Last Updated on August 14, 2019

The goal of developing an LSTM model is a final model that you can use on your sequence prediction problem.

In this post, you will discover how to finalize your model and use it to make predictions on new data.

After completing this post, you will know:

- How to train a final LSTM model.
- How to save your final LSTM model, and later load it again.
- How to make predictions on new data.

**Kick-start your project** with my new book Long Short-Term Memory Networks With Python, including *step-by-step tutorials* and the *Python source code* files for all examples.

Let’s get started.

## Step 1. Train a Final Model

### What Is a Final LSTM Model?

A final LSTM model is one that you use to make predictions on new data.

That is, given new examples of input data, you want to use the model to predict the expected output. This may be a classification (assign a label) or a regression (a real value).

The goal of your sequence prediction project is to arrive at a final model that performs the best, where “best” is defined by:

**Data**: the historical data that you have available.**Time**: the time you have to spend on the project.**Procedure**: the data preparation steps, algorithm or algorithms, and the chosen algorithm configurations.

In your project, you gather the data, spend the time you have, and discover the data preparation procedures, algorithm to use, and how to configure it.

The final model is the pinnacle of this process, the end you seek in order to start actually making predictions.

There is no such thing as a perfect model. There is only the best model that you were able to discover.

### How to Finalize an LSTM Model?

You finalize a model by applying the chosen LSTM architecture and configuration on all of your data.

There is no train and test split and no cross-validation folds. Put all of the data back together into one large training dataset and fit your model.

That’s it.

With the finalized model, you can:

- Save the model for later or operational use.
- Load the model and make predictions on new data.

For more on training a final model, see the post:

### Need help with LSTMs for Sequence Prediction?

Take my free 7-day email course and discover 6 different LSTM architectures (with code).

Click to sign-up and also get a free PDF Ebook version of the course.

## Step 2. Save Your Final Model

Keras provides an API to allow you to save your model to file.

The model is saved in HDF5 file format that efficiently stores large arrays of numbers on disk. You will need to confirm that you have the h5py Python library installed. It can be installed as follows:

1 |
sudo pip install h5py |

You can save a fit Keras model to file using the save() function on the model.

For example:

1 2 3 4 5 6 7 8 9 |
# define model model = Sequential() model.add(LSTM(...)) # compile model model.compile(...) # fit model model.fit(...) # save model to single file model.save('lstm_model.h5') |

This single file will contain the model architecture and weights. It also includes the specification of the chosen loss and optimization algorithm so that you can resume training.

The model can be loaded again (from a different script in a different Python session) using the load_model() function.

1 2 3 4 5 6 |
from keras.models import load_model # load model from single file model = load_model('lstm_model.h5') # make predictions yhat = model.predict(X, verbose=0) print(yhat) |

Below is a complete example of fitting an LSTM model, saving it to a single file and later loading it again. Although the loading of the model is in the same script, this section may be run from another script in another Python session. Running the example saves the model to the file lstm_model.h5.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 |
from keras.models import Sequential from keras.layers import Dense from keras.layers import LSTM from numpy import array from keras.models import load_model # return training data def get_train(): seq = [[0.0, 0.1], [0.1, 0.2], [0.2, 0.3], [0.3, 0.4], [0.4, 0.5]] seq = array(seq) X, y = seq[:, 0], seq[:, 1] X = X.reshape((len(X), 1, 1)) return X, y # define model model = Sequential() model.add(LSTM(10, input_shape=(1,1))) model.add(Dense(1, activation='linear')) # compile model model.compile(loss='mse', optimizer='adam') # fit model X,y = get_train() model.fit(X, y, epochs=300, shuffle=False, verbose=0) # save model to single file model.save('lstm_model.h5') # snip... # later, perhaps run from another script # load model from single file model = load_model('lstm_model.h5') # make predictions yhat = model.predict(X, verbose=0) print(yhat) |

For more on saving and loading your Keras model, see the post:

## Step 3. Make Predictions on New Data

After you have finalized your model and saved it to file, you can load it and use it to make predictions.

For example:

- On a sequence regression problem, this may be the prediction of the real value at the next time step.
- On a sequence classification problem, this may be a class outcome for a given input sequence.

Or it may be any other variation based on the specifics of your sequence prediction problem. You would like an outcome from your model (yhat) given an input sequence (X) where the true outcome for the sequence (y) is currently unknown.

You may be interested in making predictions in a production environment, as the backend to an interface, or manually. It really depends on the goals of your project.

Any data preparation performed on your training data prior to fitting your final model must also be applied to any new data prior to making predictions.

Predicting is the easy part.

It involves taking the prepared input data (X) and calling one of the Keras prediction methods on the loaded model.

Remember that the input for making a prediction (X) is only comprised of the input sequence data required to make a prediction, not all prior training data. In the case of predicting the next value in one sequence, the input sequence would be 1 sample with the fixed number of time steps and features used when you defined and fit your model.

For example, a raw prediction in the shape and scale of the activation function of the output layer can be made by calling the predict() function on the model:

1 2 3 |
X = ... model = ... yhat = model.predict(X) |

The prediction of a class index can be made by calling the predict_classes() function on the model.

1 2 3 |
X = ... model = ... yhat = model.predict_classes(X) |

The prediction of probabilities can be made by calling the predict_proba() function on the model.

1 2 3 |
X = ... model = ... yhat = model.predict_proba(X) |

For more on the life-cycle of your Keras model, see the post:

## Further Reading

This section provides more resources on the topic if you are looking go deeper.

