I'm observing the following patterns in a one-layer CNN, binary classification model:
Training loss decreases while dev loss increased with number of steps
Training accuracy increases while dev accuracy decreases with number of steps
Based on past SO questions and literature review, it seems that these patterns are indicative of over-fitting (the model performs well in training, but cannot generalize to new examples).
The graphs below illustrate the loss and accuracy with respect to the number of steps in training.
In both,
The orange line represents the summary of the dev set performance.
The blue line represents the summary of the training set performance.
Loss:
Accuracy:
Traditional remedies I've considered, and my observations about them:
Adding L2 Regularization : I've tried many coefficients of L2 regularization -- from 0.0 to 4.5; all of these tests yield a similar pattern by the 5,000th step in both loss and accuracy.
Cross validation : It seems that the role of cross-validation is widely mis-understood online. As this answer states, cross-validation is for model checking, not model building. Indeed, cross-validation would be a way to check if the model generalizes well. And actually, the graphs I show are from one fold of a 4-fold cross-validation. If I observe a similar pattern in the loss/accuracy in all the folds, what other insight does cross-validation offer other than the confirmation that the model does not generalize well?
Early stopping : This would seem the most intuitive, but the loss graph seems to indicate that the loss levels out only after a divergence in the dev set loss is observed; the starting point of this early stop, then, doesn't seem easy to decide.
Data : The amount of labeled data I have available is limited, so training on more data is not an option right now.
All this said, what I am asking is:
If the patterns observed in the loss and accuracy are indeed indicative of over-fitting, are there any other methods to counteract over-fitting that I haven't considered?
If these patterns are not indicative of over-fitting, what else could they mean?
Thanks -- any insight would be much appreciated.
I think that you are totally on the right track. Looks like classic over-fitting.
One option is adding dropout if you don't already have it. It falls into the category of regularization, but it is more commonly used now then L1 and L2 regularization.
Changing the model architcture could get better results but it's hard to say what specifically would be best. It could help to make it deeper with more layers and possibly some pooling layers. It will likely still overfit but you might get a higher accuracy on the dev set before that happens.
Getting more data may be one of the best things you could do. If you can't get more data you can try to augment the data. You can also try cleaning the data to remove noise which can help prevent the model from fitting to noise.
You may ultimately want to try setting up a hyperparameter optimization search. This, however, can take a while on neural nets which take a while to train. Make sure you remove a test set before hyper parameter tuning.
Related
I am building ANN as below:-
model=Sequential()
model.add(Flatten(input_shape=(25,)))
model.add(Dense(25,activation='relu'))
model.add(Dropout(0.8))
model.add(Dense(16,activation='relu'))
model.add(Dropout(0.8))
model.add(Dense(5,activation='relu'))
model.add(Dense(1,activation='sigmoid'))
model.compile(optimizer='adam',loss='binary_crossentropy',metrics=['accuracy'])
model.fit(xtraindata,ytraindata,epochs=50)
test_loss,test_acc=model.evaluate(xtestdata,ytestdata)
print(test_acc)
I am adding different features into the model and checking whether the newly added feature decreases or increases the accuracy but the problem is that each time I run this code with the same values I get different accuracy, sometimes it gets as low as 0.50 and so, I have few doubts and kindly answer them:-
Is the model giving different accuracy each time because in dropout reg. there are random dropouts in nodes and each time I run diff. nodes get silenced so thereby giving different accuracies i.e sometimes low and sometimes high?
How can I trust the accuracy of the model if each time it gives different accuracies? How can I know that the feature I have added has resulted in a decrement or increment of the accuracy?
If I get high accuracy and wanted to reproduce these results how do I save the parameters that the model has used?
Great questions. Answers:
I think your theory is right; it's the dropout. That's the only layer with an element of randomness each run, so it's likely the culprit. Try removing that layer, leaving everything else fixed, and run multiple times. Check if the accuracy is the same.
