I am finetuning a network. In a specific case I want to use it for regression, which works. In another case, I want to use it for classification.
For both cases I have an HDF5 file, with a label. With regression, this is just a 1-by-1 numpy array that contains a float. I thought I could use the same label for classification, after changing my EuclideanLoss layer to SoftmaxLoss. However, then I get a negative loss as so:
Iteration 19200, loss = -118232
Train net output #0: loss = 39.3188 (* 1 = 39.3188 loss)
Can you explain if, and so what, goes wrong? I do see that the training loss is about 40 (which is still terrible), but does the network still train? The negative loss just keeps on getting more negative.
UPDATE
After reading Shai's comment and answer, I have made the following changes:
- I made the num_output of my last fully connected layer 6, as I have 6 labels (used to be 1).
- I now create a one-hot vector and pass that as a label into my HDF5 dataset as follows
f['label'] = numpy.array([1, 0, 0, 0, 0, 0])
Trying to run my network now returns
Check failed: hdf_blobs_[i]->shape(0) == num (6 vs. 1)
After some research online, I reshaped the vector to a 1x6 vector. This lead to the following error:
Check failed: outer_num_ * inner_num_ == bottom[1]->count() (40 vs. 240)
Number of labels must match number of predictions; e.g., if softmax axis == 1
and prediction shape is (N, C, H, W), label count (number of labels)
must be N*H*W, with integer values in {0, 1, ..., C-1}.
My idea is to add 1 label per data set (image) and in my train.prototxt I create batches. Shouldn't this create the correct batch size?
Since you moved from regression to classification, you need to output not a scalar to compare with "label" but rather a probability vector of length num-labels to compare with the discrete class "label". You need to change num_output parameter of the layer before "SoftmaxWithLoss" from 1 to num-labels.
I believe currently you are accessing un-initialized memory and I would expect caffe to crash sooner or later in this case.
Update:
You made two changes: num_output 1-->6, and you also changed your input label from a scalar to vector.
The first change was the only one you needed for using "SoftmaxWithLossLayer".
Do not change label from a scalar to a "hot-vector".
Why?
Because "SoftmaxWithLoss" basically looks at the 6-vector prediction you output, interpret the ground-truth label as index and looks at -log(p[label]): the closer p[label] is to 1 (i.e., you predicted high probability for the expected class) the lower the loss. Making a prediction p[label] close to zero (i.e., you incorrectly predicted low probability for the expected class) then the loss grows fast.
Using a "hot-vector" as ground-truth input label, may give rise to multi-category classification (does not seems like the task you are trying to solve here). You may find this SO thread relevant to that particular case.
Related
So what I need to do is to apply an operation like
(x(i,j)-min(x)) / max(x(i,j)-min(x))
which basically converts each pixel value such that the values range between 0 and 1.
First of all, I realised that Matlab saves our image(rows * col * colour) in a 3D matrix on using imread,
Image = imread('image.jpg')
So, a simple max operation on image doesn't give me the max value of pixel and I'm not quite sure what it returns(another multidimensional array?). So I tried using something like
max_pixel = max(max(max(Image)))
I thought it worked fine. Similarly I used min thrice. My logic was that I was getting the min pixel value across all 3 colour planes.
After performing the above scaling operation I got an image which seemed to have only 0 or 1 values and no value in between which doesn't seem right. Has it got something to do with integer/float rounding off?
image = imread('cat.jpg')
maxI = max(max(max(image)))
minI = min(min(min(image)))
new_image = ((I-minI)./max(I-minI))
This gives output of only 1s and 0s which doesn't seem correct.
The other approach I'm trying is working on all colour planes separately as done here. But is that the correct way to do it?
I could also loop through all pixels but I'm assuming that will be time taking. Very new to this, any help will be great.
If you are not sure what a matlab functions returns or why, you should always do one of the following first:
Type help >functionName< or doc >functionName< in the command window, in your case: doc max. This will show you the essential must-know information of that specific function, such as what needs to be put in, and what will be output.
In the case of the max function, this yields the following results:
M = max(A) returns the maximum elements of an array.
If A is a vector, then max(A) returns the maximum of A.
If A is a matrix, then max(A) is a row vector containing the maximum
value of each column.
If A is a multidimensional array, then max(A) operates along the first
array dimension whose size does not equal 1, treating the elements as
vectors. The size of this dimension becomes 1 while the sizes of all
other dimensions remain the same. If A is an empty array whose first
dimension has zero length, then max(A) returns an empty array with the
same size as A
In other words, if you use max() on a matrix, it will output a vector that contains the maximum value of each column (the first non-singleton dimension). If you use max() on a matrix A of size m x n x 3, it will result in a matrix of maximum values of size 1 x n x 3. So this answers your question:
I'm not quite sure what it returns(another multidimensional array?)
Moving on:
I thought it worked fine. Similarly I used min thrice. My logic was that I was getting the min pixel value across all 3 colour planes.
This is correct. Alternatively, you can use max(A(:)) and min(A(:)), which is equivalent if you are just looking for the value.
And after performing the above operation I got an image which seemed to have only 0 or 1 values and no value in between which doesn't seem right. Has it got something to do with integer/float rounding off?
There is no way for us to know why this happens if you do not post a minimal, complete and verifiable example of your code. It could be that it is because your variables are of a certain type, or it could be because of an error in your calculations.
The other approach I'm trying is working on all colour planes separately as done here. But is that the correct way to do it?
This depends on what the intended end result is. Normalizing each colour (red, green, blue) seperately will result in a different result as compared to normalizing the values all at once (in 99% of cases, anyway).
You have a uint8 RGB image.
Just convert it to a double image by
I=imread('https://upload.wikimedia.org/wikipedia/commons/thumb/0/0b/Cat_poster_1.jpg/1920px-Cat_poster_1.jpg')
I=double(I)./255;
alternatively
I=im2double(I); %does the scaling if needed
Read about image data types
What are you doing wrong?
If what you want todo is convert a RGB image to [0-1] range, you are approaching the problem badly, regardless of the correctness of your MATLAB code. Let me give you an example of why:
Say you have an image with 2 colors.
A dark red (20,0,0):
A medium blue (0,0,128)
Now you want this changed to [0-1]. How do you scale it? Your suggested approach is to make the value 128->1 and either 20->20/128 or 20->1 (not relevant). However when you do this, you are changing the color! you are making the medium blue to be intense blue (maximum B channel, ) and making R way way more intense (instead of 20/255, 20/128, double brightness! ). This is bad, as this is with simple colors, but with combined RGB values you may even change the color itsef, not only the intensity. Therefore, the only correct way to convert to [0-1] range is to assume your min and max are [0, 255].
I am tuning an SVM using a for loop to search in the range of hyperparameter's space. The svm model learned contains the following fields
SVMModel: [1×1 ClassificationSVM]
C: 2
FeaturesIdx: [4 6 8]
Score: 0.0142
Question1) What is the meaning of the field 'score' and its utility?
Question2) I am tuning the BoxConstraint, C value. Let, the number of features be denoted by the variable featsize. The variable gridC will contain the search space which can start from any value say 2^-5, 2^-3, to 2^15 etc. So, gridC = 2.^(-5:2:15). I cannot understand if there is a way to select the range?
1. score had been documented in here, which says:
Classification Score
The SVM classification score for classifying observation x is the signed distance from x to the decision boundary ranging from -∞ to +∞.
A positive score for a class indicates that x is predicted to be in
that class. A negative score indicates otherwise.
In two class cases, if there are six observations, and the predict function gave us some score value called TestScore, then we could determine which class does the specific observation ascribed by:
TestScore=[-0.4497 0.4497
-0.2602 0.2602;
-0.0746 0.0746;
0.1070 -0.1070;
0.2841 -0.2841;
0.4566 -0.4566;];
[~,Classes] = max(TestScore,[],2);
In the two-class classification, we can also use find(TestScore > 0) instead, and it is clear that the first three observations are belonging to the second class, and the 4th to 6th observations are belonging to the first class.
In multiclass cases, there could be several scores > 0, but the code max(scores,[],2) is still validate. For example, we could use the code (from here, an example called Find Multiple Class Boundaries Using Binary SVM) following to determine the classes of the predict Samples.
for j = 1:numel(classes);
[~,score] = predict(SVMModels{j},Samples);
Scores(:,j) = score(:,2); % Second column contains positive-class scores
end
[~,maxScore] = max(Scores,[],2);
Then the maxScore will denote the predicted classes of each sample.
2. The BoxConstraint denotes C in the SVM model, so we can train SVMs in different hyperparameters and select the best one by something like:
gridC = 2.^(-5:2:15);
for ii=1:length(gridC)
SVModel = fitcsvm(data3,theclass,'KernelFunction','rbf',...
'BoxConstraint',gridC(ii),'ClassNames',[-1,1]);
%if (%some constraints were meet)
% %save the current SVModel
%end
end
Note: Another way to implement this is using libsvm, a fast and easy-to-use SVM toolbox, which has the interface of MATLAB.
For each input I have, I have a 49x2 matrix associated. Here's what 1 input-output couple looks like
input :
[Car1, Car2, Car3 ..., Car118]
output :
[[Label1 Label2]
[Label1 Label2]
...
[Label1 Label2]]
Where both Label1 and Label2 are LabelEncode and they have respectively 1200 and 1300 different classes.
Just to make sure this is what we call a multi-output multi-class problem?
I tried to flatten the output but I feared the model wouldn't understand that all similar Label share the same classes.
Is there a Keras layer that handle output this peculiar array shape?
Generally, multi-class problems correspond with models outputting a probability distribution over the set of classes (that is typically scored against the one-hot encoding of the actual class through cross-entropy). Now, independently of whether you are structuring it as one single output, two outputs, 49 outputs or 49 x 2 = 98 outputs, that would mean having 1,200 x 49 + 1,300 x 49 = 122,500 output units - which is not something a computer cannot handle, but maybe not the most convenient thing to have. You could try having each class output to be a single (e.g. linear) unit and round it's value to choose the label, but, unless the labels have some numerical meaning (e.g. order, sizes, etc.), that is not likely to work.
If the order of the elements in the input has some meaning (that is, shuffling it would affect the output), I think I'd approach the problem through an RNN, like an LSTM or a bidirectional LSTM model, with two outputs. Use return_sequences=True and TimeDistributed Dense softmax layers for the outputs, and for each 118-long input you'd have 118 pairs of outputs; then you can just use temporal sample weighting to drop, for example, the first 69 (or maybe do something like dropping the 35 first and the 34 last if you're using a bidirectional model) and compute the loss with the remaining 49 pairs of labellings. Or, if that makes sense for your data (maybe it doesn't), you could go with something more advanced like CTC (although Keras does not have it, I'm trying to integrate TensorFlow implementation into it without much sucess), which is also implemented in Keras (thanks #indraforyou)!.
If the order in the input has no meaning but the order of the outputs does, then you could have an RNN where your input is the original 118-long vector plus a pair of labels (each one-hot encoded), and the output is again a pair of labels (again two softmax layers). The idea would be that you get one "row" of the 49x2 output on each frame, and then you feed it back to the network along with the initial input to get the next one; at training time, you would have the input repeated 49 times along with the "previous" label (an empty label for the first one).
If there are no sequential relationships to exploit (i.e. the order of the input and the output do not have a special meaning), then the problem would only be truly represented by the initial 122,500 output units (plus all the hidden units you may need to get those right). You could also try some kind of middle ground between a regular network and a RNN where you have the two softmax outputs and, along with the 118-long vector, you include the "id" of the output that you want (e.g. as a 49-long one-hot encoded vector); if the "meaning" of each label at each of the 49 outputs is similar, or comparable, it may work.
I created a GoogleNet Model via Nvidia DIGITS with two classes (called positive and negative).
If I classify an image with DIGITS, it shows me a nice result like positive: 85.56% and negative: 14.44%.
If it pass that model it into pycaffe's classify.py with the same image, I get a result like array([[ 0.38978559, -0.06033826]], dtype=float32)
So, how do I read/interpret this result? How do I calculate the confidence levels (not sure if this is the right term) shown by DIGITS from the results shown by classify.py?
This issue led me to the solution.
As the log shows, the network produces three outputs. Classifier#classify only returns the first output. So e.g. by changing predictions = out[self.outputs[0]] to predictions = out[self.outputs[2]], I get the desired values.
Hi my question is a bit long please bare and read it till the end.
I am working on a project with 30 participants. We have two type of data set (first data set has 30 rows and 160 columns , and second data set has the same 30 rows and 200 columns as outputs=y and these outputs are independent), what i want to do is to use the first data set and predict the second data set outputs.As first data set was rectangular type and had high dimension i have used factor analysis and now have 19 factors that cover up to 98% of the variance. Now i want to use these 19 factors for predicting the outputs of the second data set.
I am using neuralnet and backpropogation and everything goes well and my results are really close to outputs.
My questions :
1- as my inputs are the factors ( they are between -1 and 1 ) and my outputs scale are between 4 to 10000 and integer , should i still scaled them before running neural network ?
2-I scaled the data ( both input and outputs ) and then predicted with neuralnet , then i check the MSE error it was so high like 6000 while my prediction and real output are so close to each other. But if i rescale the prediction and outputs then check The MSE its near zero. Is it unbiased to rescale and then check the MSE ?
3- I read that it is better to not scale the output from the beginning but if i just scale the inputs all my prediction are 1. Is it correct to not to scale the outputs ?
4- If i want to plot the ROC curve how can i do it. Because my results are never equal to real outputs ?
Thank you for reading my question
[edit#1]: There is a publication on how to produce ROC curves using neural network results
http://www.lcc.uma.es/~jja/recidiva/048.pdf
1) You can scale your values (using minmax, for example). But only scale your training data set. Save the parameters used in the scaling process (in minmax they would be the min and max values by which the data is scaled). Only then, you can scale your test data set WITH the min and max values you got from the training data set. Remember, with the test data set you are trying to mimic the process of classifying unseen data. Unseen data is scaled with your scaling parameters from the testing data set.
2) When talking about errors, do mention which data set the error was computed on. You can compute an error function (in fact, there are different error functions, one of them, the mean squared error, or MSE) on the training data set, and one for your test data set.
4) Think about this: Let's say you train a network with the testing data set,and it only has 1 neuron in the output layer . Then, you present it with the test data set. Depending on which transfer function (activation function) you use in the output layer, you will get a value for each exemplar. Let's assume you use a sigmoid transfer function, where the max and min values are 1 and 0. That means the predictions will be limited to values between 1 and 0.
Let's also say that your target labels ("truth") only contains discrete values of 0 and 1 (indicating which class the exemplar belongs to).
targetLabels=[0 1 0 0 0 1 0 ];
NNprediction=[0.2 0.8 0.1 0.3 0.4 0.7 0.2];
How do you interpret this?
You can apply a hard-limiting function such that the NNprediction vector only contains the discreet values 0 and 1. Let's say you use a threshold of 0.5:
NNprediction_thresh_0.5 = [0 1 0 0 0 1 0];
vs.
targetLabels =[0 1 0 0 0 1 0];
With this information you can compute your False Positives, FN, TP, and TN (and a bunch of additional derived metrics such as True Positive Rate = TP/(TP+FN) ).
If you had a ROC curve showing the False Negative Rate vs. True Positive Rate, this would be a single point in the plot. However, if you vary the threshold in the hard-limit function, you can get all the values you need for a complete curve.
Makes sense? See the dependencies of one process on the others?