By setting the bottom and the top blob to be the same we can tell Caffe to do "in-place" computation to preserve memory consumption.
Currently I know I can safely use in-place "BatchNorm", "Scale" and "ReLU" layers (please let me know if I'm wrong). While it seems to have some issues for other layers (this issue seems to be an example).
When to use in-place layers in Caffe?
How does it work with back-propagation?
As you well noted, in-place layers don't usually work "out of the box".
For some layers, it is quite trivial ("ReLU" and other neuron activation layers).
However, for others it requires special handling in code. For example, the implementation of "PReLU" layer has specific cache bottom_memory_ member variable that stores information needed for backprop.
You can see similar code for other layers that specifically test for if (top[0] == bottom[0]) to see if the layer is used in an "in-place" case.
Moreover, it makes little sense to have an in-place layer for which the input and output are of different shapes, thus layers such as "Convolution", "InnerProduct", "Pool" are not considered as candidates for "in-place" layers.
Related
In the PyTorch LSTM documentation it is written:
batch_first – If True, then the input and output tensors are provided
as (batch, seq, feature). Default: False
I'm wondering why they chose the default batch dimension as the second one and not the first one. for me, it is easier to imaging my data as [batch, seq, feature] than [seq, batch, feature]. The first one is more intuitive for me and the second one is counterintuitive.
I'm asking here to know if the is any reason behind this and if you can help me to have some understanding about it.
As far as I know, there is not a heavily justified answer. Nowadays, it is different from other frameworks where, as you say, the shape is more intuitive, such as Keras, but only for compatibility reasons with older versions, changing a default parameter that modifies the dimensions of a vector would probably break half of the models out there if their maintainers update to newer PyTorch versions.
Probably the idea, in the beginning, was to set the temporal dimension first to simplify the iterating process over time, so you can just do a
for t, out_t in enumerate(my_tensor)
instead of having to do less visual stuff such as accessing with my_tensor[:, i] and having to iterate in range(time).
In an answer to another question it is written:
There is an argument in favor of not using batch_first, which states
that the underlying API provided by Nvidia CUDA runs considerably
faster using batch as secondary.
But I don't know if this argument is true or not.
I understand that Batch Normalisation helps in faster training by turning the activation towards unit Gaussian distribution and thus tackling vanishing gradients problem. Batch norm acts is applied differently at training(use mean/var from each batch) and test time (use finalized running mean/var from training phase).
Instance normalisation, on the other hand, acts as contrast normalisation as mentioned in this paper https://arxiv.org/abs/1607.08022 . The authors mention that the output stylised images should be not depend on the contrast of the input content image and hence Instance normalisation helps.
But then should we not also use instance normalisation for image classification where class label should not depend on the contrast of input image. I have not seen any paper using instance normalisation in-place of batch normalisation for classification. What is the reason for that? Also, can and should batch and instance normalisation be used together. I am eager to get an intuitive as well as theoretical understanding of when to use which normalisation.
Definition
Let's begin with the strict definition of both:
Batch normalization
Instance normalization
As you can notice, they are doing the same thing, except for the number of input tensors that are normalized jointly. Batch version normalizes all images across the batch and spatial locations (in the CNN case, in the ordinary case it's different); instance version normalizes each element of the batch independently, i.e., across spatial locations only.
In other words, where batch norm computes one mean and std dev (thus making the distribution of the whole layer Gaussian), instance norm computes T of them, making each individual image distribution look Gaussian, but not jointly.
A simple analogy: during data pre-processing step, it's possible to normalize the data on per-image basis or normalize the whole data set.
Credit: the formulas are from here.
Which normalization is better?
The answer depends on the network architecture, in particular on what is done after the normalization layer. Image classification networks usually stack the feature maps together and wire them to the FC layer, which share weights across the batch (the modern way is to use the CONV layer instead of FC, but the argument still applies).
This is where the distribution nuances start to matter: the same neuron is going to receive the input from all images. If the variance across the batch is high, the gradient from the small activations will be completely suppressed by the high activations, which is exactly the problem that batch norm tries to solve. That's why it's fairly possible that per-instance normalization won't improve network convergence at all.
On the other hand, batch normalization adds extra noise to the training, because the result for a particular instance depends on the neighbor instances. As it turns out, this kind of noise may be either good and bad for the network. This is well explained in the "Weight Normalization" paper by Tim Salimans at al, which name recurrent neural networks and reinforcement learning DQNs as noise-sensitive applications. I'm not entirely sure, but I think that the same noise-sensitivity was the main issue in stylization task, which instance norm tried to fight. It would be interesting to check if weight norm performs better for this particular task.
Can you combine batch and instance normalization?
Though it makes a valid neural network, there's no practical use for it. Batch normalization noise is either helping the learning process (in this case it's preferable) or hurting it (in this case it's better to omit it). In both cases, leaving the network with one type of normalization is likely to improve the performance.
Great question and already answered nicely. Just to add: I found this visualisation From Kaiming He's Group Norm paper helpful.