### Posts

- How to Train a Final Machine Learning Model
- Save and Load Your Keras Deep Learning Models
- The 5 Step Life-Cycle for Long Short-Term Memory Models in Keras

### API

## Summary

In this post, you discovered how to finalize your model and use it to make predictions on new data.

Specifically, you learned:

- How to train a final LSTM model.
- How to save your final LSTM model, and later load it again.
- How to make predictions on new data.

Do you have any questions?

Ask your questions in the comments below and I will do my best to answer.

Thanks Jason

One question. Why should we finalize the model on the whole data. We would change the weights again right? The model which was trained on the training data is the one we tested on unseen data (test set). The new model (trained on all data) could be worse, overfitted,… not?

Learn more about how and why to finalize models here:

https://machinelearningmastery.com/train-final-machine-learning-model/

I address this exact concern.

Predict_classes is displaying single output whereas I have designed my model for 6 output.

Please help me understand why is it doing that.

Perhaps call predict() to get the probability of each of the 6 classes?

Hi,

Just wondering. Whether you selected ‘delay’ in ur LSTM model or not if yes how you selected.

You mean the number of timesteps? Arbitrarily.

I tried both keras and tensorflow. Tensorflow has more features.

It does, but it is much harder to use.

thanks, Jason.

I ran your sample code, but found the result as below, which seems not expected.

… …

Epoch 290/300

6/6 [==============================] – 0s – loss: 0.0155

Epoch 295/300

6/6 [==============================] – 0s – loss: 0.0153

Epoch 296/300

6/6 [==============================] – 0s – loss: 0.0153

Epoch 297/300

6/6 [==============================] – 0s – loss: 0.0152

Epoch 298/300

6/6 [==============================] – 0s – loss: 0.0152

Epoch 299/300

6/6 [==============================] – 0s – loss: 0.0152

Epoch 300/300

6/6 [==============================] – 0s – loss: 0.0151

[[ 0.28978038]

[ 0.31878966]

[ 0.3477335 ]

[ 0.37631655]

[ 0.4042924 ]

[ 0.43146992]]

Suppose input of X are 0 0.1 0.2 0.3 0.4 0.5, then predict value of y should be similar to 0.1 0.2… 0.6. but the result is such value as 0.28978038… … 0.43146992.

can you check more about it?

What is the problem exactly?

Hi Dr. Jason,

Thanks so much for your tutorials.

I will like to clarify what Tieliu’s question was:

He was making a reference to the example cited on this page where you demonstrated making predictions with LSTM as follows:

# return training data

def get_train():

seq = [[0.0, 0.1], [0.1, 0.2], [0.2, 0.3], [0.3, 0.4], [0.4, 0.5]]

seq = array(seq)

X, y = seq[:, 0], seq[:, 1]

X = X.reshape((len(X), 1, 1))

return X, y

# define model

model = Sequential()

model.add(LSTM(10, input_shape=(1,1)))

model.add(Dense(1, activation=’linear’))

# compile model

model.compile(loss=’mse’, optimizer=’adam’)

# fit model

X,y = get_train()

model.fit(X, y, epochs=300, shuffle=False, verbose=0)

# save model to single file

model.save(‘lstm_model.h5’)

# snip…

# later, perhaps run from another script

# load model from single file

model = load_model(‘lstm_model.h5’)

# make predictions

yhat = model.predict(X, verbose=0)

print(yhat)

When you run the code, the trained model did not make good prediction of the actual response variable y. It has the following predicted yhat values:

[[0.24346247]

[0.27623463]

[0.30942053]

[0.34286284]

[0.37640885]]

Rather than the actual y values:

[[0.1]

[0.2]

[0.3]

[0.4]

[0.5]]

In other words, if we approximate the predicted values to 2 decimal places, the model predicted 0.2, 0.3, and 0.4, correctly, but failed in predicting 0.1 and 0.5. In this situation, could we say our final model should be discarded and we then decide to train another model with a different set of procedures and configurations?

Sorry, this question is a bit longer than expected, but I wanted to clarify the initial question as there was no further conversation on this.

Sorry, I don’t follow, why would we discard the model?

Because not all the predicted values are equal to the actual values as shown between yhat and y, or am I missing some important concept of your tutorial?

No model is perfect. If perfection was possible we would not need machine learning.

Got it! Thanks so much.

BWT:How do we show or tell that a model is good? Do we just only care about the scores, for example, a regression problem?

Correct.

Hi, I want to predict for a whole record of shape like (160, 72) for single time step. How would I shape my numpy array of features for test. For more clear understanding, I have trained my model on trainX with shape (235, 1, 72) and trainY with shape (235,). Now I want to predict a single timestep but for 160 rows. How to do that?

See this post on how to reshape data for LSTMs:

https://machinelearningmastery.com/reshape-input-data-long-short-term-memory-networks-keras/

Hello Jason,

Thanks for the fantastic tutorial. It helped me trained a model for predicting energy patterns.

That said, I am interested in putting this into a real-time prediction production site and am wondering the following:

1. I am planning to use Flask to publish said trained model as an API to produce prediction in a live website. Is that an appropriate choice of tool for deployment?

2. I have reshaped and scaled my raw data to produce the model… does it mean that I will need to reshape and scale all the future data prior to be fitted into the trained model for it to output relevant predictions?

Thanks and appreciate your kind reply for my questions above!

