I am trying to train a LSTM model. Is this model suffering from overfitting?
Here is train and validation loss graph:
I am trying to train a LSTM model. Is this model suffering from overfitting?
Here is train and validation loss graph:
The model is overfitting right from epoch 10, the validation loss is increasing while the training loss is decreasing.
Dealing with such a Model:
There are many other options as well to reduce overfitting, assuming you are using Keras, visit this link.
Yes this is an overfitting problem since your curve shows point of inflection. This is a sign of very large number of epochs. In this case, model could be stopped at point of inflection or the number of training examples could be increased.
Also, Overfitting is also caused by a deep model over training data. In that case, you'll observe divergence in loss between val and train very early.
I had this issue - while training loss was decreasing, the validation loss was not decreasing. I checked and found while I was using LSTM:
(-1,1)
, I choose (0,1)
, this right there reduced my validation loss by the magnitude of one orderAnother possible cause of overfitting is improper data augmentation. If you're augmenting then make sure it's really doing what you expect.
I had a similar problem, and it turned out to be due to a bug in my Tensorflow data pipeline where I was augmenting before caching:
def get_dataset(inputfile, batchsize):
# Load the data into a TensorFlow dataset.
signals, labels = read_data_from_file(inputfile)
dataset = tf.data.Dataset.from_tensor_slices((signals, labels))
# Augment the data by dynamically tweaking each training sample on the fly.
dataset = dataset.map(
map_func=(lambda signals, labels: (tuple(tf.py_function(func=augment, inp=[signals], Tout=[tf.float32])), labels)))
# Oops! Should have called cache() before augmenting
dataset = dataset.cache()
dataset = ... # Shuffle, repeat, batch, etc.
return dataset
training_data = get_dataset("training.txt", 32)
val_data = //...
model.fit(training_data, validation_data=val_data, ...)
As a result, the training data was only being augmented for the first epoch. This caused the model to quickly overfit on the training data. Moving the augment call after cache() solved the problem.
It may be that you need to feed in more data, as well. If the model overfits, your dataset may be so small that the high capacity of the model makes it easily fit this small dataset, while not delivering out-of-sample performance. In other words, it does not learn a robust representation of the true underlying data distribution, just a representation that fits the training data very well.
Solutions as stated above:
I propose to extend your dataset (largely), which will be costly in terms of several aspects obviously, but it will also serve as a form of "regularization" and give you a more confident answer. In case you cannot gather more data, think about clever ways to augment your dataset by applying transforms, adding noise, etc to the input data (or to the network output).
It's not severe overfitting. So, here is my suggestions:
1- Simplify your network! Maybe your network is too complex for your data. If you have a small dataset or features are easy to detect, you don't need a deep network.
2- Add Dropout layers.
3- Use weight regularization. Here is the link for further information: https://keras.io/api/layers/regularizers/