The standard supervised classification setup: we have a bunch of samples, each with the correct label out of N labels. We build a NN with N outputs, transform those to probabilities with softmax, and the loss is the mean cross-entropy between each NN output and the corresponding true label, represented as a 1-hot vector with 1 in the true label and 0 elsewhere. We then optimize this loss by following its gradient. The classification error is used just to measure our model quality.
HOWEVER, I know that when doing policy gradient we can use the likelihood ratio trick, and we no longer need to use cross-entropy! our loss simply tf.gather the NN output corresponding to the correct label. E.g. this solution of OpenAI gym CartPole.
WHY can't we use the same trick when doing supervised learning? I was thinking that the reason we used cross-entropy is because it is differentiable, but apparently tf.gather is differentiable as well.
I mean - IF we measure ourselves on classification error, and we CAN optimize for classification error as it's differentiable, isn't it BETTER to also optimize for classification error instead of this weird cross-entropy proxy?
Policy gradient is using cross entropy (or KL divergence, as Ishant pointed out). For supervised learning tf.gather is really just implementational trick, nothing else. For RL on the other hand it is a must because you do not know "what would happen" if you would execute other action. Consequently you end up with high variance estimator of your gradients, something that you would like to avoid for all costs, if possible.
Going back to supervised learning though
CE(p||q) = - SUM_i q_i log p_i
Lets assume that q_i is one hot encoded, with 1 at k'th position, then
CE(p||q) = - q_k log p_k = - log p_k
So if you want, you can implement this as tf.gather, it simply does not matter. The cross-entropy is simply more generic because it handles more complex targets. In particular, in TF you have sparse cross entropy which does exactly what you describe - exploits one hot encoding, that's it. Mathematically there is no difference, there is small difference computation-wise, and there are functions doing exactly what you want.
Minimization of cross-entropy loss minimizes the KL divergence between the predicted distribution and the target distribution. Which is indeed the same as maximizing the likelihood of the predicted distribution.
Related
Focal Loss given in Tensorflow is used for class imbalance. For Binary class classification, there are a lots of codes available but for Multiclass classification, a very little help is there. I ran the code with One Hot Encoded target variables of 250 classes and it gave me results without any error.
y = pd.get_dummies(df['target']) # One hot encoded target classes
model.compile(
optimizer="adam", loss=tfa.losses.SigmoidFocalCrossEntropy(), metrics= metric
)
I just want to know whoever wrote this code or someone having enough knowledge of this code, can it be used be used for Multiclass Classification. If no then how come it did not give me errors, instead better results than CrossEntropy. Also, in other implementations like this one, the value of alpha has to be given for every class but just one value in Tensorflow's implementations.
What is the correct way to use this?
Some basics first.
Categorical Crossentropy is designed to incentivize a model a model to predict 100% for the correct label. It was designed for models that predict single-label multi-class classification - like CIFAR10 or Imagenet. Usually these models finish in a Dense layer with more than one output.
Binary Crossentropy is designed to incentivize a model to predict 100% if the label is one, or, 0% is the label is zero. Usually these models finish in a Dense layer with exactly one output.
When you apply Binary Crossentropy to a single-label multi-class classification problem, you are doing something that is mathematically valid but defines a slightly different task: you are incentivizing a single-label classification model to not only get the true label correct, but also minimize the false labels.
For example, if your target is dog, and your model predict 60% dog, CCE doesn't care if your model predicts 20% cat and 20% French horn, or, 40% cat and 0% French horn. So this is aligned with a top-1 accuracy concept.
But if you take that same model and apply BCE, and your model predictions 60% dog, BCE DOES care if your models predict 20%/20% cat/frenchhorn, vs 40%/0% cat/frenchhorn. To put it in precise terminology, the former is more "calibrated" and so it has some additional measure of goodness. However, this has little correlation to top-1 accuracy.
When you use BCE, presumably you are wasting the model's energy to focus on calibration at the expense of top-1 acc. But as you might have seen, it doesn't always work out that way. Sometimes BCE gives you superior results. I don't know that there's a clear explanation of that but I'd assume that the additional signals (in the case of Imagenet, you'll literally get 1000 times more signals) somehow creates a smoother loss value that perhaps helps smooth the gradients you receive.
