get_text_classifier fails with custom AWS_LSTM

[fabridamicelli]

Please confirm you have the latest versions of fastai, fastcore, and nbdev prior to reporting a bug (delete one): YES

Describe the bug
The function get_text_classifier from the module text.models.core which takes the argument arch (eg AWD_LSTM) throws a KeyError when a user-instantiated AWD_LSTM is passed (AWD_LSTM(vocab_sz=100, emb_sz=10, n_hid=2, n_layers=2) ).
More precisely, the lookup _model_meta[arch] fails because the custom AWS_LSTM instance is not recognized as being equal to <class 'fastai.text.models.awdlstm.AWD_LSTM'> (the key of the _model_meta[arch] dictionary).

To Reproduce
Steps to reproduce the behavior:

from fastai.text.all import *

arch = AWD_LSTM(vocab_sz=100, emb_sz=10, n_hid=10, n_layers=2)
get_text_classifier(arch=arch, vocab_sz=100, n_class=2)

The error can be clearly seen in this notebook, which can be directly accessed in colab here

Expected behavior
Function should return a SequentialRNN instance.

Error with full stack trace

---------------------------------------------------------------------------
KeyError                                  Traceback (most recent call last)
Input In [13], in <cell line: 4>()
      1 from fastai.text.all import *
      3 arch = AWD_LSTM(vocab_sz=100, emb_sz=10, n_hid=10, n_layers=2)
----> 4 get_text_classifier(arch=arch, vocab_sz=100, n_class=2)

File ~/miniconda3/envs/fai/lib/python3.10/site-packages/fastai/text/models/core.py:158, in get_text_classifier(arch, vocab_sz, n_class, seq_len, config, drop_mult, lin_ftrs, ps, pad_idx, max_len, y_range)
    144 def get_text_classifier(
    145     arch:callable, # Function or class that can generate a language model architecture
    146     vocab_sz:int, # Size of the vocabulary 
   (...)
    155     y_range:tuple=None # Tuple of (low, high) output value bounds
    156 ):
    157     "Create a text classifier from `arch` and its `config`, maybe `pretrained`"
--> 158     meta = _model_meta[arch]
    159     config = ifnone(config, meta['config_clas']).copy()
    160     for k in config.keys():

KeyError: AWD_LSTM(
  (encoder): Embedding(100, 10, padding_idx=1)
  (encoder_dp): EmbeddingDropout(
    (emb): Embedding(100, 10, padding_idx=1)
  )
  (rnns): ModuleList(
    (0): WeightDropout(
      (module): LSTM(10, 10, batch_first=True)
    )
    (1): WeightDropout(
      (module): LSTM(10, 10, batch_first=True)
    )
  )
  (input_dp): RNNDropout()
  (hidden_dps): ModuleList(
    (0): RNNDropout()
    (1): RNNDropout()
  )
)

Additional context
Forum discussion with another report from @machinatoonist (with no solution so far): https://forums.fast.ai/t/how-to-customise-vocab-sz-in-text-classifier-learner/98230.

[Salehbigdeli]

There are two problem here:

1. You need to modify the way you used the API, according to docs arch need to be a class or callable creating a model (not the model itself as you gave). something like:

from fastai.text.all import *
config = awd_lstm_clas_config.copy()
config.update(emb_sz=10, n_hid=10, n_layers=2)
get_text_classifier(arch=AWD_LSTM, n_class=2, vocab_sz=100, config=config)

2. The second problem is the API itself. I don't like the part that I created config config = awd_lstm_clas_config.copy() and the next line. I'm going to create PR to solve this issue, so you can use the API like:

from fastai.text.all import *
config = dict(emb_sz=10, n_hid=10, n_layers=2)
get_text_classifier(arch=AWD_LSTM, n_class=2, vocab_sz=100, config=config)