# What is missing from MATLAB #2 - the next decade edition

55 次查看（过去 30 天）

显示 更早的评论

Meta threads have a tendency to grow large. This has happened several times before (the wishlist threads #1 #2 #3 #4 #5, and 'What frustrates you about MATLAB?' #1 and #2).

No wonder that a thread from early 2011 has also kept growing. After just under a decade there are (at time of writing) 119 answers, making the page slow to load and navigate (especially on mobile). So after a friendly nudge; here is a new thread for the things that are missing from Matlab.

Same question: are there things you think should be possible in Matlab, but aren't? What things are possible with software packages similar to Matlab that Matlab would benefit from? (note that you can also submit an enhancement request through support, although I suspect they will be monitoring activity on this thread as well)

What should you post where?

Next Gen threads (#1): features that would break compatibility with previous versions, but would be nice to have

@anyone posting a new thread when the last one gets too large (about 50 answers seems a reasonable limit per thread), please update this list in all last threads. (if you don't have editing privileges, just post a comment asking someone to do the edit)

##### 3 个评论

### 回答（37 个）

cui
2021-2-7

编辑：cui
2021-2-8

Deep Learning toolbox:

How to visualize the changes in dlarray/weight distribution with histograms in deep learning?

Histogram displays how the trend of tensor (weight, bias, gradient, etc.) changes during the training process in the form of histogram. Developers can adjust the model structures accurately by having an in-depth understanding of the effect of each layer.

##### 6 个评论

cui
2023-1-5

cui
2021-2-7

编辑：cui
2021-4-1

Deep Learning toolbox:

- action recognition: mmaction2
- video understanding, X3D: https://github.com/facebookresearch/SlowFast ，https://github.com/kenshohara/3D-ResNets-PyTorch
- reID: https://github.com/JDAI-CV/fast-reid
- track: https://github.com/ZQPei/deep_sort_pytorch，https://votchallenge.net/challenges.html
- Metric Learning: https://github.com/KevinMusgrave/pytorch-metric-learning
- image Classification: Zen-NAS: A Zero-Shot NAS for High-Performance Deep Image Recognition
- object detection: Centernet, https://github.com/xingyizhou/CenterNet , efficientDet, https://github.com/google/automl/tree/master/efficientdet，detectron2，https://github.com/facebookresearch/detectron2 ,PointRCNN,https://github.com/sshaoshuai/PointRCNN
- face recognition: insightface, https://github.com/deepinsight/insightface, sphereface,https://github.com/wy1iu/sphereface, arcface,https://github.com/ronghuaiyang/arcface-pytorch
- Semantic segmentation, instance segmentation，https://github.com/facebookresearch/maskrcnn-benchmark , Semantic Segmentation (Most of them are difficult to implement with matlab)
- deeplearning preprocess data: https://github.com/NVIDIA/DALI
- StyleGAN: https://github.com/NVlabs/stylegan
- DeepNetwork Architecture：Mnasnet，https://github.com/tensorflow/tpu/tree/master/models/official/mnasnet , Bayesian-Neural-Networks , Graph Convolutional Networks
- similarity search: https://github.com/facebookresearch/faiss
- deepfakes: https://github.com/deepfakes/faceswap
- deepnetwork inference performance: https://github.com/microsoft/onnxruntime ,https://github.com/zuoqing1988/ZQCNN
- image Labeler app: https://www.mathworks.com/matlabcentral/answers/789159-how-does-image-labeler-app-realize-this-interactive-operation-by-writing-a-program?s_tid=srchtitle

and so on....

The above are influential applications of deep learning in various aspects, but it is difficult to reproduce in matlab. Although Matlab2019b version supports automatic differentiation mechanism, it is still difficult to implement algorithms in matlab. The efficiency of the differentiation mechanism is not high, and many operators do not support it. I tried to implement the more famous yolov3/v4 algorithm with the latest MATLAB2020a version, but it is still not satisfactory

In summary, my personal suggestions are like my personal answer above, and I hope that future versions can improve a lot!

