def predict(input_path, output_path, start_extract, end_extract):
tmp_dir = 'tmp_files/'
tmp_features = 'tmp.features'
tmp_prob = 'tmp.prob'
tmp_prediction = 'tmp.prediction'
if not os.path.exists(input_path):
print >> sys.stderr, "wav file does not exits"
return
length = utils.get_wav_file_length(input_path)
feature_file = tmp_dir + tmp_features
prob_file = tmp_dir + tmp_prob
predict_file = tmp_dir + tmp_prediction
# remove tmo dir if exists
if os.path.exists(tmp_dir):
shutil.rmtree(tmp_dir)
os.mkdir(tmp_dir)
print '\n1) Extracting features and classifying ...'
extract_features(input_path, feature_file, start_extract, end_extract)
run(feature_file, prob_file)
print '\n3) Extract Durations ...'
post_process(prob_file, predict_file)
print '\n4) Writing TextGrid file to %s ...' % output_path
create_text_grid(predict_file, output_path, length, float(start_extract))
def main(argv):
del argv # Unused.
if FLAGS.evaluate:
extract_features(FLAGS)
else:
wandb.init(config=FLAGS, sync_tensorboard=True)
set_up_train(FLAGS)
def main():
"""These are the main training settings. Set each before running
this file."""
if (len(sys.argv) == 5):
seq_length = int(sys.argv[1])
class_limit = int(sys.argv[2])
image_height = int(sys.argv[3])
image_width = int(sys.argv[4])
else:
print ("Usage: python train.py sequence_length class_limit image_height image_width")
print ("Example: python train.py 75 2 720 1280")
exit (1)
sequences_dir = os.path.join('data', 'sequences')
if not os.path.exists(sequences_dir):
os.mkdir(sequences_dir)
checkpoints_dir = os.path.join('data', 'checkpoints')
if not os.path.exists(checkpoints_dir):
os.mkdir(checkpoints_dir)
# model can be only 'lstm'
model = 'lstm'
saved_model = None # None or weights file
load_to_memory = False # pre-load the sequences into memory
batch_size = 10
nb_epoch = 50
data_type = 'features'
image_shape = (image_height, image_width, 3)
extract_features(seq_length=seq_length, class_limit=class_limit, image_shape=image_shape)
train(data_type, seq_length, model, saved_model=saved_model,
class_limit=class_limit, image_shape=image_shape,
load_to_memory=load_to_memory, batch_size=batch_size, nb_epoch=nb_epoch)
def predict(input_path, output_path, start_extract, end_extract):
tmp_dir = 'tmp_files/'
tmp_features = 'tmp.features'
tmp_prob = 'tmp.prob'
tmp_prediction = 'tmp.prediction'
if not os.path.exists(input_path):
print >>sys.stderr, "wav file does not exits"
return
length = utils.get_wav_file_length(input_path)
feature_file = tmp_dir+tmp_features
prob_file = tmp_dir+tmp_prob
predict_file = tmp_dir+tmp_prediction
# remove tmo dir if exists
if os.path.exists(tmp_dir):
shutil.rmtree(tmp_dir)
os.mkdir(tmp_dir)
print '\n1) Extracting features and classifying ...'
extract_features(input_path, feature_file, start_extract, end_extract)
run(feature_file, prob_file)
print '\n3) Extract Durations ...'
post_process(prob_file, predict_file)
print '\n4) Writing TextGrid file to %s ...' % output_path
create_text_grid(predict_file, output_path, length, float(start_extract))
def main(argv):
del argv # Unused.
if FLAGS.evaluate:
extract_features(FLAGS)
else:
wandb.init(config=FLAGS,
sync_tensorboard=True,
dir="/dfs/scratch2/prabhat8/cs236g/wandb")
set_up_train(FLAGS)
def main():
""" Training and test settings for the algorithm."""
model_name = 'lstm'
saved_model = None
cls_lmt = None
batch_size = 32
prompt = "Enter the sequence length"
seq_length = np.int64(raw_input(prompt))
if seq_length == '':
seq_length = None
prompt = "Enter the maximum frame length"
max_frames = np.int64(raw_input(prompt))
if max_frames == '':
max_frames = None
raise ValueError("Enter Valid interger for max frames")
prompt = "Enter the model_name from the list \n 1.LSTM \n 2.CNN_LSTM"
model_name = np.int64(raw_input(prompt))
if model_name == '':
model_name = None
if model_name == 'LSTM':
data_type = 'features'
img_shape = None
# Feature Extraction
extract_features(seq_length, max_frames, abspath=None)
elif model_name == 'CNN_LSTM':
data_type = 'images'
img_shape = (100, 100, 3)
else:
raise ValueError(
"Invalid Model Selected. Select from List given in the options")
# Training the algorithm
train(seq_length,
model_name,
saved_model,
data_type,
cls_lmt=None,
img_shape=img_shape,
batch_size=32,
no_epoch=100)
# Testing the algorithm
saved_model = raw_input(