Choices of tools and frameworks for production are really a decision you and your stakeholders should be making. I cannot comment in any meaningful way as I won’t be responsible for the decision.

Any data preparation applied to training data must be applied to new data fed to the model in the future.

Hi Jason,

When performing model.predict, i do get some inconsistent outputs. Does that mean my model is wrong?as far as i know, possible scenario for inconsistent outputs is if i try to re-fit the model and not during prediction. Am i missing something? hope to get some comments from you. thank you very much

from numpy.random import seed

seed(1)

from tensorflow import set_random_seed

set_random_seed(2)

i have also set the seed before running them

It looks like you’re doing all the right things.

LSTMs are a stochastic algorithm, this post will shed light on this issue:

https://machinelearningmastery.com/randomness-in-machine-learning/

This post has advice on how to get reproducible results with Keras:

https://machinelearningmastery.com/reproducible-results-neural-networks-keras/

Thank you so much for the advise.. i will think about it more..

Hi Jason,

I am so grateful for the post U shared with us.

I just want to load 3 finalized model namely RNN, CNN, LSTM in one script concurrently which have already saved as a finalized model in keras to using them in an ensemble model to gain an average result for predicting. Is it necessary to use dask data frame to load multi finalized (saved model) models?? or loading multi finalized model has the same commands as loading one finalized model?

Thank U in advanced for taking your time to replying.

A Pandas DataFrame is not required. Each model can be saved and loaded to and from separate files and used in an ensemble.

Hi Jason,

I am really appreciated about this helpful tutorial.

I train and then save a finalized cnn model in a script after that I load the finalized cnn model in another different script just used this command :(load_model(‘cnn_model.h5’) .

In fact I have a test dataset which does not have any label and I wanna gain the proability of belonging of each sample to each class by this command : (model_cnn_final.predict_proba) but gave me this error:((AttributeError: ‘Model’ object has no attribute ‘predict_proba’) and also when I applied this command:[yhat=model.predict_classes(X) ] it gave me this error :(‘Model’ object has no attribute ‘predict_classes’).

I have used the command :(yhat = model.predict(X)) and it worked fine.

what is the problem with these commands which cuase error??

how can I fixe the errors?

I believe these methods are only supported on Sequential models, you may be using the functional Model API. In that case, you may be limited to the predict() function alone, which will return probabilities in the case of a softmax activation function in the output layer.

Hi

This is super good Jason!

Thanks your writings 🙂

I’m glad it helps.

Hi Jason,

Thank you for this really helpful tutorial.

I have a question. I think I am missing something to make predictions. I don’t understand what should be the input on model.predict(X) to predict new data. Let’s say I have one year of data (sampled every hour) and I want to predict the following week. What should my X be ?

It depends on how you define your model.

If the model expects 1 year of data to make a 1 week forecast, then you must provide 1 year of data as X.

Does that help?

It does but I still don’t get it. I am following your tutorial on Mutlivariate Time Series Forecasting (https://machinelearningmastery.com/multivariate-time-series-forecasting-lstms-keras/) and I can’t find where you are defining these parameters of the model.

Hello, it is realy nice explanation, but I want to ask what verbose does ? why we assign it to 0 ?

Verbose gives output. We can turn off this output by setting it to 0.

Hi Jason, thank you for your tutorial. I’m very grateful for what I have learnt from you.

I have a question. Let’s say I have my LSTM model and it is working properly with the train and test data, so the model is ready to be used in production. The data that I have has the following features: timestamp, price and capitalization. If I want to predict tomorrow’s price, must

I provide timestamp and capitalization values? or only timestamp?

To make one prediction, it must be provided with one input sequence, as defined by your model during training.

Hi Jason, all your blogposts are super insightful. Cannot wait to read more articles.

Earlier, I had not preprocessed X and y correctly. Now I have used the “Stacked LSTM for sequence classification” referenced in keras homepage https://keras.io/getting-started/sequential-model-guide/#examples.

This is how my result looks. https://github.com/sagar-m/character-classification/blob/master/SAP.ipynb Unfortunately, I do not have ytest.txt to verify my results. Please have a look and let me know if you have any comments.

My number of classes to predict is 12, however the predicted output range in classes was 1 to 10. Not sure why and if it incorrect.

Thanks a lot again.

Thanks.

I’m eager to help and answer questions, but I don’t have the capacity to review your code, I explain more here:

https://machinelearningmastery.com/faq/single-faq/can-you-read-review-or-debug-my-code

Hi Jason, one question I have : “validation schemes are supposed to be used to estimate quality of the model. When you found the right hyper-parameters and want to get test predictions don’t forget to retrain your model using all training data.”

Once the model is trained with validation set, should I retrain the model using all training data?

Thank you.

A final model is trained on all available data.

More here:

https://machinelearningmastery.com/train-final-machine-learning-model/

Hi Jason,

I have a concern related RealTime validation if you have input test data (60 frame per ms ) and you would like to do prediction in Realtime . How you could ensure the prediction will be done. ?

On the other hand, lets give an example that, we ‘re getting input test data ( 100 x 162 ) , time stamp x features . We are fitting all the information in one sample array (1,100,162,1), then you would like to do prediction for each time instance when you receive the data. The problem the streaming dataset is too fast to be catch by model:

y_predict = model.predict_classes(test_input)

I would like to know if you have any suggestions regarfing such a problem. In term of how we could make the prediction possible in realtime streaming data for each time the input change only. ?

