How to get output from a layer during training in Keras?

Click For Summary
SUMMARY

To obtain the output from a hidden layer during training in Keras, users can leverage Keras callbacks, specifically the `LambdaCallback`. This allows for capturing the output of any layer after each training sample is evaluated. The discussion highlights the need to modify the training process to access the outputs dynamically, rather than waiting until the training is complete. The provided code demonstrates a multivariate time series prediction model using Keras, which can be adapted to include this functionality.

PREREQUISITES
  • Familiarity with Keras 2.x and TensorFlow 2.x frameworks
  • Understanding of neural network architectures, specifically CNNs
  • Knowledge of Keras callbacks and their implementation
  • Basic proficiency in Python programming and NumPy for data manipulation
NEXT STEPS
  • Implement Keras `LambdaCallback` to capture layer outputs during training
  • Explore Keras functional API for more complex model architectures
  • Learn about TensorFlow's eager execution for real-time model evaluation
  • Investigate advanced Keras techniques for monitoring training metrics
USEFUL FOR

Data scientists, machine learning engineers, and developers working with Keras who need to monitor and analyze intermediate outputs during model training.

Atr cheema
Messages
67
Reaction score
0
In Keras, the method model.fit() is used to train the neural network. How can I get the output from any hidden layer during training? Consider following code where neural network is trained to add two time series
Python: 
#multivariate data preparation
#multivariate multiple input cnn example
from numpy import array
from numpy import hstackfrom tensorflow.python.keras.models import Sequential
from tensorflow.python.keras.layers import Dense
from tensorflow.python.keras.layers import Flatten
from tensorflow.python.keras.layers.convolutional import Conv1D
from tensorflow.python.keras.layers.convolutional import MaxPooling1D#split a multivariate sequence into samples
def split_sequences(sequences, n_steps):
    X, y = list(), list()
    for i in range(len(sequences)):
        # find the end of this pattern
        end_ix = i + n_steps
        # check if we are beyond the dataset
        if end_ix > len(sequences):
            break
        # gather input and output parts of the pattern
        seq_x, seq_y = sequences[i:end_ix, :-1], sequences[end_ix-1, -1]
        X.append(seq_x)
        y.append(seq_y)
    return array(X), array(y)

#define input sequence
dataset_len = 100
is1 = array([i*10+10 for i in range(dataset_len)])
is2 = array([i*10+15 for i in range(dataset_len)])

out_seq = array([is1[i]+is2[i]
                for i in range(len(is1))])
#convert to [rows, columns] structure
is1 = is1.reshape((len(is1), 1))
is2 = is2.reshape((len(is2), 1))

out_seq = out_seq.reshape((len(out_seq), 1))
#horizontally stack columns
dataset = hstack((is1, is2,
                  #is3,is4,is5,is6,is7,is8,is9,
                  out_seq))
#print('raw input data is : \n{}\n'.format(dataset[0:7]))

#choose a number of time steps
n_steps = 3
#convert into input/output
X, y = split_sequences(dataset, n_steps)

#print('given \n{} we want to predict {}\n'.format(X[0], y[0]))

#model input must have shape [samples, timesteps, features]

#the dataset knows the number of features, e.g. 2
n_features = X.shape[2]
#define model
model = Sequential()
conv1d = Conv1D(filters=64, kernel_size=2, activation='relu', input_shape=(n_steps, n_features))
model.add(conv1d)
mp1d = MaxPooling1D(pool_size=2)
model.add(mp1d)
fl = Flatten()
model.add(fl)
model.add(Dense(50, activation='relu'))
model.add(Dense(8, activation='relu'))
model.add(Dense(4, activation='relu'))
model.add(Dense(2, activation='relu'))
model.add(Dense(1))

#print('\nCompiling model\n')

model.compile(optimizer='adam', loss='mse')
#plot_model(model, to_file='multivariate_laos.png', show_shapes=True,)

#fit model
model.fit(X, y, epochs=1000, batch_size=None, verbose=0, shuffle=False, steps_per_epoch=None)##demonstrate prediction
x_input = array([[80, 85], [90, 95], [100, 105]])
x_input = x_input.reshape((1, n_steps, n_features))
yhat = model.predict(x_input, verbose=0)
print(yhat)`

I the above code, how can I get output of last hidden layer which is in this case dense layer, but I want the output during training of neural network. This means I want output every time a 'sample' is evaluated and its corresponding loss function is calculated.
 