"Enter the absolute path for the model saved after training")
acc = test(model_name,
saved_model,
data_type,
seq_length,
cls_lmt=None,
img_shape=None,
batch_size=30)
with open('Testfile.csv', 'a+') as f:
op = model_name + ',' + str(seq_length) + ',' + str(
max_frames) + ',' + saved_model + ',' + acc
f.write(op)
return
def train():
if request.method == 'POST':
f = request.files.getlist("my_file[]")
for i in f :
i.save('raw/'+i.filename)
makeinterim()
makeprocessed()
extract_features ()
report = train_pred()
df = pd.DataFrame(report).transpose()
return render_template('classification_report.html', tables=[df.to_html()])
def main():
foldername = 'data'
X, X_test_kaggle, y, groups, lenc = load_data(foldername)
X, X_test_kaggle = extract_features(X), extract_features(X_test_kaggle)
classifiers = [
LinearDiscriminantAnalysis(),
SVC(kernel='linear', probability=True),
SVC(kernel='rbf', probability=True),
LogisticRegression(),
RandomForestClassifier(n_estimators=1000, max_depth=4),
KNeighborsClassifier(),
ExtraTreesClassifier(n_estimators=1000, max_depth=4)
]
names = [
'LDA', 'SVC(linear)', 'SVC(rbf)', 'Logistic Regression',
'Random Forest', 'KNeighbors', 'ExtraTrees'
]
scores_by_clf = defaultdict(list)
rs = GroupShuffleSplit(n_splits=100, test_size=0.2)
for trindex, tsindex in rs.split(X, y, groups):
X_train, y_train = X[trindex, :], y[trindex]
X_test, y_test = X[tsindex, :], y[tsindex]
print("Training set has classes: ", np.unique(y_train))
amount = []
for i in range(len(np.unique(y_train))):
amount.append(len(np.argwhere(y_train == np.unique(y_train)[i])))
print(amount)
print("Testing set has classes: ", np.unique(y_test))
amount = []
for i in range(len(np.unique(y_test))):
amount.append(len(np.argwhere(y_test == np.unique(y_test)[i])))
print(amount)
for i in range(len(classifiers)):
clf = classifiers[i]
pred, score, clf, proba = classify(clf,
X_train,
y_train,
X_test,
y_test,
groups=groups)
print(np.unique(pred))
print(names[i], 'score: %.3f \n' % score)
scores_by_clf[names[i]].append(score)
for clfname in scores_by_clf.keys():
print(clfname, np.mean(scores_by_clf[clfname]))
def main(args):
print('===> args:\n', args)
config_json = args.config_json
image_dir = args.image_dir
save_dir = args.save_dir
image_list_file = args.image_list_file
gpu_id = args.gpu
extract_features(config_json,
save_dir,
image_list_file,
image_dir,
gpu_id=gpu_id)
def build_features(scale_features=False):
film_list = []
feat_list = []
qual_list = []
count = 0
with open('imsdb_ratings_processed.csv', 'rb') as csvfile:
__ = csvfile.readline()
reader = csv.reader(csvfile)
for row in reader:
count += 1
print "Film No.: " + str(count)
#if count == 10:
# break
#if get_status(row) == "Good" or get_status(row) == "Bad":
file_name = file_base + row[0] + ".txt"
print file_name
scenes = format_script(file_name)
if scenes is None:
print "No file or formatting error"
continue
print "Num scenes: " + str(len(scenes))
if len(scenes) < 5:
print "Skipping..."
continue
# Segment level features
chunks = script_to_n_chunks(scenes)
features = {}
for idx, chunk in enumerate(chunks):
# change input parameter to scenes for extract features
try:
chunk_features = extract_features(chunk, idx + 1)
except:
try:
chunk_features = extract_features(chunk, idx + 1)
except:
print "Failed to extract features"
continue
features.update(chunk_features)
# Full script features
script_summary_features = get_summary_features(scenes, row[0])
final_features = {}
final_features.update(features)
final_features.update(script_summary_features)
feat_list.append(final_features)
#if get_status(row) == "Good":
qual_list.append(row[4])
#else:
#qual_list.append(0)
film_list.append(row[0])
return film_list, feat_list, qual_list
def build_features(scale_features = False):
film_list = []
feat_list = []
qual_list = []
count = 0
with open('imsdb_ratings_processed.csv', 'rb') as csvfile:
__ = csvfile.readline()
reader = csv.reader(csvfile)
for row in reader:
count += 1
print "Film No.: " + str(count)
#if count == 10:
# break
#if get_status(row) == "Good" or get_status(row) == "Bad":
file_name = file_base + row[0] + ".txt"
print file_name
scenes = format_script(file_name)
if scenes is None:
print "No file or formatting error"
continue
print "Num scenes: " + str(len(scenes))
if len(scenes) < 5:
print "Skipping..."