Source: link to article on Medium contrasting the Norms
I wanted to add more information to this question since there are some more recent works in this area. Your intuition
use instance normalisation for image classification where class label
should not depend on the contrast of input image
is partly correct. I would say that a pig in broad daylight is still a pig when the image is taken at night or at dawn. However, this does not mean using instance normalization across the network will give you better result. Here are some reasons:
Color distribution still play a role. It is more likely to be a apple than an orange if it has a lot of red.
At later layers, you can no longer imagine instance normalization acts as contrast normalization. Class specific details will emerge in deeper layers and normalizing them by instance will hurt the model's performance greatly.
IBN-Net uses both batch normalization and instance normalization in their model. They only put instance normalization in early layers and have achieved improvement in both accuracy and ability to generalize. They have open sourced code here.
IN provide visual and appearance in-variance and BN accelerate training and preserve discriminative feature.
IN is preferred in Shallow layer(starting layer of CNN) so remove appearance variation and BN is preferred in deep layers(last CNN layer) should be reduce in order to maintain discrimination.
I searched for the reason a lot but I didn't get it clear, May someone explain it in some more detail please?
In theory you do not have to attach a fully connected layer, you could have a full stack of convolutions till the very end, as long as (due to custom sizes/paddings) you end up with the correct number of output neurons (usually number of classes).
So why people usually do not do that? If one goes through the math, it will become visible that each output neuron (thus - prediction wrt. to some class) depends only on the subset of the input dimensions (pixels). This would be something among the lines of a model, which only decides whether an image is an element of class 1 depending on first few "columns" (or, depending on the architecture, rows, or some patch of the image), then whether this is class 2 on a few next columns (maybe overlapping), ..., and finally some class K depending on a few last columns. Usually data does not have this characteristic, you cannot classify image of the cat based on a few first columns and ignoring the rest.
However, if you introduce fully connected layer, you provide your model with ability to mix signals, since every single neuron has a connection to every single one in the next layer, now there is a flow of information between each input dimension (pixel location) and each output class, thus the decision is based truly on the whole image.
So intuitively you can think about these operations in terms of information flow. Convolutions are local operations, pooling are local operations. Fully connected layers are global (they can introduce any kind of dependence). This is also why convolutions work so well in domains like image analysis - due to their local nature they are much easier to train, even though mathematically they are just a subset of what fully connected layers can represent.
note
I am considering here typical use of CNNs, where kernels are small. In general one can even think of MLP as a CNN, where the kernel is of the size of the whole input with specific spacing/padding. However these are just corner cases, which are not really encountered in practise, and not really affecting the reasoning, since then they end up being MLPs. The whole point here is simple - to introduce global relations, if one can do it by using CNNs in a specific manner - then MLPs are not needed. MLPs are just one way of introducing this dependence.
Every fully connected (FC) layer has an equivalent convolutional layer (but not vice versa). Hence it is not necessary to add FC layers. They can always be replaced by convolutional layers (+ reshaping). See details.
Why do we use FC layers then?
Because (1) we are used to it (2) it is simpler. (1) is probably the reason for (2). For example, you would need to adjust the loss fuctions / the shape of the labels / add a reshape add the end if you used a convolutional layer instead of a FC layer.
I found this answer by Anil-Sharma on Quora helpful.
We can divide the whole network (for classification) into two parts:
Feature extraction:
In the conventional classification algorithms, like SVMs, we used to extract features from the data to make the classification work. The convolutional layers are serving the same purpose of feature extraction. CNNs capture better representation of data and hence we don’t need to do feature engineering.
Classification:
After feature extraction we need to classify the data into various classes, this can be done using a fully connected (FC) neural network. In place of fully connected layers, we can also use a conventional classifier like SVM. But we generally end up adding FC layers to make the model end-to-end trainable.
The CNN gives you a representation of the input image. To learn the sample classes, you should use a classifier (such as logistic regression, SVM, etc.) that learns the relationship between the learned features and the sample classes. Fully-connected layer is also a linear classifier such as logistic regression which is used for this reason.
Convolution and pooling layers extract features from image. So this layer doing some "preprocessing" of data. Fully connected layrs perform classification based on this extracted features.
Are there known methods of continuous training and graceful degradation of a neural net while it shrinks or grows in size (by number of nodes, connections, whatever)?
To the best of my memory, everything I've read about neural networks is from a static perspective. You define the net and then train it.
If there is some neural network X with N nodes (neurons, whatever), is it possible to train the network (X) so that while N increases or decreases, the network is still useful and capable of performing?
In general, changing network architecture (adding new layers, adding more neurons into existing layers) once the network was already trained makes sense and a rather common operation in Deep Learning domain. One example is the dropout - during training half of the neurons randomly get switched off completely and only remaining half participates in training during specific iteration (each iteration or 'epoch' as it often is named has different random list of switched off neurons). Another example is transfer learning - where you learn network on one set of input data, cut off part of the outcoming layers, replace them with new layers and re-learn the model on another dataset.