The alpha value of focal loss additionally penalizes very wrong predictions and lessens the penalty if your model predicts something close to the right answer - like predicting 90% cat if the ground truth is cat. This would be a shift from the original definition of CCE, based on the theory of Maximum Likelihood Estimation... which focuses on calibration... vs the normal metric most ML practitioners care about: top-1 accuracy.
Focal loss was originally designed for binary classification so the original formulation only has a single alpha value. The repo you pointed to extends the concept of Focal Loss to single-label classification and therefore there are multiple alpha values: one per class. However, by my read, it loses the additional possible smoothing effect of BCE.
Net net, for the best results, you'll want to benchmark CCE, BCE, Binary Focal Loss (out of TFA and per the original paper), and the single-label multi-class Focal Loss that you found in that repo. In general, those the discovery of those alpha values is done via guess & check, or grid search.
There's a lot of manual guessing and checking in ML unfortunately.
I'm new to Deep Learning and I saw this for the first time. Having MAE as loss function and MSE to metric. What is the purpose of this and what is gained?
(loss=tf.metrics.MeanAbsoluteError(), metrics=[tf.losses.MeanSquaredError()])
In some cases it is useful to have a loss function different from the metric you are going to evaluate.
Consider the case in which you want to denoise an image, that is you design a network that takes as input a noise image and outputs its clean version. Here, your metric might be the Peak-Signal-to-Noise Ratio (PSNR) or some sort of structural similarity (SSIM) between your output and the ground truth clean image. However, during training, you might consider different loss function, such as L1 (MAE), L2 (MSE) or even a Perceptual Loss, such as the VGG loss, because these have been proved to lead to better results than directly optimizing for PSNR or SSIM.
For multiclass classification problems, Keras and tf.keras have metrics like SparseTopKCategoricalAccuracy and TopKCategoricalAccuracy. However, if one uses loss functions like SparseCategoricalCrossentropy or CategoricalCrossentropy, they cannot achieve the max values for these two metrics.
What is a good loss function to use when one wants to maximize SparseTopKCategoricalAccuracy or TopKCategoricalAccuracy?
I understand that SparseTopKCategoricalAccuracy is not differentiable, just like Accuracy. I am trying to find a function that can approximate the smooth loss function and yield a higher number for SparseTopKCategoricalAccuracy.
CrossEntropy is not the best loss function when you deal with Top-k accuracy because cross-entropy may be prone to overfitting on small datasets or noisy labels.
As you have already pointed out, "smooth loss" functions are developed for top-k classification with SVM. To my knowledge, there is no a "off-the-shelf" loss function in Keras/TF that is best suited for top-k. However, I suggest you to try Smooth Surrogate Loss (SSL) presented in the article and implemented in Pytorch to use with deep neural networks (see Github). It derives from multi-class SVMs as SSL creates a margin between the correct top-k predictions and the incorrect ones. The training time of SSL is comparatevely the same as in the case of cross-entropy thanking to a divide-and-conquer approach and the use of polynomials (see implementation).
I have been trying to conduct a few experiments using TensorFlow Probability (TFP), and I got a few questions.
What is the proper value of the coefficient of the KL loss?
In the paper by Blundell (2015), the coefficient is set to 1/M (where M is the number of mini-batches). In the example given by TFP, the coefficient is given as 1/mnist_data.train.num_examples. Why?
As I go from 2d input to 3d images volumes, the KL loss is still significantly larger (~1k) than the cross-entropy (~1), even after dividing by mnist_data.train.num_examples. Why?
What is the guideline for getting a proper value for this coefficient? Maybe like the two-loss terms should be the same order of magnitude?
The current coefficient only takes care of the number of training samples, but not the network complexity or number of parameters in the network, which I assume the KL loss increase with the complexity of the model.