If you run it with this speed ( the prediction will be going on frem old dataset . Cant catch the new samples )

Thanks in advance

Making predictions is very fast.

Only training the model is slow, which is only done once before it is used.

Hi Jason,

Thank you for the awesome and also practical tutorial as ever been.

I face a question as that is if I wanna use predict(x_dataset) function, is it necessary to padding “x_dataset” or not??

I will be grateful if you answer the question.

Best Regard

Maryam

To make a prediction, the input data must be prepared in the same way as the training data, including lengths and transforms.

Hi.

I have one question.

I have a training set with the labels, let’s say that I play cheess and I have the historical matches with label of the winner [0 = player1, 1 = player 2]

And I want to predict if after 10/15 moves I’ll have more probability to win or lose.

How can I write the model that predict a number between 0 and 1 ( close to 0 means that I’ll win and close to 1 means i’ll lose )

Thanks !

A good approach might be to use rating systems to estimate the skill of each player and feed this into a predictive model.

Hello Jason, and many thanks for this awesome tutorial.

My preoccupation is about using the trained and tested model to predict the future. This means values after the test set.

Thank you.

I’m happy it helped.

Hi Jason,

I tried to make predictions just for one Row input data. But I like to Know should my new data be scaled or not? if yes I tried with my scaled model but I got “0” for each feature!! which way is correct? please help me.

Thank you.

Your input data must be prepared in the same way as the training data.

If the training data was scaled, then new data must be scaled using the same coefficients.

Finally an easy to understand gist on how to implement an LSTM. Thank you!

I’m happy that it helped!

Hello Jason,

I tried 6.7 code example from Long Short Term Memory Networks with Python. But its giving error.

from random import randint

from numpy import array

from numpy import argmax

from keras.models import Sequential

from keras.layers import LSTM

from keras.layers import Dense

# generate a sequence of random numbers in [0, n_features)

def generate_sequence(length, n_features):

return [randint(0, n_features-1) for _ in range(length)]

# one hot encode sequence

def one_hot_encode(sequence, n_features):

encoding = list()

for value in sequence:

vector = [0 for _ in range(n_features)]

vector[value] = 1

encoding.append(vector)

return array(encoding)

# decode a one hot encoded string

def one_hot_decode(encoded_seq):

return [argmax(vector) for vector in encoded_seq]

# generate one example for an lstm

def generate_example(length, n_features, out_index):

# generate sequence

sequence = generate_sequence(length, n_features)

# one hot encode

encoded = one_hot_encode(sequence, n_features)

print(“Shape of encoded is:”, encoded.shape)

# reshape sequence to be 3D

X = encoded.reshape((1, length, n_features))

# select output

y = encoded[out_index].reshape(1, n_features)

return X, y

# define model

length = 5

n_features = 10

out_index = 2

model = Sequential()

model.add(LSTM(25, input_shape=(length, n_features)))

model.add(Dense(n_features, activation=’softmax’))

model.compile(loss=’categorical_crossentropy’, optimizer=’adam’, metrics=[‘acc’])

model.summary()

# fit model

for i in range(10000):

X, y = generate_example(length, n_features, out_index)

model.fit(X, y, epochs=1, verbose=2)

Shape of encoded is: (1, 10)

—————————————————————————

ValueError Traceback (most recent call last)

in ()

1 # fit model

2 for i in range(10000):

—-> 3 X, y = generate_example(length, n_features, out_index)

4 model.fit(X, y, epochs=1, verbose=2)

in generate_example(length, n_features, out_index)

7 print(“Shape of encoded is:”, encoded.shape)

8 # reshape sequence to be 3D

—-> 9 X = encoded.reshape((1, length, n_features))

10 # select output

11 y = encoded[out_index].reshape(1, n_features)

ValueError: cannot reshape array of size 10 into shape (1,5,10)

It suggest that the shape of your data does not match the expectation of your model.

You can change the shape of the data or change the expectation of the model.

Hello Jason,

Thanks for the great tutorial. I just wanted to know how can i calculate computational complexity of LSTM?

Thanks in advance.

Sorry, I don’t have material on calculating the computational complexity of neural networks.

I would like to predict the category of sentence using classification ,but here i need to know while doing model predicting ,if a sentence is already given to the model for predictions even after first time i need to get you have already predicted this sentence.

Please suggest on this issue,this is useful for chatbot application,if a question was already asked in conservation , again user had asked same question i need to get yo uhave already this question to previoustime.

Perhaps add an if-statement to look up the sentence in a hashtable before passing a sentence to the model for prediction.

Sounds like engineering, not machine learning.

Thanks for your answer

X=[x =21, x1=13, ‘grassMinTemp’=15]

yhat = model.predict(X, verbose=0)

print(yhat)

would i do like this?

No, input is an array of numbers, just like training data, e.g. X = [21, 13, 15]

if multiple dependent variables in LSTM model , what should i do ?

X = [x1=12,x2=1234]

model =the path of the model

yhat = model.predict(X)

thx

I give an example here:

https://machinelearningmastery.com/how-to-develop-lstm-models-for-time-series-forecasting/

your model show’s like this error how do i solve it

Using TensorFlow backend.

