Context: let's say I have trained a CNN on datasetA and I've obtained caffeModelA.
Current situation: new pictures arrive so I can build a new dataset, datasetB
Question: would these two situations lead to same caffemodel?
merge datasetA and datasetB and train the net from scratch.
perform some fine-tuning on existing caffeModelA by training it only on datasetB (as explained here: http://caffe.berkeleyvision.org/gathered/examples/finetune_flickr_style.html)
It might seem a dumb question, but I'm not really sure about its answer. And it's really important because if the two approximations lead to same result I can save time by performing number 2.
Note: bear in mind that it's the same problem, so no need to change architecture here, I just plan to add new images to the training.
In the Flicker-style example the situation is a bit more generic. They use the weights of first layers from a model trained for a different classification task and employ it for a new task, training only a new last layer and fine-tuning the first layers a bit (by setting a low learning rate for those pretrained layers). Your case is similar but more specific, you want to use the pretrained model to train the exact architecture for the exact same task but with an extension of your data.
If your question if whether Option 1. will produce exactly the same model (all resulting weights are equal) as Option 2. Then no, most probably not.
In Option 2. the network is trained for iterations of dataset A then for dataset B then dataset A again..and so on (assuming both were just concatenated together).
While in Option 1. will have the network trained for some iterations/epochs on dataset A, then later continue learning for iterations/epochs on only dataset B and that's it. So the solver will see a different sequence of gradients in both options resulting in two different models. That's from a strict theoretical perspective.
If you ask from a practical perspective, the two options will probably end up with very similar models. How many epochs (not iterations) did you train on dataset A ? say N epochs, then you can safely go with Option 2. and train your existing model further on dataset B for the same number of epochs and same learning rate and batch size.
Related
For a class project, I designed a neural network to approximate sin(x), but ended up with a NN that just memorized my function over the data points I gave it. My NN took in x-values with a batch size of 200. Each x-value was multiplied by 200 different weights, mapping to 200 different neurons in my first layer. My first hidden layer contained 200 neurons, each one a linear combination of the x-values in the batch. My second hidden layer also contained 200 neurons, and my loss function was computed between the 200 neurons in my second layer and the 200 values of sin(x) that the input mapped to.
The problem is, my NN perfectly "approximated" sin(x) with 0 loss, but I know it wouldn't generalize to other data points.
What did I do wrong in designing this neural network, and how can I avoid memorization and instead design my NN's to "learn" about the patterns in my data?
It is same with any machine learning algorithm. You have a dataset based on which you try to learn "the" function f(x), which actually generated the data. In real life datasets, it is impossible to get the original function from the data, and therefore we approximate it using something g(x).
The main goal of any machine learning algorithm is to predict unseen data as best as possible using the function g(x).
Given a dataset D you can always train a model, which will perfectly classify all the datapoints (you can use a hashmap to get 0 error on the train set), but which is overfitting or memorization.
To avoid such things, you yourself have to make sure that the model does not memorise and learns the function. There are a few things which can be done. I am trying to write them down in an informal way (with links).
Train, Validation, Test
If you have large enough dataset, use Train, Validation, Test splits. Split the dataset in three parts. Typically 60%, 20% and 20% for Training, Validation and Test, respectively. (These numbers can vary based on need, also in case of imbalanced data, check how to get stratified partitions which preserve the class ratios in every split). Next, forget about the Test partition, keep it somewhere safe, don't touch it. Your model, will be trained using the Training partition. Once you have trained the model, evaluate the performance of the model using the Validation set. Then select another set of hyper-parameter configuration for your model (eg. number of hidden layer, learaning algorithm, other parameters etc.) and then train the model again, and evaluate based on Validation set. Keep on doing this for several such models. Then select the model, which got you the best validation score.
The role of validation set here is to check what the model has learned. If the model has overfit, then the validation scores will be very bad, and therefore in the above process you will discard those overfit models. But keep in mind, although you did not use the Validation set to train the model, directly, but the Validation set was used indirectly to select the model.
