[Bug]: index out of range in self

[lennertvandevelde]

Describe the bug

ner-dutch model throws index out of range in self error on some character combinations.

To Reproduce

Example 1

import flair
from flair.data import Sentence
from flair.models import SequenceTagger

tagger = SequenceTagger.load("flair/ner-dutch")

error_chars = []
for ichar in [chr(i) for i in range(ord('a'),ord('z')+1)]:
  for jchar in [chr(i) for i in range(ord('a'),ord('z')+1)]:
    try:
      tagger.predict(Sentence(jchar+ichar))

    except:
      error_chars.append(jchar+ichar)
print(error_chars)

OUTPUT:

['aa', 'ea', 'ia', 'oa', 'qa', 'ra', 'sa', 'ta', 'ua', 'xa', 'ya', 'cb', 'eb', 'fb', 'gb', 'hb', 'ib', 'jb', 'kb', 'lb', 'mb', 'nb', 'pb', 'qb', 'rb', 'sb', 'tb', 'ub', 'xb', 'yb', 'ec', 'fc', 'gc', 'hc', 'ic', 'jc', 'kc', 'lc', 'mc', 'nc', 'oc', 'qc', 'rc', 'tc', 'uc', 'xc', 'yc', 'ad', 'ed', 'fd', 'gd', 'hd', 'id', 'jd', 'kd', 'ld', 'md', 'nd', 'od', 'pd', 'qd', 'rd', 'sd', 'td', 'ud', 'xd', 'yd', 'ae', 'ee', 'fe', 'ie', 'ke', 'ne', 'oe', 'qe', 're', 'se', 'ue', 'xe', 'ye', 'cf', 'ef', 'ff', 'gf', 'hf', 'if', 'jf', 'kf', 'lf', 'mf', 'nf', 'pf', 'qf', 'rf', 'sf', 'tf', 'uf', 'xf', 'yf', 'ag', 'cg', 'eg', 'fg', 'gg', 'hg', 'ig', 'jg', 'kg', 'lg', 'mg', 'ng', 'og', 'pg', 'qg', 'rg', 'sg', 'tg', 'ug', 'xg', 'yg', 'ah', 'ch', 'eh', 'fh', 'gh', 'hh', 'ih', 'jh', 'kh', 'lh', 'mh', 'nh', 'oh', 'ph', 'qh', 'rh', 'sh', 'uh', 'xh', 'yh', 'ai', 'ei', 'ii', 'ji', 'ni', 'oi', 'qi', 'si', 'ti', 'xi', 'yi', 'aj', 'cj', 'ej', 'fj', 'gj', 'hj', 'ij', 'jj', 'kj', 'lj', 'mj', 'nj', 'oj', 'pj', 'qj', 'rj', 'sj', 'tj', 'uj', 'xj', 'yj', 'ak', 'ck', 'ek', 'fk', 'gk', 'hk', 'jk', 'kk', 'lk', 'mk', 'nk', 'ok', 'qk', 'rk', 'sk', 'tk', 'uk', 'xk', 'yk', 'el', 'hl', 'il', 'jl', 'll', 'ml', 'nl', 'ol', 'ql', 'rl', 'sl', 'tl', 'ul', 'xl', 'yl', 'am', 'em', 'fm', 'gm', 'hm', 'im', 'jm', 'lm', 'nm', 'pm', 'qm', 'rm', 'sm', 'tm', 'um', 'xm', 'ym', 'an', 'cn', 'fn', 'gn', 'hn', 'jn', 'kn', 'ln', 'mn', 'nn', 'pn', 'qn', 'rn', 'sn', 'tn', 'un', 'xn', 'yn', 'ao', 'eo', 'io', 'oo', 'qo', 'uo', 'xo', 'yo', 'ap', 'cp', 'ep', 'fp', 'gp', 'hp', 'ip', 'jp', 'kp', 'lp', 'mp', 'np', 'pp', 'qp', 'rp', 'tp', 'xp', 'yp', 'aq', 'cq', 'eq', 'fq', 'gq', 'hq', 'iq', 'jq', 'kq', 'lq', 'mq', 'nq', 'oq', 'pq', 'qq', 'rq', 'sq', 'tq', 'uq', 'xq', 'yq', 'ar', 'hr', 'ir', 'jr', 'lr', 'mr', 'nr', 'or', 'qr', 'rr', 'sr', 'tr', 'ur', 'xr', 'yr', 'as', 'cs', 'es', 'fs', 'gs', 'hs', 'js', 'ks', 'ls', 'ms', 'ns', 'os', 'ps', 'qs', 'rs', 'ss', 'ts', 'us', 'xs', 'ys', 'at', 'ct', 'et', 'ft', 'gt', 'ht', 'it', 'jt', 'kt', 'lt', 'mt', 'nt', 'ot', 'pt', 'qt', 'rt', 'tt', 'ut', 'xt', 'yt', 'cu', 'eu', 'iu', 'ou', 'tu', 'uu', 'xu', 'yu', 'av', 'ev', 'fv', 'gv', 'hv', 'iv', 'jv', 'kv', 'lv', 'mv', 'nv', 'ov', 'pv', 'qv', 'rv', 'sv', 'uv', 'xv', 'yv', 'aw', 'cw', 'ew', 'fw', 'gw', 'hw', 'iw', 'jw', 'lw', 'mw', 'nw', 'ow', 'pw', 'qw', 'rw', 'sw', 'tw', 'xw', 'yw', 'ax', 'cx', 'fx', 'gx', 'hx', 'ix', 'jx', 'kx', 'lx', 'mx', 'nx', 'ox', 'px', 'qx', 'rx', 'sx', 'tx', 'ux', 'xx', 'yx', 'ay', 'cy', 'ey', 'fy', 'gy', 'iy', 'jy', 'ky', 'ly', 'ny', 'oy', 'qy', 'ry', 'sy', 'ty', 'uy', 'xy', 'yy', 'az', 'cz', 'ez', 'fz', 'gz', 'hz', 'iz', 'jz', 'kz', 'lz', 'mz', 'nz', 'oz', 'pz', 'qz', 'rz', 'sz', 'tz', 'uz', 'xz', 'yz']