Traceback (most recent call last):

File “C:\Python36\lib\site-packages\tensorflow\python\pywrap_tensorflow.py”, line 58, in

from tensorflow.python.pywrap_tensorflow_internal import *

File “C:\Python36\lib\site-packages\tensorflow\python\pywrap_tensorflow_internal.py”, line 28, in

_pywrap_tensorflow_internal = swig_import_helper()

File “C:\Python36\lib\site-packages\tensorflow\python\pywrap_tensorflow_internal.py”, line 24, in swig_import_helper

_mod = imp.load_module(‘_pywrap_tensorflow_internal’, fp, pathname, description)

File “C:\Python36\lib\imp.py”, line 242, in load_module

return load_dynamic(name, filename, file)

File “C:\Python36\lib\imp.py”, line 342, in load_dynamic

return _load(spec)

ImportError: DLL load failed: The specified module could not be found.

During handling of the above exception, another exception occurred:

Traceback (most recent call last):

File “D:/lstm_model.py”, line 1, in

from keras.models import Sequential

File “C:\Python36\lib\site-packages\keras\__init__.py”, line 3, in

from . import utils

File “C:\Python36\lib\site-packages\keras\utils\__init__.py”, line 6, in

from . import conv_utils

File “C:\Python36\lib\site-packages\keras\utils\conv_utils.py”, line 9, in

from .. import backend as K

File “C:\Python36\lib\site-packages\keras\backend\__init__.py”, line 89, in

from .tensorflow_backend import *

File “C:\Python36\lib\site-packages\keras\backend\tensorflow_backend.py”, line 5, in

import tensorflow as tf

File “C:\Python36\lib\site-packages\tensorflow\__init__.py”, line 24, in

from tensorflow.python import pywrap_tensorflow # pylint: disable=unused-import

File “C:\Python36\lib\site-packages\tensorflow\python\__init__.py”, line 49, in

from tensorflow.python import pywrap_tensorflow

File “C:\Python36\lib\site-packages\tensorflow\python\pywrap_tensorflow.py”, line 74, in

raise ImportError(msg)

ImportError: Traceback (most recent call last):

File “C:\Python36\lib\site-packages\tensorflow\python\pywrap_tensorflow.py”, line 58, in

from tensorflow.python.pywrap_tensorflow_internal import *

File “C:\Python36\lib\site-packages\tensorflow\python\pywrap_tensorflow_internal.py”, line 28, in

_pywrap_tensorflow_internal = swig_import_helper()

File “C:\Python36\lib\site-packages\tensorflow\python\pywrap_tensorflow_internal.py”, line 24, in swig_import_helper

_mod = imp.load_module(‘_pywrap_tensorflow_internal’, fp, pathname, description)

File “C:\Python36\lib\imp.py”, line 242, in load_module

return load_dynamic(name, filename, file)

File “C:\Python36\lib\imp.py”, line 342, in load_dynamic

return _load(spec)

ImportError: DLL load failed: The specified module could not be found.

Failed to load the native TensorFlow runtime.

See https://www.tensorflow.org/install/errors

for some common reasons and solutions. Include the entire stack trace

above this error message when asking for help.

Sorry to hear that, perhaps tensorflow is not installed correctly?

Perhaps try re-installing?

Perhaps try theano instead?

I couldn’t found clarity about prediction in this model. My doubt is how can one predict for future values without having test data. suppose we have a univariate data set(2012 Jan-2015 Dec) how can we get values for (2016 Jan to 2016 Dec).plz help me.Thank you in advance

yhat = model.predict(X)

what will be the X?

It will be whatever your model expects as input, e.g. an array of samples.

Hello Jason,

suppose time series is 1,2,3,4,5,6,7,8…..

If we want to predict next (e.g 9th, 10th, 11th…) data points in time series using LSTM.

how to do that?

model.predict(X) What would be X?

It depends on how you have framed your problem.

If the model expects 3 inputs to predict 1 output and those 3 inputs are the 3 prior obs, then:

Does that help?

Hi Jason, sorry to middle in uninvited. I am on the same spot and to me it looks like the model indeed works as long as there is data in X without projecting any future predictions. In your example above a prediction over X = asarray([[[6],[7],[8]]]) will plot a curve following along those values, and not a projection of following ones. The model.predict(X) I look for is a prediction of values that do not exist yet and for which there is no line plot of true data. If today is October the 2nd and I want to predict the open stock price of the future 7 days (when there is no existing true value yet), how can this be done? populating X with the last x steps of the full dataset will only return a nice prediction over those last x steps which I already know… Thanks!

You can perform a multi-step prediction. E.g. you design a model that outputs n predictions.

See thus tutorial to get started:

https://machinelearningmastery.com/how-to-develop-lstm-models-for-time-series-forecasting/

Thank you Jason. I have not gone that far yet and will review your tutorial. Great work!

Thanks.

Hi jason ,

For exemple if i want to predict the value of a currency then my seq will be dates ?

No, the observations at the dates.

Hello Jason!

To reduce variance, I used the average of 3 predictions using 3 different models created with your “fit_lstm” function in another tutorial and I get better results.

forecast1 = model1.predict(test_reshaped, batch_size=batch_size)

forecast2 = model2.predict(test_reshaped, batch_size=batch_size)

forecast3 = model3.predict(test_reshaped, batch_size=batch_size)

However, can I save this into ONE model for future use?

Nice work.

You can create a multi-input model that combines the predictions from the 3 models.

