From 40b3e79e78f8de37d459649f7d2313433e3320fa Mon Sep 17 00:00:00 2001
From: Devashish Lohani <devashish.lohani@univ-lyon2.fr>
Date: Fri, 7 Aug 2020 16:04:02 +0200
Subject: [PATCH] Generic variable names, comments & code cleaning
---
data_management.py | 48 ++++-------
h5py_init.py | 202 +++++++++++++++------------------------------
img_exp.py | 4 +-
models.py | 45 ++--------
seq_exp.py | 39 ++++-----
util.py | 13 +--
6 files changed, 112 insertions(+), 239 deletions(-)
diff --git a/data_management.py b/data_management.py
index a401f1e..bb81624 100644
--- a/data_management.py
+++ b/data_management.py
@@ -10,11 +10,7 @@ import sys
from h5py_init import init_videos, init_data_by_class
root_drive = '.'
-#if not os.path.isdir(root_drive):
-# print('Using Sharcnet equivalent of root_drive')
-# root_drive = '/home/jjniatsl/project/jjniatsl/Fall-Data'
-
-def get_windowed_data(dset = 'Thermal', ADL_only = True, win_len = 8, img_width = 64, img_height = 64, avoid_vid =None, save_np=False):
+def get_windowed_data(dset = 'Thermal_Fall', ADL_only = True, win_len = 8, img_width = 64, img_height = 64, avoid_vid =None, save_np=False):
'''
Creates windowed version of dset data. avoid a video if required!
@@ -40,7 +36,7 @@ def get_windowed_data(dset = 'Thermal', ADL_only = True, win_len = 8, img_width
print(data_dict)
if ADL_only == True:
data_dict = dict((key, value) for key, value in data_dict.items() if
- ('adl' in key or 'ADL' in key or 'NA' in key) and (key != avoid_vid))
+ ('ADL' in key or 'NA' in key) and (key != avoid_vid))
print(data_dict)
vids_win = create_windowed_arr_per_vid(vids_dict = data_dict, \
stride = 1, \
@@ -62,11 +58,11 @@ def get_windowed_data(dset = 'Thermal', ADL_only = True, win_len = 8, img_width
return vids_win
-def init_windowed_arr(dset = 'Thermal', ADL_only = True, win_len = 8, img_width = 64, img_height = 64):
+def init_windowed_arr(dset = 'Thermal_Fall', ADL_only = True, win_len = 8, img_width = 64, img_height = 64):
'''
Creates windowed version of dset data. Saves windowed array to
- 'npData/ADL_data-proc-win_{}.npy'.format(train_or_test, dset, win_len), vids_win)
+ npData folder as npy file
Params:
str dset: dataset to use
@@ -81,29 +77,28 @@ def init_windowed_arr(dset = 'Thermal', ADL_only = True, win_len = 8, img_width
if not os.path.isfile(master_path):
print('initializing h5py..')
- init_videos(img_width = img_width, img_height = img_height, raw = False, dset = dset)
+ init_videos(img_width=img_width, img_height=img_height, raw=False, dset=dset)
with h5py.File(master_path, 'r') as hf:
data_dict = hf[dset + '/Processed/Split_by_video']
-
if ADL_only == True:
- data_dict = dict((key,value) for\
- key, value in data_dict.items() if 'adl' in key or 'ADL' in key or 'NA' in key) #Get only ADL vids
+ # Get only normal behaviour vids
+ data_dict = dict((key,value) for key, value in data_dict.items() if 'ADL' in key or 'NA' in key)
- vids_win = create_windowed_arr_per_vid\
- (vids_dict = data_dict, \
- stride = 1, \
- win_len = win_len,\
- img_width= img_width,\
- img_height= img_height)
+ vids_win = create_windowed_arr_per_vid(
+ vids_dict=data_dict,
+ stride=1,
+ win_len=win_len,
+ img_width=img_width,
+ img_height=img_height)
if ADL_only == True:
save_path = root_drive + '/npData/{}/'.format(dset)
if not os.path.isdir(save_path):
os.makedirs(save_path)
- save_path = save_path + 'ADL_data-proc-win_{}.npy'.format(win_len)
+ save_path = save_path + 'training_data-imgdim_{}x{}-win_{}.npy'.format(img_width, img_height, win_len)
print('saving data to ', save_path)
np.save(save_path, vids_win)
@@ -257,20 +252,15 @@ def create_windowed_arr_per_vid(vids_dict, stride, win_len, img_width, img_heigh
'''
returns windows made of all videos
- Assumes vids_dict is h5py structure, ie. vids_dict = hf['Data_2017/UR/Raw/Split_by_video']