Cross validation. This article explains how it works, but the gist is that it is a statistical technique that trains and checks the accuracy of multiple runs of your model, all with different slices of data. The average accuracy of all runs is used. So highs and lows will be averaged to a true(ish) accuracy. That being said, if your model has inconsistent results by just varying dropout, it's an indicator that when you move the model to production and use real data, it will perform poorly.
Keras api has a method model.save("model_name") to save models. You can use keras.models.load_models("model_name") to get it back. As I said in point 2 though; if your model is so finicky that some trainings drastically affect accuracy, then even if you train and get good accuracy, it probably won't be useful on new data. So when you say "If I get high accuracy and wanted to reproduce these results", really you shouldn't be thinking along these lines. Instead, try to get consistently high training accuracy.
I just wanted to ask a quick question. I understand that val_loss and train_loss is insufficient to tell if the model is overfitting. However, i wish to use it as a rough gauge by monitoring if the val_loss is increasing. As i use SGD optimiser, i seem to have 2 different trends based on the smoothing value. Which should i use? Blue is val_loss and Orange is train_loss.
From smoothing = 0.999, both seems to be decreasing but from smoothing = 0.927, val_loss seems to be increasing. Thank you for reading!
Also, when is a good time to decrease the learning rate? Is it directly before the model overfits?
Smoothing = 0.999
Smoothing = 0.927
In my experience with DL as applied to CNNs, overfitting is tied more to the difference in train/val accuracies/losses rather than just one or the other. In your graphs, it's clear that the difference in loss is increasing as time goes on, showing that your model does not generalize well to the dataset, and hence shows signs of overfitting. It would also help for you to track classification accuracy on train and val datasets if possible--this will show you the generalization error which acts as a similar metric but might show more visible effects.
Dropping the learning rate once the loss starts to even out and overfitting begins is a good idea; however you may find better gains for your generalization if you adjust the net's complexity to better fit the dataset first. For such overfitting, a modest decrease in complexity may help--use the difference in train/val losses and accuracies to confirm.
I am building a predictive model where I want to know can I predict whether a package will be delivered on time (Binary Yes / No), in the event that the package is not delivered on time, I wish to be able to predict by when it will be delivered in categories of <7days, <14days, <21days >28days after expected date.
I have built and tested a model for binary classification and have got an f Score of 0.92, which is satisfactory for my needs. However, when I train my categorical model, I start to see training accuracy and validation accuracy diverge (training accuracy is much better than validation accuracy). This is a sign of overfitting.
However, I have tried regularization and different values, plus using dropout and different values, and the validation accuracy never gets above 0.7. My total training set is of ~10k examples, ~3k validation, and whilst the catgorical spread is not equal there are sufficient examples of each category (I think). I am using a NN and have increased / decreased both layers and activations and still no joy
Any thoughts on where to go next. Thanks
Because you are using NN, introduce dropout layers. See if it can help to reduce the overfitting problem. And also checkout this How to choose the number of hidden layers and nodes in a feedforward neural network?
The more complex the network (hidden layers, number of neurons in them), also contribute to overfitting problem
The approach we have chosen is to carry out a linear regression with the expected duration as target variable. We have excluded some outliers, and then taken the differences between the actual and predicted days. We then max'd and min'd the difference and we now have a prediction with a tolerable range. We will keep working on the other techniques to see if we can improve. Thanks to everyone who suggested ideas
Most of my code is based on this article and the issue I'm asking about is evident there, but also in my own testing. It is a sequential model with LSTM layers.
Here is a plotted prediction over real data from a model that was trained with around 20 small data sets for one epoch.
Here is another plot but this time with a model trained on more data for 10 epochs.
What causes this and how can I fix it? Also that first link I sent shows the same result at the bottom - 1 epoch does great and 3500 epochs is terrible.
Furthermore, when I run a training session for the higher data count but with only 1 epoch, I get identical results to the second plot.
What could be causing this issue?
A few questions:
Is this graph for training data or validation data?
Do you consider it better because:
The graph seems cool?
You actually have a better "loss" value?
If so, was it training loss?
Or validation loss?
