def sequence_truth(
file_names=['training1', 'training2', 'training3', 'training4']):
df_out = DataFrame()
for file_name in file_names:
df = skeletion_from_archive_cached(file_name)
df = df[['frame', 'gesture', 'sample_id']].dropna().drop_duplicates()
df_out = pd.concat((df_out, df))
return df_out
def agg_movement_intervals(
file_name,
has_labels=False,
train_on=['training1', 'training2', 'training3', 'training4']):
from models import movement_interval
df = skeletion_from_archive_cached(file_name)
df = preprocessed_skeleton(file_name,
demain=True,
keep_only_top_40=False,
train_id=False,
drop_lower_joints=True,
dummy_gesture=False)
df_interval = movement_interval(train_on=train_on)
df_merge = pd.merge(df,
df_interval[['sample_id', 'frame', 'movement']],
on=['frame', 'sample_id'],
how='left')
df_merge.sort(['sample_id', 'frame'], inplace=True)
df_merge.fillna(method='ffill', inplace=True)
df_merge.fillna(method='bfill', inplace=True)
df_merge['interval'] = (df_merge.movement.shift(1) !=
df_merge.movement).astype(int).cumsum()
# delete break frames tagged with gesture
df_merge = df_merge[(df_merge.movement != 0)]
# unify gestures in interval to most frequent one
def smooth_gesture(df):
df['gesture'] = df.gesture.value_counts().index[0]
return df
agg_functions = ['median', 'var', 'min', 'max']
if has_labels:
df_merge = df_merge.groupby(['interval']).apply(smooth_gesture)
# drop intervalls without movement or if labelt as break
df_merge = df_merge[(df_merge['gesture'] != 'break')
& (df_merge['movement'] == 1)]
y_gest = df_merge[['sample_id', 'gesture',
'interval']].drop_duplicates()
y_gest['gesture'] = np.array(
[gesture_to_id[g] for g in y_gest.gesture])
df_merge = df_merge.drop(['movement', 'frame'], axis=1)
df_agg = df_merge.groupby(['sample_id', 'JointType', 'interval'
]).agg(agg_functions).unstack('JointType')
df_agg['_count'] = df_merge.groupby(['sample_id', 'JointType',
'interval']).agg({
'x_p': 'count'
}).unstack('JointType')[('x_p',
'HandRight')]
if has_labels:
return df_agg, y_gest
return df_agg, None
def sequence_truth(file_names=['training1',
'training2',
'training3',
'training4']):
df_out = DataFrame()
for file_name in file_names:
df = skeletion_from_archive_cached(file_name)
df = df[['frame', 'gesture', 'sample_id']].dropna().drop_duplicates()
df_out = pd.concat((df_out, df))
return df_out
def agg_movement_intervals(file_name, has_labels=False,
train_on=['training1','training2', 'training3', 'training4']):
from models import movement_interval
df = skeletion_from_archive_cached(file_name)
df = preprocessed_skeleton(file_name, demain=True, keep_only_top_40=False,
train_id=False, drop_lower_joints=True, dummy_gesture=False)
df_interval = movement_interval(train_on=train_on)
df_merge = pd.merge(df, df_interval[['sample_id', 'frame', 'movement']], on=['frame', 'sample_id'], how='left')
df_merge.sort(['sample_id', 'frame'], inplace=True)
df_merge.fillna(method='ffill', inplace=True)
df_merge.fillna(method='bfill', inplace=True)
df_merge['interval'] = (df_merge.movement.shift(1) != df_merge.movement).astype(int).cumsum()
# delete break frames tagged with gesture
df_merge = df_merge[(df_merge.movement != 0)]
# unify gestures in interval to most frequent one
def smooth_gesture(df):
df['gesture'] = df.gesture.value_counts().index[0]
return df
agg_functions=['median', 'var', 'min', 'max']
if has_labels:
df_merge = df_merge.groupby(['interval']).apply(smooth_gesture)
# drop intervalls without movement or if labelt as break
df_merge = df_merge[(df_merge['gesture'] != 'break') & (df_merge['movement'] == 1)]
y_gest = df_merge[['sample_id', 'gesture', 'interval']].drop_duplicates()
y_gest['gesture'] = np.array([gesture_to_id[g] for g in y_gest.gesture])
df_merge = df_merge.drop(['movement', 'frame'], axis=1)
df_agg = df_merge.groupby(['sample_id', 'JointType', 'interval']).agg(agg_functions).unstack('JointType')
