I’m designing a custom training procedure for a CNN that is different from backpropagation in that I use manual update rules for layers or sets of layers. I’m studying my gradient for two types of layers: “conv + actfun + maxpool”, and “conv + actfun + avgpool”, which are identical layers except the last action is a different pooling type.
I compared the two layer types with identical data dimension sizes to see the time differences between maxpool and avgpool, both in the forward pass and the backwards pass of the pooling layers. All other steps in calculating the gradient were exactly the same between the two layers, and showed the same time costs in the two layers. But when looking at time costs specifically of the pooling operations’ forward and backwards passes, I get significantly different times (average of 5000 runs of the gradient, each measurement is in milliseconds):
gradient step | AvgPool | MaxPool | Difference
————————–|———|———|———-
pooling (forward pass) | 0.4165 | 38.6316 | +38.2151
unpooling (backward pass) | 9.9468 | 46.1667 | +36.2199

For reference, all my data arrays are dlarrays on the GPU (gpuArrays in dlarrays), all single precision, and the pooling operations convert 32 by 32 feature maps (across 2 channels and 16384 batch size) to 16 by 16 feature maps (of same # channels and batch size), so just a 2 by 2 pooling operation.
You can see here that the maxpool forward pass (using “maxpool” function) is about 92 times slower than the avgpool forward pass (using “avgpool”), and the maxpool backward pass (using “maxunpool”) is about 4.6 times slower than the avgpool backward pass (using a custom “avgunpool” function that Anthropic’s Claude had to create for me, since matlab has no “avgunpool”).
These results are extremely suspect to me. For the forwards pass, comparing matlab’s built in "maxpool" to built in "avgpool" functions gives a 92x difference, but searching online people seem to instead claim that training max pooling is supposed to be faster than avg pooling, which contradicts the results here.
For simplicity, see the code example below that runs just "maxpool" and "avgpool" only (no other functions) and compares their times:

function analyze_pooling_timing()

% GPU setup
g = gpuDevice();
fprintf(‘GPU: %sn’, g.Name);

% Parameters matching your test
H_in = 32; W_in = 32; C_in = 3; C_out = 2;
N = 16384; % batch size. Try N = 32 small or N = 16384 big
kH = 3; kW = 3;

pool_params.pool_size = [2, 2];
pool_params.pool_stride = [2, 2];
pool_params.pool_padding = 0;

conv_params.stride = [1, 1];
conv_params.padding = ‘same’;
conv_params.dilation = [1, 1];

% Initialize data
Wj = dlarray(gpuArray(single(randn(kH, kW, C_in, C_out) * 0.01)), ‘SSCU’);
Bj = dlarray(gpuArray(single(zeros(C_out, 1))), ‘C’);
Fjmin1 = dlarray(gpuArray(single(randn(H_in, W_in, C_in, N))), ‘SSCB’);

% Number of iterations for averaging
num_iter = 100;
fprintf(‘Running %d iterations for each timing measurement…nn’, num_iter);

%% setup everything in forward pass before the pooling:
% Forward convolution
Sj = dlconv(Fjmin1, Wj, Bj, …
‘Stride’, conv_params.stride, …
‘Padding’, conv_params.padding, …
‘DilationFactor’, conv_params.dilation);
% activation function (and derivative)
Oj = max(Sj, 0); Fprimej = sign(Oj);

%% Time AVERAGE POOLING
fprintf(‘=== AVERAGE POOLING (conv_af_ap) ===n’);
times_ap = struct();

for iter = 1:num_iter

% Average pooling
tic;
Oj_pooled = avgpool(Oj, pool_params.pool_size, …
‘Stride’, pool_params.pool_stride, …
‘Padding’, pool_params.pool_padding);
wait(g);
times_ap.pooling(iter) = toc;

end

%% Time MAX POOLING
fprintf(‘n=== MAX POOLING (conv_af_mp) ===n’);
times_mp = struct();

for iter = 1:num_iter

% Max pooling with indices
tic;
[Oj_pooled, max_indices] = maxpool(Oj, pool_params.pool_size, …
‘Stride’, pool_params.pool_stride, …
‘Padding’, pool_params.pool_padding);
wait(g);
times_mp.pooling(iter) = toc;

end

%% Compute statistics and display results
fprintf(‘n=== TIMING RESULTS (milliseconds) ===n’);
fprintf(‘%-25s %12s %12s %12sn’, ‘Step’, ‘AvgPool’, ‘MaxPool’, ‘Difference’);
fprintf(‘%sn’, repmat(‘-‘, 1, 65));

steps_common = { ‘pooling’};

total_ap = 0;
total_mp = 0;

for i = 1:length(steps_common)
step = steps_common{i};
if isfield(times_ap, step) && isfield(times_mp, step)
mean_ap = mean(times_ap.(step)) * 1000; % times 1000 to convert seconds to milliseconds
mean_mp = mean(times_mp.(step)) * 1000; % times 1000 to convert seconds to milliseconds
total_ap = total_ap + mean_ap;
total_mp = total_mp + mean_mp;
diff = mean_mp – mean_ap;
fprintf(‘%-25s %12.4f %12.4f %+12.4fn’, step, mean_ap, mean_mp, diff);
end
end

fprintf(‘%sn’, repmat(‘-‘, 1, 65));
%fprintf(‘%-25s %12.4f %12.4f %+12.4fn’, ‘TOTAL’, total_ap, total_mp, total_mp – total_ap);
fprintf(‘%-25s %12s %12s %12.2fxn’, ‘Speedup’, ”, ”, total_mp/total_ap);

end

Now you can see one main difference between my calling of maxpool and avgpool: for avgpool I only have 1 output (the pooled values), but with maxpool I have 2 outputs (the pooled values, and the index locations of these max values).

This is important because if I replaced the call to maxpool with only requesting the pooled values (1 output), maxpool is faster as expected:

>> analyze_pooling_timing
GPU: NVIDIA GeForce RTX 5080
Running 1000 iterations for each timing measurement…

=== AVERAGE POOLING (conv_af_ap) ===

=== MAX POOLING (conv_af_mp) ===

=== TIMING RESULTS (milliseconds) ===
Step AvgPool MaxPool Difference
—————————————————————–
pooling 0.4358 0.2330 -0.2028
—————————————————————–
max/avg 0.53x
>>

However if I call maxpool like here with 2 outputs (pooled values and indices), or with 3 outputs (pooled values, indices, and inputSize), this is where maxpool is extremely slow:

>> analyze_pooling_timing
GPU: NVIDIA GeForce RTX 5080
Running 1000 iterations for each timing measurement…

=== AVERAGE POOLING (conv_af_ap) ===

=== MAX POOLING (conv_af_mp) ===

=== TIMING RESULTS (milliseconds) ===
Step AvgPool MaxPool Difference
—————————————————————–
pooling 0.4153 38.9818 +38.5665
—————————————————————–
max/avg 93.86x
>>

So it appears that this is the reason why maxpool is so much slower: requesting the maxpool function to also output the indices leads to a massive slowdown.
Unfortunately, the indices are needed to later differentiate (backwards pass) the maxpool layer… so I need the indices…
I’d assume that whenever someone wants to train a CNN in matlab using a maxpool layer, they would have to call maxpool with indices, and thus I’d expect a similar slowdown…
Can anyone here comment on this? My application needs me to train maxpool layers, so I need to both forward pass and backwards pass through them, thus I need these indices. But it appears that matlab’s version of maxpool may not be the best implementation of a maxpool operation… maybe some inefficient processes regarding obtaining the indices?I’m designing a custom training procedure for a CNN that is different from backpropagation in that I use manual update rules for layers or sets of layers. I’m studying my gradient for two types of layers: “conv + actfun + maxpool”, and “conv + actfun + avgpool”, which are identical layers except the last action is a different pooling type.
I compared the two layer types with identical data dimension sizes to see the time differences between maxpool and avgpool, both in the forward pass and the backwards pass of the pooling layers. All other steps in calculating the gradient were exactly the same between the two layers, and showed the same time costs in the two layers. But when looking at time costs specifically of the pooling operations’ forward and backwards passes, I get significantly different times (average of 5000 runs of the gradient, each measurement is in milliseconds):
gradient step | AvgPool | MaxPool | Difference
————————–|———|———|———-
pooling (forward pass) | 0.4165 | 38.6316 | +38.2151
unpooling (backward pass) | 9.9468 | 46.1667 | +36.2199

For reference, all my data arrays are dlarrays on the GPU (gpuArrays in dlarrays), all single precision, and the pooling operations convert 32 by 32 feature maps (across 2 channels and 16384 batch size) to 16 by 16 feature maps (of same # channels and batch size), so just a 2 by 2 pooling operation.
You can see here that the maxpool forward pass (using “maxpool” function) is about 92 times slower than the avgpool forward pass (using “avgpool”), and the maxpool backward pass (using “maxunpool”) is about 4.6 times slower than the avgpool backward pass (using a custom “avgunpool” function that Anthropic’s Claude had to create for me, since matlab has no “avgunpool”).
These results are extremely suspect to me. For the forwards pass, comparing matlab’s built in "maxpool" to built in "avgpool" functions gives a 92x difference, but searching online people seem to instead claim that training max pooling is supposed to be faster than avg pooling, which contradicts the results here.
For simplicity, see the code example below that runs just "maxpool" and "avgpool" only (no other functions) and compares their times:

function analyze_pooling_timing()

% GPU setup
g = gpuDevice();
fprintf(‘GPU: %sn’, g.Name);

% Parameters matching your test
H_in = 32; W_in = 32; C_in = 3; C_out = 2;
N = 16384; % batch size. Try N = 32 small or N = 16384 big
kH = 3; kW = 3;

pool_params.pool_size = [2, 2];
pool_params.pool_stride = [2, 2];
pool_params.pool_padding = 0;

conv_params.stride = [1, 1];
conv_params.padding = ‘same’;
conv_params.dilation = [1, 1];

% Initialize data
Wj = dlarray(gpuArray(single(randn(kH, kW, C_in, C_out) * 0.01)), ‘SSCU’);
Bj = dlarray(gpuArray(single(zeros(C_out, 1))), ‘C’);
Fjmin1 = dlarray(gpuArray(single(randn(H_in, W_in, C_in, N))), ‘SSCB’);

% Number of iterations for averaging
num_iter = 100;
fprintf(‘Running %d iterations for each timing measurement…nn’, num_iter);

%% setup everything in forward pass before the pooling:
% Forward convolution
Sj = dlconv(Fjmin1, Wj, Bj, …
‘Stride’, conv_params.stride, …
‘Padding’, conv_params.padding, …
‘DilationFactor’, conv_params.dilation);
% activation function (and derivative)
Oj = max(Sj, 0); Fprimej = sign(Oj);

%% Time AVERAGE POOLING
fprintf(‘=== AVERAGE POOLING (conv_af_ap) ===n’);
times_ap = struct();

for iter = 1:num_iter

% Average pooling
tic;
Oj_pooled = avgpool(Oj, pool_params.pool_size, …
‘Stride’, pool_params.pool_stride, …
‘Padding’, pool_params.pool_padding);
wait(g);
times_ap.pooling(iter) = toc;

end

%% Time MAX POOLING
fprintf(‘n=== MAX POOLING (conv_af_mp) ===n’);
times_mp = struct();

for iter = 1:num_iter

% Max pooling with indices
tic;
[Oj_pooled, max_indices] = maxpool(Oj, pool_params.pool_size, …
‘Stride’, pool_params.pool_stride, …
‘Padding’, pool_params.pool_padding);
wait(g);
times_mp.pooling(iter) = toc;

end

%% Compute statistics and display results
fprintf(‘n=== TIMING RESULTS (milliseconds) ===n’);
fprintf(‘%-25s %12s %12s %12sn’, ‘Step’, ‘AvgPool’, ‘MaxPool’, ‘Difference’);
fprintf(‘%sn’, repmat(‘-‘, 1, 65));

steps_common = { ‘pooling’};

total_ap = 0;
total_mp = 0;

for i = 1:length(steps_common)
step = steps_common{i};
if isfield(times_ap, step) && isfield(times_mp, step)
mean_ap = mean(times_ap.(step)) * 1000; % times 1000 to convert seconds to milliseconds
mean_mp = mean(times_mp.(step)) * 1000; % times 1000 to convert seconds to milliseconds
total_ap = total_ap + mean_ap;
total_mp = total_mp + mean_mp;
diff = mean_mp – mean_ap;
fprintf(‘%-25s %12.4f %12.4f %+12.4fn’, step, mean_ap, mean_mp, diff);
end
end

fprintf(‘%sn’, repmat(‘-‘, 1, 65));
%fprintf(‘%-25s %12.4f %12.4f %+12.4fn’, ‘TOTAL’, total_ap, total_mp, total_mp – total_ap);
fprintf(‘%-25s %12s %12s %12.2fxn’, ‘Speedup’, ”, ”, total_mp/total_ap);

end

Now you can see one main difference between my calling of maxpool and avgpool: for avgpool I only have 1 output (the pooled values), but with maxpool I have 2 outputs (the pooled values, and the index locations of these max values).

This is important because if I replaced the call to maxpool with only requesting the pooled values (1 output), maxpool is faster as expected:

>> analyze_pooling_timing
GPU: NVIDIA GeForce RTX 5080
Running 1000 iterations for each timing measurement…

=== AVERAGE POOLING (conv_af_ap) ===

=== MAX POOLING (conv_af_mp) ===

=== TIMING RESULTS (milliseconds) ===
Step AvgPool MaxPool Difference
—————————————————————–
pooling 0.4358 0.2330 -0.2028
—————————————————————–
max/avg 0.53x
>>

However if I call maxpool like here with 2 outputs (pooled values and indices), or with 3 outputs (pooled values, indices, and inputSize), this is where maxpool is extremely slow:

>> analyze_pooling_timing
GPU: NVIDIA GeForce RTX 5080
Running 1000 iterations for each timing measurement…

=== AVERAGE POOLING (conv_af_ap) ===

=== MAX POOLING (conv_af_mp) ===

=== TIMING RESULTS (milliseconds) ===
Step AvgPool MaxPool Difference
—————————————————————–
pooling 0.4153 38.9818 +38.5665
—————————————————————–
max/avg 93.86x
>>

So it appears that this is the reason why maxpool is so much slower: requesting the maxpool function to also output the indices leads to a massive slowdown.
Unfortunately, the indices are needed to later differentiate (backwards pass) the maxpool layer… so I need the indices…
I’d assume that whenever someone wants to train a CNN in matlab using a maxpool layer, they would have to call maxpool with indices, and thus I’d expect a similar slowdown…
Can anyone here comment on this? My application needs me to train maxpool layers, so I need to both forward pass and backwards pass through them, thus I need these indices. But it appears that matlab’s version of maxpool may not be the best implementation of a maxpool operation… maybe some inefficient processes regarding obtaining the indices? I’m designing a custom training procedure for a CNN that is different from backpropagation in that I use manual update rules for layers or sets of layers. I’m studying my gradient for two types of layers: “conv + actfun + maxpool”, and “conv + actfun + avgpool”, which are identical layers except the last action is a different pooling type.
I compared the two layer types with identical data dimension sizes to see the time differences between maxpool and avgpool, both in the forward pass and the backwards pass of the pooling layers. All other steps in calculating the gradient were exactly the same between the two layers, and showed the same time costs in the two layers. But when looking at time costs specifically of the pooling operations’ forward and backwards passes, I get significantly different times (average of 5000 runs of the gradient, each measurement is in milliseconds):
gradient step | AvgPool | MaxPool | Difference
————————–|———|———|———-
pooling (forward pass) | 0.4165 | 38.6316 | +38.2151
unpooling (backward pass) | 9.9468 | 46.1667 | +36.2199

For reference, all my data arrays are dlarrays on the GPU (gpuArrays in dlarrays), all single precision, and the pooling operations convert 32 by 32 feature maps (across 2 channels and 16384 batch size) to 16 by 16 feature maps (of same # channels and batch size), so just a 2 by 2 pooling operation.
You can see here that the maxpool forward pass (using “maxpool” function) is about 92 times slower than the avgpool forward pass (using “avgpool”), and the maxpool backward pass (using “maxunpool”) is about 4.6 times slower than the avgpool backward pass (using a custom “avgunpool” function that Anthropic’s Claude had to create for me, since matlab has no “avgunpool”).
These results are extremely suspect to me. For the forwards pass, comparing matlab’s built in "maxpool" to built in "avgpool" functions gives a 92x difference, but searching online people seem to instead claim that training max pooling is supposed to be faster than avg pooling, which contradicts the results here.
For simplicity, see the code example below that runs just "maxpool" and "avgpool" only (no other functions) and compares their times:

function analyze_pooling_timing()

% GPU setup
g = gpuDevice();
fprintf(‘GPU: %sn’, g.Name);

% Parameters matching your test
H_in = 32; W_in = 32; C_in = 3; C_out = 2;
N = 16384; % batch size. Try N = 32 small or N = 16384 big
kH = 3; kW = 3;

pool_params.pool_size = [2, 2];
pool_params.pool_stride = [2, 2];
pool_params.pool_padding = 0;

conv_params.stride = [1, 1];
conv_params.padding = ‘same’;
conv_params.dilation = [1, 1];

% Initialize data
Wj = dlarray(gpuArray(single(randn(kH, kW, C_in, C_out) * 0.01)), ‘SSCU’);
Bj = dlarray(gpuArray(single(zeros(C_out, 1))), ‘C’);
Fjmin1 = dlarray(gpuArray(single(randn(H_in, W_in, C_in, N))), ‘SSCB’);

% Number of iterations for averaging
num_iter = 100;
fprintf(‘Running %d iterations for each timing measurement…nn’, num_iter);

%% setup everything in forward pass before the pooling:
% Forward convolution
Sj = dlconv(Fjmin1, Wj, Bj, …
‘Stride’, conv_params.stride, …
‘Padding’, conv_params.padding, …
‘DilationFactor’, conv_params.dilation);
% activation function (and derivative)
Oj = max(Sj, 0); Fprimej = sign(Oj);

%% Time AVERAGE POOLING
fprintf(‘=== AVERAGE POOLING (conv_af_ap) ===n’);
times_ap = struct();

for iter = 1:num_iter

% Average pooling
tic;
Oj_pooled = avgpool(Oj, pool_params.pool_size, …
‘Stride’, pool_params.pool_stride, …
‘Padding’, pool_params.pool_padding);
wait(g);
times_ap.pooling(iter) = toc;

end

%% Time MAX POOLING
fprintf(‘n=== MAX POOLING (conv_af_mp) ===n’);
times_mp = struct();

for iter = 1:num_iter

% Max pooling with indices
tic;
[Oj_pooled, max_indices] = maxpool(Oj, pool_params.pool_size, …
‘Stride’, pool_params.pool_stride, …
‘Padding’, pool_params.pool_padding);
wait(g);
times_mp.pooling(iter) = toc;

end

%% Compute statistics and display results
fprintf(‘n=== TIMING RESULTS (milliseconds) ===n’);
fprintf(‘%-25s %12s %12s %12sn’, ‘Step’, ‘AvgPool’, ‘MaxPool’, ‘Difference’);
fprintf(‘%sn’, repmat(‘-‘, 1, 65));

steps_common = { ‘pooling’};

total_ap = 0;
total_mp = 0;

for i = 1:length(steps_common)
step = steps_common{i};
if isfield(times_ap, step) && isfield(times_mp, step)
mean_ap = mean(times_ap.(step)) * 1000; % times 1000 to convert seconds to milliseconds
mean_mp = mean(times_mp.(step)) * 1000; % times 1000 to convert seconds to milliseconds
total_ap = total_ap + mean_ap;
total_mp = total_mp + mean_mp;
diff = mean_mp – mean_ap;
fprintf(‘%-25s %12.4f %12.4f %+12.4fn’, step, mean_ap, mean_mp, diff);
end
end

fprintf(‘%sn’, repmat(‘-‘, 1, 65));
%fprintf(‘%-25s %12.4f %12.4f %+12.4fn’, ‘TOTAL’, total_ap, total_mp, total_mp – total_ap);
fprintf(‘%-25s %12s %12s %12.2fxn’, ‘Speedup’, ”, ”, total_mp/total_ap);

end

Now you can see one main difference between my calling of maxpool and avgpool: for avgpool I only have 1 output (the pooled values), but with maxpool I have 2 outputs (the pooled values, and the index locations of these max values).

This is important because if I replaced the call to maxpool with only requesting the pooled values (1 output), maxpool is faster as expected:

>> analyze_pooling_timing
GPU: NVIDIA GeForce RTX 5080
Running 1000 iterations for each timing measurement…

=== AVERAGE POOLING (conv_af_ap) ===

=== MAX POOLING (conv_af_mp) ===

=== TIMING RESULTS (milliseconds) ===
Step AvgPool MaxPool Difference
—————————————————————–
pooling 0.4358 0.2330 -0.2028
—————————————————————–
max/avg 0.53x
>>

However if I call maxpool like here with 2 outputs (pooled values and indices), or with 3 outputs (pooled values, indices, and inputSize), this is where maxpool is extremely slow:

>> analyze_pooling_timing
GPU: NVIDIA GeForce RTX 5080
Running 1000 iterations for each timing measurement…

=== AVERAGE POOLING (conv_af_ap) ===

=== MAX POOLING (conv_af_mp) ===

=== TIMING RESULTS (milliseconds) ===
Step AvgPool MaxPool Difference
—————————————————————–
pooling 0.4153 38.9818 +38.5665
—————————————————————–
max/avg 93.86x
>>

So it appears that this is the reason why maxpool is so much slower: requesting the maxpool function to also output the indices leads to a massive slowdown.
Unfortunately, the indices are needed to later differentiate (backwards pass) the maxpool layer… so I need the indices…
I’d assume that whenever someone wants to train a CNN in matlab using a maxpool layer, they would have to call maxpool with indices, and thus I’d expect a similar slowdown…
Can anyone here comment on this? My application needs me to train maxpool layers, so I need to both forward pass and backwards pass through them, thus I need these indices. But it appears that matlab’s version of maxpool may not be the best implementation of a maxpool operation… maybe some inefficient processes regarding obtaining the indices? deep learning, maxpool, avgpool, indexing MATLAB Answers — New Questions

​

I have some data that I am displaying in my script and have outputted it in 28 different tables. My script is quite long so I am just posting a screenshot for now.

Where is says var 2 I want to change to the storey number. So for table 1 I want storey 1, for table 2, storey 2 etc…

If you can see for the main output i am running a loop and using the matrix called "Int_force". I also have made the row headers as [Axial Force, Shear Force, Bending Moment. Thus, I am hoping someone can help me so that I can change the column headers to my desired wish? I am not sure how I would change the loop to accomplish this.
Many thanks,
ScottI have some data that I am displaying in my script and have outputted it in 28 different tables. My script is quite long so I am just posting a screenshot for now.

Where is says var 2 I want to change to the storey number. So for table 1 I want storey 1, for table 2, storey 2 etc…

If you can see for the main output i am running a loop and using the matrix called "Int_force". I also have made the row headers as [Axial Force, Shear Force, Bending Moment. Thus, I am hoping someone can help me so that I can change the column headers to my desired wish? I am not sure how I would change the loop to accomplish this.
Many thanks,
Scott I have some data that I am displaying in my script and have outputted it in 28 different tables. My script is quite long so I am just posting a screenshot for now.

Where is says var 2 I want to change to the storey number. So for table 1 I want storey 1, for table 2, storey 2 etc…

If you can see for the main output i am running a loop and using the matrix called "Int_force". I also have made the row headers as [Axial Force, Shear Force, Bending Moment. Thus, I am hoping someone can help me so that I can change the column headers to my desired wish? I am not sure how I would change the loop to accomplish this.
Many thanks,
Scott tables, for loop, matrices MATLAB Answers — New Questions

​

Is there any way to change the width of the edit controls in a livescript, either globally or on a per-control basis?
The edit box is a bit small when using it to enter Matlab code etc. and I would like to have a bit of control over it’s width.
The screenshot below shows the limited width of the edit box,

It would be nice if it was longer.Is there any way to change the width of the edit controls in a livescript, either globally or on a per-control basis?
The edit box is a bit small when using it to enter Matlab code etc. and I would like to have a bit of control over it’s width.
The screenshot below shows the limited width of the edit box,

It would be nice if it was longer. Is there any way to change the width of the edit controls in a livescript, either globally or on a per-control basis?
The edit box is a bit small when using it to enter Matlab code etc. and I would like to have a bit of control over it’s width.
The screenshot below shows the limited width of the edit box,

It would be nice if it was longer. livescript, mlx MATLAB Answers — New Questions

​

I’m trying to pick up the zoom event from an active figure. I have a figure with the world’s coastlines, and a number of events and their geographical coordinates. I Want to make a 2d density map of the events over the coastline. The map should be smoothed with a gaussian filter with a width that is dependent on the level of zoom, such that when I zoom in I get a more.dtailed view. Has anyone implemented something similar already? Even a minimal working example would be great.I’m trying to pick up the zoom event from an active figure. I have a figure with the world’s coastlines, and a number of events and their geographical coordinates. I Want to make a 2d density map of the events over the coastline. The map should be smoothed with a gaussian filter with a width that is dependent on the level of zoom, such that when I zoom in I get a more.dtailed view. Has anyone implemented something similar already? Even a minimal working example would be great. I’m trying to pick up the zoom event from an active figure. I have a figure with the world’s coastlines, and a number of events and their geographical coordinates. I Want to make a 2d density map of the events over the coastline. The map should be smoothed with a gaussian filter with a width that is dependent on the level of zoom, such that when I zoom in I get a more.dtailed view. Has anyone implemented something similar already? Even a minimal working example would be great. zoom, callbacks, histogram, density MATLAB Answers — New Questions

​

Hello All,
I am trying to simulate the NPUSCH BLER example and would like to understand the hidden assumptions when we disable perfect channel estimation for this example? The simulation results in BLER = 0 for SNR = -25dB when SCS=3.75kHz for 1RU and 1 subcarrier. I would like to use a custom channel (say Rician) for example. I have made the changes accordingly, but it does not seem to work. Please could you help me figure out what is happening in the background?
Update: I realised that noiseEst is always 0 when ‘perfectChannelEstimate=false’ and this is giving me wrong results for lower SNR.

Thanks in advanceHello All,
I am trying to simulate the NPUSCH BLER example and would like to understand the hidden assumptions when we disable perfect channel estimation for this example? The simulation results in BLER = 0 for SNR = -25dB when SCS=3.75kHz for 1RU and 1 subcarrier. I would like to use a custom channel (say Rician) for example. I have made the changes accordingly, but it does not seem to work. Please could you help me figure out what is happening in the background?
Update: I realised that noiseEst is always 0 when ‘perfectChannelEstimate=false’ and this is giving me wrong results for lower SNR.

Thanks in advance Hello All,
I am trying to simulate the NPUSCH BLER example and would like to understand the hidden assumptions when we disable perfect channel estimation for this example? The simulation results in BLER = 0 for SNR = -25dB when SCS=3.75kHz for 1RU and 1 subcarrier. I would like to use a custom channel (say Rician) for example. I have made the changes accordingly, but it does not seem to work. Please could you help me figure out what is happening in the background?
Update: I realised that noiseEst is always 0 when ‘perfectChannelEstimate=false’ and this is giving me wrong results for lower SNR.

Thanks in advance nbiot, npusch, bler MATLAB Answers — New Questions

​

Hi everyone, I aim to build a model of my system using real-time data. My system is a MIMO (Multiple Input Multiple Output) system, where the inputs are RPM and rudder angle, and the outputs are linear velocity and angular velocity. Also the system is cross-coupled, so both inputs effect both outputs and I think that there is nonlinear model. I want to implemented system identification model. I collected data for train and validation. Then, I implemented filter and prepared for system identification. I tried different models such as TF, State Space and Nonlinear models. I see that best models estimated by Nonlinear ARX model. But I don’t know how can I find the best model fit because there are many options in Nonlinear ARX model window. For example there are many options in Nonlinear Function bar such as Wavelet, Sigmoid, Neural, Gaussian etc… So how can i find best model fit. Do I need to write code about grid search may be it finds best possible trying different conditions? Do you have any recommendation?Hi everyone, I aim to build a model of my system using real-time data. My system is a MIMO (Multiple Input Multiple Output) system, where the inputs are RPM and rudder angle, and the outputs are linear velocity and angular velocity. Also the system is cross-coupled, so both inputs effect both outputs and I think that there is nonlinear model. I want to implemented system identification model. I collected data for train and validation. Then, I implemented filter and prepared for system identification. I tried different models such as TF, State Space and Nonlinear models. I see that best models estimated by Nonlinear ARX model. But I don’t know how can I find the best model fit because there are many options in Nonlinear ARX model window. For example there are many options in Nonlinear Function bar such as Wavelet, Sigmoid, Neural, Gaussian etc… So how can i find best model fit. Do I need to write code about grid search may be it finds best possible trying different conditions? Do you have any recommendation? Hi everyone, I aim to build a model of my system using real-time data. My system is a MIMO (Multiple Input Multiple Output) system, where the inputs are RPM and rudder angle, and the outputs are linear velocity and angular velocity. Also the system is cross-coupled, so both inputs effect both outputs and I think that there is nonlinear model. I want to implemented system identification model. I collected data for train and validation. Then, I implemented filter and prepared for system identification. I tried different models such as TF, State Space and Nonlinear models. I see that best models estimated by Nonlinear ARX model. But I don’t know how can I find the best model fit because there are many options in Nonlinear ARX model window. For example there are many options in Nonlinear Function bar such as Wavelet, Sigmoid, Neural, Gaussian etc… So how can i find best model fit. Do I need to write code about grid search may be it finds best possible trying different conditions? Do you have any recommendation? model, system MATLAB Answers — New Questions

​

Hi there,
I am using ode15s to resolve a DAE system. Since there are 300 variables x1(t)~x300(t) in my case, so I feel it is a little bit inconvenient to define them by writing
syms x1(t) … x300(t)
one by one. Are there any efficient codes to define them?

Many thanks!Hi there,
I am using ode15s to resolve a DAE system. Since there are 300 variables x1(t)~x300(t) in my case, so I feel it is a little bit inconvenient to define them by writing
syms x1(t) … x300(t)
one by one. Are there any efficient codes to define them?

Many thanks! Hi there,
I am using ode15s to resolve a DAE system. Since there are 300 variables x1(t)~x300(t) in my case, so I feel it is a little bit inconvenient to define them by writing
syms x1(t) … x300(t)
one by one. Are there any efficient codes to define them?

Many thanks! function calls, ode, differential equations MATLAB Answers — New Questions

​

I am working with Surface EMG data and need to use a 4th order butterworth bandpass filter but don’t have the signal processing toolbox. The cutoff frequency range is from 20 to 450 hz. I am also trying to understand how to come up with the coefficients using this bandpass filter and where I can apply that if I’m using that filtered signal to measure the MVIC of the data.
Thank youI am working with Surface EMG data and need to use a 4th order butterworth bandpass filter but don’t have the signal processing toolbox. The cutoff frequency range is from 20 to 450 hz. I am also trying to understand how to come up with the coefficients using this bandpass filter and where I can apply that if I’m using that filtered signal to measure the MVIC of the data.
Thank you I am working with Surface EMG data and need to use a 4th order butterworth bandpass filter but don’t have the signal processing toolbox. The cutoff frequency range is from 20 to 450 hz. I am also trying to understand how to come up with the coefficients using this bandpass filter and where I can apply that if I’m using that filtered signal to measure the MVIC of the data.
Thank you filter, signal processing MATLAB Answers — New Questions

​

I’ve always thought that indexing with an empty array on the LHS of an = is a no-op for the variable on the LHS (there could be side effects of the evaluation of the RHS).
For example
% Case 1
A = [1 2;3 4];
A([]) = 5
But, assigning a non-scalar on the RHS result in an error
% Case 2
try
A([]) = [5,6];
catch ME
ME.message
end
Why doesn’t Case 1 result in an error insofar as it also had a different number of elements on the left and right sides?
Using the deletion operator on the RHS with an empty index on the LHS reshapes the matrix
% Case 3
A = [1 2;3 4];
A([]) = []
Why does a request to delete nothing from A change A?
But that doesn’t happen if A is column vector
% Case 4
A = [1;2;3;4];
A([]) = []
Deletion with a non-zero dimension of an empty index also reshapes the LHS
% Case 5
A = [1 2;3 4];
A(double.empty(0,1)) = []
A = [1 2; 3 4];
A(double.empty(1,0)) = []
I suppose Cases 3 and 5 are related to deleting a specific (non-empty) element from a matrix
% Case 6
A = [1 2;3 4];
A(3) = []
But I’m surprised that the result is not a column vector.
Are any/all of these results following a general rule? Pointers to relevant documentation would be welcome. I couldn’t find any.I’ve always thought that indexing with an empty array on the LHS of an = is a no-op for the variable on the LHS (there could be side effects of the evaluation of the RHS).
For example
% Case 1
A = [1 2;3 4];
A([]) = 5
But, assigning a non-scalar on the RHS result in an error
% Case 2
try
A([]) = [5,6];
catch ME
ME.message
end
Why doesn’t Case 1 result in an error insofar as it also had a different number of elements on the left and right sides?
Using the deletion operator on the RHS with an empty index on the LHS reshapes the matrix
% Case 3
A = [1 2;3 4];
A([]) = []
Why does a request to delete nothing from A change A?
But that doesn’t happen if A is column vector
% Case 4
A = [1;2;3;4];
A([]) = []
Deletion with a non-zero dimension of an empty index also reshapes the LHS
% Case 5
A = [1 2;3 4];
A(double.empty(0,1)) = []
A = [1 2; 3 4];
A(double.empty(1,0)) = []
I suppose Cases 3 and 5 are related to deleting a specific (non-empty) element from a matrix
% Case 6
A = [1 2;3 4];
A(3) = []
But I’m surprised that the result is not a column vector.
Are any/all of these results following a general rule? Pointers to relevant documentation would be welcome. I couldn’t find any. I’ve always thought that indexing with an empty array on the LHS of an = is a no-op for the variable on the LHS (there could be side effects of the evaluation of the RHS).
For example
% Case 1
A = [1 2;3 4];
A([]) = 5
But, assigning a non-scalar on the RHS result in an error
% Case 2
try
A([]) = [5,6];
catch ME
ME.message
end
Why doesn’t Case 1 result in an error insofar as it also had a different number of elements on the left and right sides?
Using the deletion operator on the RHS with an empty index on the LHS reshapes the matrix
% Case 3
A = [1 2;3 4];
A([]) = []
Why does a request to delete nothing from A change A?
But that doesn’t happen if A is column vector
% Case 4
A = [1;2;3;4];
A([]) = []
Deletion with a non-zero dimension of an empty index also reshapes the LHS
% Case 5
A = [1 2;3 4];
A(double.empty(0,1)) = []
A = [1 2; 3 4];
A(double.empty(1,0)) = []
I suppose Cases 3 and 5 are related to deleting a specific (non-empty) element from a matrix
% Case 6
A = [1 2;3 4];
A(3) = []
But I’m surprised that the result is not a column vector.
Are any/all of these results following a general rule? Pointers to relevant documentation would be welcome. I couldn’t find any. empty index, deletion, assignment MATLAB Answers — New Questions

​