I am trying to implement a neural network with the KL loss, without using keras.model.losses, as some software production and hardware support limitation. I am trying to train my model with TF 1.10 and TFP 0.3.0., the issue is that for tf<=1.14, tf.keras.model does not support tf.layers inside the Keras model, so I can't use my original model straight away. Is there a way to get the KL loss, not from model.losses, but from layers or weights of the network in a TF construct?
Is batch normalization or group normalization still helpful in Bayesian deep learning?
In the paper by Blundell (2015), the coefficient is set to 1/M (where M is the number of mini-batches). In the example given by TFP, the coefficient is given as 1/mnist_data.train.num_examples. Why?
In the BBB paper eq. 8, they refer to M being the number of mini-batches. To be consistent with the non-stochastic gradient learning, it should be scaled by the number of mini-batches which is what is done by Graves. Another alternative is that done in eq. 9, where they scale it by \pi_i, where the sum of all the values in the set {\pi} sum to one.
In the TFP example, it does look like the num_examples is the total number of independent samples within the training set, which is much larger than the number of batches. This is goes by a few names, such as Safe Bayes or Tempering. Have a look at sec. 8 of this paper for some more discussion about the use of tempering within Bayesian inference and it's suitability.
As I go from 2d input to 3d images volumes, the KL loss is still significantly larger (~1k) than the cross-entropy (~1), even after dividing by mnist_data.train.num_examples. Why?
The ELBO will always be larger than just your cross-entropy (which defines your likelihood). Have a look at how the KL divergence term in the ELBO is found. (and a full mean-field approach where each weight/parameter is assumed to be independent).
Since the assumed posterior is factorised (assume each parameter is independent), can write the joint distribution as a product. This means when you take the log when you are computing the KL between the approx. posterior and the prior, you can write it as a sum of the KL terms between each parameter. Since the KL is >= 0, for each parameter you add to your model you will be adding another positive term to your ELBO. This is likely why your loss is so much more for your 3D model, likely because there is more parameters.
Another reason this could occur is if you have less data (your M is smaller, than the KL term is weighted less).
What is the guideline for getting a proper value for this coefficient? Maybe like the two-loss terms should be the same order of magnitude?
I am unsure of any specific guideline, for training you are interested primarily in the gradients. A large loss does not mean a large gradient. Have a look at the gradients contributed by the negative log likelihood and the KL term in your ELBO. If the KL term is too large, you probably need a more informative prior or more data (you could simply scale the KL term but this feels a bit yucky for the Bayesian in me).
The current coefficient only takes care of the number of training samples, but not the network complexity or the number of parameters in the network, which I assume the KL loss increase with the complexity of the model.
Yes, as stated before, in general, more parameters == greater ELBO (for a mean-field approach as used in Bayes by Backprop).
I am trying to implement a neural network with the KL loss, without using keras.model.losses, as some software production and hardware support limitation. I am trying to train my model with TF 1.10 and TFP 0.3.0., the issue is that for tf<=1.14, tf.keras.model does not support tf.layers inside the Keras model, so I can't use my original model straight away. Is there a way to get the KL loss, not from model.losses, but from layers or weights of the network in a TF construct?
I am unsure about the best way to tackle this part of it. I would be cautious about going to older versions where it isn't explicitly supported. They put those warnings/exceptions in for a reason.
Is batch normalization or group normalization still helpful in Bayesian deep learning?
For variational inference (as done in Bayes by Backprop) Batchnorm is fine. For sampling methods such as MCMC, Batch normalization is no longer suitable. Have a look at https://arxiv.org/pdf/1908.03491v1.pdf for info on suitability for batch norm with sampling methods for approx. Bayesian inference.
This question already has answers here:
What are logits? What is the difference between softmax and softmax_cross_entropy_with_logits?
(8 answers)
Closed 2 years ago.
In the following TensorFlow function, we must feed the activation of artificial neurons in the final layer. That I understand. But I don't understand why it is called logits? Isn't that a mathematical function?
loss_function = tf.nn.softmax_cross_entropy_with_logits(
logits = last_layer,
labels = target_output
)
Logits is an overloaded term which can mean many different things:
In Math, Logit is a function that maps probabilities ([0, 1]) to R ((-inf, inf))
Probability of 0.5 corresponds to a logit of 0. Negative logit correspond to probabilities less than 0.5, positive to > 0.5.
