def main():
# initialize an empty list
list_of_csv = []
# arg: set on config(csv file)
# put all args in the list
for arg in sys.argv[1:]:
list_of_csv.append(arg)
#iterate over the list
for csv_file in list_of_csv:
#save the content of csv file in a df
df = pd.read_csv(csv_file)
#links = df['youtube_id']
#file without .cvs extention
folder_name = get_filename_without_extension(csv_file)
# The folder where the downloaded videos will be saved
video_downloaded_path = os.path.join(UNTRIMMED_VIDEOS, folder_name)
# The folder where the downloaded videos will be trimmed and saved
trimmed_video_path = os.path.join(TRIMMED_VIDEOS, folder_name)
# make foloder according to given path if already not exist
EnsurePath(video_downloaded_path)
EnsurePath(trimmed_video_path)
#iterate over the df
index = 0
for index, row in df.iterrows():
try:
#make the string of youtube link
link = "https://www.youtube.com/watch?v=" + row['youtube_id']
# object creation using YouTube which was imported in the beginning
yt = YouTube(link)
startTime = row['time_start']
endTime = row['time_end']
except Exception as e:
print(str(e)) # to handle exception
stream = yt.streams.filter(file_extension='mp4').first()
noError = True
try:
# downloading the video and save as index(0.mp4,1.mp4)
fname_index = str(index)
downloaded_path = stream.download(
output_path=video_downloaded_path, filename=fname_index)
except Exception as e:
print(str(e))
noError = False
finally:
if noError:
#If there is noError then only go for trimming the video
file_name = os.path.basename(downloaded_path)
target_path = os.path.join(trimmed_video_path, file_name)
command_name = "ffmpeg -i " + downloaded_path + " -ss " + str(
startTime) + " -to " + str(
endTime) + " -c copy " + target_path
print(command_name)
os.system(command_name)
index = index + 1
def save_npy(list_of_folders):
'''
For each action folder in the list of folders save the numpy array files for further processing
:param list_of_folders:
:return:
'''
for folders in list_of_folders:
action_base_name = os.path.basename(folders)
parent_action_folder_path = os.path.join(EXTRACTED_FEATURES_FOLDER,
action_base_name)
EnsurePath(parent_action_folder_path)
onlyfolders = [
f for f in listdir(folders) if not isfile(join(folders, f))
]
model = VGG19Extractor()
for numbered_folder in onlyfolders:
images = get_frames_for_sample(source_folder=folders,
numbered_folder=numbered_folder)
path = os.path.join(parent_action_folder_path,
str(numbered_folder))
if os.path.isfile(path + '.npy'):
continue
frames = RescaleList(images, 60)
if frames is not None:
sequence = []
for image in frames:
features = model.extract(image)
sequence.append(features)
# Save the sequence.
np.save(path, sequence)
def saved_npy(list_of_folders):
for folders in list_of_folders:
action_base_name = os.path.basename(folders)
parent_action_folder_path = os.path.join(EXTRACTED_FEATURES_FOLDER,
action_base_name)
EnsurePath(parent_action_folder_path)
onlyfolders = [
f for f in listdir(folders) if not isfile(join(folders, f))
]
model = InceptionV3Extractor()
for numbered_folder in onlyfolders:
images = get_frames_for_sample(source_folder=folders,
numbered_folder=numbered_folder)
path = os.path.join(parent_action_folder_path,
str(numbered_folder))
if os.path.isfile(path + '.npy'):
continue
frames = RescaleList(images, 100)
if frames is not None:
sequence = []
for image in frames:
features = model.extract(image)
sequence.append(features)
# Save the sequence.
np.save(path, sequence)
def main():
#initialize an empty list
list_of_folder_with_videos = []
#arg: set on config(csv file)
#put all args in the list
for arg in sys.argv[1:]:
list_of_folder_with_videos.append(arg)
#Extract frames
for folders in list_of_folder_with_videos:
main_folder_name = os.path.basename(folders)
frames_extracted_main_folder = os.path.join(EXTRACTED_FRAMES_FOLDER,
main_folder_name)
EnsurePath(frames_extracted_main_folder)
onlyfiles = [f for f in listdir(folders) if isfile(join(folders, f))]
for file_name in onlyfiles:
folder_base_name = os.path.splitext(file_name)[0]
new_folder_path = os.path.join(frames_extracted_main_folder,
folder_base_name)
EnsurePath(new_folder_path)
source_file_path = os.path.join(folders, file_name)
vidcap = cv2.VideoCapture(source_file_path)
# #Create the directory into for the file
video_length = int(vidcap.get(cv2.CAP_PROP_FRAME_COUNT)) - 1
print("Number of frames: ", video_length)
count = 0
print("Converting video..\n")
# Start converting the video
while vidcap.isOpened():
# Extract the frame
ret, frame = vidcap.read()
# Write the results back to output location.
cv2.imwrite(new_folder_path + "/%#05d.jpg" % (count + 1),
frame)
count = count + 1
# If there are no more frames left
if (count > (video_length - 1)):
# Log the time again
# Release the feed
vidcap.release()
# Print stats
print("Done extracting frames.\n%d frames extracted" %
count)
#print("It took %d seconds forconversion." % (time_end - time_start))
break
def main():
#load the data and its classes
list_of_folder_with_extracted_feature = []
for arg in sys.argv[1:]:
list_of_folder_with_extracted_feature.append(arg)
X, y = [], []
index = 0
for folders in list_of_folder_with_extracted_feature:
onlyfiles = [
f for f in listdir(folders) if join(folders, f).endswith(".npy")
]
for files in onlyfiles:
path = os.path.join(folders, files)
img_data = np.load(path)
X.append(img_data)
output_list = []
for i in range(len(list_of_folder_with_extracted_feature)):
if (i == index):
output_list.append(1)
else:
output_list.append(0)
y.append(output_list)
index = index + 1
X = np.array(X)
y = np.array(y)
print(X.shape)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2)
input_shape = (60, 4096)
model = Sequential()
model = Sequential()
model.add(
LSTM(
1024, # kernel_regularizer=l1(0.01), recurrent_regularizer=l1(0.01), bias_regularizer=l1(0.01),
return_sequences=False,
input_shape=input_shape,
dropout=0.5))
model.add(BatchNormalization())
# model.add(Dense(512, activation='tanh'))
# model.add(Dropout(0.5))
model.add(Dense(256, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(2, activation='softmax'))
optimizer = Adam(lr=1e-5, decay=1e-6)
model.compile(loss='categorical_crossentropy',
optimizer=optimizer,
metrics=['accuracy'])
history = model.fit(X_train,
y_train,
batch_size=100,
validation_data=(X_test, y_test),
verbose=1,
epochs=150)
#Now its time to save the model
RESULT_FOLDER_TS = os.path.join(
RESULT_FOLDER,
datetime.now().strftime('%Y-%m-%d_%H-%M-%S'))
EnsurePath(RESULT_FOLDER_TS)
model_folder_path = os.path.join(RESULT_FOLDER_TS, 'model_dir')
EnsurePath(model_folder_path)
model_json = model.to_json()
json_path = os.path.join(model_folder_path, "model_num.json")
with open(json_path, "w") as json_file:
json_file.write(model_json)
weight_path = os.path.join(model_folder_path, "model_weight.h5")
model.save_weights(weight_path)
entire_model_path = os.path.join(model_folder_path, "entire_model.h5")
model.save(entire_model_path)
history_file_path = os.path.join(RESULT_FOLDER_TS, 'value_history.json')
with open(history_file_path, 'w') as f:
json.dump(str(history.history), f)
[test_loss, test_acc] = model.evaluate(X_test, y_test)
print("Evaluation result on Test Data : Loss = {}, accuracy = {}".format(
test_loss, test_acc))
y_pred = model.predict_classes(X_test)
y_pred = to_categorical(y_pred)
cm = confusion_matrix(y_test.argmax(axis=1), y_pred.argmax(axis=1))
print("confusion_matrix")
print(cm)
print("Accuracy", cm.diagonal() / cm.sum(axis=1))
print(
"Precision",
precision_score(y_test.argmax(axis=1),
y_pred.argmax(axis=1),
average=None))
print(
"Recall",
recall_score(y_test.argmax(axis=1),
y_pred.argmax(axis=1),
average=None))
print("f1_score",
f1_score(y_test.argmax(axis=1), y_pred.argmax(axis=1), average=None))
test_eval_dict = {}
test_eval_dict['accuracy'] = (cm.diagonal() / cm.sum(axis=1)).tolist()
test_eval_dict['precision'] = (precision_score(y_test.argmax(axis=1),
y_pred.argmax(axis=1),
average=None)).tolist()
test_eval_dict['recall'] = (recall_score(y_test.argmax(axis=1),
y_pred.argmax(axis=1),
average=None)).tolist()
test_eval_dict['f1_score'] = (f1_score(y_test.argmax(axis=1),
y_pred.argmax(axis=1),
average=None)).tolist()
test_eval_dict[
'description_of_model'] = 'Add some description about the model here'
test_eval_file_path = os.path.join(RESULT_FOLDER_TS, 'test_eval.json')
with open(test_eval_file_path, 'w') as fp:
json.dump(test_eval_dict, fp)
model_summary_file_path = os.path.join(RESULT_FOLDER_TS,
'model_summary.txt')
with open(model_summary_file_path, 'w') as f:
with redirect_stdout(f):
model.summary()
plt.ylabel('Loss', fontsize=16)
plt.title('Loss Curves', fontsize=16)
# Plot the Accuracy Curves
plt.figure(figsize=[8, 6])
plt.plot(history.history['accuracy'], 'r', linewidth=3.0)
plt.plot(history.history['val_accuracy'], 'b', linewidth=3.0)
plt.legend(['Training Accuracy', 'Validation Accuracy'], fontsize=18)
plt.xlabel('Epochs ', fontsize=16)
plt.ylabel('Accuracy', fontsize=16)
plt.title('Accuracy Curves', fontsize=16)
##SAVE EVERYTHING
RESULT_FOLDER_TS = os.path.join(RESULT_FOLDER,
datetime.now().strftime('%Y-%m-%d_%H-%M-%S'))
EnsurePath(RESULT_FOLDER_TS)
model_folder_path = os.path.join(RESULT_FOLDER_TS, 'model_dir')
EnsurePath(model_folder_path)
model_json = model.to_json()
json_path = os.path.join(model_folder_path, "model_num.json")
with open(json_path, "w") as json_file:
json_file.write(model_json)
weight_path = os.path.join(model_folder_path, "model_weight.h5")
model.save_weights(weight_path)
entire_model_path = os.path.join(model_folder_path, "entire_model.h5")
model.save(entire_model_path)
history_file_path = os.path.join(RESULT_FOLDER_TS, 'value_history.json')
def clean_tmp_dir():
rmtree(TESTING_DIR)
EnsurePath(TESTING_DIR)
##Clean the directory
def clean_tmp_dir():
rmtree(TESTING_DIR)
EnsurePath(TESTING_DIR)
for sample_file in test_video_list:
total_time = 0
start_time = time.time()
clean_tmp_dir()
##Extract images from videos
IMG_DIR = os.path.join(TESTING_DIR, 'extracted_frames')
EnsurePath(IMG_DIR)
vidcap = cv2.VideoCapture(sample_file)
# #Create the directory into for the file
video_length = int(vidcap.get(cv2.CAP_PROP_FRAME_COUNT)) - 1
#print("Number of frames: ", video_length)
count = 0
#print("Converting video..\n")
# Start converting the video
while vidcap.isOpened():
# Extract the frame
ret, frame = vidcap.read()
# Write the results back to output location.
cv2.imwrite(IMG_DIR + "/%#05d.jpg" % (count + 1), frame)
count = count + 1
# If there are no more frames left
