If you want to add a LSTM layer instead of regular ANN you can use this as model and change the layer size of info model = tf.keras.Sequential() model.add(hub_layer) model.add(tf.keras.layers.Reshape( target_shape=(128 , 1 ) )) model.add(tf.keras.layers.Bidirectional(tf.keras.layers.LSTM(32, return_sequences=True))) model.add(tf.keras.layers.Bidirectional(tf.keras.layers.LSTM(16))) for units in [128, 64 ]: model.add(tf.keras.layers.Dense(units, activation='relu')) model.add(tf.keras.layers.Dropout(0.3)) model.add(tf.keras.layers.Dense(6, activation='softmax')) model.summary()
If you are using Neural Network one reason you might get low accuracy is due to low volume of input data. If your dataset size is not huge better to stay with normal ML technique as shown in my previous video. Mostly to increase accuracy that is so low you might have to collect more data or just keep layers to minimum and see if it helps
Ajith.. I already have a tensorflow NLP classifier deployment deploy along with scaling it using k8s - ruclips.net/p/PL3N9eeOlCrP4VXtFJTjmGsqI-Emk2keVL Are you looking for anything in specific?
Hi - great video. Any chance you can share the Jupyter notebook file? Also - assuming I'm having a (relatively) small dataset (~17000 samples) can I still use this aproach?
sir your video has been great for me. sir I have one question now that model is trained i want to pass a real data and check for output please can you tell me the code for it. As we have taken tarain_data_f and not just x_train how to write code for model checking. sir plz waiting for a reply.
Here you go - github.com/srivatsan88/Natural-Language-Processing/blob/master/Text_classification_Tensorflow_Multiclass.ipynb dataset link is within notebook
Thank you for sharing. Looking for small example of end to end process of a ml project ( including deployment). That would help me alot. If you have already have t, pls share the link.
Very nice video just a doubt... what should be the code if we want to test a single complaint and get its product/target.....like a function where we pass a single complaint string and it returns the product/label
Adwait you can pass single complaint as well here but in case if you are looking for exposing it as Flask then you can load the model and take single instance of prediction as well. I have done it for another model for same data but not this one - ruclips.net/video/-F0CRcaNeao/видео.html
@@AIEngineeringLife Thanks but before deploying if I want to test it what should be the code because when I am passing single string in model.predict() it's giving error saying list index out of range
Hello, thank you very much for your video. But i have a question if you can help me. how to do with multiple text features and one output label please?
Can I also pass in a string tensor: object when using a GloVe embedding layer or is the ability to consume strings directly (without tokenizer) a special feature of the hub embedding layer? Is there a reference documentation for this direct way of consuming text without having to tokenize it first?
Where do i find syntax to convert an imported tf dataset from slices into a list of text to tokenize it? In your From Scratch video, you tokenized a tfds in built dataset but how to tokenize custom imported dataset like in this video? I could not find it in the docs.
You can still do the same as finally both tfds and tf dataset data type is same. Check this function that can help you with it - www.tensorflow.org/api_docs/python/tf/keras/preprocessing/text/Tokenizer
why have you not removed XXXXX values which are there? How to deal with such values? In which cases to use upsampling and in which cases use down sampling? When to use synthetic points?
You can tokenize outside like I did in my previous video to clean the text and then use that to create embedding before feeding into TF model. My previous videos here - ruclips.net/video/EHt_x8r1exU/видео.html Upsampling I am not that sure as in high dimension space it might add noise as well to data. Downsampling is good when your data is highly skewed like say 99% to 1%
class_weight error `class_weight` not supported for 3+ dimensional targets. When you are running : history = model.fit(train_data_f, epochs=4, validation_data=test_data_f, verbose=1, class_weight = weights)
When I type this, class_weights = list(class_weight.compute_class_weight('balanced', np.unique(df['Priority']),df['Priority'])) It gives me this error, TypeError Traceback (most recent call last) in () ----> 1 class_weights = list(class_weight.compute_class_weight('balanced', np.unique(df['Priority']),df['Priority'])) TypeError: compute_class_weight() takes 1 positional argument but 3 were given Why it is?
Somewhere dimensions of you input data seem to be problematic. Here is my code have you checked it to verify - github.com/srivatsan88/Natural-Language-Processing/blob/master/Text_classification_Tensorflow_Multiclass.ipynb
Hi sir..Can you please give us insights and models that can use how transactions were categorized in real time, with an example..To be more precise, if I have transaction description and category code in the train data..
