def model(data, test_image_vector, distance_metric="cosine"):
"""
Perform Modeling on data by sending the image through the pipeline. The test vector
is a random vector used to verify if code is functioning. Its value is assigned at the
bottom of this file
Input: final data , test vector to get results ,distanc metric to be used
Output: predicted values for test vector
"""
model = NearestNeighbors(n_neighbors=10, metric="cosine", algorithm="brute")
model.fit(data)
pickle_this(model, "2knn_model.pkl")
# Predict K nearest neighbors for the given vector
predicted_indices = model.kneighbors(test_image_vector, return_distance=False)
return predicted_indices
def vectorize(data_path, no_of_images, indicator):
'''
Takes the path, total number of folders(not images coz folders may be empty)
in path and creates vector matrix. Generlly called in Model.py file
Input: Path to files, Total Images, print step(how many images till notification)
Output: Final dataset to be used in modeling after PCA ,Index dictionary
'''
# Since some of the images may not be loaded(dont exist or other errors) we
# start a dictionary to keep track of which ones do and which ones dont
index_dict = {}
# dict_counter keeps a count of number of images successfully loaded
dict_counter = 0
predata = []
for i in xrange(no_of_images):
image_name = data_path + "{0}/{0}.jpg".format(i)
if os.path.exists(image_name):
# Get features
features = preprocess(image_name)
# Forming three matrices from all the images. these will be
# combined later
predata.append(features)
index_dict[dict_counter] = i
dict_counter += 1
else:
#index_dict[i] = "Load Error:No image"
pass
if i % indicator == 0:
print "%d images finished" % i
# Convert the two lists to np arrays
data = np.array(predata)
print data
print">>>>>>>>>>>>>>", data.shape
pickle_this(data, "Final_Feature_Matrix_For_PCA.pkl")
pickle_this(index_dict, "Image_model_Index_dict.pkl")
# Perform PCA
data = data_pca_pipeline(data)
return data, index_dict
def pca(
data,
n_components=100,
filename="pca_image_model.pkl",
plot=False,
mayipickle=False):
'''
Since Standard scalar expects <=2 dimensions but colored images have 3 we will be
convert them to gray if not gray already
Input: Gray image, number of components you wish to use, filename for pickled pca
object, plot boolean to do the scree plot, boolean for deciding to pickle object or not
Output: Features from image
'''
# Since standard scalar works with floats we change dtype here.
# Else we get a warning
data = data.astype("float64")
scale = StandardScaler()
img_data_scaled = scale.fit_transform(data)
print "Scaling the data", img_data_scaled
pickle_this(scale, "SS_model.pkl")
print "Performing PCA on scaled data"
# features
pca_model = PCA(n_components)
pca_data = pca_model.fit_transform(img_data_scaled)
if mayipickle:
pickle_this(pca_model, filename)
# print ">>>>>>>>>>>>>after>>>>>>>", pca_data.shape
if plot:
# PLot to see how much variance does each principal component explain
scree_plot(pca_model)
plt.show()
variance_array = pca_model.explained_variance_ratio_
return pca_data, variance_array
