mindspore-lab / mindnlp Goto Github PK

fail to run "python setup.py bdist_wheel"; error information is as folloews:
.....
running install_scripts
error: [WinError 5] 拒绝访问。: 'build\bdist.win-amd64\wheel\mindnlp-0.0.1.dist-info\dependency_links.txt'

When using the squad dataset and bidaf model to train the QA task (question_answer.py), the training progress always maintains at 0% because of the slow speed of nn.LSTM.

To solve the slow speed of nn.LSTM (issue), when I'm training the model I change nn.LSTM into LSTM written by lvyufeng. However, an error occured when training for a period time. (And I once used the validation set of squad dataset for training for a try, which can trains successfully without any problem)
Here is the error message:

"Trainer" does not take into account the case where "loss" is already defined in the model, and there is no need to pass "loss_fn" to "Trainer".

"trainer" does not support Bi-LSTM+CRF model (custom loss is in the model layer)

BiDAF Model will generate two outputs. Error occured when Trainer was called.
Parts of the code:

Error information:

Using "TextBaseDataset.build_vocab" to construct "vocab", and then using "text.Lookup(vocab)" will report an error.

Update build version from 0.0.1 to 0.1.1.

Using the CoNLL2000Chunking dataset and the Bi-LSTM+CRF model to train the Chunking task loss will appear "nan".

in GRAPH_MODE， CRF compilation is too slow if sequence length is larger than 100

Some Datasets have their own embedding tables which are npy tpye files. However, the loading of npy files is missing in the "embedding" API.