[Bug] error in typing_metric.py about adaseq HOT 10 OPEN

Hi, could you please provide more details about this issue? e.g., the config file you run, the environments, and screenshots. These will help us find the problem, thanks.

from adaseq.

Please see below for the config file (only the source and target emb files are changed according to the instruction). I'm not sure what information I should provide regarding the environment (except pip freeze above) and screenshots? Thanks.

config file:
experiment:
exp_dir: experiments/
exp_name: ufet
seed: 17

task: entity-typing

dataset:
data_file:
train: 'https://www.modelscope.cn/api/v1/datasets/izhx404/ufet/repo/files?Revision=master&FilePath=train.json'
valid: 'https://www.modelscope.cn/api/v1/datasets/izhx404/ufet/repo/files?Revision=master&FilePath=dev.json'
test: 'https://www.modelscope.cn/api/v1/datasets/izhx404/ufet/repo/files?Revision=master&FilePath=test.json'
tokenizer: blank
lower: true
labels: 'https://www.modelscope.cn/api/v1/datasets/izhx404/ufet/repo/files?Revision=master&FilePath=labels.txt'

preprocessor:
type: multilabel-concat-typing-preprocessor
model_dir: roberta-large
max_length: 150

data_collator: MultiLabelConcatTypingDataCollatorWithPadding

model:
type: multilabel-concat-typing-model
embedder:
model_name_or_path: roberta-large
drop_special_tokens: false
dropout: 0
decoder:
type: pairwise-crf
label_emb_type: glove
label_emb_dim: 300
source_emb_file_path: None
target_emb_dir: /home/averie/name-entity-recognition/experiments/adaseq/glove_embeds # TODO
target_emb_name: glove.300.emb
pairwise_factor: 70
mfvi_iteration: 4
two_potential: false
sign_trick: true
loss_function: WBCE
pos_weight: 4

train:
max_epochs: 30
dataloader:
batch_size_per_gpu: 4
optimizer:
type: AdamW
lr: 2.0e-5
lr_scheduler:
type: cosine
warmup_rate: 0.1 # when choose concat typing model, default to use cosine_linear_with_warmup
options:
by_epoch: false
hooks:
- type: "CheckpointHook"
interval: 100
- type: "BestCkptSaverHook"
save_file_name: "best_model.pt"

evaluation:
dataloader:
batch_size_per_gpu: 32
metrics: typing-metric

from adaseq.

could you successfully run the default npcrf example?

from adaseq.

how do i run the default example? it requires PATH_TO_DIR to be replaced, which is what I did.

from adaseq.

decoder:
  type: pairwise-crf
  label_emb_type: glove
  label_emb_dim: 300
  source_emb_file_path: ${PATH_TO_DIR}/glove.6B.300d.txt
  target_emb_dir: ${PATH_TO_DIR}  # TODO
  target_emb_name: glove.300.emb
  pairwise_factor: 70
  mfvi_iteration: 4
  two_potential: false
  sign_trick: true

It seems that your configuration is incorrect, the above shows the default configuration. The glove path can be downloaded from the official stanford website: https://nlp.stanford.edu/data/glove.6B.zip.
The source_emb_file_path should be the absolute path to for example the glove.6B.300d.txt, and the target_emb_dir, is the directory that you want to store the label embedding matrix named with target_emb_name.
In other word, the label embedding is preprocessed from ${YOUR_SRC_EMB_DIR}/glove.6B.300d.txt, and saved to ${YOUR_TGT_EMB_SAVE_DIR}/glove.300.emb

from adaseq.

thank you for your answer. according to the readme in the NPCRF directory:

"NPCRF requires static label embeddings, the preprocessed label embeddings (from GloVe for EN, Tencent for ZH) can be downloaded here: UFET, CFET, and you can place them in yoru folder and run the following config: (you need to reset your target_emb_dir in the config). Or you can provide the path of the glove embedding file (e.g., /path/to/your/glove.6B.300d.txt) and the code will generate label embedding for you."

so i cannot use the glove.300.emb given in this description?

from adaseq.

Could you please give a screenshot of the error message?
And, can you successfully run the model when you create embedding from the glove source?

from adaseq.

I just tried the glove source, it resulted in the same error (see screenshot)

from adaseq.

oops, it seems that the bug is caused by the latest update of adaseq in the typing metric. No problem occurs in the training and loading label embeddings, a quick fix could be downgrading the adaseq to 0.6.2 and modelscope to 1.4.2.
We will fix the bug later.

from adaseq.

thanks for the reply. i did pip install adaseq==0.6.2 and pip install modelscope==1.4.2 but still getting the same error. am I missing something?

from adaseq.