[BUG] AttributeError: 'tuple' object has no attribute 'layer'

[mahdisnapp]

first of all thanks for your impressive work and effort. I faced a problem when using your package in google colab. while I want to use bert embedding with this command:
bert_embedding = BERTEmbedding(bert_model_path,
task=kashgari.CLASSIFICATION,
sequence_length=128)
I got this error:
AttributeError: 'tuple' object has no attribute 'layer'
thanks in advance for your help.

[BrikerMan]

Are you working on a multi-label classification task? I might there is a small issue in the documents.

[BrikerMan]

Long story short, if you import keras_bert from setting the environment variable, it will cause this issue. You can fix it by importing Kashgari at the top or add the two lines at the top. Basically, remember to import kashgari first.

import os
os.environ['TF_KERAS'] = '1'

[mahdisnapp]

thanks for response, to answer your question, I am working on multi-class classification not multi-label classification,
base on your suggestion I will import kashgari first then import keras_bert and let you know if it is fixed.
Important note: I want to use tokenizer from keras_bert.
again thanks .

[BrikerMan]

I am working on to make it easier to access the keras_bert tokenizer. Will release 5.0.1 today. Check out this commit 22ad76f

[mahdisnapp]

thanks for your help, problem has been solved. I noticed a note in your tutorial for BERT embeddings, you tokenized sentences with BERT tokenizer but did not pass them into your neural net model, you just feed the raw text as input data.
tokenizer = Tokenizer(token_dict) sentences_tokenized = [] for sentence in sentences: sentence_tokenized = tokenizer.tokenize(sentence) sentences_tokenized.append(sentence_tokenized)

but then
train_x, train_y = sentences[:2], labels[:2] validate_x, validate_y = sentences[2:], labels[2:]

[BrikerMan]

Fixed the typo and released the 0.5.1 version, it is much easier to use BERT tokenizer. Check it out, document

[mahdisnapp]

dear BrikerMan
now we have other main error, after update, we got this error
tokenizer = bert_embed.tokenizer
'BERTEmbedding' object has no attribute 'tokenizer'

[BrikerMan]

You have updated to 0.5.1 version? mahdisnapp <notifications@github.com>于2019年7月15日 周一19:42写道：

…

dear BrikerMan now we have other main error, after update, we got this error tokenizer = bert_embed.tokenizer 'BERTEmbedding' object has no attribute 'tokenizer' — You are receiving this because you modified the open/close state. Reply to this email directly, view it on GitHub <#152?email_source=notifications&email_token=ACHPKS2CETBDQJK3U76RDV3P7RPDDA5CNFSM4IDUIGQ2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODZ5N6PI#issuecomment-511369021>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ACHPKS4VTGMUGTIGUSO2K4DP7RPDDANCNFSM4IDUIGQQ> .

[mahdisnapp]

yes I updated my kashgari just while you informed me.

[BrikerMan]

I just tried and it works. Please make sure you have imported the latest version if you started a kernel like jupyter before updating, it will require a kernel restart to import latest version. You can confirm by

print(kashgari.__version__)

Here is my demo: colab link

[mahdisnapp]

errors have been solved, thank you again.

[mahdisnapp]

dear BrikerMan,
your packages worked well and I have developed my models with success until the error occurred today,
when I tried to fit model I got this error:
mdel.fit(x_train, y_train, x_validate, y_validate, epochs = 5, batch_size = 32)

Keras: operands could not be broadcast together with shapes

[BrikerMan]

dear BrikerMan,
your packages worked well and I have developed my models with success until the error occurred today,
when I tried to fit model I got this error:
mdel.fit(x_train, y_train, x_validate, y_validate, epochs = 5, batch_size = 32)

Keras: operands could not be broadcast together with shapes

I need a little bit more details with this bug. Please provide the model config and data samples.

[mahdisnapp]

I used BiLSTM model from kashgari.tasks.classification, my input data (x_train and x_validate) are tokenized texts and outpus (y_train, y_validate) are labels,
note : i used BERTEmbedding and it's tokenization method
my training size is 450 and validate size is 50

[BrikerMan]

Maybe you could provide the code, how you build the model, and sample data. Or just give me a colab notebook which could reproduce the issue.

[BrikerMan]

It seems the issue with data, please run this code.

print(x_train[0])
print(x_validate[0])
print(y_train[0])
print(y_validate[0])

[mahdisnapp]

there were no problem with shape of training and validation sets, I provide link to my colab
colab

[BrikerMan]

there were no problem with shape of training and validation sets, I provide link to my colab
colab

I might know the issue, try to convert data to list rather than np.array.

mdel.fit(list(x_train), list(y_train), list(x_validate),list(y_validate), epochs = 5, batch_size = 32)

[mahdisnapp]

yes your solution worked, may I ask what is the reason the model did not support numpy array as an input?
again thanks for your patience and kindness to help us.

[BrikerMan]

I need to add all your corpus together to build token dict and label dict. So here is one important line

Kashgari/kashgari/tasks/base_model.py

Line 117 in 63f0991

self.embedding.analyze_corpus(x_train + x_validate, y_train + y_validate)

But when you add two numpy array, they must be able to broadcast

I have tried mainly on my build in corpus and haven't tested on numpy array input. So, this is a bug and will fix in the next version by converting x_train and x_validate to list before adding them.

[BrikerMan]

@mahdisnapp Maybe you guys could help me write some blog about how you use Kashgari~ That means a lot. I hope more people know about this project.

[mahdisnapp]

Dear BrikerMan
your work is very impressive and I want to help you to improve kashgari performance and add new features to it. at the moment I could help with bugs or add other models. if you need help just let me know. and if you want I could mail you for further talks

[BrikerMan]

Dear BrikerMan
your work is very impressive and I want to help you to improve kashgari performance and add new features to it. at the moment I could help with bugs or add other models. if you need help just let me know. and if you want I could mail you for further talks

Let's start with fixing bugs and documents, also feel free to add new features and create pull request. here is the contributing guide.

[alexwwang]

This may be a bug induced by tf.keras. In tf.keras, they haven’t done type check or explicit conversion from other types to list in some method implementations.

…

On Tue, Jul 16, 2019 at 20:27 Eliyar Eziz ***@***.***> wrote: Dear BrikerMan your work is very impressive and I want to help you to improve kashgari performance and add new features to it. at the moment I could help with bugs or add other models. if you need help just let me know. and if you want I could mail you for further talks Let's start with fixing bugs and documents, also feel free to add new features and create pull request. here is the contributing guide <https://kashgari.bmio.net/about/contributing/>. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#152?email_source=notifications&email_token=AAGRFKU7OOBCXLQSPSX2LQ3P7W5BRA5CNFSM4IDUIGQ2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD2AVZ5I#issuecomment-511794421>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAGRFKQ742WCHTWS3TGSIVTP7W5BRANCNFSM4IDUIGQQ> .