In ML, it can be
the vector of raw (non-normalized) predictions that a classification
model generates, which is ordinarily then passed to a normalization
function. If the model is solving a multi-class classification
problem, logits typically become an input to the softmax function. The
softmax function then generates a vector of (normalized) probabilities
with one value for each possible class.
Logits also sometimes refer to the element-wise inverse of the sigmoid function.
Just adding this clarification so that anyone who scrolls down this much can at least gets it right, since there are so many wrong answers upvoted.
Diansheng's answer and JakeJ's answer get it right.
A new answer posted by Shital Shah is an even better and more complete answer.
Yes, logit as a mathematical function in statistics, but the logit used in context of neural networks is different. Statistical logit doesn't even make any sense here.
I couldn't find a formal definition anywhere, but logit basically means:
The raw predictions which come out of the last layer of the neural network.
1. This is the very tensor on which you apply the argmax function to get the predicted class.
2. This is the very tensor which you feed into the softmax function to get the probabilities for the predicted classes.
Also, from a tutorial on official tensorflow website:
Logits Layer
The final layer in our neural network is the logits layer, which will return the raw values for our predictions. We create a dense layer with 10 neurons (one for each target class 0–9), with linear activation (the default):
logits = tf.layers.dense(inputs=dropout, units=10)
If you are still confused, the situation is like this:
raw_predictions = neural_net(input_layer)
predicted_class_index_by_raw = argmax(raw_predictions)
probabilities = softmax(raw_predictions)
predicted_class_index_by_prob = argmax(probabilities)
where, predicted_class_index_by_raw and predicted_class_index_by_prob will be equal.
Another name for raw_predictions in the above code is logit.
As for the why logit... I have no idea. Sorry.
[Edit: See this answer for the historical motivations behind the term.]
Trivia
Although, if you want to, you can apply statistical logit to probabilities that come out of the softmax function.
If the probability of a certain class is p,
Then the log-odds of that class is L = logit(p).
Also, the probability of that class can be recovered as p = sigmoid(L), using the sigmoid function.
Not very useful to calculate log-odds though.
Summary
In context of deep learning the logits layer means the layer that feeds in to softmax (or other such normalization). The output of the softmax are the probabilities for the classification task and its input is logits layer. The logits layer typically produces values from -infinity to +infinity and the softmax layer transforms it to values from 0 to 1.
Historical Context
Where does this term comes from? In 1930s and 40s, several people were trying to adapt linear regression to the problem of predicting probabilities. However linear regression produces output from -infinity to +infinity while for probabilities our desired output is 0 to 1. One way to do this is by somehow mapping the probabilities 0 to 1 to -infinity to +infinity and then use linear regression as usual. One such mapping is cumulative normal distribution that was used by Chester Ittner Bliss in 1934 and he called this "probit" model, short for "probability unit". However this function is computationally expensive while lacking some of the desirable properties for multi-class classification. In 1944 Joseph Berkson used the function log(p/(1-p)) to do this mapping and called it logit, short for "logistic unit". The term logistic regression derived from this as well.
The Confusion
Unfortunately the term logits is abused in deep learning. From pure mathematical perspective logit is a function that performs above mapping. In deep learning people started calling the layer "logits layer" that feeds in to logit function. Then people started calling the output values of this layer "logit" creating the confusion with logit the function.
TensorFlow Code
Unfortunately TensorFlow code further adds in to confusion by names like tf.nn.softmax_cross_entropy_with_logits. What does logits mean here? It just means the input of the function is supposed to be the output of last neuron layer as described above. The _with_logits suffix is redundant, confusing and pointless. Functions should be named without regards to such very specific contexts because they are simply mathematical operations that can be performed on values derived from many other domains. In fact TensorFlow has another similar function sparse_softmax_cross_entropy where they fortunately forgot to add _with_logits suffix creating inconsistency and adding in to confusion. PyTorch on the other hand simply names its function without these kind of suffixes.