Perhaps this would be a good starting point:

https://machinelearningmastery.com/stacking-ensemble-for-deep-learning-neural-networks/

hey jason i need to ask how to predict the next 1000 data from the model?

what i meant is if i already save the model with 8000 window size and total of data is 12000 and when i want to use the model with 8000 data it has to have 8000 window size and how do i predict the 4000 if i can’t create the shape?

def create_dataset(dataset, look_back = 1):

data_X, data_Y = [], []

for i in range(len(dataset) – look_back – 1):

a = dataset[i:(i + look_back), 0]

data_X.append(a)

data_Y.append(dataset[i + look_back, 0])

return(np.array(data_X), np.array(data_Y))

can you please help me how can i predict the next 4000 cause i keep getting error on the shape.

thank you

Sorry, I don’t follow.

Perhaps you can elaborate?

You can develop a model to predict 1000 steps at once, or use a recursive strategy.

I list some approaches here:

https://machinelearningmastery.com/faq/single-faq/how-do-you-use-lstms-for-multi-step-time-series-forecasting

Hello Jason,

The model I developed performs very well on the test set. However for prediction its performance drops. Is it an overfitting problem? Could you go please tell me how to counter this problem?

Could it be the fact that I’m using LSTM for a small dataset (<300 data points)?

Perhaps.

The tutorials here will help you to diagnose issues with your model:

https://machinelearningmastery.com/learning-curves-for-diagnosing-machine-learning-model-performance/

Hi Jason,

I am trying to forecast into the future using my LSTM and my training data contains features that are unavailable for future X inputs (e.g. Stock price). What are some ways to deal with this issue and still be able to make predictions for the future?

Thanks.

Frame the prediction problem with the data that you do have. E.g. how you intend to use the model should dictate how the problem is framed.

Hi Jason,

I have tried to implement this tutorial in a real life case & predicted 100 future time steps (in a univariate scenario).

As a next step I have included some seasonal dummies & other predictors. In order to predict 100 future time steps I have used the following code:

#future unknown predictions: in this case, test_set doesn’t exist

future_pred_count = 100 #let’s predict 100 new steps

model.reset_states() #always reset states when inputting a new sequence

#first, let set the model’s states (it’s important for it to know the previous trends)

predictions = model.predict(previous_inputs) #this creates states

#future predictions

future = []

currentStep = predictions[:,-1:,:] #last step from the previous prediction

for i in range(future_pred_count):

currentStep = model.predict(currentStep) #get the next step

future.append(currentStep) #store the future steps

#after processing a sequence, reset the states for safety

model.reset_states()

But I am getting an error IndexError: too many indices for array. It would be great if you can help.

Perhaps start with this simpler tutorial:

https://machinelearningmastery.com/how-to-develop-lstm-models-for-time-series-forecasting/

Please ignore my last question

Can you please help me in predicting the sales of future dates using the LSTM model trained in the given article. It would be very helpful.

https://towardsdatascience.com/predicting-sales-611cb5a252de

In this article, prediction has been done on the time period given in the dataset with all the features available but i want to predict it for future dates.

Looking towards your reply.

Perhaps if you have a question about that post, you can contact the authors?

Generally, you can learn how to make out of sample forecasts here:

https://machinelearningmastery.com/make-sample-forecasts-arima-python/

Hi Jason,

If I have 3 files in my train data and 1 file in my testing data.

and in the file, there is supervised data which has viscosity and temperature at time t and t-1 timestamp. Can I predict the temp and log viscosity on time for the test set also?

My data is being made with respect to time and I have to predict at this time step this will be the viscosity and that will be the temperature. I am able to do it in one file by diving the data into 75 perc training and 25 perc in testing. But what If I want to train my model on 3 files like this and test on 1 file. Is it possible with time-series data`?

I believe so, perhaps experiment and see what is viable?

Hello,

How do i split time series data set to predict for next 7 days

This tutorial will help:

https://machinelearningmastery.com/convert-time-series-supervised-learning-problem-python/

Hi Jason,

An LSTM model gives different output for different runs unless you assign a seed. In that case how should one decide on which model to be used (basically which weights) ?

Correct.

Use the average performance of the model over multiple training runs.

Alright the way you have demonstrated in one of the articles. Thanks Jason.

Thanks.

hi Jason,

this is the code i used to make a prediction out of my saved lstm model.

the dataset is one row of inputs with the header and index column which is:

0 0 0 0 0 0 0 0 0 26.1 5.201

i want to predict the last column upto 2 time steps. (t and t+1) i wrote the lstm model code accordingly.

prediction code:

dataset = read_csv(‘predict.csv’, header=0, index_col=0)

dataset.columns = [‘Ambewela’, ‘Annfield’, ‘Campion’,’Helboda’,’Holmwood’,’Hatton Police Station’, ‘Labukelle’,’Sandringham’,’Watawala’, ‘El – Nino’, ‘Inflow’]

dataset.index.name = ‘Date’

print(“#################test1###############”)

print(dataset.head())

values = dataset.values

groups = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10]

# normalize features

scaler = MinMaxScaler(feature_range=(0, 1))

scaled = scaler.fit_transform(values)

print(“#################test6###############”)

print(scaled)

days=1

test_X = scaled[:days, :]

print(“#################test9###############”)

print(test_X)

test_X = test_X.reshape((test_X.shape[0], 1, test_X.shape[1]))

# load model from single file

model = load_model(‘modelTest.h5’)

yhat = model.predict(test_X)

print(“#################test10###############”)

print(yhat)

#test_X = test_X.reshape((test_X.shape[0], test_X.shape[2]))

inv_yhat = concatenate((yhat, test_X), axis=1)

inv_yhat = scaler.inverse_transform(inv_yhat)

inv_yhat = inv_yhat[:,0]

print(“#################test11###############”)

print(inv_yhat)

however i noticed that the 0,1 min max scaler values of the dataset is all zeros.