Once you have selected a final model based on Validation set. Now take out your Test set, as if you just got new dataset from real life, which no one has ever seen. The prediction of the model on this Test set will be an indication how well your model has "learned" as it is now trying to predict datapoints which it has never seen (directly or indirectly).
It is key to not go back and tune your model based on the Test score. This is because once you do this, the Test set will start contributing to your mode.
Crossvalidation and bootstrap sampling
On the other hand, if your dataset is small. You can use bootstrap sampling, or k-fold cross-validation. These ideas are similar. For example, for k-fold cross-validation, if k=5, then you split the dataset in 5 parts (also be carefull about stratified sampling). Let's name the parts a,b,c,d,e. Use the partitions [a,b,c,d] to train and get the prediction scores on [e] only. Next, use the partitions [a,b,c,e] and use the prediction scores on [d] only, and continue 5 times, where each time, you keep one partition alone and train the model with the other 4. After this, take an average of these scores. This is indicative of that your model might perform if it sees new data. It is also a good practice to do this multiple times and perform an average. For example, for smaller datasets, perform a 10 time 10-folds cross-validation, which will give a pretty stable score (depending on the dataset) which will be indicative of the prediction performance.
Bootstrap sampling is similar, but you need to sample the same number of datapoints (depends) with replacement from the dataset and use this sample to train. This set will have some datapoints repeated (as it was a sample with replacement). Then use the missing datapoins from the training dataset to evaluate the model. Perform this multiple times and average the performance.
Others
Other ways are to incorporate regularisation techniques in the classifier cost function itself. For example in Support Vector Machines, the cost function enforces conditions such that the decision boundary maintains a "margin" or a gap between two class regions. In neural networks one can also do similar things (although it is not same as in SVM).
In neural network you can use early stopping to stop the training. What this does, is train on the Train dataset, but at each epoch, it evaluates the performance on the Validation dataset. If the model starts to overfit from a specific epoch, then the error for Training dataset will keep on decreasing, but the error of the Validation dataset will start increasing, indicating that your model is overfitting. Based on this one can stop training.
A large dataset from real world tends not to overfit too much (citation needed). Also, if you have too many parameters in your model (to many hidden units and layers), and if the model is unnecessarily complex, it will tend to overfit. A model with lesser pameter will never overfit (though can underfit, if parameters are too low).
In the case of you sin function task, the neural net has to overfit, as it is ... the sin function. These tests can really help debug and experiment with your code.
Another important note, if you try to do a Train, Validation, Test, or k-fold crossvalidation on the data generated by the sin function dataset, then splitting it in the "usual" way will not work as in this case we are dealing with a time-series, and for those cases, one can use techniques mentioned here
First of all, I think it's a great project to approximate sin(x). It would be great if you could share the snippet or some additional details so that we could pin point the exact problem.
However, I think that the problem is that you are overfitting the data hence you are not able to generalize well to other data points.
Few tricks that might work,
Get more training points
Go for regularization
Add a test set so that you know whether you are overfitting or not.
Keep in mind that 0 loss or 100% accuracy is mostly not good on training set.
I have to solve 2 class classification problem.
I have 2 classifiers that output probabilities. Both of them are neural networks of different architecture.
Those 2 classifiers are trained and saved into 2 files.
Now I want to build meta classifier that will take probabilities as input and learn weights of those 2 classifiers.
So it will automatically decide how much should I "trust" each of my classifiers.
This model is described here:
http://rasbt.github.io/mlxtend/user_guide/classifier/StackingClassifier/#stackingclassifier
I plan to use mlxtend library, but it seems that StackingClassifier refits models.
I do not want to refit because it takes very huge amount of time.
From the other side I understand that refitting is necessary to "coordinate" work of each classifier and "tune" the whole system.
What should I do in such situation?
I won't talk about mlxtend because I haven't worked with it but I'll tell you the general idea.
You don't have to refit these models to the training set but you have to refit them to parts of it so you can create out-of-fold predictions.
