caffe batchnorm layer affect base_lr? - neural-network

I saw the following code under caffe framework. The whole code is trying to write caffe train_val.prototxt and solver.prototxt.
# Use different initial learning rate.
if use_batchnorm:
base_lr = 0.0004
else:
base_lr = 0.00004
Why is the base learning rate different?

You could try if this paper helps.
They say that if you don't use normalization, you need to train "more carefully", meaning using a lower learning rate.
Skimming the first pages, I could imagine it works like this:
For some nonlinearities, there's a 'good input value range', and batch norm brings values into that range. High input values are bad and lead to saturation (little slope in the function and "vanishing gradients").
So, if you don't normalize, you need to make smaller steps - a lower learning rate - to avoid 'jumping' into weights that lead to high values within the net. And also be more careful how you initialize weights. I guess if you use ReLus, that's not so much of a problem. But please correct me if someone else has had different experiences with ReLus.

Related

why is my Neural Network stuck at high loss value after the first epochs

I'm doing regression using Neural Networks. It should be a simple task for NN to do, I have 10 features and 1 output that I want to predict.I’m using pytorch for my project but my Model is not learning well. the loss start with a very high value (40000), then after the first 5-10 epochs the loss decrease rapidly to 6000-7000 and then it stuck there, no matter what I make. I tried even to change to skorch instead of pytorch so that I can use cross validation functionality but that also didn’t help. I tried different optimizers and added layers and neurons to the network but that didn’t help, it stuck at 6000 which is a very high loss value. I’m doing regression here, I have 10 features and I’m trying to predict one continuous value. that should be easy to do that’s why it is confusing me more.
here is my network:
I tried here all the possibilities from making more complex architectures like adding layers and units to batch normalization, changing activations etc.. nothing have worked
class BearingNetwork(nn.Module):
def __init__(self, n_features=X.shape[1], n_out=1):
super().__init__()
self.model = nn.Sequential(
nn.Linear(n_features, 512),
nn.BatchNorm1d(512),
nn.LeakyReLU(),
nn.Linear(512, 64),
nn.BatchNorm1d(64),
nn.LeakyReLU(),
nn.Linear(64, n_out),
# nn.LeakyReLU(),
# nn.Linear(256, 128),
# nn.LeakyReLU(),
# nn.Linear(128, 64),
# nn.LeakyReLU(),
# nn.Linear(64, n_out)
)
def forward(self, x):
out = self.model(x)
return out
and here are my settings:
using skorch is easier than pytorch. here I'm monitoring also the R2 metric and I made RMSE as a custom metric to also monitor the performance of my model. I also tried the amsgrad for Adam but that didn't help.
R2 = EpochScoring(r2_score, lower_is_better=False, name='R2')
explained_var_score = EpochScoring(EVS, lower_is_better=False, name='EVS Metric')
custom_score = make_scorer(RMSE)
rmse = EpochScoring(custom_score, lower_is_better=True, name='rmse')
bearing_nn = NeuralNetRegressor(
BearingNetwork,
criterion=nn.MSELoss,
optimizer=optim.Adam,
optimizer__amsgrad=True,
max_epochs=5000,
batch_size=128,
lr=0.001,
train_split=skorch.dataset.CVSplit(10),
callbacks=[R2, explained_var_score, rmse, Checkpoint(), EarlyStopping(patience=100)],
device=device
)
I also standardize the Input values.
my Input have the shape:
torch.Size([39006, 10])
and shape of output is:
torch.Size([39006, 1])
I’m using 128 as my Batch_size but I also tried other values like 32, 64, 512 and even 1024. Although normalizing output is not necessary but I also tried that and It didn’t work when I predict values, the loss is high. Please someone help me on this, I would appreciate every helpful advice. I ll also add a screenshot of my training and val losses and metrics over epochs to visualize how the loss is decreasing in the first 5 epochs and then it stays like forever at the value 6000 which is a very high value for a loss.
considering that your training and dev loss are decreasing over time, it seems like your model is training correctly. With respect to your worry regarding your training and dev loss values, this is entirely dependent on the scale of your target values (how big are your target values?) and the metric used to compute the training and dev losses. If your target values are big and you want smaller train and dev loss values, you can normalise the target values.
From what I gather with respect to your experiments as well as your R2 scores, it seems that you are looking for a solution in the wrong area. To me, it seems like your features aren't strong enough considering that your R2 scores are low, which could mean that you have a data quality issue. This would also explain why your architecture tuning has not improved your model's performance as it is not your model that is the issue. So if I were you, I would think about what new useful features I could add and see if that helps. In machine learning, the general rule is that models are only as good as the data that they are trained on. I hope this helps!
The metric you should be looking at is R^2, not the magnitude of the loss function. The purpose of a loss function is just to let the optimizer know if it's going in the right direction--it's not a measure of fit that's comparable across data sets and learning setups. That's what R^2 is for.
Your R^2 scores show that you're explaining around a third of the total variance in the output, which is often a very good result for a data set with only 10 features. Actually, given the shape of your data, it's more likely that your hidden layers are considerably larger than necessary and risk over fitting.
To really evaluate this model, you'd need to know (1) how the R^2 score compares to simpler regression approaches like OLS and (2) why you should have any confidence that more than 30% of the output variance should be captured by the input variables.
For #1, at least the R^2 shouldn't be worse. As for #2, consider the canonical digit categorization example. We know that all the information necessary to recognize digits with very high accuracy (i.e. R^2 approaching 1) because humans can do it. That's not necessarily the case with other data sets, because there are important sources of variance that aren't captured in the source data.
As your loss decreases from 40000 to 6000, that means your NN model has learnt the prevalent relation but not all of them. You can aid this learning by transforming the predictor variables and then feeding them as derived ones to your model and see if that helps. You can try performing step wise addition of features to your NN model, by adding the most influential predictors first. At every iteration evaluate the model performance (i.e. training loss).
If first step doesn't help and as you are open to other approaches, Presuming your data's dynamics, Gaussian process Regression or Quantile regression should help as these methods are free from assumptions like linear regression techniques. Also it should help to explore different aspects of relationship between your independent and dependent variable.

Neural Network - Working with a imbalanced dataset

I am working on a Classification problem with 2 labels : 0 and 1. My training dataset is a very imbalanced dataset (and so will be the test set considering my problem).
The proportion of the imbalanced dataset is 1000:4 , with label '0' appearing 250 times more than label '1'. However, I have a lot of training samples : around 23 millions. So I should get around 100 000 samples for the label '1'.
Considering the big number of training samples I have, I didn't consider SVM. I also read about SMOTE for Random Forests. However, I was wondering whether NN could be efficient to handle this kind of imbalanced dataset with a large dataset ?
Also, as I am using Tensorflow to design the model, which characteristics should/could I tune to be able to handle this imbalanced situation ?
Thanks for your help !
Paul
Update :
Considering the number of answers, and that they are quite similar, I will answer all of them here, as a common answer.
1) I tried during this weekend the 1st option, increasing the cost for the positive label. Actually, with less unbalanced proportion (like 1/10, on another dataset), this seems to help a bit to get a better result, or at least to 'bias' the precision/recall scores proportion.
However, for my situation,
It seems to be very sensitive to the alpha number. With alpha = 250, which is the proportion of the unbalanced dataset, I have a precision of 0.006 and a recall score of 0.83, but the model is predicting way too many 1 that it should be - around 0.50 of label '1' ...
With alpha = 100, the model predicts only '0'. I guess I'll have to do some 'tuning' for this alpha parameter :/
I'll take a look at this function from TF too as I did it manually for now : tf.nn.weighted_cross_entropy_with_logitsthat
2) I will try to de-unbalance the dataset but I am afraid that I will lose a lot of info doing that, as I have millions of samples but only ~ 100k positive samples.
3) Using a smaller batch size seems indeed a good idea. I'll try it !
There are usually two common ways for imbanlanced dataset:
Online sampling as mentioned above. In each iteration you sample a class-balanced batch from the training set.
Re-weight the cost of two classes respectively. You'd want to give the loss on the dominant class a smaller weight. For example this is used in the paper Holistically-Nested Edge Detection
I will expand a bit on chasep's answer.
If you are using a neural network followed by softmax+cross-entropy or Hinge Loss you can as #chasep255 mentionned make it more costly for the network to misclassify the example that appear the less.
To do that simply split the cost into two parts and put more weights on the class that have fewer examples.
For simplicity if you say that the dominant class is labelled negative (neg) for softmax and the other the positive (pos) (for Hinge you could exactly the same):
L=L_{neg}+L_{pos} =>L=L_{neg}+\alpha*L_{pos}
With \alpha greater than 1.
Which would translate in tensorflow for the case of cross-entropy where the positives are labelled [1, 0] and the negatives [0,1] to something like :
cross_entropy_mean=-tf.reduce_mean(targets*tf.log(y_out)*tf.constant([alpha, 1.]))
Whatismore by digging a bit into Tensorflow API you seem to have a tensorflow function tf.nn.weighted_cross_entropy_with_logitsthat implements it did not read the details but look fairly straightforward.
Another way if you train your algorithm with mini-batch SGD would be make batches with a fixed proportion of positives.
I would go with the first option as it is slightly easier to do with TF.
One thing I might try is weighting the samples differently when calculating the cost. For instance maybe divide the cost by 250 if the expected result is a 0 and leave it alone if the expected result is a one. This way the more rare samples have more of an impact. You could also simply try training it without any changes and see if the nnet just happens to work. I would make sure to use a large batch size though so you always get at least one of the rare samples in each batch.
Yes - neural network could help in your case. There are at least two approaches to such problem:
Leave your set not changed but decrease the size of batch and number of epochs. Apparently this might help better than keeping the batch size big. From my experience - in the beginning network is adjusting its weights to assign the most probable class to every example but after many epochs it will start to adjust itself to increase performance on all dataset. Using cross-entropy will give you additional information about probability of assigning 1 to a given example (assuming your network has sufficient capacity).
Balance your dataset and adjust your score during evaluation phase using Bayes rule:score_of_class_k ~ score_from_model_for_class_k / original_percentage_of_class_k.
You may reweight your classes in the cost function (as mentioned in one of the answers). Important thing then is to also reweight your scores in your final answer.
I'd suggest a slightly different approach. When it comes to image data, the deep learning community has already come up with a few ways to augment data. Similar to image augmentation, you could try to generate fake data to "balance" your dataset. The approach I tried was to use a Variational Autoencoder and then sample from the underlying distribution to generate fake data for the class you want. I tried it and the results are looking pretty cool: https://lschmiddey.github.io/fastpages_/2021/03/17/data-augmentation-tabular-data.html

In what order should we tune hyperparameters in Neural Networks?

I have a quite simple ANN using Tensorflow and AdamOptimizer for a regression problem and I am now at the point to tune all the hyperparameters.
For now, I saw many different hyperparameters that I have to tune :
Learning rate : initial learning rate, learning rate decay
The AdamOptimizer needs 4 arguments (learning-rate, beta1, beta2, epsilon) so we need to tune them - at least epsilon
batch-size
nb of iterations
Lambda L2-regularization parameter
Number of neurons, number of layers
what kind of activation function for the hidden layers, for the output layer
dropout parameter
I have 2 questions :
1) Do you see any other hyperparameter I might have forgotten ?
2) For now, my tuning is quite "manual" and I am not sure I am not doing everything in a proper way.
Is there a special order to tune the parameters ? E.g learning rate first, then batch size, then ...
I am not sure that all these parameters are independent - in fact, I am quite sure that some of them are not. Which ones are clearly independent and which ones are clearly not independent ? Should we then tune them together ?
Is there any paper or article which talks about properly tuning all the parameters in a special order ?
EDIT :
Here are the graphs I got for different initial learning rates, batch sizes and regularization parameters. The purple curve is completely weird for me... Because the cost decreases like way slowly that the others, but it got stuck at a lower accuracy rate. Is it possible that the model is stuck in a local minimum ?
Accuracy
Cost
For the learning rate, I used the decay :
LR(t) = LRI/sqrt(epoch)
Thanks for your help !
Paul
My general order is:
Batch size, as it will largely affect the training time of future experiments.
Architecture of the network:
Number of neurons in the network
Number of layers
Rest (dropout, L2 reg, etc.)
Dependencies:
I'd assume that the optimal values of
learning rate and batch size
learning rate and number of neurons
number of neurons and number of layers
strongly depend on each other. I am not an expert on that field though.
As for your hyperparameters:
For the Adam optimizer: "Recommended values in the paper are eps = 1e-8, beta1 = 0.9, beta2 = 0.999." (source)
For the learning rate with Adam and RMSProp, I found values around 0.001 to be optimal for most problems.
As an alternative to Adam, you can also use RMSProp, which reduces the memory footprint by up to 33%. See this answer for more details.
You could also tune the initial weight values (see All you need is a good init). Although, the Xavier initializer seems to be a good way to prevent having to tune the weight inits.
I don't tune the number of iterations / epochs as a hyperparameter. I train the net until its validation error converges. However, I give each run a time budget.
Get Tensorboard running. Plot the error there. You'll need to create subdirectories in the path where TB looks for the data to plot. I do that subdir creation in the script. So I change a parameter in the script, give the trial a name there, run it, and plot all the trials in the same chart. You'll very soon get a feel for the most effective settings for your graph and data.
For parameters that are less important you can probably just pick a reasonable value and stick with it.
Like you said, the optimal values of these parameters all depend on each other. The easiest thing to do is to define a reasonable range of values for each hyperparameter. Then randomly sample a parameter from each range and train a model with that setting. Repeat this a bunch of times and then pick the best model. If you are lucky you will be able to analyze which hyperparameter settings worked best and make some conclusions from that.
I don't know any tool specific for tensorflow, but the best strategy is to first start with the basic hyperparameters such as learning rate of 0.01, 0.001, weight_decay of 0.005, 0.0005. And then tune them. Doing it manually will take a lot of time, if you are using caffe, following is the best option that will take the hyperparameters from a set of input values and will give you the best set.
https://github.com/kuz/caffe-with-spearmint
for more information, you can follow this tutorial as well:
http://fastml.com/optimizing-hyperparams-with-hyperopt/
For number of layers, What I suggest you to do is first make smaller network and increase the data, and after you have sufficient data, increase the model complexity.
Before you begin:
Set batch size to maximal (or maximal power of 2) that works on your hardware. Simply increase it until you get a CUDA error (or system RAM usage > 90%).
Set regularizes to low values.
The architecture and exact numbers of neurons and layers - use known architectures as inspirations and adjust them to your specific performance requirements: more layers and neurons -> possibly a stronger, but slower model.
Then, if you want to do it one by one, I would go like this:
Tune learning rate in a wide range.
Tune other parameters of the optimizer.
Tune regularizes (dropout, L2 etc).
Fine tune learning rate - it's the most important hyper-parameter.

Accuracy in Caffe keeps on 0.1 and does not change

Through all training process, accuracy is 0.1. What am I doing wrong?
Model, solver and part of log here:
https://gist.github.com/yutkin/3a147ebbb9b293697010
Topology in png format:
P.S. I am using the latest version of Caffe and g2.2xlarge instance on AWS.
You're working on CIFAR-10 dataset which has 10 classes. When the training of a network commences, the first guess is usually random due to which your accuracy is 1/N, where N is the number of classes. In your case it is 1/10, i.e., 0.1. If your accuracy stays the same over time it implies that your network isn't learning anything. This may happen due to a large learning rate. The basic idea of training a network is that you calculate the loss and propagate it back. The gradients are multiplied with the learning rate and added to the current weights and biases. If the learning rate is too big you may overshoot the local minima every time. If it is too small, the convergence will be slow. I see that your base_lr here is 0.01. As far as my experience goes, this is somewhat large. You may want to keep it at 0.001 in the beginning and then go on reducing it by a factor of 10 whenever you observe that the accuracy is not improving. But then anything below 0.00001 usually doesn't make much of a difference. The trick is to observe the progress of the training and make parameter changes as and when required.
I know the thread is quite old but maybe my answer helps somebody. I experienced the same problem with an accuracy like a random guess.
What helped was to set the number of outputs of the last layer before the accuracy layer to the number of labels.
In your case that should be the ip2 layer. Open the model definition of your net and set num_outputs to the number of labels.
See Section 4.4 for more information: A Practical Introduction to Deep Learning with Caffe and Python

Why does my neural network trained on MNIST data set not predict 7 and 9 correctly?

I'm using Matlab ( github code repository ). The details of the network are:
Hidden units: 100 ( variable )
Epochs : 500
Batch size: 100
The weights are being updated using Back propagation algorithm.
I've been able to recognize 0,1,2,3,4,5,6,8 which I have drawn in photoshop.
However 7,9 are not recognized, but upon running on the test set I get only 749/10000 wrong and it correctly classifies 9251/10000.
Any idea what might be wrong? Because it is learning and based on the test set results its learning correctly.
I don't see anything downright incorrect in your code, but there is a lot that can be improved:
You use this to set the initial weights:
hiddenWeights = rand(hiddenUnits,inputVectorSize);
outputWeights = rand(outputVectorSize,hiddenUnits);
hiddenWeights = hiddenWeights./size(hiddenWeights, 2);
outputWeights = outputWeights./size(outputWeights, 2);
This will make your weights very small I think. Not only that, but you will have no negative values, so you'll throw away half of the sigmoid's range of values. I suggest you try:
weights = 2*rand(x, y) - 1
Which will generate random numbers in [-1, 1]. You can then try dividing this interval to get smaller weights (try dividing by the sqrt of the size).
You use this as the output delta:
outputDelta = dactivation(outputActualInput).*(outputVector - targetVector) % (tk-yk)*f'(yin)
Multiplying by the derivative is done if you use the square loss function. For log loss (which is usually the one used in classification), you should have just outputVector - targetVector. It might not make that big of a difference, but you might want to try.
You say in the comments that the network doesn't detect your own sevens and nines. This can suggest overfitting on the MNIST data. To address this, you'll need to add some form of regularization to your network: either weight decay or dropout.
You should try different learning rates as well, if you haven't already.
You don't seem to have any bias neurons. Each layer, except the output layer, should have a neuron that only returns the value 1 to the next layer. You can implement this by adding another feature to your input data that is always 1.
MNIST is a big data set for which better algorithms are still being researched. Your networks is very basic, small, with no regularization, no bias neurons and no improvements to classic gradient descent. It's not surprising that it's not working too well: you'll likely need a more complex network for better results.
Nothing to do with neural nets or your code,
but this picture of KNN-nearest digits shows that some MNIST digits
are simply hard to recognize: