There are 241 directories. There 15 variabels in blacknames, I think blacknames are notes?
black_names = ['4', '8', '8_b_n', '8_b_r', '16', '16_b_n', '16_b_r',
'32', '32_b_n', '32_b_r', 'a_4', 'a_8', 'a_16', 'a_32', 'chord']When I train my model using the whole dataset nothing matches. The directories are labels.
None of the directories in the published dataset are in blacknames.
But there are similar labels usually named with label_[number] for example
there is no 4 directory but there is
predctions on 04.PNG:
cd_1.sav
[ c1/2 d1/4 e1/4 f1/4 g1/4 a1/4 b1/4 c2/4 d2/4 e2/4 f2/4 g2/4 a2/4 b2/4 f1/8 f1/8]
Processing new image data/testing_data/04.PNG...
most common: 20
Writing predictions: ['c1/2', 'd1/4', 'e1/4', 'f1/4', 'g1/4', 'a1/4', 'b1/4', 'c2/4', 'd2/4', 'e2/4', 'f2/4', 'g2/4', 'a2/4', 'b2/4', 'f1/8', 'f1/8']
og.sav
[ c1/4 d1/4 e1/4 f1/4 g1/4 a1/4 b1/4 c2/4 d2/4 e2/4 f2/4 g2/4 a2/4 b2/4]
Processing new image data/testing_data/04.PNG...
Writing predictions: ['c1/4', 'd1/4', 'e1/4', 'f1/4', 'g1/4', 'a1/4', 'b1/4', 'c2/4', 'd2/4', 'e2/4', 'f2/4', 'g2/4', 'a2/4', 'b2/4']
The problem seems to be with detecting bars
The traing data is either over fit, improperly labeld or something else. The model trained with the full dataset don't work...
black_names = ['4', '8', '8_b_n', '8_b_r', '16', '16_b_n', '16_b_r',
'32', '32_b_n', '32_b_r', 'a_4', 'a_4_075', 'a_8', 'a_16', 'a_32', 'chord']
# or
black_names = os.listdir('data/training_data/')The model will work if I properly include all of the blacknotes labels from the
directories.
There are 102 directories in the training data I have downloaded,
There are 241 directories in the dataset posted on google drive.
That means I'm missing 139 feature labels and 57% of the dataset.
my understanding of a histogram is just a "bar graph" of the frequency
of elements in an array. For a histogram of a collored image this
would be the counts of red, green and blue pixels in an image although
I don't know at what threashold we count a pixel value to be red, green
or blue. For this project we are not using a histogram of colored pixels,
all of our dataset is grey scale and we convert images into greyscale
before getting predictions from the model. We can see that
our image is the return value of extract_hsv_histogram() which applies
openCVs Histogram of Oriented Gradients method to it. A histogram
of of grey scale images is mesuring the intencity of a pixel on
a scale from 0 to 255, rather than the color values. The "oriented"
part has something to do with the direction the intensity flows from
one part to another pixel. I belive this has something to do with
finding edges of pixels in images but I don't understand the math
behind it (which involves some trig).
These are some images that describe what I mean
As you can see the arrows on the image are pointing to where the intensity either increases or decreases (i don't know whih one) but the arrow would be where the "oriented" comes from and the direction of intencity would be where the "gradient" comes from.
I belive all of this means whatever hog is doing, it relates to finding the edges of objects in an image. If we know the edges we can make out shapes?
A big question I have is, how does the model know it's being trained on a hog dataset rather than the other standard formats of data?
The reson run_experiment is named run_experiment is because
it looks like the original developer was testing out multiple
different models and "feature sets".
These models are fetched from load_classifiers() as a dictionary
of different sklean models.
classifiers = {
'SVM': svm.LinearSVC(random_state=random_seed),
'KNN': KNeighborsClassifier(n_neighbors=7),
'NN': MLPClassifier(activation='relu',
hidden_layer_sizes=(200,),
max_iter=10000, alpha=1e-4,
solver='adam', verbose=20,
tol=1e-8, random_state=1,
learning_rate_init=.0001,
learning_rate='adaptive')
}The only one actually used in the final product was 'NN' which is a Multi Layer Perseptron Classifier, which is an old term for Nural Network.
the second paramater taken by run_experiment is feature_set='hog'. HOG
stands for Histogram of Oriented Gradients.
Before feeding the images to the model for training the dataset is loaded with
this feature_set flag set as hog
features, labels = load_dataset(feature_set, dir_names) features is actually an array of ndarrays, these arrays corispond to images
from the dataset. labels are just the name of directories corisopding to the
images.
load_dataset() itterates through the dataset, converts each image to an
ndarray and passes it to extract_features() with the argument of the image and
a string specifiying the feature set. which is allways hog
extract_hog_features then calls extract_hog_features which does the
Histogram of Oriented Gradients thing which I don't understand.
We can simplify all this removing the dynamic parts.
- return only a
sklearnmodle and random seadfromload_classifiers - refactor
load_datasetto always usehogfeature extraction
but something tells me I should at least try to use the other feature extractors and models to see how they do.
rotation seems to be the biggest problem