+ Assumes vids_dict is h5py structure
data set must contain atleast win_len frames
'''
vid_list = [len(vid['Data'][:]) for vid in list(vids_dict.values())]
-
num_windowed = sum([int( np.floor((val-win_len)/stride) ) + 1 for val in vid_list])
-
output_shape = (num_windowed, win_len, img_width, img_height, 1)
-
print(output_shape)
-
total = np.zeros(output_shape)
-
i = 0
for vid, name in zip(vids_dict.values(), vids_dict.keys()):
print('windowing vid at', name)
@@ -293,14 +283,12 @@ def create_windowed_arr(arr, stride, win_len):
img_width, img_height = arr.shape[1], arr.shape[2]
- output_length = int( np.floor((len(arr) - win_len) / stride) ) + 1
+ output_length = int(np.floor((len(arr) - win_len) / stride)) + 1
output_shape = (output_length, win_len, img_width, img_height, 1)
-
total = np.zeros(output_shape)
-
i = 0
while i < output_length:
- next_chunk = np.array( [arr[i+j] for j in range(win_len)] )
+ next_chunk = np.array([arr[i+j] for j in range(win_len)])
# Can use np.arange if want to use time step i.e. np.arange(0,win_len,dt)
total[i] = next_chunk
i = i + stride
@@ -310,7 +298,7 @@ def create_windowed_arr(arr, stride, win_len):
def load_data(split_by_vid_or_class = 'Split_by_vid', raw = False, img_width = 64, \
- img_height = 64, vid_class = 'NonFall', dset = 'Thermal'):
+ img_height = 64, vid_class = 'NonFall', dset = 'Thermal_Dummy'):
"""
Note :to use this function, need to have downloaded h5py for dset,
diff --git a/h5py_init.py b/h5py_init.py
index ed4ec80..dfc7ae5 100644
--- a/h5py_init.py
+++ b/h5py_init.py
@@ -8,80 +8,50 @@ import sys
'''
Note, these functions will not work without setting up the directories of video frames as shown in get_dir_lists.
-Alternatively, contact me to get access to the final h5Py datasets, which this code procudes.
'''
-root_drive = '.' #Current dir for now
-
-#if not os.path.isdir(root_drive):
-# print('Using Sharcnet equivalent of root_drive')
-# root_drive = '/home/jjniatsl/project/jjniatsl/Datasets'
+root_drive = '.'
def get_dir_lists(dset):
'''
- Gets videos (ADl, Fall) directory path list
-
+ Gets videos directory path list
Params:
str dset: dataset to be loaded
Returns:
- paths to ADL and Fall videos
+ paths to ADL/NA and Fall/Intrusion videos
'''
- path_Fall = root_drive + '/Datasets/{}/Fall/Fall*'.format(dset)
- path_ADL = root_drive + '/Datasets/{}/NonFall/ADL*'.format(dset)
-
- # Update path_Fall & ADL if some dataset in arranged in some particular way in folders!
- if dset == 'IDD':
- path_Fall = root_drive + '/Datasets/{}/Intrusion/Intru*'.format(dset)
- path_ADL = root_drive + '/Datasets/{}/NonIntrusion/NA*'.format(dset)
-
- elif dset == 'Thermal-Dummy':
- path_Fall = root_drive + '/Datasets/Thermal-Dummy/Fall/Fall*'
- path_ADL = root_drive + '/Datasets/Thermal-Dummy/NonFall/ADL*'
-
- elif dset == 'Thermal':
- path_Fall = root_drive + '/Datasets/Thermal/Fall/Fall*'
- path_ADL = root_drive + '/Datasets/Thermal/NonFall/ADL*'
-
- elif dset == 'UR':
- path_Fall = root_drive + '/Datasets/UR_Kinect/Fall/original/Fall*'
- path_ADL = root_drive + '/Datasets/UR_Kinect/NonFall/original/adl*'
-
- elif dset == 'UR-Filled':
- path_Fall = root_drive + '/Datasets/UR_Kinect/Fall/filled/Fall*'
- path_ADL = root_drive + '/Datasets/UR_Kinect/NonFall/filled/adl*'
+ path_abnormal_vids = root_drive + '/Datasets/{}/Fall/Fall*'.format(dset)
+ path_normal_vids = root_drive + '/Datasets/{}/NonFall/ADL*'.format(dset)
- elif dset == 'SDU':
- path_Fall = root_drive + '/Datasets/SDUFall/Fall/Fall*/Depth'
- path_ADL = root_drive + '/Datasets/SDUFall/NonFall/ADL*/Depth'
-
- elif dset == 'SDU-Filled':
- path_Fall = root_drive + '/Datasets/SDUFall/Fall/Fall*/Filled'
- path_ADL = root_drive + '/Datasets/SDUFall/NonFall/ADL*/Filled'
+ # Update paths if some dataset is arranged in some particular way in folders!
