def run_model(which_set, classifier, batch_size, model_axes, data_axes=('b', 0, 1, 'c')):
dset = SeqDataset(which_set=which_set)
#indices = dset.get_filtered_indices(perturbations=['0'], flips=[False])
indices = range(dset.numSamples)
#indices = [2,3,4]
#print indices
targets_ints = []
stats = []
fileName = []
for n in indices:
features, targets, fname = dset.get_clip(n)
misclass = []
frame_misclass = []
feature = features.reshape(len(features), 48, 48, 1)
target = None
if targets is not None:
target = basic_7emotion_names.index(targets.lower().replace("angry", "anger"))
#feature = feature / 255.
#feature = feature.astype('float32')
if data_axes != model_axes:
feature = feature.transpose(*[data_axes.index(axis) for axis in model_axes])
num_samples = feature.shape[3]
predictions = []
for i in range(num_samples / batch_size):
# TODO FIX ME, after grayscale
predictions.append(classifier(feature[0,:,:,i*batch_size:(i+1)*batch_size][numpy.newaxis,:,:,:])) #XXX:numpy.newaxis
# for modulo we pad with garbage
if batch_size > num_samples:
modulo = batch_size - num_samples
else:
modulo = num_samples % batch_size
if modulo != 0:
# TODO FIX ME, after grayscale
shape = [1, feature.shape[1], feature.shape[2], modulo]
padding = numpy.ones((shape)).astype('float32')
# TODO FIX ME, after grayscale
feature = numpy.concatenate((feature[0,:,:,(num_samples/batch_size) * batch_size:][numpy.newaxis,:,:,:], padding), axis = 3) #XXX:axis,numpy.newaxis
predictions.append(classifier(feature)[:batch_size - modulo])
targets_ints.append(target)
predictions = numpy.concatenate(predictions, axis=0)
misclass.append(accuracy(predictions, target))
frame_misclass.append(framewise_accuracy(predictions, target))
stats.append(get_stats(predictions, target))
fileName.append(fname)
# error = numpy.sum(misclass) / float(len(features))
# print "clip wise: ", error, 1-error
# frame_misclass = numpy.concatenate(frame_misclass)
# error = frame_misclass.sum() / float(len(frame_misclass))
# print "frame wise: ", error, 1-error
return numpy.vstack(stats), targets_ints, fileName
def __init__(self, emotion, size = (96, 96), num_channels = 1, img_per_seq = 3):
if isinstance(emotion, str):
emotion = basic_7emotion_names.index(emotion.lower())
sets = glob.glob('/data/lisa/data/faces/EmotiW/preproc/static/*/*_' + str(emotion) + '.npy')
sets.sort()
self.images = []
for s in sets:
img = numpy.memmap(s, mode='r', dtype='float32')
img = img.view()
img.shape = (len(img)/(size[0]*size[1]*num_channels), size[0], size[1], num_channels)
self.images.append(img)
self.emotion = emotion
self.ips = img_per_seq
def load_data(which_set, source, preproc=None, size=None, prep=''):
if source == 'original':
dataset = AFEW2ImageSequenceDataset(preload_facetubes=False,
preproc=preproc,
size=size)
train_idx, val_idx = dataset.get_standard_train_test_splits()[0]
if which_set == 'train':
data_idx = train_idx
elif which_set == 'valid':
data_idx = val_idx
else:
raise AssertionError
features = []
clip_ids = []
targets = []
for idx in data_idx:
fts = dataset.get_facetubes(idx)
tgt = basic_7emotion_names.index(dataset.get_label(idx))
for ft in fts:
features.append(ft)
clip_ids.append(idx)
targets.append(tgt)
return features, clip_ids, targets
elif source == 'samira':
path = "/data/lisa/data/faces/EmotiW/preproc/samira/KGL-AFEW/afew2_{}{}.pkl".format(
which_set, prep)
data = serial.load(path)
return data['data_x'], data['clip_ids'], data['data_y']
elif source == 'samira_iso':
path = "/data/lisa/data/faces/EmotiW/preproc/samira/KGLIS-AFEWIS/afew2_{}.pkl".format(
which_set)
data = serial.load(path)
return data['data_x'], data['clip_ids'], data['data_y']
else:
raise ValueError("Unknow source")
def __init__(self, emotion, size = (48, 48), num_channels = 3, img_per_seq = 3):
if isinstance(emotion, str):
self.emotion = basic_7emotion_names.index(emotion)
emotion = emotion[0].upper() + emotion[1:]
else:
self.emotion = emotion
emotion = basic_7emotion_names[emotion]
files = glob.glob('/data/lisa/data/faces/EmotiW/preproc/seq/*/'+emotion+'/*.npy')
files.sort()
self.seq = []
self.lgts = []
for f in files:
seq = numpy.memmap(f, mode='r', dtype='float32')
lgt = (len(seq)/(size[0]*size[1]*num_channels))