To better explain why it makes sense lets step back for a moment. In deep networks, where you have lots of hidden layers each layer learns some abstraction from the incoming data. Each additional layer uses abstract representations learned by previous layer and builds upon them, combining such abstraction to form a higher level of the data representation. For instance, you could be trying to classify the images with DNN. First layer will learn rather simple concepts from images - like edges or points in data. Next layer could combine this simple concepts to learn primitives - like triangles or circles of squares. Next layer could drive it further and combine this primitives to represent some objects which you could find in images, like 'a car' or 'a house'and using softmax it calculates the probabilities of the answer you are looking for (what to actually output). I need to mention that these facts and learned representations could be actually checked. You could visualize the activation of your hidden layer and see what it learned. For example this was done with google's project 'inceptionism'. With that in mind let's get back to what I mentioned earlier.
Dropout is used to improve generalization of the network. It forces each neuron to 'not be so sure' that some pieces of the information from the previous layer will be available and makes it to try to learn the representations relying on less favorable and informative pieces of abstractions from previous layer. It forces it to consider all of the representations from previous layer to make decisions instead of putting all of its weight into couple of neurons it 'likes most of all'. By doing this the network is usually better prepared to new data where the input will be different from the training set.
Q: "As far as you're aware is the quality of the stored knowledge (whatever training has done to the net) still usable following the dropout? Maybe random halves could be substituted by random 10ths with a single 10th dropping, that might result in less knowledge loss during the transition period."
A: Unfortunately I can't properly answer why precisely half of the neurons is switched off and not 10% (or any other number). Maybe there is an explanation but I haven't seen it. In general it just works and that's it.
Also I need to mention that the task of dropout is to ensure that each neuron doesn't consider just several of the neurons from previous layer and is ready to make some decision even if neurons which usually helped it to make correct decision are not available. This is used for generalization only and helps the network to better cope with the data it haven't seen previously, nothing else is achieved with a dropout.
Now let's consider Transfer Learning again. Consider that you have a network with 4 layers. You train it to recognize specific objects in pictures (cat, dog, table, car etc). Than you cut off last layer, replace it with three additional layers and now you train the resulting 6-layered network on a dataset which, for instance, wrights short sentences about what is shown on this image ('a cat is on the car', 'house with windows and tree nearby' etc). What we did with such operation? Our original 4-layer network was capable to understand if some specific object is in the image we feed it with. Its first 3 layers learned good representations of the images - first layer learned about possible edges or points or some extremely primitive geometric shapes in images. Second layer learned some more elaborate geometric figures like 'circle' or 'square'. Last layer knows how to combine them to form some higher level objects - 'car', 'cat', 'house'. Now, we could just re-use this good representation which we learned in different domain and just add several more layers. Each of them will use abstractions from last (3rd) layer of original network and learn how combine them to create meaningful descriptions of images. While you will perform learning on new dataset with images as input and sentences as output it will adjust first 3 layers which we got from original network but these adjustments will be mostly minor, while 3 new layers will be adjusted by learning significantly. What we achieve with transfer learning is:
1) We can learn a much better data representations. We could create a network which is very good at specific task and than build upon that network to perform something different.
2) We can save training time - first layers of network will already be trained well enough so that your layers which are closer to output already get a rather good data representations. So the training should finish much faster using pre-trained first layers.
So the bottom line is that pre-training some network and than re-using part or whole network in another network makes perfect sense and is not something uncommon.
This is something I have seen in the likes of this video...
https://youtu.be/qv6UVOQ0F44
There are links to further resources in the video description.
And is based on a process called NEAT. Neuro Evolution of Augmenting Topologies.
It uses a genetic algorithm and evolutionary process to design and evolve a neural net from scratch with no prior assumptions of structure or complexity of the neural net.
I believe this is what you are looking for.
Would someone be able to explain to me or point me to some resources of why (or situations where) more than one hidden layer would be necessary or useful in a neural network?
Basically more layers allow more functions to be represented. The standard book for AI courses, "Artificial Intelligence, A Modern Approach" by Russell and Norvig, goes into some detail of why multiple layers matter in Chapter 20.
One important point is that with a sufficiently large single hidden layer, you can represent every continuous function, but you will need at least 2 layers to be able to represent every discontinuous function.
In practice, though, a single layer is enough at least 99% of the time.
That's more similar to the way the brain works (which might not necessarily be a computational advantage, but a lot of people are researching NN to gain insight about the way the mind works, rather than to solve real world problems.
Its easier to achieve some kinds of invariance using more layers. For example, an image classifier that works regardless of where in the image the object is found, or the object's size. see Bouvrie, J. , L. Rosasco, and T. Poggio. "On Invariance in Hierarchical Models". Advances in Neural Information Processing Systems (NIPS) 22, 2009.
Each layer effectively raises the potential "complexity" of adaptation in an exponential fashion (as opposed to a multiplicative fashion of adding more nodes to a single layer).