def populate_gmms(sample_vids, GMM_OUT, k_gmm, sample_size=400000, PCA=False):
"""
sample_size is the number of IDTFs that we sample from the total_lines number of IDTFs
that were computed previously.
GMM_OUT is the output file to save the list of GMMs.
Saves the GMMs in the GMM_OUT file as the gmm_list attribute.
Returns the list of gmms.
"""
#total_lines = 2488317, sample size=1500000
#total_line (30 classes) = 686680, sample size = 400000
total_lines = total_IDTF_lines()
print total_lines
sample_size = min(total_lines, sample_size)
sample_indices = random.sample(xrange(total_lines), sample_size)
sample_indices.sort()
sample_descriptors = IDT_feature.list_descriptors_sampled(
GMM_dir, sample_vids, sample_indices)
bm_list = IDT_feature.bm_descriptors(sample_descriptors)
#Construct gmm models for each of the different descriptor types.
gmm_list = [gmm_model(bm, k_gmm, PCA=PCA) for bm in bm_list]
print "gmm_list complete"
np.savez(GMM_OUT, gmm_list=gmm_list)
print "Save gmm_list.npz"
return gmm_list
def populate_gmms(PROJ_DIR,TMP_FEATURES,k_gmm,GMM_OUT,sample_size=1500000, PCA=False):
"""
sample_size is the number of IDTFs that we sample from the total_lines number of IDTFs
that were computed previously.
Saves the GMMs in the GMM_OUT file as the gmm_list attribute.
PROJ_DIR is the directory of the project (directory where we save the gmm_list)
TMP_FEATURES is the directory of the temporary .feature vectors
Returns the list of gmms.
"""
feature_list = [filename for filename in os.listdir(TMP_FEATURES) if filename.endswith('.features')]
#total_lines = 2488317
total_lines = total_IDTF_lines(TMP_FEATURES)
print "Total IDTFs constructed", total_lines
sample_size = min(total_lines,sample_size)
sample_indices = random.sample(xrange(total_lines),sample_size)
sample_indices.sort()
sample_descriptors = IDT_feature.list_descriptors_sampled(TMP_FEATURES, feature_list, sample_indices)
bm_list = IDT_feature.bm_descriptors(sample_descriptors)
#Construct gmm models for each of the different descriptor types.
gmm_list = [gmm_model(bm, k_gmm, PCA=PCA) for bm in bm_list]
np.savez(GMM_OUT, gmm_list=gmm_list)
return gmm_list
def populate_gmms(sample_vids, GMM_OUT, k_gmm, sample_size=1500000, PCA=False):
"""
sample_size is the number of IDTFs that we sample from the total_lines number of IDTFs
that were computed previously.
GMM_OUT is the output file to save the list of GMMs.
Saves the GMMs in the GMM_OUT file as the gmm_list attribute.
Returns the list of gmms.
"""
#total_lines = 2488317
total_lines = total_IDTF_lines()
print total_lines
sample_size = min(total_lines,sample_size)
sample_indices = random.sample(xrange(total_lines),sample_size)
sample_indices.sort()
sample_descriptors = IDT_feature.list_descriptors_sampled(GMM_dir, sample_vids, sample_indices)
bm_list = IDT_feature.bm_descriptors(sample_descriptors)
#Construct gmm models for each of the different descriptor types.
gmm_list = [gmm_model(bm, k_gmm, PCA=PCA) for bm in bm_list]
np.savez(GMM_OUT, gmm_list=gmm_list)
return gmm_list
def populate_gmms(IDT_DIR, sample_vids, gmm_file, k_gmm, sample_size=1500000, PCA=False):
"""
sample_size is the number of IDTFs that we sample from the total_lines number of IDTFs
that were computed previously.
gmm_file is the output file to save the list of GMMs.
Saves the GMMs in the gmm_file file as the gmm_list attribute.
Returns the list of gmms.
"""
nr_vids = len(sample_vids)
nr_samples_pvid = int(np.ceil(sample_size/nr_vids))
sample_descriptors = IDT_feature.list_descriptors_sampled(IDT_DIR, sample_vids, nr_samples_pvid)
bm_list = IDT_feature.bm_descriptors(sample_descriptors)
# save all sampled descriptors for learning gmm
bm_file = os.path.join(os.path.dirname(gmm_file), 'bm_descriptors_%d' % (sample_size,))
np.savez(bm_file, bm_list=bm_list)
# why sqrt? just like root sift! already done!!
# bm_list[0] = bm_list[0]
# bm_list[1] = np.sqrt(bm_list[1])
# bm_list[2] = np.sqrt(bm_list[2])
# bm_list[3] = np.sqrt(bm_list[3])
# Construct gmm models for each of the different descriptor types.
gmm_list = [gmm_model(bm, k_gmm, PCA=PCA) for bm in bm_list]
np.savez(gmm_file, gmm_list=gmm_list)
return gmm_list
def populate_gmms(sample_vids, GMM_OUT, k_gmm, sample_size=256000, PCA=False):
"""
sample_size is the number of IDTFs that we sample from the total_lines number of IDTFs
that were computed previously.
GMM_OUT is the output file to save the list of GMMs.
Saves the GMMs in the GMM_OUT file as the gmm_list attribute.
Returns the list of gmms.
