [training.batcher] [training.batcher.size] [training.logger] [training.optimizer] [training.score_weights] [pretraining] [initialize] [initialize.components] [initialize.tokenizer]` Here is what i tried to use for transformers: [system] [nlp] [components] [components.spancat] [components.spancat.model] [components.spancat.model.reducer] [components.spancat.model.scorer] [components.spancat.model.tok2vec] [components.spancat.suggester] [components.transformer] [components.transformer.model] [components.transformer.model.get_spans] [components.transformer.model.grad_scaler_config] [components.transformer.model.tokenizer_config] [components.transformer.model.transformer_config] [corpora] [corpora.dev] [corpora.train] [training] [training.batcher] [training.logger] [training.optimizer] [training.optimizer.learn_rate] [training.score_weights] [pretraining] [initialize] [initialize.components] [initialize.tokenizer] |
Beta Was this translation helpful? Give feedback.
-
|
@adrianeboyd Nevermind, figured out I needed to update training.score_weights in the config. Thanks |
Beta Was this translation helpful? Give feedback.
I was going to suggest using
init config, but I see that it doesn't have a default transformer option forspancatyet (sincespancatis still experimental, it's not as integrated into the defaults). A transformer config for thener_spancatdemo project could look like this: