Layer probability always 1 - caffe

I'm using caffe/example/mnist network to classify numbers. When I give the network a picture of number it seems working ok. But when I give the network a picture not a number, the mnist trained network softmax layer gives the probabilities, which always has one probability 1 and others 0, like:
[0,0,0,...,1,0,0,0].
I think it should be something like:
[0,0.1,0.2,...,0.4,0.1,0.2],
in which case I can say that this should not be a number. What is the problem?

It is hard to know what to expect since it hasn't been trained on non-numbers and it has to give a result that sums to 1. By using Softmax, you are telling the network that there is a number, while showing it a non-number. You cannot look at its output to then determine whether it is a number or not.
Furthermore, the training data for MNIST is very stereotypical and not good for generalization. The foreground numbers are always 255 and the background is always 0. The average value will be much closer to 0, since there are more background pixels. Simply presenting an image with a mean pixel value of 100 could bias the prediction to numbers that typically have more pixels (like 8, perhaps). You can only expect to the network to generalize to similar types of stimulus. For your task, you should do quite a lot of data augmentation.
You want to allow the probability to be zeros for all numbers, which you can do by using the cross-entropy loss. This would also allow the probability to also sum to more than 1 (maximum 10). You could also try adding another class for "non-number" with the Softmax, however, then you should present non-number stimuli and number stimuli that is more similar to natural stimuli (so that it is not trivially separable).

Related

Would this be a valid Implementation of an ordinal CrossEntropy?

Would this be a valid implementation of a cross entropy loss that takes the ordinal structure of the GT y into consideration? y_hat is the prediction from a neural network.
ce_loss = F.cross_entropy(y_hat, y, reduction="none")
distance_weight = torch.abs(y_hat.argmax(1) - y) + 1
ordinal_ce_loss = torch.mean(distance_weight * ce_loss)
I'll attempt to answer this question by first fully defining the task, since the question is a bit sparse on details.
I have a set of ordinal classes (e.g. first, second, third, fourth,
etc.) and I would like to predict the class of each data example from
among this set. I would like to define an entropy-based loss-function
for this problem. I would like this loss function to weight the loss
between a predicted class torch.argmax(y_hat) and the true class y
according to the ordinal distance between the two classes. Does the
given loss expression accomplish this?
Short answer: sure, it is "valid". You've roughly implemented L1-norm ordinal class weighting. I'd question whether this is truly the correct weighting strategy for this problem.
For instance, consider that for a true label n, the bin n response is weighted by 1, but the bin n+1 and n-1 responses are weighted by 2. This means that a lot more emphasis will be placed on NOT predicting false positives than on correctly predicting true positives, which may imbue your model with some strange bias.
It also means that examples on the edge will result in a larger total sum of weights, meaning that you'll be weighting examples where the true label is say "first" or "last" more highly than the intermediate classes. (Say you have 5 classes: 1,2,3,4,5. A true label of 1 will require distance_weight of [1,2,3,4,5], the sum of which is 15. A true label of 3 will require distance_weight of [3,2,1,2,3], the sum of which is 11.
In general, classification problems and entropy-based losses are underpinned by the assumption that no set of classes or categories is any more or less related than any other set of classes. In essence, the input data is embedded into an orthogonal feature space where each class represents one vector in the basis. This is quite plainly a bad assumption in your case, meaning that this embedding space is probably not particularly elegant: thus, you have to correct for it with sort of a hack-y weight fix. And in general, this assumption of class non-correlation is probably not true in a great many classification problems (consider e.g. the classic ImageNet classification problem, wherein the class pairs [bus,car], and [bus,zebra] are treated as equally dissimilar. But this is probably a digression into the inherent lack of usefulness of strict ontological structuring of information which is outside the scope of this answer...)
Long Answer: I'd highly suggest moving into a space where the ordinal value you care about is instead expressed in a continuous space. (In the first, second, third example, you might for instance output a continuous value over the range [1,max_place]. This allows you to benefit from loss functions that already capture well the notion that predictions closer in an ordered space are better than predictions farther away in an ordered space (e.g. MSE, Smooth-L1, etc.)
Let's consider one more time the case of the [first,second,third,etc.] ordinal class example, and say that we are trying to predict the places of a set of runners in a race. Consider two races, one in which the first place runner wins by 30% relative to the second place runner, and the second in which the first place runner wins by only 1%. This nuance is entirely discarded by the ordinal discrete classification. In essence, the selection of an ordinal set of classes truncates the amount of information conveyed in the prediction, which means not only that the final prediction is less useful, but also that the loss function encodes this strange truncation and binarization, which is then reflected (perhaps harmfully) in the learned model. This problem could likely be much more elegantly solved by regressing the finishing position, or perhaps instead by regressing the finishing time, of each athlete, and then performing the final ordinal classification into places OUTSIDE of the network training.
In conclusion, you might expect a well-trained ordinal classifier to produce essentially a normal distribution of responses across the class bins, with the distribution peak on the true value: a binned discretization of a space that almost certainly could, and likely should, be treated as a continuous space.