+ if dset == 'Thermal_Intrusion':
+ path_abnormal_vids = root_drive + '/Datasets/{}/Intrusion/Intru*'.format(dset)
+ path_normal_vids = root_drive + '/Datasets/{}/NonIntrusion/NA*'.format(dset)
- print(path_Fall, path_ADL)
+ print(path_normal_vids, path_abnormal_vids)
# glob returns non-ordered list of all pathnames matching a specified pattern
- vid_dir_list_Fall = glob.glob(path_Fall)
- vid_dir_list_ADL = glob.glob(path_ADL)
+ normal_vids_dir = glob.glob(path_normal_vids)
+ abnormal_vids_dir = glob.glob(path_abnormal_vids)
- if len(vid_dir_list_Fall) == 0:
- print('no Fall vids found')
+ if len(abnormal_vids_dir) == 0:
+ print('no Fall/Intru vids found')
- if len(vid_dir_list_ADL) == 0:
- print('no ADL vids found')
+ if len(normal_vids_dir) == 0:
+ print('no ADL/NA vids found')
- return vid_dir_list_ADL, vid_dir_list_Fall
+ return normal_vids_dir, abnormal_vids_dir
-def init_videos(img_width = 64, img_height = 64, raw = False, dset = 'Thermal'):
+def init_videos(img_width = 64, img_height = 64, raw = False, dset = 'Thermal_Fall'):
'''
Creates or overwrites h5py group corresponding to root_path (in body),
for the h5py file located at
- 'N:/FallDetection/Datasets/H5Data/Data_set-{}-imgdim{}x{}.h5'.format(dset, img_width, img_height)
+ 'H5Data/Data_set-{}-imgdim{}x{}.h5'.format(dset, img_width, img_height)
For info on h5py: http://docs.h5py.org/en/stable/quick.html#quick
@@ -89,12 +59,12 @@ def init_videos(img_width = 64, img_height = 64, raw = False, dset = 'Thermal'):
Processed (or Raw)
Split_by_video
- ADL1
+ ADL1 or NA1
Data
<HDF5 dataset "Data": shape (1397, 4096), type "<f8">
Labels
<HDF5 dataset "Labels": shape (1397,), type "<i4">
- ADL2
+ ADL2 or NA2
Data
<HDF5 dataset "Data": shape (3203, 4096), type "<f8">
Labels
@@ -103,13 +73,13 @@ def init_videos(img_width = 64, img_height = 64, raw = False, dset = 'Thermal'):
.
.
- ADL{N}
+ ADL{N} or NA{N}
Data
<HDF5 dataset "Data": shape (3203, 4096), type "<f8">
Labels
<HDF5 dataset "Labels": shape (3203,), type "<i4">
- Fall1
+ Fall1 or Intru1
Data
<HDF5 dataset "Data": shape (49, 4096), type "<f8">
Labels
@@ -117,14 +87,14 @@ def init_videos(img_width = 64, img_height = 64, raw = False, dset = 'Thermal'):
.
.
.
- Fall{M}
+ Fall{M} or Intru{M}
Data
<HDF5 dataset "Data": shape (49, 4096), type "<f8">
Labels
<HDF5 dataset "Labels": shape (49,), type "<i4">
- where N is number of ADL videos, and M is number of Fall videos.
+ where N is number of ADL/NA videos, and M is number of Fall/Intrusion videos.
Params:
bool raw: if true, data will be not processed (mean centering and intensity scaling)
@@ -133,11 +103,11 @@ def init_videos(img_width = 64, img_height = 64, raw = False, dset = 'Thermal'):
str dset: dataset to be loaded
'''
- path = root_drive + '/H5Data/Data_set-{}-imgdim{}x{}.h5'.format( dset, img_width, img_height)
+ path = root_drive + '/H5Data/Data_set-{}-imgdim{}x{}.h5'.format(dset, img_width, img_height)
- vid_dir_list_0, vid_dir_list_1 = get_dir_lists(dset) # Dir of ADL, Fall videos
+ normal_vids_dir, abnormal_vids_dir = get_dir_lists(dset) # Directories of ADL/NA, Fall/Intrusion videos
- if len(vid_dir_list_0) == 0 and len(vid_dir_list_1) == 0:
+ if len(normal_vids_dir) == 0 and len(abnormal_vids_dir) == 0:
print('no videos found, make sure video files are placed in Datasets folder, terminating...')
sys.exit()
@@ -148,19 +118,19 @@ def init_videos(img_width = 64, img_height = 64, raw = False, dset = 'Thermal'):
print('creating data at root_path', root_path)
- def init_videos_helper(root_path): #Nested to keep scope
+ def init_videos_helper(root_path):
with h5py.File(path, 'a') as hf:
root = hf.create_group(root_path)
- for vid_dir in vid_dir_list_1:
- # Create hf group for each Fall video with frames and labels for frame
- init_vid(vid_dir = vid_dir, vid_class = 1, img_width = img_width,
- img_height = img_height, hf = root, raw = raw, dset = dset)
+ for vid_dir in abnormal_vids_dir:
+ # Create hf group for each Fall/Intrusion video with frames and labels for frame
+ init_vid(vid_dir=vid_dir, vid_class=1, img_width=img_width,
+ img_height=img_height, hf=root, raw=raw, dset=dset)
- for vid_dir in vid_dir_list_0:
- # Create hf group for each ADL video with frames and labels for frame
- init_vid(vid_dir = vid_dir, vid_class = 0, img_width = img_width,
- img_height = img_height, hf = root, raw = raw, dset = dset)
+ for vid_dir in normal_vids_dir:
+ # Create hf group for each ADL/NA video with frames and labels for frame
+ init_vid(vid_dir=vid_dir, vid_class=0, img_width=img_width,
+ img_height=img_height, hf=root, raw=raw, dset=dset)
if os.path.isfile(path):
# If .h5 already exists
@@ -169,7 +139,6 @@ def init_videos(img_width = 64, img_height = 64, raw = False, dset = 'Thermal'):
print('video h5py file exists, deleting old group {}, creating new'.format(root_path))
del hf[root_path]
hf.close()
-
init_videos_helper(root_path)
else:
@@ -183,56 +152,47 @@ def init_videos(img_width = 64, img_height = 64, raw = False, dset = 'Thermal'):
init_videos_helper(root_path)
-def init_vid(vid_dir = None, vid_class = None, img_width = 32, img_height = 32,
- hf = None, raw = False, dset = 'Thermal'):
+def init_vid(vid_dir = None, vid_class = None, img_width = 64, img_height = 64,
+ hf = None, raw = False, dset = 'Thermal_Fall'):
'''
Creates hf group with vid_dir name to put Data and Labels for each frame
- Processes all frames in video in ascending order & fetches labels for Fall vids from csv
+ Processes all frames in video in ascending order & fetches labels for Fall/Intrusion vids from csv if asked
Params:
- str vid_dir: path to vid dir of frames to be initialzied
- int vid_class: 1 for Fall, 0 for NonFall
+ str vid_dir: path to vid dir of frames to be initialized
+ int vid_class: 1 for Fall/Intrusion, 0 for NonFall/NonIntrusion
h5py group: group within which new group is nested
-
'''
print('initializing vid at', vid_dir)
#--Get all frames from a video reading frames in ↑ order & processing each frame
- # data.shape = (#frames, img_ht, img_wdth, 1)
- data = create_img_data_set(fpath = vid_dir, ht = img_height, wd = img_width, raw = raw,
- sort = True, dset = dset)
-
- labels = np.zeros(len(data)) # Labels intialized to 0 for each frame
-
- if dset == 'SDU' or dset == 'SDU-Filled':
- vid_dir_name = os.path.basename(os.path.dirname(vid_dir))
- else:
- vid_dir_name = os.path.basename(vid_dir)
-
- print('vid_dir_name', vid_dir_name) # eg: Fall31 (vid folder name)
- grp = hf.create_group(vid_dir_name) # Create folder name group eg. Fall31
-
- # Fetch fall label start end from csv if Fall vid, assign to labels array
- if (vid_dir_name in ['Fall' + str(i) for i in range(201)]) or (vid_dir_name in ['Intru' + str(i) for i in range(250)]): # 201 is max fall index across all vids
- print('setting fall start')
- Fall_start, Fall_stop = get_fall_indeces(vid_dir_name, dset)
- labels[Fall_start-1:Fall_stop] = 1 # Fall_start:Fall_stop + 1
+ data = create_img_data_set(fpath=vid_dir, ht=img_height, wd=img_width, raw=raw, sort=True, dset=dset)
+ labels = np.zeros(len(data)) # Labels initialized to 0 for each frame
+
+ vid_dir_name = os.path.basename(vid_dir)
+ print('vid_dir_name', vid_dir_name) # eg: ADL2 or Fall31 (vid folder name)
+ grp = hf.create_group(vid_dir_name) # Create folder name group eg. ADL2 or Fall31
+
+ # If abnormal video, get fall/intrusion labels for each frame
+ if vid_class == 1:
+ print('setting fall/intrusion start & end')
+ abnormality_start, abnormality_stop = get_abnormal_indices(vid_dir_name, dset)
+ labels[abnormality_start-1:abnormality_stop] = 1
grp['Labels'] = labels
grp['Data'] = data
-def get_fall_indeces(Fall_name, dset):
+def get_abnormal_indices(dir_name, dset):
- # Get Fall start, end indices from Labels.csv
+ # Get Fall/Intrusion start, end indices from Labels.csv
root_dir = './Datasets/'
labels_dir = root_dir + '/{}/Labels.csv'.format(dset)
import pandas as pd
- my_data = pd.read_csv(labels_dir, sep=',', header = 0, index_col = 0)
-
- start, stop = my_data.loc[Fall_name][:2]
+ my_data = pd.read_csv(labels_dir, sep=',', header=0, index_col=0)
+ start, stop = my_data.loc[dir_name][:2]
print('start, stop', start, stop)
return int(start), int(stop)
@@ -240,41 +200,19 @@ def get_fall_indeces(Fall_name, dset):
def sort_frames(frames, dset):
+ # Frames are organized as: FALL_1-0001.jpg, FALL_1-0003.jpg, FALL_1-0002.jpg,...
# Sorts the list frames in the ascending order acc to dset type
- if dset == 'SDU' or dset == 'SDU-Filled': #TODO remove try except, failing to sort shoudl stop!
- print('sorting SDU frames...')
-
- #try:
- frames = sorted(frames, key = lambda x: int(os.path.basename(x).split('.')[0])) #SDU
- # except ValueError:
- # print('failed to sort SDU vid frames')
- # pass
- elif dset == 'UR' or dset == 'UR-Filled' or dset == 'Thermal' or dset == 'Thermal-Intrusion' or dset == 'IDD':
- print('sorting UR or Thermal frames...')
- try:
- frames = sorted(frames, key = lambda x: int(x.split('-')[-1].split('.')[0]))
- except ValueError:
- print('failed to sort UR vid frames')
- return
-
- elif dset == 'TST':
- try:
- frames = sorted(frames, key = lambda x: int(x.split('_')[-1].split('.')[0]))
- except ValueError:
- print('failed to sort vid frames, trying again....')
- pass
-
- elif dset == 'FallFree' or dset == 'FallFree-Filled':
- try:
- frames = sorted(frames, key = lambda x: int(x.split('_')[2]))
- except ValueError:
- print('failed to sort vid frames, trying again....')
- pass
+ print('sorting frames...')
+ try:
+ frames = sorted(frames, key=lambda x: int(x.split('-')[-1].split('.')[0]))
+ except ValueError:
+ print('failed to sort vid frames')
+ return
return frames
-def create_img_data_set(fpath, ht = 64, wd = 64, raw = False, sort = True, dset = 'Thermal'):
+def create_img_data_set(fpath, ht = 64, wd = 64, raw = False, sort = True, dset = 'Thermal_Fall'):
'''
Creates data set of all images located at fpath. Sorts images if asked!