Cool graph
The early graph seems interesting, indeed, but take a close look at it:
I clearly see huge predicted valleys where the expected data should be a peak
Is this really better? It sounds like a random wave that is completely out of phase, meaning that a straight line would indeed represent a better loss than this.
Take a look a the "training loss", this is what can surely tell you if your model is better or not.
If this is the case and your model isn't reaching the desired output, then you should probably make a more capable model (more layers, more units, a different method, etc.). But be aware that many datasets are simply too random to be learned, no matter how good the model.
Overfitting - Training loss gets better, but validation loss gets worse
In case you actually have a better training loss. Ok, so your model is indeed getting better.
Are you plotting training data? - Then this straight line is actually better than a wave out of phase
Are you plotting validation data?
What is happening with the validation loss? Better or worse?
If your "validation" loss is getting worse, your model is overfitting. It's memorizing the training data instead of learning generally. You need a less capable model, or a lot of "dropout".
Often, there is an optimal point where the validation loss stops going down, while the training loss keeps going down. This is the point to stop training if you're overfitting. Read about the EarlyStopping callback in keras documentation.
Bad learning rate - Training loss is going up indefinitely
If your training loss is going up, then you've got a real problem there, either a bug, a badly prepared calculation somewhere if you're using custom layers, or simply a learning rate that is too big.
Reduce the learning rate (divide it by 10, or 100), create and compile a "new" model and restart training.
Another problem?
Then you need to detail your question properly.
I have a set of data in a .tsv file available here. I have written several classifiers to decide whether a given website is ephemeral or evergreen.
Now, I want to make them better. I know from speaking with people that my classifier is 'overfitting' the data; what I am looking for is a solid way to prove this so that the next time I write a classifier I will be able to run a test and see if I am overfitting or underfitting.
What is the best way of doing this? I am open to all suggestion!
I've spent literally weeks googling this topic and found no canonical or trusted ways to do this effectively, so any response will be appreciated. I will be putting a bounty on this question.
Edit:
Let's assume my clasifier spits out a .tsv containing :
the website UID<tab>the likelihood it is to be ephemeral or evergreen, 0 being ephemeral, 1 being evergreen<tab>whether the page is ephemeral or evergreen
The most simple way to check your classifier "efficiency" is to perform a cross validation:
Take your data, lets call them X
Split X into K batches of equal sizes
For each i=1 to K:
Train your classifier on all batches but i'th
Test on i'th
Return the average result
One more important aspect - if your classifier uses any parameters, some constants, thresholds etc. which are not trained, but rather given by the user you cannot just select the ones giving the best results in the above procedure. This has to be somehow automatized in the "Train your classifier on all batches but i'th". In other words - you cannot use the testing data to fit any parameters to your model. Once done this, there are four possible outcomes:
Training error is low but is much lower than testing error - overfitting
Both errors are low - ok
Both errors are high - underfitting
Training error is high but testing is low - error in implementation or very small dataset
There are many ways that people try to handle overfitting:
Cross-validation, you might also see it mentioned as x-validation
see lejlot's post for details
choose a simpler model
linear classifiers have a high bias because the model must be linear but lower variance in the optimal solution because of the high bias. This means that you wouldn't expect to see much difference in the final model given a large number of random training samples.
Regularization is a common practice to combat overfitting.
It is generally done by adding a term to the minimization function
Typically this term is the sum of squares of the model's weights because it is easy to differentiate.
Generally there is a constant C associated with the regularization term. Tuning this constant will increase / decrease the effect of regularization. A high weight applied to regularization generally helps with overfitting. C should always be greater or equal to zero. (Note: some training packages apply 1/C as the regularization weight. In this case, the close C gets to zero the greater weight is applied to regularization)
Regardless of the specifics, regularization works by reducing the variance in a model by biasing it to solutions with low regularization weight.
Finally, boosting is a method of training that mysteriously/magically does not overfit. Not sure if anyone has discovered why, but it is a process of combining high bias low variance simple learns into a high variance low bias model. Its pretty slick.