what's the meaning of 'parameterize' in deep learning?

no
13
what's the meaning of 'parameterize' in deep learning? As shown in the photo, does it means the matrix 'A' can be changed by the optimization during training?
Yes, when something can be parameterized it means that gradients can be calculated.
This means that the (dE/dw) which means the derivative of Error with respect to weight can be calculated (i.e it must be differentiable) and subtracted from the model weights along with obviously a learning_rate and other params being included depending on the optimizer.
What the paper is saying is that if you make a binary matrix a weight and then find the gradient (dE/dw) of that weight with respect to a loss and then make an update on the binary matrix through backpropagation, there is not really an activation function (which by requirement must be differentiable) that can keep the values discrete (like 0 and 1) but rather you will end up with continous values (like these decimal values).
Therefore it is saying since the idea of having binary values be weights and for them to be back-propagated in a way where the weights + activation function also yields an updated weight matrix that is also binary is difficult, another solution like the Bernoulli Distribution is used instead to initialize parameters of a model.
Hope this helps,

Both validation loss and accuracy are increasing using a pre-trained VGG-16

So, I'm doing a 4 label x-ray images classification on around 12600 images:
Class1:4000
Class2:3616
Class3:1345
Class4:4000
I'm using VGG-16 architecture pertained on the imageNet dataset with cross-entrpy and SGD and a batch size of 32 and a learning rate of 1e-3 running on pytorch
[[749., 6., 50., 2.],
[ 5., 707., 9., 1.],
[ 56., 8., 752., 0.],
[ 4., 1., 0., 243.]]
I know since both train loss/acc are relatively 0/1 the model is overfitting, though I'm surprised that the val acc is still around 0.9!
How to properly interpret that and what causing it and how to prevent it?
I know it's something like because the accuracy is the argmax of softmax like the actual predictions are getting lower and lower but the argmax always stays the same, but I'm really confused about it! I even let it train for +64 epochs same results flat acc while loss increases gradually!
PS. I have seen other questions with answers and didn't really get an explanation
I think your question already says about what is going on. Your model is overfitting as you have also figured out. Now, as you are training more your model slowly becoming more specialized to the train set and loosing the the capability to generalize gradually. So the softmax probabilities are getting more and more flat. But still it is showing more or less the same accuracy for validation set as still now the correct class has at least slightly more probability than the others. So in my opinion there can be some possible reasons for this:
Your train set and validation set may not be from the same distribution.
Your validation set doesn't cover all cases need to be evaluated, it probably contains similar types of images but they do not differ too much. So, when the model can identify one, it can identify many of them from the validation set. If you add more heterogeneous images in validation set, you will no longer see such a large accuracy in validation set.
Similarly, we can say your train set has images which are heterogeneous i.e, they have a lot of variations, and the validation set is covering only a few varieties, so as training goes on, those minorities are getting less priority as the model yet to have many things to learn and generalize. This can happen if you augment your train-set and your model finds the validation set is relatively easier initially (until overfitting), but as training goes on the model gets lost itself while learning a lot of augmented varieties available in the train set. In this case don't make the augmentation too much wild. Think, if the augmented images are still realistic or not. Do augmentation on images as long as they remain realistic and each type of these images' variations occupy enough representative examples in the train set. Don't include unnecessary situations in augmentation those will never occur in reality, as these unrealistic examples will just increase burden on the model than doing any help.

How to deal with ordinal labels in keras?

I have data with integer target class in the range 1-5 where one is the lowest and five the highest. In this case, should I consider it as regression problem and have one node in the output layer?
My way of handling it is:
1- first I convert the labels to binary class matrix
labels = to_categorical(np.asarray(labels))
2- in the output layer, I have five nodes
main_output = Dense(5, activation='sigmoid', name='main_output')(x)
3- I use 'categorical_crossentropy with mean_squared_error when compiling
model.compile(optimizer='rmsprop',loss='categorical_crossentropy',metrics=['mean_squared_error'],loss_weights=[0.2])
Also, can anyone tells me: what is the difference between using categorical_accuracy and 'mean_squared_error in this case?
Regression and classification are vastly different things. If you reimagine this as a regression task than the difference of predicting 2 when the ground truth is 4 will be rated more than if you predict 3 instead of 4. If you have class like car, animal, person you do not care for the ranking between those classes. Predicting car is just as wrong as animal, iff the image shows a person.
Metrics do not impact your learning at all. It is just something that is computed additionally to the loss to show the performance of the model. Here the accuracy makes sense, because this is mostly the metric that we care about. Mean squared error does not tell you how well your model performs. If you get something like 0.0015 mean squared error it sounds good, but it is hard to visualize just how well this performs. In contrast using accuracy and achieving 95% accuracy for example is meaningful.
One last thing you should use softmax instead of sigmoid as your final output to get a probability distribution in your final layer. Softmax will output percentages for every class that sum up to 1. Then crossentropy calculates the difference of the probability distribution of your network output and the ground truth.