@@ -292,15 +230,12 @@ def create_img_data_set(fpath, ht = 64, wd = 64, raw = False, sort = True, dset
'''
fpath = fpath.replace('\\', '/')
-
# Get all frames inside folder or folders with jpg or png extension
frames = glob.glob(fpath+'/*.jpg') + glob.glob(fpath+'/*.png')
if sort == True:
frames = sort_frames(frames, dset)
- #print("\n".join(frames)) #Use this to check if sorted
-
data = np.zeros((frames.__len__(), ht, wd, 1)) # (frame_length, 64, 64, 1)
for x, i in zip(frames, range(0, frames.__len__())):
@@ -313,16 +248,11 @@ def create_img_data_set(fpath, ht = 64, wd = 64, raw = False, sort = True, dset
if raw == False:
# Image Processing
-
img = img - np.mean(img) # Mean Centering
img = img.astype('float32') / 255. # Rescaling
-
data[i, :, :, :] = img
- # data = data.reshape((len(data), np.prod(data.shape[1:]))) #Flatten the images
-
print('data.shape', data.shape)
-
return data
def init_data_by_class(vid_class = 'NonFall', dset = 'Thermal',\
diff --git a/img_exp.py b/img_exp.py
index fdedd3a..fbf2217 100644
--- a/img_exp.py
+++ b/img_exp.py
@@ -64,11 +64,11 @@ class ImgExp:
'''
Saves the model of the experiment to './Models/self.dset/self.model_name'
'''
- #save_string = self.exp_name #Do this again incase info added to str based on data load(ie after init)
+
if self.hor_flip == True:
self.model_name = self.model_name + '-hor_flip'
- base = './Models/{}/'.format(self.dset)
+ base = './Models/{}'.format(self.dset)
if not os.path.isdir(base):
os.makedirs(base)
diff --git a/models.py b/models.py
index 551a44e..93044a7 100644
--- a/models.py
+++ b/models.py
@@ -11,36 +11,30 @@ from keras import backend as K
"""
Defining Keras models as functions which return model object, aswell as model name and mode type strs.
All models take take img_width and img_height ints, which correpsond to dimensions of images passed to models.
-
"""
+
def DSTCAE_C3D(img_width, img_height, win_length):
input_shape = (win_length, img_width, img_height, 1)
- input_window = Input(shape = input_shape)
+ input_window = Input(shape=input_shape)
temp_pool = 2
x = Conv3D(16, (5, 3, 3), activation='relu', padding='same')(input_window)
x = MaxPooling3D((1, 2, 2), padding='same')(x)
-
x = Conv3D(8, (5, 3, 3), activation='relu', padding='same')(x)
- x = MaxPooling3D((temp_pool, 2, 2), padding='same')(x) #4
-
+ x = MaxPooling3D((temp_pool, 2, 2), padding='same')(x)
x = Dropout(0.25)(x)
-
x = Conv3D(8, (5, 3, 3), activation='relu', padding='same')(x)
- encoded = MaxPooling3D((temp_pool, 2, 2), padding='same')(x) #2
+ encoded = MaxPooling3D((temp_pool, 2, 2), padding='same')(x)
# at this point the representation is (2, 8, 8) i.e. 128-dimensional
x = Conv3D(8, (5, 3, 3), activation='relu', padding='same')(encoded)
- x = UpSampling3D((temp_pool, 2, 2))(x) #4
-
+ x = UpSampling3D((temp_pool, 2, 2))(x)
x = Conv3D(8, (5, 3, 3), activation='relu', padding='same')(x)
- x = UpSampling3D((temp_pool, 2, 2))(x) #8
-
+ x = UpSampling3D((temp_pool, 2, 2))(x)
x = Conv3D(16, (5, 3, 3), activation='relu', padding = 'same')(x)
x = UpSampling3D((1, 2, 2))(x)
-
decoded = Conv3D(1, (5, 3, 3), activation='tanh', padding='same')(x)
autoencoder = Model(input_window, decoded)
@@ -49,7 +43,6 @@ def DSTCAE_C3D(img_width, img_height, win_length):
model_type = 'conv'
model_name = 'DSTCAE_C3D'
model = autoencoder
-
return model, model_name, model_type
@@ -66,23 +59,15 @@ def DSTCAE_UpSampling(img_width, img_height, win_length):
x = Conv3D(16, (temp_depth, 3, 3), activation='relu', padding='same')(input_window)
x = MaxPooling3D((2, 2, 2), padding='same')(x)
-
- # x = Conv3D(8, (temp_depth, 3, 3), activation='relu', padding='same')(x)
- # x = MaxPooling3D((temp_pool, 2, 2), padding='same')(x)
-
x = Dropout(0.25)(x)
-
x = Conv3D(8, (temp_depth, 3, 3), activation='relu', padding='same')(x)
encoded = MaxPooling3D((temp_pool, 2, 2), padding='same')(x)
-
# at this point the representation is (2, 16, 16) i.e. 128-dimensional
x = Conv3D(8, (temp_depth, 3, 3), activation='relu', padding='same')(encoded)
x = UpSampling3D((temp_pool, 2, 2))(x)
-
x = Conv3D(16, (temp_depth, 3, 3), activation='relu', padding='same')(x)
x = UpSampling3D((temp_pool, 2, 2))(x)
-
decoded = Conv3D(1, (temp_depth, 3, 3), activation='tanh', padding='same')(x)
autoencoder = Model(input_window, decoded)
@@ -91,14 +76,12 @@ def DSTCAE_UpSampling(img_width, img_height, win_length):
model_type = 'conv'
model_name = 'DSTCAE_UpSamp'
model = autoencoder
-
return model, model_name, model_type
def DSTCAE_Deconv(img_width, img_height, win_length):
"""
int win_length: Length of window of frames
-
Replace Upsampling with Deconv
"""
@@ -110,16 +93,12 @@ def DSTCAE_Deconv(img_width, img_height, win_length):
x = Conv3D(16, (temp_depth, 3,3), activation='relu', padding='same')(input_window)
x = MaxPooling3D((2, 2, 2), padding='same')(x)
- #x = Conv3D(8, (temp_depth, 3, 3), activation='relu', padding='same')(x)
- #x = MaxPooling3D((temp_pool, 2, 2), padding='same')(x)
x = Dropout(0.25)(x)
-
x = Conv3D(8, (temp_depth, 3, 3), activation='relu', padding='same')(x)
encoded = MaxPooling3D((temp_pool, 2, 2), padding='same')(x)
x = Deconvolution3D(8, (temp_depth, 3, 3), strides = (2,2,2), activation='relu', padding='same')(encoded)
x = Deconvolution3D(16, (temp_depth, 3, 3), strides = (2,2,2), activation='relu', padding='same')(x)
-
decoded = Conv3D(1, (temp_depth, 3, 3), activation='tanh', padding='same')(x)
autoencoder = Model(input_window, decoded)
@@ -128,14 +107,4 @@ def DSTCAE_Deconv(img_width, img_height, win_length):
model_type = 'conv'
model_name = 'DSTCAE_Deconv'
model = autoencoder
-
- return model, model_name, model_type
-
-
-import numpy as np
-if __name__ == "__main__":
- model = dummy_3d(64,64,2)
- print(model.summary())
- # dummy = np.ones((1,8,64,64,1))*255
- # #dummy = dummy- np.mean(dummy)
-
+ return model, model_name, model_type
\ No newline at end of file
diff --git a/seq_exp.py b/seq_exp.py
index d9ecb0c..324710e 100644
--- a/seq_exp.py
+++ b/seq_exp.py
@@ -44,28 +44,27 @@ class SeqExp(ImgExp):
pre_load=None, initial_epoch=0, epochs=1, dset='Thermal',
win_len=8, hor_flip=False, img_width=64, img_height=64):
- ImgExp.__init__(self, model=model, img_width=img_width, \
- img_height=img_height, model_name=model_name, \
- batch_size=batch_size, model_type=model_type, \
- pre_load=pre_load, initial_epoch=initial_epoch, \
+ ImgExp.__init__(self, model=model, img_width=img_width,
+ img_height=img_height, model_name=model_name,
+ batch_size=batch_size, model_type=model_type,
+ pre_load=pre_load, initial_epoch=initial_epoch,
epochs=epochs, hor_flip=hor_flip, dset=dset)
self.win_len = win_len
- def set_train_data(self, raw=False, mmap_mode=None): # TODO init windows from h5py if no npData found
+ def set_train_data(self, raw=False, mmap_mode=None):
'''
loads or initializes windowed train data, and sets self.train_data accordingly
'''
- to_load = root_drive + '/npData/{}/ADL_data-proc-win_{}.npy'.format(self.dset, self.win_len)
+ to_load = root_drive + '/npData/{}/training_data-imgdim_{}x{}-win_{}.npy'.format(self.dset, self.img_width, self.img_height, self.win_len)
if os.path.isfile(to_load):
- print('npData found, loading..')