Specifically, split your training data in a few pieces (usually 3 to 10). Keep one piece (i.e. fold) as validation data and train both models on the other folds. Then, predict the probabilities for the validation data using both models. Repeat the procedure treating each fold as a validation set. In the end, you should have the probabilities for all data points in the training set.
Then, you can train a meta-classifier using these probabilities and the ground truth labels. You can use the trained meta-classifier on your new data.
I trained GoogLeNet model from scratch. But it didn't give me the promising results.
As an alternative, I would like to do fine tuning of GoogLeNet model on my dataset. Does anyone know what are the steps should I follow?
Assuming you are trying to do image classification. These should be the steps for finetuning a model:
1. Classification layer
The original classification layer "loss3/classifier" outputs predictions for 1000 classes (it's mum_output is set to 1000). You'll need to replace it with a new layer with appropriate num_output. Replacing the classification layer:
Change layer's name (so that when you read the original weights from caffemodel file there will be no conflict with the weights of this layer).
Change num_output to the right number of output classes you are trying to predict.
Note that you need to change ALL classification layers. Usually there is only one, but GoogLeNet happens to have three: "loss1/classifier", "loss2/classifier" and "loss3/classifier".
2. Data
You need to make a new training dataset with the new labels you want to fine tune to. See, for example, this post on how to make an lmdb dataset.
3. How extensive a finetuning you want?
When finetuning a model, you can train ALL model's weights or choose to fix some weights (usually filters of the lower/deeper layers) and train only the weights of the top-most layers. This choice is up to you and it ususally depends on the amount of training data available (the more examples you have the more weights you can afford to finetune).
Each layer (that holds trainable parameters) has param { lr_mult: XX }. This coefficient determines how susceptible these weights to SGD updates. Setting param { lr_mult: 0 } means you FIX the weights of this layer and they will not be changed during the training process.
Edit your train_val.prototxt accordingly.
4. Run caffe
Run caffe train but supply it with caffemodel weights as an initial weights:
~$ $CAFFE_ROOT/build/tools/caffe train -solver /path/to/solver.ptototxt -weights /path/to/orig_googlenet_weights.caffemodel
Fine-tuning is a very useful trick to achieve a promising accuracy compared to past manual feature. #Shai already posted a good tutorial for fine-tuning the Googlenet using Caffe, so I just want to give some recommends and tricks for fine-tuning for general cases.
In most of time, we face a task classification problem that new dataset (e.g. Oxford 102 flower dataset or Cat&Dog) has following four common situations CS231n:
New dataset is small and similar to original dataset.
New dataset is small but is different to original dataset (Most common cases)
New dataset is large and similar to original dataset.
New dataset is large but is different to original dataset.
In practice, most of time we do not have enough data to train the network from scratch, but may be enough for pre-trained model. Whatever which cases I mentions above only thing we must care about is that do we have enough data to train the CNN?
If yes, we can train the CNN from scratch. However, in practice it is still beneficial to initialize the weight from pre-trained model.
If no, we need to check whether data is very different from original datasets? If it is very similar, we can just fine-tune the fully connected neural network or fine-tune with SVM. However, If it is very different from original dataset, we may need to fine-tune the convolutional neural network to improve the generalization.
We all know that the objective function of SVM is iteratively trained. In order to continue training, at least we can store all the variables used in the iterations if we want to continue on the same training dataset.
While, if we want to train on a slightly different dataset, what should we do to make full use of the previously trained model? Or does this kind of thought make sense? I think it is quite reasonable if we train a K-means model. But I am not sure if it still makes sense for the SVM problem.
There are some literature on this topic:
alpha-seeding, in which the training data is divided into chunks. After you train a SVM on the ith chunk, you take those and use them to train your SVM with the (i+1)th chunk.
Incremental SVM serves as an online learning in which you update a classifier with new examples rather than retrain the entire data set.
SVM heavy package with online SVM training as well.
What you are describing is what an online learning algorithm does and unfortunately the classic definition for SVM is done in a batch fashion.