understanding getByteTimeDomainData and getByteFrequencyData in web audio

The documentation for both of these methods are both very generic wherever I look. I would like to know what exactly I'm looking at with the returned arrays I'm getting from each method.
For getByteTimeDomainData, what time period is covered with each pass? I believe most oscopes cover a 32 millisecond span for each pass. Is that what is covered here as well? For the actual element values themselves, the range seems to be 0 - 255. Is this equivalent to -1 - +1 volts?
For getByteFrequencyData the frequencies covered is based on the sampling rate, so each index is an actual frequency, but what about the actual element values themselves? Is there a dB range that is equivalent to the values returned in the returned array?
getByteTimeDomainData (and the newer getFloatTimeDomainData) return an array of the size you requested - its frequencyBinCount, which is calculated as half of the requested fftSize. That array is, of course, at the current sampleRate exposed on the AudioContext, so if it's the default 2048 fftSize, frequencyBinCount will be 1024, and if your device is running at 44.1kHz, that will equate to around 23ms of data.
The byte values do range between 0-255, and yes, that maps to -1 to +1, so 128 is zero. (It's not volts, but full-range unitless values.)
If you use getFloatFrequencyData, the values returned are in dB; if you use the Byte version, the values are mapped based on minDecibels/maxDecibels (see the minDecibels/maxDecibels description).
Mozilla 's documentation describes the difference between getFloatTimeDomainData and getFloatFrequencyData, which I summarize below. Mozilla docs reference the Web Audio
experiment ; the voice-change-o-matic. The voice-change-o-matic illustrates the conceptual difference to me (it only works in my Firefox browser; it does not work in my Chrome browser).
TimeDomain/getFloatTimeDomainData
TimeDomain functions are over some span of time.
We often visualize TimeDomain data using oscilloscopes.
In other words:
we visualize TimeDomain data with a line chart,
where the x-axis (aka the "original domain") is time
and the y axis is a measure of a signal (aka the "amplitude").
Change the voice-change-o-matic "visualizer setting" to Sinewave to
see getFloatTimeDomainData(...)
Frequency/getFloatFrequencyData
Frequency functions (GetByteFrequencyData) are at a point in time; i.e. right now; "the current frequency data"
We sometimes see these in mp3 players/ "winamp bargraph style" music players (aka "equalizer" visualizations).
In other words:
we visualize Frequency data with a bar graph
where the x-axis (aka "domain") are frequencies or frequency bands
and the y-axis is the strength of each frequency band
Change the voice-change-o-matic "visualizer setting" to Frequency bars to see getFloatFrequencyData(...)
Fourier Transform (aka Fast Fourier Transform/FFT)
Another way to think about "time domain vs frequency" is shown the diagram below, from Fast Fourier Transform wikipedia
getFloatTimeDomainData gives you the chart on on the top (x-axis is Time)
getFloatFrequencyData gives you the chart on the bottom (x-axis is Frequency)
a Fast Fourier Transform (FFT) converts the Time Domain data into Frequency data, in other words, FFT converts the first chart to the second chart.
cwilso has it backwards.
the time data array is the longer one (fftSize), and the frequency data array is the shorter one (half that, frequencyBinCount).
fftSize of 2048 at the usual sample rate of 44.1kHz means each sample has 1/44100 duration, you have 2048 samples at hand, and thus are covering a duration of 2048/44100 seconds, which 46 milliseconds, not 23 milliseconds. The frequencyBinCount is indeed 1024, but that refers to the frequency domain (as the name suggests), not the time domain, and it the computation 1024/44100, in this context, is about as meaningful as adding your birth date to the fftSize.
A little math illustrating what's happening: Fourier transform is a 'vector space isomorphism', that is, a mapping going bijectively (i.e., reversible) between 2 vector spaces of the same dimension; the 'time domain' and the 'frequency domain.' The vector space dimension we have here (in both cases) is fftSize.
So where does the 'half' come from? The frequency domain coefficients 'count double'. Either because they 'actually are' complex numbers, or because you have the 'sin' and the 'cos' flavor. Or, because you have a 'magnitude' and a 'phase', which you'll understand if you know how complex numbers work. (Those are 3 ways to say the same in a different jargon, so to speak.)
I don't know why the API only gives us half of the relevant numbers when it comes to frequency - I can only guess. And my guess is that those are the 'magnitude' numbers, and the 'phase' numbers are thrown out. The reason that this is my guess is that in applications, magnitude is far more important than phase. Still, I'm quite surprised that the API throws out information, and I'd be glad if some expert who actually knows (and isn't guessing) can confirm that it's indeed the magnitude. Or - even better (I love to learn) - correct me.