+ print('Training (npData) found, loading..')
self.train_data = np.load(to_load, mmap_mode=mmap_mode)
else:
- print('npData not found, initializing..')
-
+ print('Training (npData) not found, initializing..')
self.train_data = init_windowed_arr(dset=self.dset, ADL_only=True,
win_len=self.win_len,
img_width=self.img_width,
@@ -95,9 +94,9 @@ class SeqExp(ImgExp):
def train(self, sample_weight=None):
"""
- trains a sequential autoencoder on windowed data(sequences of
- contiguous frames are reconstucted)
- """
+ trains a sequential autoencoder on windowed data(sequences of
+ contiguous frames are reconstructed)
+ """
model_name = self.model_name
base = './Checkpoints/{}'.format(self.dset)
@@ -108,8 +107,7 @@ class SeqExp(ImgExp):
if not os.path.isdir(base_logs):
os.mkdir(base_logs)
- checkpointer = ModelCheckpoint(filepath=base + '/' + model_name + '-' + \
- '{epoch:03d}-{loss:.3f}.hdf5', period=100, verbose=1)
+ checkpointer = ModelCheckpoint(filepath=base + '/' + model_name + '-' + '{epoch:03d}-{loss:.3f}.hdf5', period=100, verbose=1)
timestamp = time.time()
print('./Checkpoints/' + model_name + '-' + '.{epoch:03d}-{loss:.3f}.hdf5')
@@ -118,25 +116,20 @@ class SeqExp(ImgExp):
print(callbacks_list)
print(csv_logger)
- self.model.fit(self.train_data, self.train_data, epochs = self.epochs,
- batch_size = self.batch_size, verbose = 2,
- callbacks = callbacks_list, sample_weight = sample_weight)
+ self.model.fit(self.train_data, self.train_data, epochs=self.epochs,
+ batch_size=self.batch_size, verbose=2,
+ callbacks=callbacks_list, sample_weight=sample_weight)
self.save_exp()
def init_flipped_by_win(self, to_load_flip):
if os.path.isfile(to_load_flip):
data_flip = np.load(to_load_flip)
- data_flip = data_flip.reshape(len(data_flip), self.train_data.shape[1],
- self.train_data.shape[2],
- self.train_data.shape[3], 1)
-
+ data_flip = data_flip.reshape(len(data_flip), self.train_data.shape[1], self.train_data.shape[2], self.train_data.shape[3], 1)
return data_flip
-
else:
print('creating flipped by window data..')
data_flip = flip_windowed_arr(self.train_data)
-
return data_flip
def get_MSE(self, test_data, agg_type='r_sigma'):
diff --git a/util.py b/util.py
index 30a5ec4..033be7f 100644
--- a/util.py
+++ b/util.py
@@ -37,16 +37,9 @@ def threshold(predictions=None, t=0.5):
def get_output(labels, predictions, get_thres=False, to_plot=False, data_option=None, t=0.5, pos_label=1,
- dset='Thermal', model_name='dst', dir_name='None'):
+ dset='Thermal_Fall', model_name='dst', dir_name='None'):
# Calculate AUROC, AUPR and optimal threshold if asked for!
- # predicted_classes = threshold(predictions, t) # not useful t=0.5
- # true_classes = labels
- #
- # # create confusion matrix (2 classes)
- # conf_mat = confusion_matrix(y_true=true_classes, y_pred=predicted_classes)
- # # report = classification_report(true_classes, predicted_classes)
- # g_mean = geometric_mean_score(labels, predicted_classes)
AUROC = []
AUPR = []
@@ -296,9 +289,9 @@ def play_frames(frames, decoded_frames=[], labels=[]):
def generate_test_vid_names(data_dict, dset):
- ##--generates test folder names
+ '''--generates test folder names'''
- if dset == 'Thermal_Fall':
+ if dset == 'Thermal_Fall' or 'Thermal_Dummy':
vid_base_name = 'Fall'
elif dset == 'Thermal_Intrusion':
vid_base_name = 'Intru'
--
GitLab