this is the output i got:

[1 rows x 11 columns]

#################test6###############

[[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]]

#################test9###############

[[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]]

#################test10###############

[[0.24231862]]

Traceback (most recent call last):

File “”, line 1, in

runfile(‘F:/documents/Final Year Project/2 python test code/load model.py’, wdir=’F:/documents/Final Year Project/2 python test code’)

File “F:\software\anaconda 3\lib\site-packages\spyder_kernels\customize\spydercustomize.py”, line 786, in runfile

execfile(filename, namespace)

File “F:\software\anaconda 3\lib\site-packages\spyder_kernels\customize\spydercustomize.py”, line 110, in execfile

exec(compile(f.read(), filename, ‘exec’), namespace)

File “F:/documents/Final Year Project/2 python test code/load model.py”, line 60, in

inv_yhat = concatenate((yhat, test_X), axis=1)

ValueError: all the input arrays must have same number of dimensions

i wanna know whats wrong with the minmax scaler and why i am only getting one output?

This is a common question that I answer here:

https://machinelearningmastery.com/faq/single-faq/can-you-read-review-or-debug-my-code

Hi Dr. Jason

Thank you for the tutorial, it is really interested. I am just wondering is it possible to have the output of the prediction using LSTM not as a crisp value? In other words, if I have to predict 5 different classes, I want to have the results as degree of belonging such as 0.6 for class one, 0.2 for class two and so on.

Thanks in advance

Yes, you can predict probabilities using predict() instead of predict_classes().

See examples here:

https://machinelearningmastery.com/how-to-make-classification-and-regression-predictions-for-deep-learning-models-in-keras/

Hi Jason

Thank you very much for answering my question, really appreciated that. I am also wondering if the same predict probabilities could be done using CNN instead of LSTM and getting the same output vector?

Thanks in advance

Yes.

hi Dr.. sorry i am very new

where did you define ythat exactly i mean i have the model i want to predict with one sentence

my model is seq2seq model

yhat is the out put of the model when you call predict()

The output is whatever you trained your model to do.

hi Dr. jason

i am still searching though the internet and trying to find the solution but i can not find it

i have a machine translation seq2seq build with keras and now i want to make a prediction

but i am getting a weird error and i do not know how and exactly what is happening with the shape of my model .

code:

model = load_model(‘model.h5’)

single_x_test = [‘how are you’]

model.predict(np.array(single_x_test, ndmin=0))

error:

model = load_model(‘model.h5’)

single_x_test = [‘how are you’]

model.predict(np.array(single_x_test, ndmin=0))

I don’t see an error. Perhaps try posting your code to stackoverflow.

Hi Jason,

I appreciate your work. Its quite helpful. I have trained and tested a CNN-LSTM model but I am having problem in using the saved .h5 model file for real time predictions with the help of webcam. If you could please guide about that.

Regards

Thanks!

Yes, see this:

https://machinelearningmastery.com/how-to-make-classification-and-regression-predictions-for-deep-learning-models-in-keras/

Hey Jason, your tutorials are very impressive!!

i have a simple question, do you have any example on Predicting the future on a timeseries problem but beyond the dataset (i mean generating new dates that we dont have in the dataset)

Thanks again for all your explanations!!

The above tutorial explores this exactly.

Hi Jason,

1, You used time_steps = 1 for above example. Now, i have new data point (for example

0.5) i have to reshape to [1,1,1] and predict. Is that correct?

2, One more question: above example, you don’t tranform any data. I mean with some train

data like stock prices, we have to scale data such Minmax (rang 0 to 1) or Standard. Now

we have present price (ex: 118,12 not in range(0,1)). How can we predict tomorow price?

Thanks !

No, the input shape should match the expectation of the model.

Yes, it can be a good idea to scale data prior to modeling, and any scaling applied to the training dataset should be applied to new data.

Yes, but with time step equal 1 mean that we predict next day, and sample has shape [1, 1, 1]. So how can we scale data with only one point?

I don’t understand, sorry. Perhaps you can rephrase your question?

Sorry Jason,

I mean If the model expects 1 input to predict 1 output, so:

1. Is it impossible?

2. For example X_train = [ [6], [7], [8] ] we want result Predict Y_train = [ [7], [8], [9] ], but at step of processing data, we scale data X_train = scaler.tranform(X_train). Now, we have new data a = [9], how can we predict?

Thanks!

If the model expects 3 inputs, then during training and during inference, you must provide 3 inputs.

No, i mean X_train has shape [n,1,1]. I scale X_train by standard tranform and want to predict output with [n,1,1] shape. Now with new data [1,1,1], how to predict?

Call it directly, e.g.

PLease how do I use the model to forecast future dates

You can make a forecast by calling model.predict(), for more help see this:

https://machinelearningmastery.com/make-predictions-long-short-term-memory-models-keras/

And this:

https://machinelearningmastery.com/how-to-make-classification-and-regression-predictions-for-deep-learning-models-in-keras/

Hey Jason,

very nice tutorial and explanation.

I’m using a multivariate LSTM approach with multiple LSTM layers including selfattention for time series data. I have 60 time stamps, 12 variables and about 30k samples. I’m using the data to predict the gain for time stamp 65. The gain is defined as the percentage difference between the closing value at time stamp 60 and the closing value at time stamp 65. This is defined as the target variable. In this case LSTM is not really used to predict the future rather then a specific target variable. Therefore, I think I don’t use the full potential of LSTM.