Sorry Hari.. Did not follow the question completely. Are you asking about real time inference of this model?. If so I have similar videos where i have shown in this playlist - ruclips.net/p/PL3N9eeOlCrP5PlN1jwOB3jVZE6nYTVswk
AIEngineering I am not referring to this video/lecture actually...Asking you, in general how can we do transaction categorisation will be done in ML/DL?
basically all tf functions complies into a tensorflow graph when we call model save. If it is regular python functions we have to re-import it again separately apart from the model itself
How to handle a highly imbalanced multi-class (64 classes) image dataset in which some classes don't occur and in which some classes only occur 1 time?
Eddy.. Very Low occurrence classes are difficult to handle in multi class classification. Ideally it is good to collect more data. Being text data here upsampling and all might not be efficient
Sure Malik.. I try to but once I get started kind of gets too fast. Will try to slow it down in future. 0.9x speed has benefited others who have watched my old videos
If you want to add a LSTM layer instead of regular ANN you can use this as model and change the layer size of info
model = tf.keras.Sequential()
model.add(hub_layer)
model.add(tf.keras.layers.Reshape( target_shape=(128 , 1 ) ))
model.add(tf.keras.layers.Bidirectional(tf.keras.layers.LSTM(32, return_sequences=True)))
model.add(tf.keras.layers.Bidirectional(tf.keras.layers.LSTM(16)))
for units in [128, 64 ]:
model.add(tf.keras.layers.Dense(units, activation='relu'))
model.add(tf.keras.layers.Dropout(0.3))
model.add(tf.keras.layers.Dense(6, activation='softmax'))
model.summary()
Thanks again and again
Please how can we improve the model accuracy ? i had 0.4 !!
If you are using Neural Network one reason you might get low accuracy is due to low volume of input data. If your dataset size is not huge better to stay with normal ML technique as shown in my previous video. Mostly to increase accuracy that is so low you might have to collect more data or just keep layers to minimum and see if it helps
@@AIEngineeringLife Yes thats it, thank you
@@AIEngineeringLife Can we try Language Models when we have less amount of data ?
This is HUGE.....I am indebted to you in terms of gratitude. Appreciate your efforts !
Awesome tutorial!! Really appreciate the hard work you put into making this tutorial
Hi, Thanks for all your contents and knowledge sharing. very informative.
Very nice and detailed video
Indeed the right one for me to work . Thanks a ton for making such vedios
Thanks a ton . Please conclude this series by industrial deployment demo.
Ajith.. I already have a tensorflow NLP classifier deployment deploy along with scaling it using k8s - ruclips.net/p/PL3N9eeOlCrP4VXtFJTjmGsqI-Emk2keVL
Are you looking for anything in specific?
No no exactly what I was looking for. Thanks
Such a nice video. Very well explained bro!!
Excellent tutorial
Thank you so much for this!
Hi - great video.
Any chance you can share the Jupyter notebook file?
Also - assuming I'm having a (relatively) small dataset (~17000 samples) can I still use this aproach?
sir your video has been great for me. sir I have one question now that model is trained i want to pass a real data and check for output please can you tell me the code for it. As we have taken tarain_data_f and not just x_train how to write code for model checking. sir plz waiting for a reply.
Thanks a lot. Great Explanation. Could you share the notebook and dataset link
Here you go - github.com/srivatsan88/Natural-Language-Processing/blob/master/Text_classification_Tensorflow_Multiclass.ipynb
dataset link is within notebook
Thank you for sharing. Looking for small example of end to end process of a ml project ( including deployment). That would help me alot. If you have already have t, pls share the link.
@@srikanth1107262 I have multiple this is one of it where I cover end to end
ruclips.net/p/PL3N9eeOlCrP4VXtFJTjmGsqI-Emk2keVL
Very nice video just a doubt... what should be the code if we want to test a single complaint and get its product/target.....like a function where we pass a single complaint string and it returns the product/label
Adwait you can pass single complaint as well here but in case if you are looking for exposing it as Flask then you can load the model and take single instance of prediction as well. I have done it for another model for same data but not this one - ruclips.net/video/-F0CRcaNeao/видео.html
@@AIEngineeringLife Thanks but before deploying if I want to test it what should be the code because when I am passing single string in model.predict() it's giving error saying list index out of range
@
AIEngineering
getting this error while running show batch: TypeError: Signature mismatch. Keys must be dtype , got .
Can you pls solve this.
I'm missing how to tokenize and embedding since I was doing it in a spanish dataset and I can't use the google reviews dataset...
Can you share the code how use CNN instead of regular NN , Like you LSTM model code.