Reference
The Logit/Probit lecture slides is one of the best resource to understand logit. I have also updated Wikipedia article with some of above information.
Logit is a function that maps probabilities [0, 1] to [-inf, +inf].
Softmax is a function that maps [-inf, +inf] to [0, 1] similar as Sigmoid. But Softmax also normalizes the sum of the values(output vector) to be 1.
Tensorflow "with logit": It means that you are applying a softmax function to logit numbers to normalize it. The input_vector/logit is not normalized and can scale from [-inf, inf].
This normalization is used for multiclass classification problems. And for multilabel classification problems sigmoid normalization is used i.e. tf.nn.sigmoid_cross_entropy_with_logits
Personal understanding, in TensorFlow domain, logits are the values to be used as input to softmax. I came to this understanding based on this tensorflow tutorial.
https://www.tensorflow.org/tutorials/layers
Although it is true that logit is a function in maths(especially in statistics), I don't think that's the same 'logit' you are looking at. In the book Deep Learning by Ian Goodfellow, he mentioned,
The function σ−1(x) is called the logit in statistics, but this term
is more rarely used in machine learning. σ−1(x) stands for the
inverse function of logistic sigmoid function.
In TensorFlow, it is frequently seen as the name of last layer. In Chapter 10 of the book Hands-on Machine Learning with Scikit-learn and TensorFLow by Aurélien Géron, I came across this paragraph, which stated logits layer clearly.
note that logits is the output of the neural network before going
through the softmax activation function: for optimization reasons, we
will handle the softmax computation later.
That is to say, although we use softmax as the activation function in the last layer in our design, for ease of computation, we take out logits separately. This is because it is more efficient to calculate softmax and cross-entropy loss together. Remember that cross-entropy is a cost function, not used in forward propagation.
(FOMOsapiens).
If you check math Logit function, it converts real space from [0,1] interval to infinity [-inf, inf].
Sigmoid and softmax will do exactly the opposite thing. They will convert the [-inf, inf] real space to [0, 1] real space.
This is why, in machine learning we may use logit before sigmoid and softmax function (since they match).
And this is why "we may call" anything in machine learning that goes in front of sigmoid or softmax function the logit.
Here is G. Hinton video using this term.
Here is a concise answer for future readers. Tensorflow's logit is defined as the output of a neuron without applying activation function:
logit = w*x + b,
x: input, w: weight, b: bias. That's it.
The following is irrelevant to this question.
For historical lectures, read other answers. Hats off to Tensorflow's "creatively" confusing naming convention. In PyTorch, there is only one CrossEntropyLoss and it accepts un-activated outputs. Convolutions, matrix multiplications and activations are same level operations. The design is much more modular and less confusing. This is one of the reasons why I switched from Tensorflow to PyTorch.
logits
The vector of raw (non-normalized) predictions that a classification model generates, which is ordinarily then passed to a normalization function. If the model is solving a multi-class classification problem, logits typically become an input to the softmax function. The softmax function then generates a vector of (normalized) probabilities with one value for each possible class.
In addition, logits sometimes refer to the element-wise inverse of the sigmoid function. For more information, see tf.nn.sigmoid_cross_entropy_with_logits.
official tensorflow documentation
They are basically the fullest learned model you can get from the network, before it's been squashed down to apply to only the number of classes we are interested in. Check out how some researchers use them to train a shallow neural net based on what a deep network has learned: https://arxiv.org/pdf/1312.6184.pdf
It's kind of like how when learning a subject in detail, you will learn a great many minor points, but then when teaching a student, you will try to compress it to the simplest case. If the student now tried to teach, it'd be quite difficult, but would be able to describe it just well enough to use the language.
The logit (/ˈloʊdʒɪt/ LOH-jit) function is the inverse of the sigmoidal "logistic" function or logistic transform used in mathematics, especially in statistics. When the function's variable represents a probability p, the logit function gives the log-odds, or the logarithm of the odds p/(1 − p).
See here: https://en.wikipedia.org/wiki/Logit