以上都为深度学习在各个方面有影响力的应用，但是在matlab中复现困难，虽然Matlab2019b版本支持自动微分机制，但仍然不易在matlab实现算法，微分机制效率不高，很多operators也不支持。。。

总之，我的个人建议就像上面的个人回答建议一样，我希望将来的版本可以有所改善！

##### 1 个评论

cui
2021-9-24

I have been following the latest development progress of "deeplearning toolbox", and although there are many new features listed, most of them are superficial improvements with very limited substantive support for the "operator", which is still very inadequate for experienced and in-depth researchers.

I hope that mathworks will seriously develop basic modules for the toolbox, instead of small features that are not very important to attract beginners.

我一直在关注“deeplearning toolbox”的最新开发进展，虽然有很多新的功能项被列出来，但这些大部分都是表面上改进，实质上的支持operator非常有限，对于有经验的深入研究者来说还显现的非常不足。

dpb
2020-9-10

##### 0 个评论

Walter Roberson
2020-9-13

A "soft interrupt" -- that is, an interrupt that can be handled with try/catch .

For example I am running a long calculation at the moment that iterates several times, and each iteration can take more than an hour. I don't always have the patience to wait through as many iterations as I asked for. For my purposes, the output of the last successful iteration would be "good enough" if I were to ask to interrupt the function.

I can control-C, but that interrupts the function completely, losing all of the outputs, and losing the function variables.

If I could somehow "soft interrupt" and have it return the current values of all variables, then that would be good enough for my current task. But the generalization of that would be the ability to catch a soft interrupt in a try/catch so that the code could make whatever final summariziation it needed to in order to create usable output variables. Furthermore, even though my own routine might be happy to return the "last good" values of the output variables, any soft-interrupt I requested might well get received while some lower-level routine had control that did not know about soft-interrupts, so I would want interrupt (with no return values) to apply to those layers until control reached my handler for a clean run-down.

##### 7 个评论

Sindar
2020-10-3

Michael Foote
2021-10-20

In the code analyzer, there is no easy way to find what #ok directive controls a particular flagged issue.

(And no overall list of the available #ok directives, at least that I can find. Even the enable/disable messages in preferences don't show them.)

##### 2 个评论

Steven Lord
2022-9-22

Shae Morgan
2020-7-31

facet_wrap or facet_grid (or general ggplot2 functionality) version of subplots, or some altered, simpler customizability for subplotting subsets of data.

gramm is an available toolbox, but it'd be nice to have it built in

##### 0 个评论

Rafael S.T. Vieira
2020-7-31

编辑：Rafael S.T. Vieira
2020-7-31

I would love to have a command TeX2sym and sym2TeX. With it, we could bring formulas from LaTeX and run them in MATLAB and the other way around. It is a tedious work doing it by hand, and I could bet most MATLAB users do it or will do it eventually.

Another useful feature to add would be arbitrary-precision arithmetic...languages such as python and Java (with Bignum), allows unlimited precision when dealing with integers at least. Granted that It is slow, but I believe that the MATLAB team could do something better. And I honestly feel like MATLAB is missing a feature by not having it...even some competing software has it.

And finally I was considering buying the Text Analytics Toolbox, and it would be nice, if it could have a grammar/spell checker. Even if it is not as advanced or doesn't contain all words. With live scripts, we can write an interactive document, so it would be nice if MATLAB could correct our spelling (even if it required a ToolBox to do it).

##### 5 个评论

Rafael S.T. Vieira
2020-8-1

编辑：Rafael S.T. Vieira
2020-8-1

Thank you for your interest, Walter and Rik. I believe that It is best to teach ourselves to code in a particular way, which will allow MATLAB to convert it to sym, than to be doing the task by hand every time.

Operator precedence could take care some of these issues. If \sin x + y was the input, then left-to-right precedence could dictate the output to be sin(x) + y. And to obtain sin(x+y), we would have to code in LaTeX \sin{x+y} or \sin(x+y).

Implied multiplication is indeed very common in mathematics and latex. On the other hand, it is also almost ubiquitous that variables are written just as single letters (especially if we are writing implied multiplication).

Finally, a command like tex2sym does not need to contemplate every math package imho, just some set of commands and macros. MATLAB could even return its best guess, and let us do the remainder of the task. Of course, ideally, we would just copy the contents from environments like $$ and \[ ... \] and paste them into MATLAB for a tex2sym conversion.