continue
# Segment level features
chunks = script_to_n_chunks(scenes)
features = {}
for idx, chunk in enumerate(chunks):
# change input parameter to scenes for extract features
try:
chunk_features = extract_features(chunk, idx + 1)
except:
try:
chunk_features = extract_features(chunk, idx + 1)
except:
print "Failed to extract features"
continue
features.update(chunk_features)
# Full script features
script_summary_features = get_summary_features(scenes, row[0])
final_features = {}
final_features.update(features)
final_features.update(script_summary_features)
feat_list.append(final_features)
#if get_status(row) == "Good":
qual_list.append(row[4])
#else:
#qual_list.append(0)
film_list.append(row[0])
return film_list, feat_list, qual_list
def training(data, classifier, one_vs_all, apply_pca, tag_dict):
ind_tags_phrases = individual_phrase_tags(data, tag_dict)
ind_phrases, ind_tags, phrase_position = [], [], []
for i in ind_tags_phrases:
ind_phrases.append(i[0].split(',')[1])
ind_tags.append(i[0].split(',')[0])
phrase_position.append(i[1])
print('train', len(ind_phrases))
train_features = extract_features(ind_phrases, phrase_position, 'training')
write_features_labels('./', ind_phrases, train_features, ind_tags,
'train_features')
pca, scaler = '', ''
if apply_pca == 'y':
scaler = StandardScaler()
scaler.fit(train_features)
train_features = scaler.transform(train_features)
pca = PCA(0.95)
pca.fit(train_features)
train_features = pca.transform(train_features)
model = OneVsRestClassifier(RandomForestClassifier(n_estimators=100),
n_jobs=-1)
model.fit(train_features, ind_tags)
pred_tags = model.predict(train_features)
accuracy = metrics.accuracy_score(ind_tags, pred_tags)
return model, accuracy, pca, scaler
def evaluate_windows_of_size(self, windows, window_size):
X = []
for w in windows:
window_yuv = self.cropped_frame_yuv[int(w.y1):int(w.y2),
int(w.x1):int(w.x2)]
if w.width() != window_size or w.height() != window_size:
window_yuv = cv2.resize(window_yuv, (window_size, window_size))
ppc = 8 if self.use_hires_classifier else 16
X.append(extract_features(window_yuv, window_size, ppc=ppc))
X = np.array(X)
windows = np.array(windows)
normalized_feature_vector = self.scaler[window_size].transform(X)
score = self.classifier[window_size].predict(normalized_feature_vector)
pos_window_indexes = np.where(score == 1.0)[0]
pos_windows = windows[pos_window_indexes]
pos_window_scores = score[pos_window_indexes]
result = []
self.evaluated_windows.extend(windows)
for r, score in zip(pos_windows, pos_window_scores):
result.append((r, score))
if self.save_false_positives and self.is_false_positive_candidate(
r):
window_img = crop_img(self.cropped_frame, r.x1, r.y1, r.x2,
r.y2)
save_img(
window_img, "%s/%d/%04d-%04d" %
(self.false_positive_dir_name, window_size,
self.frame_count, self.false_positive_count))
self.false_positive_count += 1
return result
def BoltMotionObjToFeatureObj(all_bolt_data):
"""
Pull out PCA components from all data
For each object - pull out features and store in feature_obj
with the same structure as all_bolt_data
Dictionary - "tap", "slide", "slow_slide",
"thermal_hold", "squeeze"
"""
# DO PCA Calculations here
# Store in feature class object
all_features_obj_dict = dict();
for motion_name in all_bolt_data:
trial_list = all_bolt_data.get(motion_name)
print motion_name
feature_list = list()
# For all objects
for trial in trial_list:
bolt_feature_obj = extract_features.extract_features(trial)
feature_list.append(bolt_feature_obj)
# Store all of the objects away
all_features_obj_dict[motion_name] = feature_list
return all_features_obj_dict
def compute_probability_vector(self, bolt_obj):
if bolt_obj.state == bolt_obj.TAP:
# Store results as they come in
self.adjective_vectors = dict()
self.all_motion_results = dict()
# Store dictionary of strings
self.state_string = {bolt_obj.DISABLED:'disabled',
bolt_obj.THERMAL_HOLD:'thermal_hold',
bolt_obj.SLIDE:'slide',
bolt_obj.SQUEEZE:'squeeze',
bolt_obj.TAP:'tap',
bolt_obj.DONE:'done',
bolt_obj.SLIDE_FAST:'slide_fast',
bolt_obj.CENTER_GRIPPER:'center_gripper'
}
current_motion = self.state_string[bolt_obj.state]
# Create feature vector
self.bolt_object = bolt_obj
utilities.normalize_data(self.bolt_object)
self.bolt_feature_object = extract_features.extract_features(self.bolt_object, self.pca_model[current_motion])
# Create a dictionary to store the results in
for adj in self.all_classifiers:
results, prob = utilities.compute_adjective_probability_score(self.all_classifiers[adj], self.bolt_feature_object, self.feature_list, adj, self.scaler_dict)
# Store off adjective probabilities for ensemble
if adj in self.adjective_vectors:
pass
else:
self.adjective_vectors[adj] = list()
self.adjective_vectors[adj].append(prob)
# Store classifier score based on best motion
best_motion = self.best_motion_dict[adj][1]
if current_motion == best_motion:
rospy.loginfo("Best Motion is: %s" % best_motion)
self.all_motion_results[adj] = results
print len(self.adjective_vectors[adj])
if len(self.adjective_vectors[adj]) == 5:
ensembled_results = dict()
print self.adjective_vectors
#for adj in self.adjective_vectors:
# ensembled_results[adj] = self.ensemble_classifiers[adj].predict(self.adjective_vectors[adj])
# Store off the adjectives that returned true
adjectives_found = []