Hello, thank you very much for your video. But i have a question if you can help me. how to do with multiple text features and one output label please?
You can give both text features as well by creating a embedding of the other feature
Can I also pass in a string tensor:
object when using a GloVe embedding layer or is the ability to consume strings directly (without tokenizer) a special feature of the hub embedding layer?
Is there a reference documentation for this direct way of consuming text without having to tokenize it first?
It is a special feature of hub models. For others you might have to encode it before passing
Where do i find syntax to convert an imported tf dataset from slices into a list of text to tokenize it? In your From Scratch video, you tokenized a tfds in built dataset but how to tokenize custom imported dataset like in this video? I could not find it in the docs.
You can still do the same as finally both tfds and tf dataset data type is same. Check this function that can help you with it - www.tensorflow.org/api_docs/python/tf/keras/preprocessing/text/Tokenizer
Excellent!
Can you share the guthub code path?
Here you go - github.com/srivatsan88/Natural-Language-Processing/blob/master/Text_classification_Tensorflow_Multiclass.ipynb
I need input_shape to pass to build() function to build this model when using LSTM model ..... what is input_shape?
Can u make vedios for keyword extraction which makes much meaningful 3 grams keyphrases?
Thanks!
why have you not removed XXXXX values which are there? How to deal with such values? In which cases to use upsampling and in which cases use down sampling? When to use synthetic points?
You can tokenize outside like I did in my previous video to clean the text and then use that to create embedding before feeding into TF model. My previous videos here - ruclips.net/video/EHt_x8r1exU/видео.html
Upsampling I am not that sure as in high dimension space it might add noise as well to data. Downsampling is good when your data is highly skewed like say 99% to 1%
class_weight error
`class_weight` not supported for 3+ dimensional targets.
When you are running :
history = model.fit(train_data_f,
epochs=4,
validation_data=test_data_f,
verbose=1,
class_weight = weights)
I get the same error, but when I rerun the previous 5 blocks the error disappears. Need to implement sample weights to correctly fix this issue though
Can I get link to this Colab file ?
Bert implementation and fine-tuning how to do
Please let me know where can I find this Code file sir
When I type this,
class_weights = list(class_weight.compute_class_weight('balanced', np.unique(df['Priority']),df['Priority']))
It gives me this error,
TypeError Traceback (most recent call last)
in ()
----> 1 class_weights = list(class_weight.compute_class_weight('balanced', np.unique(df['Priority']),df['Priority']))
TypeError: compute_class_weight() takes 1 positional argument but 3 were given
Why it is?
hi i am getting the following error ValueError: `class_weight` not supported for 3+ dimensional targets. for passing class weights...any help on this
Somewhere dimensions of you input data seem to be problematic. Here is my code have you checked it to verify - github.com/srivatsan88/Natural-Language-Processing/blob/master/Text_classification_Tensorflow_Multiclass.ipynb
Hi sir..Can you please give us insights and models that can use how transactions were categorized in real time, with an example..To be more precise, if I have transaction description and category code in the train data..
Sorry Hari.. Did not follow the question completely. Are you asking about real time inference of this model?. If so I have similar videos where i have shown in this playlist - ruclips.net/p/PL3N9eeOlCrP5PlN1jwOB3jVZE6nYTVswk
AIEngineering I am not referring to this video/lecture actually...Asking you, in general how can we do transaction categorisation will be done in ML/DL?
could you add on why do we add this? @tf.function
basically all tf functions complies into a tensorflow graph when we call model save. If it is regular python functions we have to re-import it again separately apart from the model itself
@@AIEngineeringLife Thanks. Yes I think i got an idea.
can we leverage LSTM model for binary classification?
Yes just change the model loss function
@@AIEngineeringLife is there any social media handle where I can reach you directly
@@shubhamsingh8578 I am on LinkedIn.Check my channel home page for my linkedin profile link
Can you please provide notebook link
For same model I am getting 99% accuracy. Is it overfitting? Please answer
How to handle a highly imbalanced multi-class (64 classes) image dataset in which some classes don't occur and in which some classes only occur 1 time?
Eddy.. Very Low occurrence classes are difficult to handle in multi class classification. Ideally it is good to collect more data. Being text data here upsampling and all might not be efficient
can u share the code sir
Aur y havent you preprocessed the data
Why are you talking so fast? Breathe!
Sure Malik.. I try to but once I get started kind of gets too fast. Will try to slow it down in future. 0.9x speed has benefited others who have watched my old videos
@@AIEngineeringLife Well, 0.9 speed doesn't help *you*. We can't learn anything if you pass out! ;-)