Walter Roberson
2020-9-9

I know I've said it before, but it's still missing and still important:

We need a way to gather all of the outputs of function call into a cell in the middle of an expression .

I know this may be tricky to implement internally. There are internal rules that are hard to work out, that have to do with how many outputs to request from functions. For example,

[A, B] = cellfun(@C, D, 'uniform', 0)

somehow passes knowledge of "two outputs" to C -- for example if you were to use @max then A would be a cell array of the maxima and B would be a cell array of the indices. The situation can get more complicated than this, and figuring out all the cases can make your head hurt. But we do know that any expression C(D(E)) that D(E) will be evaluated asking for exactly one output and that would be passed into C... but the knowledge of multiple outputs would be passed to C and yet not D.

The number of outputs to use is not inherently clear. If for example you call ode45 and ask to gather the outputs, are you asking for the common TOUT, YOUT case, or the full TOUT,YOUT,TE,YE,IE ? There are some cases where extra outputs can be expensive to calculate, so even though an operation that gathered "maximum" outputs might be useful, it is not always desireable, so ability to select the number would be useful.

Then there are issues with, for example, deal(), where you can have any number of outputs with just a single input:

[A,B,C,D] = deal(123)

would initialize A, B, C, and D all to 123. So if you ask to gather "all" of the outputs from deal(123), that number is not well defined.

Working these things out is not trivial -- but it is a really missing bit of the language.

There might be an opportunity for a syntax such as {}name -- e.g.,

arrayfun(@(X0) {}ode45(@f, tspan, X0), x0s)

meaning to gather all of the outputs of the call. At present, {} is not valid before a name.

##### 2 个评论

Walter Roberson
2020-9-10

Bruno,

Imagine that you want to implement

[temp{1:3}] = ndgrid(-1:.1:1);

C = cell2mat(cellfun(@(M) M(:), temp, 'uniform', 0);

That is, you want a single array in which each of the columns is one of the outputs of ndgrid(). And you want to do it as a single expression.

ndgrid() does not have a fixed number of outputs -- it is not like sin() with one fixed outputs, or max with two fixed outputs. nargout(@ndgrid) is -1 -- in other words the declaration is like

function varargout = ndgrid(varargin)

If you wanted to capture all of the outputs of max() then you could query ndgrid(@max) to get 2:

function outputs = gather_outputs(f, varargin)

n = nargout(f);

[outputs{1:n}] = f(varargin{:});

end

and you could

gather_outputs(@max, rand(3,5))

and this would be fine for gathering the two outputs of max into a single cell array.

But if we try to

gather_outputs(@deal, [])

then nargout(@deal) is -1 and if we said "Okay, take the absolute value of that -1" then we would be doing

[outputs{1:1}] = deal([])

which would give you just {[]} as the output.

This shows that you cannot just look at nargout() of the function you are invoking, such as @max or @deal .

Can we just look at nargout of the overall expression to determine the number of outputs to use for deal? No,

C = cell2mat(cellfun(@(M) M(:), gather_outputs(@ndgrid, -1:.1:1), 'uniform', 0));

would at best tell you that nargout is 1 (the C variable). You need something else to tell you the number of outputs you want to get -- something like

function outputs = gather_n_outputs(f, n, varargin)

[outputs{1:n}] = f(varargin{:});

end

and then you could

C = cell2mat(cellfun(@(M) M(:), gather_n_outputs(@ndgrid, 3, -1:.1:1), 'uniform', 0));

The request is to build this kind of facility in to MATLAB instead of having to write true functions like gather_n_outputs and have to pass function handles into them. Some syntax like

C = cell2mat(cellfun(@(M) M(:), {3}ndgrid(-1:.1:1), 'uniform',0));

where the hypothethical new syntax {3} indicates that you are requesting 3 outputs and that you want them gathered in a cell array. A common alternate syntax would be {} to request all of the outputs, like

C = cellfun(@(V) V.^2, {}max(rand(3,5)), 'uniform', 0)

which would hypothetically gather both (all) outputs of the max() call into a cell array that would then be available for further processing.

If you just used

C = cellfun(@(V) V.^2, max(rand(3,5)), 'uniform', 0)

then the knowledge of the single output would be carried through into the cellfun, which would tell max() to only emit a single output, so you would not get the second output processed. And the first output would be numeric not cell, so you would need arrayfun instead of cellfun...)

cui
2021-2-7

编辑：cui
2021-2-8

MATLAB

I very much hope that the official version will strengthen the readstruct function in the future! Lack of complete uniformity to support more format requirements.

##### 0 个评论

Jim Svensson
2021-10-24

Fix the semantics of "clear".

"clear all" does not clear all, but "clear classes" does. Go figure. I want a way to clear classes and only classes.

##### 1 个评论

Jim Svensson
2021-10-24

Seth Wagenman
2020-8-31

##### 2 个评论

Steven Lord
2022-9-22

Sindar
2020-11-5

Ok, this is pretty minor, but:

For common functions that return values as the first output and indices as the second, it would be nice if there was a direct way of getting the indices, so I could do things like this:

x = [1 2 5 4 5];

y = 1:5;

y_xsorted = y(sortInd(x));

% or

y_xunique = y(unique(x,'ia'));

instead of needing to create temporary variables:

x = [1 2 5 4 3];

y = 1:5;

[~,idx] = sort(x);

y_xsorted = y(idx);

[~,idx] = unique(x);

y_xunique = y(idx);

I know I could make wrappers myself, but this seems like a case where a builtin function could potentially be noticeably optimized. (And, is more straightforward than rebuilding the whole output system)

##### 0 个评论

Tim
2020-11-5

The following changes to the internal volume renderer would make Matlab much more useful for volume visualization:

- True RGB volume visualization (vs. just scalar data + colormap)
- Independent, decoupled specification of alpha values and intensity values
- The ability to combine volume images with standard axes and axes objects like points, lines, patches, etc.

##### 1 个评论

Walter Roberson
2022-8-24

Mario Malic
2020-11-13

##### 0 个评论

Bruno Luong
2020-11-17

编辑：Bruno Luong
2020-11-17

No big deal but I wish SIGN(X) function could return 1 for X=0 instead of 0. May be implementing an option to not break the compatibility.

I just rarely use SIGN because of this exception choice.

##### 19 个评论

Camilo Andres Manrique Escobar
2021-5-6

##### 0 个评论

Jim Svensson
2021-10-24

Maintain dimensions when getting a field of a class.

If Data is a 3x2x4 array of objects/structs with a field "foo", then Data.foo should be an 3x2x4 array of the type of "foo". Not a 1x24 array, or is it the other way around.

##### 3 个评论

Jim Svensson
2021-10-24

Yes you are right, now I remember. Always have to do something like "foo = [Data.foo]".

gwoo
2021-10-26

Being able to access the properties/methods/fields of an object/struct even after indexing into one would be a big deal.

Currently, if you have a struct A, with fields size and count, you could do A. then tab to show the options for the fields. But if you do A(1). you cannot now tab to show the options for the fields. Same with an object. So the discoverability is gone.

I want to be able to index into an object/struct and still have the available fields/properties/methods for that parent show up.

##### 2 个评论

Bruno Luong
2021-10-27

编辑：Bruno Luong
2021-10-27

Probably it breaks compatibility, but it would make functions/operators svd, *, ', .' dealing with n-d arrays similar to pagesvd, pagemtimes, ... followed by appropriated reshape.

Extend the pagexxx with matrix left/right division, lu, qr, chol, eig, etc...

##### 12 个评论

Bruno Luong
2023-2-9

编辑：Adam Danz
2023-2-9

@Adam Not only the floating point rounding can create different numerically result, but what matters even more is that the floating point counts when multiply matrices can be different.

Consider the product

A*B*x

with

- A: p x m,
- B: m x n,
- x: m x 1

with p = 100, n = 100, m = 1000,

It is much cheaper to perform

A*(B*x) than (A*B)*x.

And in general the less operations is carried out, the less round-off has effect on the final result.

dpb
2022-11-27

read/writetable from/to Excel workbooks should be able to return/write the comment (now called something else, I forget what) field and the formula associated with the cell as well as the value. While one can write COM to do so, it then takes either using COM entirely (can be a lot of work) or have to use both high-level and COM on the same file; either of which options isn't ideal. Sure, one could forego Excel entirely since have MATLAB, but unfortunately can't always have what we want in that regards.

One can mung on the old xlsread/write routine and add stuff into it, but it's also kinda' klunky and while more convenient in some ways as compared to writing all COM totally from scratch isn't nearly as convenient as would be a higher-level interface with the other current toolset.

There's a hook to a function call in xlsread although I was never able to get it to work to do either of the above; because the Excel COM object isn't exposed to the function to be able to make direct access from that function.

##### 0 个评论

dpb
2023-1-1

Maybe there's a way I've not found, but I wish arrayfun and cellfun would have automagic argument expansion so one could pass other arguments to the anonymous function without having to replicate them manually to match the others. This would add greatly to the convenience in using either; the present working example happens to be building a set of target range expressions to stuff into a cell array that will be written to Excel although that really has nothing to do with the request/enhancement, just happens to be current time was frustrated that there's no way to pass constants to the anonymous functions to allow them to be generalized.

Example:

Building a variably-sized workbook where it is desirable that the sums over a section of the sheet be formulas rather than the current fixed constant value of the data as the sheet will subsequently be modified by hand; the tool is to build the original working pattern by compending various data sources and arranging for the end user...

It boils down to the point at which one has a set of row range indices and a set of columns over which to build the formula and insert into the cell array which is subsequently written to the workbook. That code looks something like

% anonymous function that builds Excel =SUM(r1:r2) expression for given row range, column

xlsSumRange=@(r1,r2,c)strcat('=SUM(',xlsAddr(r1,c),':',xlsAddr(r2,c),')');

% typical use

col=xlsCol2Col('G'); % another internal translation layer of local array position from Excel column

cOut(isTotal,col)=arrayfun(@(r1,r2,c)xlsSumRange(r1,r2,c),RS1,RS2,repmat('G',numel(RS)-1,1),'UniformOutput',0);

col=xlsCol2Col('K'); % another internal translation layer of local array position from Excel column

cOut(isTotal,col)=arrayfun(@(r1,r2,c)xlsSumRange(r1,r2,c),RS1,RS2,repmat('K',numel(RS)-1,1),'UniformOutput',0);

This works, but the expression "repmat('G',numel(RS)-1,1)" needed is really inconvenient and clutters up the code legibility greatly. I've had any number of similar case in the past where an anonymous function is useful shorthand but then to use more than once requires a workaround like the above.

The alternative is to redefine the anonymous function dynamically and embed the constant inside it for the given invocation.

To visualize, an example output for the above for one invocation looks like for the call with column 'G'

K>> arrayfun(@(r1,r2,c)xlsSumRange(r1,r2,c),RS1,RS2,repmat('K',numel(RS)-1,1),'UniformOutput',0);

ans =

3×1 cell array

{'=SUM($G$3:$G$17)' }

{'=SUM($G$22:$G$372)' }

{'=SUM($G$377:$G$414)'}

K>>

The above could be in a loop over the number of columns instead of using the explicit columns, but was part of still rearranging the output file structure at the time...and, haven't taken the time to clean it all up yet...

##### 8 个评论

Andrew Janke
2023-1-4

+1 on this, @dpb. This "implicit scalar expansion for (array|cell)fun args" is a use case I've had occasional cause to want, and haven't been able to figure out a good way to do, except for ditching arrayfun and Functional Programming style and inverting the logic/control-flow to just write a vectorized or "regular Matlab" version of the function I'm trying to do.

In other use cases, as long as I don't care about method-call overhead, I've been able to wrap the value to be expanded in an "infinite copies" array, like this SPAM thing here:

classdef SPAM

% "spam" a value to every element of an infinite virtual array

properties

Value

end

methods

function this = SPAM(x)

this.Value = x;

end

function out = subsref(this, S) %#ok<INUSD>

out = this.Value;

end

end

end

But that doesn't seem to work with cellfun/arrayfun, because it looks like they do an explicit "size(...)" test on their inputs before trying to index in to them.

>> x = magic(3);

>> z = arrayfun(@plus, x, SPAM(420))

Error using arrayfun

All of the input arguments must be of the same size and shape.

Previous inputs had size 3 in dimension 1. Input #3 has size 1

>> spam = SPAM(420);

>> z = NaN(size(x)); for i = 1:numel(z); z(i) = x(i) + spam(i); end

>> z

z =

428 421 426

423 425 427

424 429 422

>>

Can't say I blame Matlab for this, though. I like explicit size-conformity checks; catches bugs more often than it prevents features.

dpb
2023-2-8

编辑：dpb
2023-2-8

##### 1 个评论

Andrew Janke
2023-2-8

Adam
2023-2-16

编辑：Adam
2023-2-16

When defining Abstract methods on a class, I would like to be able to fix the arguments on the abstract class.

Something like this would be usefull:

classdef AbstractSuperClass

methods (Abstract)

function AbstractMethod(self,input1,input2)

arguments

self

input1 (1,1) double

input2 (1,1) double

end

% no function body because it is abstract

end

end

end

Right now, I always end writing two versions of the same method to avoid the need to copy the arguments validation block to the subclass:

classdef AbstractSuperClass

methods

function Method(self,input1,input2)

arguments

self

input1 (1,1) double

input2 (1,1) double

end

abstract_version_of_method(self,input1,input2)

end

end

methods (Abstract,access=protected)

abstract_version_of_method(self,input1,input2)

end

end

but this feels wrong to do it and results in programming errors because I forget what the input signature of the method is, the above syntax would be more elegant and would allow for checking whether the input signature of the implementation of the method follows the definition in the abstract class.

With the new arguments (output), I would even be able to fix the output type of my abstract function

##### 1 个评论

Andrew Janke
2023-2-17

Mikhail
2020-8-2

Some (well, most) dynamic programming languages allow for an experienced developer to have an insight look into how their code is actually executed by the runtime environment.

They are able to see the parse tree and the compiled byte code.

They are able to see what parts of code are JIT-compiled, and how exactly the are compiled. When something is not compiled, they are able to see why.

The developer doesn't have to guess whether a particular optimization has kicked in or not. They know for sure how each and every object in their code is handled (whether it is CoW, passed by reference, passed by value).

I'd love to see these capabilities in MATLAB, too.

dpb
2020-9-6

编辑：dpb
2020-9-6

A ready-built insertion method for tables, arrays, etc., so don't have to do all the under-the-hood indexing manually...given an array of the right size and matching index, if had table (say), then

tMyTable=insert(tMyTable,indices,tNewData);

ideally, an option with keyword 'Empty', T|F would be available as well.

Maybe I'm missing some magic pixie dust, but I've been unable to figure out a way to this without physically moving a section at a time to insert the new row which means working from rear to front to avoid changing the indices to the insertion points or by catenating the pieces starting from the front.

##### 7 个评论

dpb
2022-9-22

编辑：dpb
2022-9-23

No, the feature/extension doesn't yet exist; the 'Before/After' position is one value, whether it is provided as index, variable name or logical vector. The logical vector, while a vector, can have only one true position.

I'm suggesting N new variables should be able to be added, each at some specified position -- use case/example is to add the column of means following each variable, for example. As is, they all have to go in one place and then get rearranged or added one-at-a-time.

Walter Roberson
2020-10-12

A "select" clause for readtable() and kin.

For example one user only wanted to read the rows in which one particular variable had a particular value

It could at some point be implemented in terms of a rowfun() type function that got passed the variables for a row and could make arbitrary decisions based upon the row contents.

However an earlier stage could potentially have pairs of variable names (or numbers) and a vector or cell array of values, in which the selection code did an ismember(). Such a facility could be further improved if there were a "sorted" option (so the code could figure out when to give up looking -- any one value should no longer be looked for if a larger value were encountered). Or even "grouped", which would not imply sorted as such but would imply that when you find a value that all the instances of that value that exist will be in adjacent rows and so as soon as you detect a change you can know to give up looking for that value.

##### 0 个评论

Bruno Luong
2020-11-6

##### 0 个评论

Bruno Luong
2020-11-6