amaas / stanford_dl_ex Goto Github PK

Line 22 of grad_check.m
fprintf('% 5d % 6d % 15g % 15f % 15f % 15f\n', ... i,j,error,g(j),g_est,f);
should be
fprintf('% 5d % 6d % 15g % 15f % 15f % 15f\n', ... i,j,error,g_est,g(j),f);

Hi,

I think the initialisation of the bias in cnnExercise.m should b:
b = rand(numFilters, 1);

instead of

b = rand(numFilters);

because in the cnnConvolve exercise, the bias should be of shape (numFilters,1)

Hi! Thank you for making and maintaining this tutorial! I'm reading through your Softmax Regression Tutorial, and I have a question about the following excerpt of the section entitled Properties of softmax regression parameterization:

Indeed, rather than optimizing over the K⋅n parameters (θ(1),θ(2),…,θ(K)) (where θ(k)∈ℜ^n), one can instead set θ(K)=0⃗ and optimize only with respect to the K⋅n remaining parameters.

Should the second part of this sentence instead read "optimize only with respect to the (K - 1)⋅n remaining parameters"?

Thank you!

http://ufldl.stanford.edu/tutorial/supervised/DebuggingBiasAndVariance/

I was expecting a discussion of learning curves like in the Coursera course on Machine Learning. Is this also "left to be written", like the Optimizers and Objectives subsection?

Hello,

Thank you for this great tutorial and all the work you have put into. This is a question not a issue, but I do not know how to just ask a question.

I can I find out what the columns in the ex1/house.data are (room number, square footage, etc). Is there a version of the file somewhere with headers or could you point me to where you originally got the data?

Thank you.

Does anybody have an example of the "sane results" mentioned in runSoftICA.m? Looks like image captions aren't supported, so I'll limit things to one image per post for legibility... For completeness, here is my source code...

Here's what I am getting for the default parameter values (10000 patches).

http://ufldl.stanford.edu/tutorial/supervised/DebuggingGradientChecking/

Find-Replace {\rm EPSILON} -> {\epsilon} ? I guess the "\rm" means it was deliberate, but it's still pretty distracting/jarring to see a programming-style constant in equations...

when running octave ex1a_linreg.m get the following error

error: 'lbfgsAddC' undefined near line 21 column 3
error: called from:
error:   /home/~~~~/stanford_dl_ex/common/minFunc_2012/minFunc/lbfgsAdd.m at line 21, column 3
error:   /home/~~~~/stanford_dl_ex/common/minFunc_2012/minFunc/minFunc.m at line 571, column 50
error:   /home/~~~~/stanford_dl_ex/ex1/ex1a_linreg.m at line 47, column 7

solve it with modifying ex1a_linreg.m line 46 to

options = struct('MaxIter', 200,'useMex',0);

All the math on the site are broken http://ufldl.stanford.edu/tutorial/ because mathjax updated their servers. Could you please update the page to use https://cdn.mathjax.org/ ... instead?

During running the "multilayer_supervised" in matlab, I got an error as below:

Reference to non-existent field 'W'.
Error in stack2params (line 31)
assert(mod(size(stack{d+1}.W, 2), size(stack{d}.W, 1)) == 0, ...
Error in supervised_dnn_cost (line 85)
[grad] = stack2params(gradStack);
Error in minFunc (line 314)
[f,g] = funObj(x,varargin{:});
Error in run_train (line 45)
[opt_params,opt_value,exitflag,output] = minFunc(@supervised_dnn_cost,...

It seems that the error came from the ''stack2params.m''. But I do not find out where the error is in the code. I used the ''stack2params.m'' directly and do not make any modification. Is there any modification be needed? Could anyone help me with this out? Thank you so much.

The very first tutorial at

http://ufldl.stanford.edu/tutorial/supervised/LinearRegression/

has text missing at the end of the second paragraph. It ends with:
For short, we will denote the [missing text]

http://ufldl.stanford.edu/tutorial/supervised/ExerciseSupervisedNeuralNetwork/

The equation for delta should not have a sum over i, since a different set of delta values is computed for each training example.

Also, the RHS is a scalar and not a vector, so the LHS should have a matching subscript k.

This was all noticed previously by @civilstat

I do not find the function that named as “minFunc“.Who could tell me where is the function. Thanks a lot

Last sentence of section "Properties of softmax regression parameterization"

"Indeed, rather than optimizing over the K⋅n parameters...optimize only with respect to the K⋅n remaining parameters."

The second "K⋅n" should be "(K-1)⋅n"

images = loadMNISTImages('../common/train-images-idx3-ubyte');

when I run the cnnExercise.m, gives the following error:
error usage loadMNISTImages (line 6)
Could not open ../common/train-images-idx3-ubyte
error cnnExercise (line 26)
images =
loadMNISTImages('../common/train-images-idx3-ubyte');

I think there is a problem in the formulation to compute the derivative of W and b in this tutorial. Isn't the W of layer l comes from error in layer l and activation in layer l-1? But the formulation suggests W in layer l comes from error in layer l+1 and activation in layer l.

I think the right one should look like this

The same goes to b. Or maybe I just misunderstood this, if so, please point out, thanks!