"""
# total_lines = 3000
# total_lines = 158638780-787-1323 # ucf101
total_lines = 8081693 # something
# print('Counting all IDTF lines')
# total_lines = total_IDTF_lines()
print('All lines: ', str(total_lines))
sample_size = min(total_lines, sample_size)
sample_indices = random.sample(xrange(total_lines), sample_size)
sample_indices.sort()
sample_descriptors = IDT_feature.list_descriptors_sampled(
GMM_dir, sample_vids, sample_indices)
bm_list = IDT_feature.bm_descriptors(sample_descriptors)
#Construct gmm models for each of the different descriptor types.
gmm_list = [gmm_model(bm, k_gmm, PCA=PCA) for bm in bm_list]
np.savez(GMM_OUT, gmm_list=gmm_list)
return gmm_list
def processVideo(vid,IDT_DIR,FV_DIR,gmm_list):
"""
Extracts the IDTFs, constructs a Fisher Vector, and saves the Fisher Vector at FV_DIR
output_file: the full path to the newly constructed fisher vector.
gmm_list: a list of gmms
"""
input_file = os.path.join(IDT_DIR, vid.split('.')[0]+'.bin')
output_file = os.path.join(FV_DIR, vid.split('.')[0]+'.fv')
if not os.path.exists(input_file):
print '%s IDT Feature does not exist!' % vid
return False
if os.path.exists(output_file+'.mat'):
print '%s Fisher Vector exists, skip!' % vid
return False
video_desc = IDT_feature.vid_descriptors(IDT_feature.read_IDTF_file(input_file))
computeFV.create_fisher_vector(gmm_list, video_desc, output_file)
return True
def processVideo(vid, IDT_DIR, FV_DIR, gmm_list):
"""
Extracts the IDTFs, constructs a Fisher Vector, and saves the Fisher Vector at FV_DIR
output_file: the full path to the newly constructed fisher vector.
gmm_list: a list of gmms
"""
input_file = os.path.join(IDT_DIR, vid.split('.')[0] + '.bin')
output_file = os.path.join(FV_DIR, vid.split('.')[0] + '.fv')
if not os.path.exists(input_file):
print '%s IDT Feature does not exist!' % vid
return False
if os.path.exists(output_file + '.mat'):
print '%s Fisher Vector exists, skip!' % vid
return False
video_desc = IDT_feature.vid_descriptors(
IDT_feature.read_IDTF_file(input_file))
computeFV.create_fisher_vector(gmm_list, video_desc, output_file)
return True
def populate_gmms(IDT_DIR,
sample_vids,
gmm_file,
k_gmm,
sample_size=1500000,
PCA=False):
"""
sample_size is the number of IDTFs that we sample from the total_lines number of IDTFs
that were computed previously.
gmm_file is the output file to save the list of GMMs.
Saves the GMMs in the gmm_file file as the gmm_list attribute.
Returns the list of gmms.
"""
nr_vids = len(sample_vids)
nr_samples_pvid = int(np.ceil(sample_size / nr_vids))
sample_descriptors = IDT_feature.list_descriptors_sampled(
IDT_DIR, sample_vids, nr_samples_pvid)
bm_list = IDT_feature.bm_descriptors(sample_descriptors)
# save all sampled descriptors for learning gmm
bm_file = os.path.join(os.path.dirname(gmm_file),
'bm_descriptors_%d' % (sample_size, ))
np.savez(bm_file, bm_list=bm_list)
# why sqrt? just like root sift! already done!!
# bm_list[0] = bm_list[0]
# bm_list[1] = np.sqrt(bm_list[1])
# bm_list[2] = np.sqrt(bm_list[2])
# bm_list[3] = np.sqrt(bm_list[3])
# Construct gmm models for each of the different descriptor types.
gmm_list = [gmm_model(bm, k_gmm, PCA=PCA) for bm in bm_list]
np.savez(gmm_file, gmm_list=gmm_list)
return gmm_list
else:
svm = classify_library.load_model('../data/models/svm_nopca.sav')
gmm_list = np.load(gmm_list + ".npz")['gmm_list']
index_class = np.load(class_index)['index_class']
index_class = index_class[()]
points = [] # a list of IDT features.
frame_lim = frame_step
for line in sys.stdin:
if line[0] != '[': # avoid getting info message as data
frame = int(line.split()[0])
if frame_lim <= frame:
frame_lim = frame_lim + frame_step
# print frame_lim<=frame
if points != []:
video_desc = IDT_feature.vid_descriptors(points)
fish = computeFV.create_fisher_vector_unsaved(
gmm_list, video_desc)
fish = np.array(fish).reshape(1, -1)
if args.no_pca:
result = svm.predict(fish)
else:
fish_pca = pca.transform(fish)
result = svm.predict(fish_pca)
print '\n' + 'RESULT: ' + OKGREEN + BOLD + index_class[
result[0]] + ENDC + '\n'
points = []
points.append(IDT_feature.IDTFeature(line))
import numpy as np
from yael import ynumpy
import IDT_feature
from tempfile import TemporaryFile
import argparse
import computeFV
"""
computes a Fisher vector given an input stream of IDTFs
Usage:
stream_of_IDTFs | python computeFVstream.py fisher_path gmm_list
./DenseTrackStab video_file | python computeFVstream.py fisher_path gmm_list
"""
#The input is a stream of IDTFs associated with a single video.
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument("fisher_path", help="File to save the output Fisher Vector", type=str)
parser.add_argument("gmm_list", help="File of saved list of GMMs", type=str)
args = parser.parse_args()
gmm_list = np.load(args.gmm_list+".npz")['gmm_list']
points = [] # a list of IDT features.
for line in sys.stdin:
points.append(IDT_feature.IDTFeature(line))
video_desc = IDT_feature.vid_descriptors(points)
computeFV.create_fisher_vector(gmm_list, video_desc, args.fisher_path)