Technology news on Phys.org
Borg said:
Maybe this will help?
https://stackoverflow.com/questions/41711190/keras-how-to-get-the-output-of-each-layer
All the answers in that post provide the way to get output once the training is finished but I want the output during the training process simultaneously. This might require some modification in model.fit method in training.py file but I am unable to find a solution so far. Can you tell me which line in Keras source code actually runs the tensorflow session during execution of model.fit() method. .
 
I don't know - I'm pretty new at Keras as well. I know that it's mainly a wrapper around Tensorflow to make life easier but, if you want something else, you may have to build your model without it. Maybe @jedishrfu knows a way.
 
The only thing i can think of is to manipulate your training using one element at a time.
 
jedishrfu said:
The only thing i can think of is to manipulate your training using one element at a time.
Can you implement this for this example code? Thanks.
 
Atr cheema said:
Can you implement this for this example code? Thanks.

Nope, the task is left to the student. :-)
 
Atr cheema said:
In Keras, the method model.fit() is used to train the neural network. How can I get the output from any hidden layer during training? Consider following code where neural network is trained to add two time series
Python: 
#multivariate data preparation
#multivariate multiple input cnn example
from numpy import array
from numpy import hstackfrom tensorflow.python.keras.models import Sequential
from tensorflow.python.keras.layers import Dense
from tensorflow.python.keras.layers import Flatten
from tensorflow.python.keras.layers.convolutional import Conv1D
from tensorflow.python.keras.layers.convolutional import MaxPooling1D#split a multivariate sequence into samples
def split_sequences(sequences, n_steps):
    X, y = list(), list()
    for i in range(len(sequences)):
        # find the end of this pattern
        end_ix = i + n_steps
        # check if we are beyond the dataset
        if end_ix > len(sequences):
            break
        # gather input and output parts of the pattern
        seq_x, seq_y = sequences[i:end_ix, :-1], sequences[end_ix-1, -1]
        X.append(seq_x)
        y.append(seq_y)
    return array(X), array(y)

#define input sequence
dataset_len = 100
is1 = array([i*10+10 for i in range(dataset_len)])
is2 = array([i*10+15 for i in range(dataset_len)])

out_seq = array([is1[i]+is2[i]
                for i in range(len(is1))])
#convert to [rows, columns] structure
is1 = is1.reshape((len(is1), 1))
is2 = is2.reshape((len(is2), 1))

out_seq = out_seq.reshape((len(out_seq), 1))
#horizontally stack columns
dataset = hstack((is1, is2,
                  #is3,is4,is5,is6,is7,is8,is9,
                  out_seq))
#print('raw input data is : \n{}\n'.format(dataset[0:7]))

#choose a number of time steps
n_steps = 3
#convert into input/output
X, y = split_sequences(dataset, n_steps)

#print('given \n{} we want to predict {}\n'.format(X[0], y[0]))

#model input must have shape [samples, timesteps, features]

#the dataset knows the number of features, e.g. 2
n_features = X.shape[2]
#define model
model = Sequential()
conv1d = Conv1D(filters=64, kernel_size=2, activation='relu', input_shape=(n_steps, n_features))
model.add(conv1d)
mp1d = MaxPooling1D(pool_size=2)
model.add(mp1d)
fl = Flatten()
model.add(fl)
model.add(Dense(50, activation='relu'))
model.add(Dense(8, activation='relu'))
model.add(Dense(4, activation='relu'))
model.add(Dense(2, activation='relu'))
model.add(Dense(1))

#print('\nCompiling model\n')

model.compile(optimizer='adam', loss='mse')
#plot_model(model, to_file='multivariate_laos.png', show_shapes=True,)

#fit model
model.fit(X, y, epochs=1000, batch_size=None, verbose=0, shuffle=False, steps_per_epoch=None)##demonstrate prediction
x_input = array([[80, 85], [90, 95], [100, 105]])
x_input = x_input.reshape((1, n_steps, n_features))
yhat = model.predict(x_input, verbose=0)
print(yhat)`

I the above code, how can I get output of last hidden layer which is in this case dense layer, but I want the output during training of neural network. This means I want output every time a 'sample' is evaluated and its corresponding loss function is calculated.
Hi Atr cheema,
I have the same problem, did you find a solution for your question?