for adj in self.all_motion_results:
if self.all_motion_results[adj] == 1:
adjectives_found.append(adj)
print "Results from max classification"
print self.all_motion_results
print str(adjectives_found)
self.adjectives_pub.publish(str(adjectives_found))
def BoltMotionObjToFeatureObj(all_bolt_data, electrode_pca_dict):
"""
For each object - pull out features and store in feature_obj
with the same structure as all_bolt_data
Dictionary - "tap", "slide", "slow_slide",
"thermal_hold", "squeeze"
"""
# Store in feature class object
all_features_obj_dict = dict();
for motion_name in all_bolt_data:
trial_list = all_bolt_data.get(motion_name)
print motion_name
feature_list = list()
# For all objects
for trial in trial_list:
bolt_feature_obj = extract_features.extract_features(trial, electrode_pca_dict[motion_name])
feature_list.append(bolt_feature_obj)
# Store all of the objects away
all_features_obj_dict[motion_name] = feature_list
return all_features_obj_dict
def cascaded_retrieval(queries, ord_features, feature_dict, k_list,
metric_dict):
queries = preprocess_sketches(queries)
results = []
for query in queries:
img_indices = np.arange(len(feature_dict[ord_features[0]]))
for i, feature in enumerate(ord_features):
sketch_features = extract_features(feature, query)
if feature != "sift":
image_features = np.asarray(feature_dict[feature])
image_features = image_features[img_indices]
sketch_features = np.array(sketch_features).reshape(1, -1)
# print(f"Image: {np.array(image_features).shape}")
# print(f"Sketch: {np.array(sketch_features).shape}")
else:
image_features = np.asarray(feature_dict[feature])
image_features = [image_features[idx] for idx in img_indices]
sketch_features = [sketch_features]
distances = compute_distances(image_features, sketch_features,
metric_dict[feature])
top_results = get_top_results(k_list[i], distances)
img_indices = img_indices[top_results[0]]
results.append(img_indices)
return results
def train_model():
for files in file_paths:
# Each speaker will have 1 features array
features = np.asarray(())
for filepath in files:
print("Training: " + filepath)
try:
sr, audio = read(filepath)
vector = extract_features(audio, sr)
except Exception as e:
print(e)
continue
if features.size == 0:
features = vector
else:
try:
features = np.vstack((features, vector))
except:
print("ValueError: Shape does not match")
# gmm
gmm = GMM(n_components=16,
max_iter=200,
covariance_type='diag',
n_init=3)
gmm.fit(features)
# export trained model
picklefile = model + os.path.basename(filepath).split('_')[0] + ".gmm"
with open(picklefile, 'wb') as gmm_file:
pickle.dump(gmm, gmm_file)
print('successfully modeling for speaker:', picklefile,
" with data point = ", features.shape)
def process_file(img):
"""process single image, extract features and emotions
the process create the user model as well.
Args:
img (EmotionalImage): image data
"""
assert isinstance(img, Image)
if str(img.name).find('json') > -1:
return
user = get_user(os.path.join(img.path, 'meta.json'))
filePath = os.path.join(img.path, img.name)
# logging.info("---------------Processsing----------------", img.name)
print("---------------Processsing----------------", img.name)
try:
features = extract_features(filePath)
except:
logging.exception('Somthing went wrong with feature extraction')
return
try:
emotions = predict_emotions(features)
except:
logging.exception('Somthing went wrong with emotions extraction')
return
uuid1 = uuid.uuid4()
emImage = EmotionalImage(uuid1, img.name, img.path, features, emotions, "",
"", "")
# TODO: fix that, currently add one image at a time, not a functional issue but can be imroved.
user.images.append(emImage)
user.save()
def compare_features(filepath):
bboxes, face_features = extract_features(filepath)
if len(bboxes) == 0:
return []
names, allfeatures = load_features()
results = []
for bbox, feature in zip(bboxes, face_features):
min = sys.float_info.max
max = 0.0
nearest = None
furthest = None
for n, f in zip(names, allfeatures):
dist = scipy.spatial.distance.cosine(feature, f)
if dist < min:
nearest = n
min = dist
if dist > max:
furthest = n
max = dist
results.append({
'face': bbox,
'nearest': {
'name': nearest,
'distance': min
},
'furthest': {
'name': furthest,
'distance': max
}
})
return results
def BoltMotionObjToFeatureObj(all_bolt_data, electrode_pca_dict):
"""
For each object - pull out features and store in feature_obj
with the same structure as all_bolt_data
Dictionary - "tap", "slide", "slow_slide",
"thermal_hold", "squeeze"
"""
# Store in feature class object
all_features_obj_dict = dict()
for motion_name in all_bolt_data:
trial_list = all_bolt_data.get(motion_name)
print motion_name
feature_list = list()
# For all objects
for trial in trial_list:
bolt_feature_obj = extract_features.extract_features(
trial, electrode_pca_dict[motion_name])
feature_list.append(bolt_feature_obj)
# Store all of the objects away
all_features_obj_dict[motion_name] = feature_list
return all_features_obj_dict
def train(data_folder=TRAIN_FOLDER,
labels_file=LABELS_FILE,
model_prefix=MODEL_PREFIX,
cdf_prediction=PREDICT_CDF,
parameters=PARAMETERS):
"""Train ensemble."""