Example 2

tagger.predict(Sentence("eedaflegging"))

OUTPUT

IndexError: index out of range in self

Expected behaivor

Expected the pipeline to return entities.

Logs and Stack traces

No response

Screenshots

No response

Additional Context

Same Issue as: https://github.com/flairNLP/flair/issues/2813

Environment

Versions:

Flair

0.12

Pytorch

1.13.1+cu116

Transformers

4.26.1

GPU

True

[helpmefindaname]

Hi @lennertvandevelde looks like this happens, because the vocabulary of the bertje-embeddings, was updated after release see here.

@alanakbik I suppose we should just retrain the model.

In the meantime you can just use a hotfix by adding the embeddings for the tokens that wer added afterwards using the following code:

from flair.models import SequenceTagger
from flair.embeddings import TransformerWordEmbeddings
from torch import nn
import torch

tagger = SequenceTagger.load("flair/ner-dutch")
embeddings = TransformerWordEmbeddings('GroNLP/bert-base-dutch-cased')
new_embedding_tensor = torch.cat([tagger.embeddings.model.get_input_embeddings().weight, embeddings.model.get_input_embeddings().weight[tagger.embeddings.model.get_input_embeddings().num_embeddings:-1]])
new_input_embeddings = nn.Embedding.from_pretrained(new_embedding_tensor , freeze=False)
tagger.embeddings.model.set_input_embeddings(new_input_embeddings)
tagger.embeddings.base_model_name="GroNLP/bert-base-dutch-cased"
tagger.save("ner-dutch-fixed.pt")

[Nickyboo1194]

#nDD v nevzvvbqlekvl m s das Z umm Vvelvet

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.