However, there are several solvers for SVM that produces quasy optimal hypothesis to the underneath optimization problem in an online learning way. In particular my favourite is Pegasos-SVM which can find a good near optimal solution in linear time:
http://ttic.uchicago.edu/~nati/Publications/PegasosMPB.pdf
In general this doesn't make sense. SVM training is an optimization process with regard to every training set vector. Each training vector has an associated coefficient, which as a result is either 0 (irrelevant) or > 0 (support vector). Adding another training vector imposes another, different, optimization problem.
The only way to reuse information from previous training I can think of is to choose support vectors from the previous training and add them to the new training set. I'm not sure, but this probably will negatively affect generalization - VC dimension of an SVM is related to the number of support vectors, so adding previous support vectors to the new dataset is likely to increase the support vector count.
Apparently, there are more possibilities, as noted in lennon310's answer.
I am considering using random forest for a classification problem. The data comes in sequences. I plan to use first N(500) to train the classifier. Then, use the classifier to classify the data after that. It will make mistakes and the mistakes sometimes can be recorded.
My question is: can I use those mis-classified data to retrain the original classifier and how? If I simply add the mis-classified ones to original training set with size N, then the importance of the mis-classified ones will be exaggerated as the corrected classified ones are ignored. Do I have to retrain the classifier using all data? What other classifiers can do this kind of learning?
What you describe is a basic version of the Boosting meta-algorithm.
It's better if your underlying learner have a natural way to handle samples weights. I have not tried boosting random forests (generally boosting is used on individual shallow decision trees with a depth limit between 1 and 3) but that might work but will likely be very CPU intensive.
Alternatively you can train several independent boosted decision stumps in parallel with different PRNG seed values and then aggregate the final decision function as you would do with a random forests (e.g. voting or averaging class probability assignments).
If you are using Python, you should have a look at the scikit-learn documentation on the topic.
Disclaimer: I am a scikit-learn contributor.
Here is my understanding of your problem.
You have a dataset and create two subdata set with it say, training dataset and evaluation dataset. How can you use the evaluation dataset to improve classification performance ?
The point of this probleme is'nt to find a better classifier but to find a good way for the evaluation, then have a good classifier in the production environnement.
Evaluation purpose
As the evaluation dataset has been tag for evaluation there is now way yo do this. You must use another way for training and evaluation.
A common way to do is cross-validation;
Randomize your samples in your dataset. Create ten partitions from your initial dataset. Then do ten iteration of the following :
Take all partitions but the n-th for training and do the evaluation with the n-th.
After this take the median of the errors of the ten run.
This will give you the errors rate of yours classifiers.
The least run give you the worst case.
Production purpose
(no more evaluation)
You don't care anymore of evaluation. So take all yours samples of all your dataset and give it for training to your classifier (re-run a complet simple training). The result can be use in production environnement, but can't be evaluate any more with any of yours data. The result is as best as the worst case in previous partitions set.
Flow sample processing
(production or learning)
When you are in a flow where new samples are produce over time. You will face case where some sample correct errors case. This is the wanted behavior because we want the system to
improve itself. If you just correct in place the leaf in errors, after some times your
classifier will have nothing in common with the original random forest. You will be doing
a form of greedy learning, like meta taboo search. Clearly we don't wanna this.
If we try to reprocess all the dataset + the new sample every time a new sample is available we will experiment terrible low latency. The solution is like human, sometime
a background process run (when service is on low usage), and all data get a complet
re-learning; and at the end swap old and new classifier.
Sometime the sleep time is too short for a complet re-learning. So you have to use node computing clusturing like that. It cost lot of developpement because you probably need to re-write the algorithms; but at that time you already have the bigest computer you could have found.
note : Swap process is very important to master. You should already have it in your production plan. What do you do if you want to change algorithms? backup? benchmark? power-cut? etc...
I would simply add the new data and retrain the classifier periodically if it weren't too expensive.
A simple way to keep things in balance is to add weights.
If you weigh all positive samples by 1/n_positive and all negative samples by 1/n_negative ( including all the new negative samples you're getting ), then you don't have to worry about the classifier getting out of balance.