random.seed(SEED)
_, sys_lab, dia_lab, features = \
extract_features.extract_features(data_folder, labels_file)
assert len(sys_lab) == len(dia_lab) == features.shape[0]
print 'total data', features.shape[0]
print 'num features', features.shape[1]
random.seed(SEED)
test_prop = TEST_PROP
test_size = int(features.shape[0] * test_prop)
shuffled_indices = random.shuffle(range(features.shape[0]))
test_sys_lab = sys_lab[:test_size]
test_dia_lab = dia_lab[:test_size]
test_features = features[:test_size]
train_sys_lab = sys_lab[test_size:]
train_dia_lab = dia_lab[test_size:]
train_features = features[test_size:]
print 'train size', train_features.shape[0]
print 'test size', test_features.shape[0]
systole_prefix = '%s_sys' % model_prefix
diastole_prefix = '%s_dia' % model_prefix
print 'training systole model'
_train(train_features, train_sys_lab, test_features, test_sys_lab,
model_prefix=systole_prefix,
cdf_prediction=cdf_prediction,
parameters=parameters)
print 'training diastole model'
_train(train_features, train_dia_lab, test_features, test_dia_lab,
model_prefix=diastole_prefix,
cdf_prediction=cdf_prediction,
parameters=parameters)
print 'final systole evaluation on test set'
sys_crps = evaluate(test_features, test_sys_lab,
model_prefix=systole_prefix,
cdf_prediction=cdf_prediction)
print 'final diastole evaluation on test set'
dia_crps = evaluate(test_features, test_dia_lab,
model_prefix=diastole_prefix,
cdf_prediction=cdf_prediction)
print 'test sys crps', sys_crps
print 'test dia crps', dia_crps
print 'overall test crps', (sys_crps + dia_crps) * 0.5
def evaluate_window(self, window):
window_size = window.height()
window_yuv = self.cropped_frame_yuv[int(window.y1):int(window.y2),
int(window.x1):int(window.x2)]
X = np.array(extract_features(window_yuv, window_size))
normalized_feature_vector = self.scaler[window_size].transform(X)
return self.classifier[window_size].predict(
normalized_feature_vector)[0]
def define_segments(QLINK_URLS, UNKNOWN_URLS, QUOTA):
global quota
global clusterizer
global df
global classificator
global toexist
global standartizator
quota = []
y = np.empty(1000)
threshold = 90
nclusters = 15
df = pd.DataFrame()
df = df.append(exf.extract_features(QLINK_URLS), ignore_index=True)
df = df.append(exf.extract_features(UNKNOWN_URLS), ignore_index=True)
cnts = df.count()
df = df.fillna(0)
toexist = []
for i in xrange(len(cnts)):
if cnts[i] > threshold:
toexist.append(df.columns[i])
y[:500] = 1
y[500:] = 0
X = df[toexist].values
# X1 = TSNE().fit_transform(X)
# plt.scatter(X1[:, 0], X1[:, 1], c=y*8, cmap=plt.cm.get_cmap("jet", 10), s=1)
# plt.colorbar(ticks=range(10))
# plt.clim(-0.5, 9.5)
# plt.show()
standartizator = StandardScaler()
X = standartizator.fit_transform(X)
clusterizer = KMeans(n_clusters=nclusters)
clusterizer.fit(X)
#classificator = LDA(solver='lsqr').fit(X, y)
#classificator = SVC()
classificator = KNeighborsClassifier()
classificator.fit(X, y)
qlInCluster = []
for i in xrange(nclusters):
qlInCluster.append(sum(y[clusterizer.labels_ == i]))
quota.append(90 * qlInCluster[i] + QUOTA / 100)
def daily_batch():
print "**** [Step 1] Clean old data and prepare directories ****"
try:
shutil.rmtree('data/')
os.makedirs('data/')
os.makedirs('output/')
except:
pass
print "**** [Step 2] Crawl news articles ****"
grd.get_raw_data()
print "**** [Step 3] Compute features for each news articles ****"
ef.extract_features(mode="batch")
print "**** [Step 4] Compute the related news ****"
print "(may take long time)"
fv, id_list = grn.get_feature_vectors(mode="batch")
grn.ANN(fv, id_list)
print "**** [Step 5] Loading data to Redis ****"
ftr.load_data(mode="batch")
print "DONE!"
def input_fn():
# Load and parse dataset
dataset = tf.data.TFRecordDataset(filename, compression_type='GZIP')
corpus = dataset.map(exp_TFR.decode, num_parallel_calls=8)
corpus = corpus.shuffle(5000000)
# Build the dictionary
# Extract the top 60000 most common words to include in our embedding vector
vocab_file_path = "Dataset/Vocabulary/vocabulary.txt"
vocab_size = 60000
# Gather together all the unique words and index them with a unique integer value
# Loop through every word in the dataset and assign it to the unique integer word identified.
# Any words not within the top 60000 most common words will be marked with "-1" and replace with "UNK" token
# Load the dictionary populated by keys corresponding to each unique word
table = tf.contrib.lookup.index_table_from_file(
vocabulary_file=vocab_file_path,
vocab_size=vocab_size,
key_column_index=1,
delimiter=' ')
# Create a reverse_table that allows us to look up a word based on its unique integer identifier,
# rather than looking up the identifier based on the word.
# reverse_table = tf.contrib.lookup.index_to_string_table_from_file(vocabulary_file=vocab_file_path,
# vocab_size=vocab_size,
# value_column_index=1,
# delimiter=' ')
# Load ocean dictionary
ocean_dict_file_path = "Dataset/Vocabulary/ocean_dict_filtered.txt"
ocean_dict_size = 634 # 636 before (deleted 2 adjective)
# Ocean lookup-table
ocean_table = tf.contrib.lookup.index_table_from_file(
vocabulary_file=ocean_dict_file_path,
vocab_size=ocean_dict_size,
key_column_index=0,
delimiter='\t')
# Extract labels and features and generate dataset
dataset = corpus.map(
lambda ids, text: extract_features(ids, text, table, ocean_table),
num_parallel_calls=8)
# Delete all the sentences without adjective
# dataset = dataset.filter(lambda features, ocean_vector: tf.reduce_all(tf.is_finite(ocean_vector)))
dataset = dataset.filter(lambda features, ocean_value: tf.reduce_all(
tf.is_finite(ocean_value)))
dataset = dataset.batch(batch_size=500)
return dataset
def main():
"""These are the main training settings. Set each before running
this file."""
#os.environ['CUDA_VISIBLE_DEVICES'] = "0"
start = time.time()
os.environ['TF_CPP_MIN_LOG_LEVEL'] = "2"
if (len(sys.argv) == 5):
seq_length = int(sys.argv[1])
class_limit = int(sys.argv[2])
image_height = int(sys.argv[3])
image_width = int(sys.argv[4])
else:
print ("Usage: python train.py sequence_length class_limit image_height image_width")
print ("Example: python train.py 75 2 720 1280")
exit (1)
sequences_dir = os.path.join('data', 'sequences')
if not os.path.exists(sequences_dir):
os.mkdir(sequences_dir)
checkpoints_dir = os.path.join('data', 'checkpoints')
if not os.path.exists(checkpoints_dir):
os.mkdir(checkpoints_dir)
# model can be only 'lstm'
model = 'lstm'
saved_model = None # None or weights file
load_to_memory = False # pre-load the sequences into memory
batch_size = 16
nb_epoch = 50
data_type = 'features'
image_shape = (image_height, image_width, 3)
extract_features(seq_length=seq_length, class_limit=class_limit, image_shape=image_shape)
train(data_type, seq_length, model, saved_model=saved_model,
class_limit=class_limit, image_shape=image_shape,
load_to_memory=load_to_memory, batch_size=batch_size, nb_epoch=nb_epoch)
print('time required {:0.3f}'.format(time.time() - start))
def main():
# Load the data, get features, scale data
foldername = 'data'
X, X_test_kaggle, y, groups, lenc = load_data(foldername)
X, X_test_kaggle = extract_features(X), extract_features(X_test_kaggle)
print('the shape of X: ' + str(X.shape))
scaler = StandardScaler()
X = scaler.fit_transform(X)
X_test_kaggle = scaler.transform(X_test_kaggle)
classifiers = [RandomForestClassifier(n_estimators=500, max_depth=4)]
predictions, scores, classifiers, probabilities = classify_multi(classifiers, X, y, X_test_kaggle)
print(predictions)
# Submission stuff
labels = lenc.inverse_transform(predictions[:,0].astype(int))
with open('results/submission.csv', "w") as fp:
fp.write("# Id,Surface\n")
for i in range(len(labels)):
fp.write("%d,%s\n"%(i, labels[i]))
def predict(wav_path, textgrid_path, start_extract, end_extract):
tmp_feature_file = generate_tmp_filename('features')
tmp_prob_file = generate_tmp_filename('prob')
tmp_predict_file = generate_tmp_filename('prediction')
if not os.path.exists(wav_path):
print >> sys.stderr, "wav file %s does not exits" % wav_path
return
length = utils.get_wav_file_length(wav_path)
print '\n1) Extracting features and classifying ...'
extract_features(wav_path, tmp_feature_file, start_extract, end_extract)
run(tmp_feature_file, tmp_prob_file)
print '\n3) Extract Durations ...'
post_process(tmp_prob_file, tmp_predict_file)
print '\n4) Writing TextGrid file to %s ...' % textgrid_path
create_text_grid(tmp_predict_file, textgrid_path, length, float(start_extract))
# remove leftovers
os.remove(tmp_feature_file)
os.remove(tmp_prob_file)
os.remove(tmp_predict_file)
def main():
import sys
classifier = cv2.SVM()
classifier.load(sys.argv[1])
im = cv2.imread(sys.argv[2])
im = cv2.GaussianBlur(im, (3, 3), 0)
imshow_large(__file__, im)
cv2.waitKey()
unrotated = unrotate(im)
#
# Check if image is right way up, correct otherwise
#
imshow_large(__file__, unrotated)
key = cv2.waitKey()
if key & 0xff == ord("r"):
unrotated = cv2.flip(cv2.flip(unrotated, 0), 1)
imshow_large(__file__, unrotated)
cv2.waitKey()
binarized = binarize(unrotated)
rois = get_rois(binarized)
results = {}
grayscale = cv2.cvtColor(unrotated, cv.CV_BGR2GRAY)
for (x, y, width, height) in rois:
roi = grayscale[y:y + height, x:x + width]
vec = extract_features(roi)
label = classifier.predict(vec)
results[(x, y, width, height)] = "01234567890X"[int(label)]
scale = SCREEN_HEIGHT / unrotated.shape[0]
unrotated = cv2.cvtColor(grayscale, cv.CV_GRAY2BGR)
scaled = cv2.resize(unrotated, (0, 0), fx=scale, fy=scale)
for roi in rois:
x = int(roi[0] * scale)
y = int(roi[1] * scale)
width = int(roi[2] * scale)
height = int(roi[3] * scale)
cv2.rectangle(scaled, (x, y), (x + width, y + height), (0, 255, 0, 0),
1)
if results[roi] == "X":
continue
cv2.putText(scaled, results[roi], (x, y), cv.CV_FONT_HERSHEY_SIMPLEX,
1, (0, 255, 0, 0))
cv2.imshow(__file__, scaled)
cv2.waitKey()
def test_classifier_parameters(classifier = 'classifier_name', infile = 'features.npz', subcats=True, norm = False):
"""
main function to determine number of features to be used for classification and classifier hyperparameters
classifier - classifier to be used, Naive Bayes or LDA
infile - .csv file with the extracted features
norm - Boolean parameter whether to normalize the Naive Bayes for unbalanced class sizes
"""
# extract numpy arrays, lists and dictionaries from the features.npz
extract_features(subcats=subcats)
features_file = np.load(infile)
features, featurenames, categoryids, categories = features_file['features'], features_file['featurenames'], \
features_file['categoryids'], features_file['categories'].item()
labels = categoryids[0,:]
features = features.T
categoryids = categoryids.T
if classifier == 'Naive Bayes':
# if True, do the normalization for unbalanced class sizes
if norm:
features = normalize(features, categoryids, categories)
classify('nb', features, labels, categories)
else:
classify('lda', features, labels, categories)
def main():
import sys
classifier = cv2.SVM()
classifier.load(sys.argv[1])
im = cv2.imread(sys.argv[2])
im = cv2.GaussianBlur(im, (3,3), 0)
imshow_large(__file__, im)
cv2.waitKey()
unrotated = unrotate(im)
#
# Check if image is right way up, correct otherwise
#
imshow_large(__file__, unrotated)
key = cv2.waitKey()
if key & 0xff == ord("r"):
unrotated = cv2.flip(cv2.flip(unrotated, 0), 1)
imshow_large(__file__, unrotated)
cv2.waitKey()
binarized = binarize(unrotated)
rois = get_rois(binarized)
results = {}
grayscale = cv2.cvtColor(unrotated, cv.CV_BGR2GRAY)
for (x,y,width,height) in rois:
roi = grayscale[y:y+height,x:x+width]
vec = extract_features(roi)
label = classifier.predict(vec)
results[(x,y,width,height)] = "01234567890X"[int(label)]
scale = SCREEN_HEIGHT/unrotated.shape[0]
unrotated = cv2.cvtColor(grayscale, cv.CV_GRAY2BGR)
scaled = cv2.resize(unrotated, (0,0), fx=scale, fy=scale)
for roi in rois:
x = int(roi[0]*scale)
y = int(roi[1]*scale)
width = int(roi[2]*scale)
height = int(roi[3]*scale)
cv2.rectangle(scaled, (x,y), (x+width, y+height), (0,255,0,0), 1)
if results[roi] == "X":
continue
cv2.putText(scaled, results[roi], (x, y), cv.CV_FONT_HERSHEY_SIMPLEX, 1, (0, 255, 0, 0))
cv2.imshow(__file__, scaled)
cv2.waitKey()
def testing(data, model, pca, scaler, apply_pca):
ind_tags_phrases = individual_phrase_tags(data)
ind_phrases, ind_tags, phrase_position = [], [], []
for i in ind_tags_phrases:
ind_phrases.append(i[0].split(',')[1])
ind_tags.append(i[0].split(',')[0])
phrase_position.append(i[1])
test_features = extract_features(ind_phrases, phrase_position, 'testing')
if apply_pca == 'y':
test_features = scaler.transform(test_features)
test_features = pca.transform(test_features)
pred_tags = model.predict(test_features)
accuracy = metrics.accuracy_score(ind_tags, pred_tags)
return accuracy
def fetch_url(url):
global quota
global clusterizer
global df
global standartizator
d = exf.extract_features([url], toexist)[0].values()
d = standartizator.transform([d])[0]
cls = clusterizer.predict([d])[0]
if classificator.predict_proba([d])[0][1] > 0.7:
quota[cls] -= 1
return True
if quota[cls] > 0:
quota[cls] -= 1
return True
else:
return False
def get_test_data(self):
# loads the features array
file_list = os.listdir(os.path.join(self.test_path, 'video'))
# with open(os.path.join(self.test_path, 'testing.txt')) as testing_file:
# lines = testing_file.readlines()
# file_name = lines[self.num].strip()
file_name = file_list[self.num]
path = os.path.join(self.test_path, 'feat', file_name + '.npy')
if os.path.exists(path):
f = np.load(path)
else:
model = extract_features.model_cnn_load()
f = extract_features.extract_features(file_name, model)
if self.num < len(file_list):
self.num += 1
else:
self.num = 0
return f, file_name
def calculateFoldability(seq,sec_str):
#---------------------------#
#seq = 'GUAAGUCGGGGACCUCUUAAGAUGAGAGACUUCUGAACCGGGUCAGGAUCGGAAGAUAGCAGCCCUAAGGAAAGGCCUUUUGUGCUAAGAGUCUUCUCUGACUUAC'
#sec_str = '..(.((((((((.(((((.(((((((((..((((.....((((....(((....)))....))))...))))....))))))).)))))))...)))))))))...'
##-- -------------------------
#---------------------------#
real_RNA_loc = "RNASTRAND_real_feature_space.csv"
folder_simulation_result = "/home/dhrumil/Desktop/Lab/RNAWebsite_v1/RNASTRAND_extract_feature/"
#---------------------------#
df_pred = extract_features.extract_features(seq,sec_str)
clf = bulid_model(real_RNA_loc,folder_simulation_result)
foldability = pred_foldability(df_pred,clf)
fe,mfe,mfe_str = pred_fe(seq,sec_str)
print('sequence:',seq)
print('secondary structure:',sec_str)
return foldability,fe,mfe,mfe_str
def validate(data_folder=VALIDATE_FOLDER,
model_prefix=MODEL_PREFIX,
cdf_prediction=PREDICT_CDF):
"""Use ensemble to predict CDFs on validation set."""