Lets say I would like to tell the model that 11 of these variables influence lets say the closing value and then use LSTM to predict the closing value for time stamp 65. How can I implement that? In short, what would be the synthax to tell the model that 11 specific variables influence the closing value and based on that, what would be the prediction of the closing value for time stamp 65?

Thanks in advance,

Daniel

Thanks.

You don’t really. You provide the data to the model that you believe is most relevant and let it learn what most influences the target and in what way. This is called function approximation:

https://machinelearningmastery.com/neural-networks-are-function-approximators/

Good

Thanks!

Hi Jason ,

Thanks for great your post. How can I see the future when the true outcome for the sequence is currently unknown.I couldn’t find anything about this question .

For example ; I have a dataset between on 2018-2019 but ı want to see the predicts on 2020.

Yes, the model predicts observations for the future and you use them in the application directly.

If you are having trouble making a prediction, see this example:

https://machinelearningmastery.com/make-predictions-long-short-term-memory-models-keras/

Thanks for your answer , Jason.

I

You’re welcome.

Thnks for your post, but how to trian the model with GUI?

You can write your own GUI I guess.

Hi,I have a dataset.

At time 0- 3 rows 4 columns

At time 1-3 rows 4 columns

I want to predict

At time 2

The value of last row,last column element

Perhaps this will help you prepare your data:

https://machinelearningmastery.com/convert-time-series-supervised-learning-problem-python/

Hi Jason,

Thanks for your post. Your teaching of deep learning is really useful!

I am implementing a RNN model for predicting the output based on the previous output (1 time step) and 3 input.

In your post “Multivariate Time Series Forecasting with LSTMs in Keras”, you taught us to use the previous output as one of the input. However, our output “variable 1” is blank at the beginning. We would not know the output before the prediction.

In this case, how can we prepare the input and output for prediction?

For example, var1 is output, var2, var3 and var4 are input.

var1[t-1], var2[t-1], var3[t-1] and var4[t-1] are used for predicting var1[t]

However, I have only the initial value of var1 which is 8*10^6.

How can I update the value of var1[t-1] with the predicted output during the prediction?

Without update, all the rows of var1 will be NAN (not a number) except the first row with initial value. My model can predict nothing with NAN.

For scaling (0~1), it is also a problem to have only the initial value.

My RNN_save_model.py is simply like your code in “Multivariate Time Series Forecasting with LSTMs in Keras”. There is no problem for training and testing as var1 (output) is already known in the training data.

Thanks for your time!

The first row must be removed as their is no prior observation.

Thanks for response.

My question may be not clear enough.

I am confused that how can I feedback output to input using RNN.

I need var1[t-1], var2[t-1], var3[t-1] and var4[t-1] to predict var1[t]

However, var1[t-1] is not known until var1[t-1] is predicted using var1[t-2].

I only have the initial value var1[0]. var1[1], var1[2], var[3], ……. are unknown until the prediction is done.

var2, var3 and var4 are already known.

My goal is to predict var1 using LSTM.

You can do it manually one sample at a time. E.g. predict then construct the next sample, then predict. This is called a recursive model.

Thanks for response!

If I am going to use a recursive model, can I keep the same method of training and testing before saving the model?

Should I train and test the model using another method?

I am using the same method to train and test my model like this post: https://machinelearningmastery.com/multivariate-time-series-forecasting-lstms-keras/#comment-540267

You can assume that I am using the same code in the above link to make prediction using the save and load method in this post.

I would recommend evaluating the model in an identical manner in which you intend to you use.

Hi Jason, thanks for your suggestion

I have tried to make prediction with LSTM model and feedback the prior predicted output to current input.

However, I cannot invert scaling for forecast as I have only the initial value (8*10^6) for the output. I used it to predict the next output and use the next output to predict the further output. There is other known input data like temperature and humidity.

Let’s call the output “stiffness”. stiffness(t) is the output and stiffness(t-1) is one of the input to predict stiffness(t).

The initial value of stiffness is scale to 0 in the scaling between 0 and 1 as only the first sample exists among 10000 samples at the beginning.

At a result, after the prediction and invert scaling, it always give me a horizontal line of 8*10^6.

How can I turn the scale between 0 and 1 back to the original scaling in this case?

One thought is run the transform manually to understand what is going on and ensure you are applying it on the data in the same form both times.

Hi Mr.Brownlee,

Thank you for the excellent article. I have tried following the steps mentioned here and a few others which i have come across in other places and built a model but I have encountered some issues which i have mentioned in the following question on stack overflow:

https://stackoverflow.com/questions/63318474/training-output-drops-to-0-after-fixed-timesteps-

and-again-retrains-in-lstm-mode

I’d appreciate it if you could have a look and help me out.

Thanks and Regards,

Gopal.

This is a common question that I answer here:

https://machinelearningmastery.com/faq/single-faq/can-you-comment-on-my-stackoverflow-question

Hi Jason! Awesome tutorial. I have a question, how can I try to reuse KERAS if my training data set is in this format :

[

[2, 5, 6, 15, 22],

[12, 8, 2, 33, 44],

…

]

in general 5 numbers, 1-49 each of them.

Perhaps this will help:

https://machinelearningmastery.com/faq/single-faq/what-is-the-difference-between-samples-timesteps-and-features-for-lstm-input