df_agg['_count'] = df_merge.groupby(['sample_id', 'JointType', 'interval']).agg({'x_p': 'count'}).unstack('JointType')[('x_p', 'HandRight')]
if has_labels:
return df_agg, y_gest
return df_agg, None
def preprocessed_skeleton(file_name,
demain=True,
keep_only_top_40=True,
train_id=True,
drop_lower_joints=True,
dummy_gesture=False,
window_shift=1,
window_length=40):
df = skeletion_from_archive_cached(file_name)
if demain:
def demean(arr):
num_cols = [
'w_r', 'x_p', 'x_pix', 'x_r', 'y_p', 'y_pix', 'y_r', 'z_p',
'z_r'
]
arr[num_cols] = arr[num_cols] - arr[num_cols].mean()
return arr #- arr.mean()
df = df.groupby('sample_id').apply(demean)
if train_id:
def add_train_id(df):
df = df.sort('frame')
gestures = iter(df.gesture)
last = gestures.next()
count = 0
gest_i = []
gest_i.append(count)
for gest in gestures:
if gest == last:
gest_i.append(str(count))
last = gest
else:
count += 1
gest_i.append(str(count))
last = gest
df['gesture_nr'] = np.array(gest_i)
return df
df = df.groupby('sample_id').apply(add_train_id)
if keep_only_top_40:
def get_top(arr, n=window_length, column='frame'):
start_frame = arr.frame.min()
return arr[(arr.frame < start_frame + n) |
(arr.gesture == 'break')]
# make sure each gesture appears only once in each sequence!!!
df = df.groupby(df.gesture + df.sample_id + df.gesture_nr,
group_keys=False).apply(get_top)
if drop_lower_joints:
joints_to_drop = [
'KneeLeft', 'AnkleLeft', 'FootLeft', 'HipRight', 'KneeRight',
'AnkleRight', 'FootRight'
]
for joint in joints_to_drop:
df = df[df.JointType != joint]
if dummy_gesture:
def add_dummy_gestures(df,
window_length=window_length,
window_shift=window_shift):
start_ = df.frame.min()
end_ = df.frame.max()
w_start = start_
w_end = w_start + window_length
windows = DataFrame()
while w_end < end_:
df_w = df[(df.frame >= w_start) & (df.frame <= w_end)]
df_w['dummy_gesture'] = w_start
windows = pd.concat([windows, df_w], axis=0)
w_start += window_shift
w_end += window_shift
return windows
df = df.groupby('sample_id',
group_keys=False).apply(add_dummy_gestures)
return df
def preprocessed_skeleton(file_name, demain=True, keep_only_top_40=True,
train_id=True, drop_lower_joints=True, dummy_gesture=False,
window_shift=1, window_length=40):
df = skeletion_from_archive_cached(file_name)
if demain:
def demean(arr):
num_cols = ['w_r', 'x_p', 'x_pix', 'x_r', 'y_p', 'y_pix',
'y_r', 'z_p', 'z_r']
arr[num_cols] = arr[num_cols] - arr[num_cols].mean()
return arr #- arr.mean()
df = df.groupby('sample_id').apply(demean)
if train_id:
def add_train_id(df):
df = df.sort('frame')
gestures = iter(df.gesture)
last = gestures.next()
count = 0
gest_i = []
gest_i.append(count)
for gest in gestures:
if gest == last:
gest_i.append(str(count))
last = gest
else:
count += 1
gest_i.append(str(count))
last = gest
df['gesture_nr'] = np.array(gest_i)
return df
df = df.groupby('sample_id').apply(add_train_id)
if keep_only_top_40:
def get_top(arr, n=window_length, column='frame'):
start_frame = arr.frame.min()
return arr[ (arr.frame < start_frame + n) | (arr.gesture == 'break')]
# make sure each gesture appears only once in each sequence!!!
df = df.groupby(df.gesture + df.sample_id + df.gesture_nr,
group_keys=False).apply(get_top)
if drop_lower_joints:
joints_to_drop = ['KneeLeft', 'AnkleLeft',
'FootLeft', 'HipRight', 'KneeRight', 'AnkleRight', 'FootRight']
for joint in joints_to_drop:
df = df[df.JointType != joint]
if dummy_gesture:
def add_dummy_gestures(df, window_length=window_length,
window_shift=window_shift):
start_ = df.frame.min()
end_ = df.frame.max()
w_start = start_
w_end = w_start + window_length
windows = DataFrame()
while w_end < end_:
df_w = df[(df.frame >= w_start) & (df.frame <= w_end)]
df_w['dummy_gesture'] = w_start
windows = pd.concat([windows, df_w], axis=0)
w_start += window_shift
w_end += window_shift
return windows
df= df.groupby('sample_id', group_keys=False).apply(add_dummy_gestures)
return df