study_ids, sys_lab, dia_lab, features = \
extract_features.extract_features(data_folder)
assert study_ids == tuple(range(501, 701))
assert len(sys_lab) == len(dia_lab) == features.shape[0]
print 'total data', features.shape[0]
print 'num features', features.shape[1]
systole_prefix = '%s_sys' % model_prefix
diastole_prefix = '%s_dia' % model_prefix
systole_cdfs = predict_cdfs(features, systole_prefix, cdf_prediction)
diastole_cdfs = predict_cdfs(features, diastole_prefix, cdf_prediction)
return systole_cdfs, diastole_cdfs
#print "\tProcessing: " + names[i*2][0];
f1 = INPUT_PATH + names[i*2][2]
f2 = INPUT_PATH + names[i*2+1][2]
name = names[i*2][0];
out = CHECK_PATH + name + '.fea.res';
check_f = CHECK_PATH + name + '.fea';
if name not in test_result:
test_result[name] = (0,0,0, False);
if os.path.exists(check_f) != True:
print >> sys.stderr, "Output test file was not created";
continue;
if os.path.exists(out):
os.remove(out);
t1 = time.time()
extract_features.extract_features(f1, f2, out);
t2 = time.time()
if os.path.exists(out)!= True:
print >> sys.stderr, "Output test file was not created";
continue;
result = read_feas(out)
pattern = read_feas(check_f)
t = test_result[name][0] +(t2-t1);
w = test_result[name][1] + compare_results(result, pattern);
c = test_result[name][2] + float(len(result)) / float(len(pattern));
test_result[name] = (t, w, c , True );
count_fails=0;
for name in test_result.keys():
model = model_from_json(json_string)
model.load_weights("model_weights/super_awesome_merged_model_weights.hdf5")
model.compile(loss='categorical_crossentropy',
optimizer='adam',
metrics=['accuracy']
)
# Parse song
if len(sys.argv) < 2:
print("missing parameter")
else:
filename = sys.argv[1]
song_folder = os.path.dirname(os.path.realpath(filename))#should get the directory to the file
if os.path.isdir(filename):
batch_thirty_seconds(song_folder)
extract_features(song_folder)
else:
thirty_seconds(filename)
print("File split. Now extracting features.")
extract_features(song_folder)
print("Extracted features.")
keyword_2 = "mfcc_coefficients"
x2 = []
for root, dirs, files in os.walk(song_folder, topdown=False):
for name in files:
if re.search(keyword_2+".csv",name):
song_path = (os.path.join(root,name))
song_features = genfromtxt(song_path, delimiter=",")
# Train with 3/4 of the sample + 1/4 of the error
#
error_count = sum([e.shape[0] for e in error])
keep_count = max(error_count * 3, npts_sampled - error_count)
if coords.shape[0] > keep_count:
coords = coords[r.permutation(coords.shape[0])[:keep_count], :]
coords = np.vstack([coords] + error)
p = r.permutation(coords.shape[0])[:npts_sampled]
coords = coords[p, :]
coords = coords[np.lexsort(coords.transpose())]
if ("training_features" in tiffcvt.h5_file.keys() and
(tiffcvt.h5_file["training_features"].shape[1] != extract_features.n_features or
tiffcvt.h5_file["training_features"].shape[0] != npts_sampled)):
del tiffcvt.h5_file["training_features"]
del tiffcvt.h5_file["training_classification"]
tf = tiffcvt.h5_file.require_dataset("training_features",
(npts_sampled, extract_features.n_features),
np.float32)
tc = tiffcvt.h5_file.require_dataset("training_classification",
(npts_sampled, ), np.uint32)
for i in range(0, coords.shape[0], 1024):
my_slice = slice(i, min(i+1024, coords.shape[0]))
tf[my_slice,:] = extract_features.extract_features(
img, blur_img, coords[my_slice,:])
tc[my_slice] = labels[coords[my_slice,0],
coords[my_slice,1],
coords[my_slice,2]]
print "Finished %d of %d" % (i+1024, npts_sampled)
tiffcvt.h5_file.close()
def align_and_tune(self):
extract_features(self.original_audio_folder, self.original_features_folder, [features[0]], "jams")
tune_wavs(self.original_audio_folder, self.tuned_audio_folder, self.original_features_folder)
parser.add_argument('-v', '--verbose',
dest='verbose', action='store_true', default=False,
help='turn on verbose message output')
options = parser.parse_args()
# establish MongoDB connection
collection = myutils.get_mongodb_collection(options.hosts, options.database)
# load models for each label
models = test.load_models(collection['models'], ast.literal_eval(options.model))
cursor = myutils.get_mysql_connection(options.host, options.db).cursor()
# contruct the testing set from the MediaWiki table
vectors = []
for ent in wikilove_revs.get_entries(cursor, options.start, options.end, options.window, options.limit, newest=True):
features = extract_features.extract_features({'entry': {'content': {'added': [ent.others.message], 'removed':[]},
'comment': ''}})
vector = myutils.map_key_dict(int, extract_features.extract_vector(features, options.bits))
if ent.receiver_id != ent.sender_id:
vectors.append(myutils.entry_t(ent, features, vector))
labels = sorted(models.keys())
vecs = [x.vector for x in vectors]
predictions = [[[] for y in xrange(0, len(labels))] for x in xrange(0,len(vectors))]
for (n,lname) in enumerate(labels):
lab,_,val = liblinear.linearutil.predict([0]*len(vecs), vecs, models[lname], '-b 1')
for (i,(pred,score)) in enumerate(zip(lab,val)):
predictions[i][n] = score[1] # get the confidence for the label being 'True'
print >>options.output, '<style type="text/css">.prediction{text-align: right;} td{vertical-align: top;} li{border: 1px solid; list-style: none inside; margin: 0.2em;} ul{padding: 0;} blockquote{ font: normal italic 100% serif; }</style>'
print >>options.output, '<body style="background: #EEE;">Generated at %s.' % str(datetime.now())
def realign_and_separate_and_analyze(self):
extract_features(self.tuned_audio_folder, self.tuned_features_folder, [features[0]], "jams")
separate_channels(self.tuned_audio_folder, self.channels_audio_folder)
copy_features_of_separated_channels(self.tuned_features_folder, self.channels_features_folder)
extract_features(self.channels_audio_folder, self.channels_features_folder, features[1:], "jams")