seq.shape = (lgt, size[0], size[1], num_channels)
self.seq.append(seq)
self.ips = img_per_seq
self.size = size
self.num_channels = num_channels
def __init__(self, emotion, size=(48, 48), num_channels=3, img_per_seq=3):
if isinstance(emotion, str):
self.emotion = basic_7emotion_names.index(emotion)
emotion = emotion[0].upper() + emotion[1:]
else:
self.emotion = emotion
emotion = basic_7emotion_names[emotion]
files = glob.glob('/data/lisa/data/faces/EmotiW/preproc/seq/*/' +
emotion + '/*.npy')
files.sort()
self.seq = []
self.lgts = []
for f in files:
seq = numpy.memmap(f, mode='r', dtype='float32')
lgt = (len(seq) / (size[0] * size[1] * num_channels))
seq.shape = (lgt, size[0], size[1], num_channels)
self.seq.append(seq)
self.ips = img_per_seq
self.size = size
self.num_channels = num_channels
def load_data(which_set,source, preproc=None, size=None, prep = ''):
if source == 'original':
dataset = AFEW2ImageSequenceDataset(preload_facetubes=False,
preproc=preproc, size=size)
train_idx, val_idx = dataset.get_standard_train_test_splits()[0]
if which_set == 'train':
data_idx = train_idx
elif which_set == 'valid':
data_idx = val_idx
else:
raise AssertionError
features = []
clip_ids = []
targets = []
for idx in data_idx:
fts = dataset.get_facetubes(idx)
tgt = basic_7emotion_names.index(dataset.get_label(idx))
for ft in fts:
features.append(ft)
clip_ids.append(idx)
targets.append(tgt)
return features, clip_ids, targets
elif source == 'samira':
path = "/data/lisa/data/faces/EmotiW/preproc/samira/KGL-AFEW/afew2_{}{}.pkl".format(which_set, prep)
data = serial.load(path)
return data['data_x'], data['clip_ids'], data['data_y']
elif source == 'samira_iso':
path = "/data/lisa/data/faces/EmotiW/preproc/samira/KGLIS-AFEWIS/afew2_{}.pkl".format(which_set)
data = serial.load(path)
return data['data_x'], data['clip_ids'], data['data_y']
else:
raise ValueError("Unknow source")
def __init__(self, which_set, sequence_length = 3, preload_facetubes=True,
batch_size = None, preproc=[], size=(96, 96),
greyscale = False):
if which_set == 'train':
which_set = 'Train'
elif which_set == 'valid':
which_set = 'Val'
if which_set not in ['Train', 'Val']:
raise ValueError(
"Unrecognized value for 'which_set': %s. "
"Valid values are 'Train' and 'Val'." % which_set)
if not preload_facetubes:
raise NotImplementedError(
"For now, we need to preload all facetubes")
dataset = AFEW2ImageSequenceDataset(preload_facetubes=False,
preproc=preproc,
size=size)
self.dataset = dataset
train_idx, val_idx = dataset.get_standard_train_test_splits()[0]
if which_set == 'Train':
data_idx = train_idx
elif which_set == 'Val':
data_idx = val_idx
else:
raise AssertionError
if preload_facetubes:
_features = []
_clip_ids = []
_targets = []
for idx in data_idx:
fts = dataset.get_facetubes(idx)
tgt = basic_7emotion_names.index(dataset.get_label(idx))
for ft in fts:
temp = []
for frame in ft:
if greyscale:
temp.append(cv2.cvtColor(frame, cv.CV_BGR2GRAY))
else:
temp.append(frame)
ft = numpy.array(temp)
_features.append(ft)
_clip_ids.append(idx)
_targets.append(tgt)
features = []
#self.clip_ids = []
targets = []
count = 0
for feat, clip_id, target in zip(_features, _clip_ids, _targets):
# duplicate frames at the end if it's not modulo of sequence_length
modulo = feat.shape[0] % sequence_length
#if modulo != 0:
# TODO return a warning here
#feat = numpy.concatenate((feat, feat[-modulo,:,:,:][None,:,:,:]))
for i in xrange(feat.shape[0] - sequence_length + 1):
features.append(feat[i:i+sequence_length,:,:])
assert len(features[-1]) == sequence_length
count += 1
#self.clip_ids.append(clip_id)
targets.append(target)
self.n_samples = count
feat_shape = features[0].shape
features = numpy.concatenate(features)
features = features.reshape((self.n_samples, sequence_length * numpy.product(feat_shape[1:])))
one_hot = numpy.zeros((self.n_samples, 7), dtype = 'float32')
for i in xrange(self.n_samples):
one_hot[i, targets[i]] = 1.
targets = one_hot
'''
if batch_size is not None and self.n_samples % batch_size != 0:
warnings.warn("since batch size is forced adding some duplicate data, be carefull when comparing results. fixed batch size is needed usually for convolution networks")
self.n_samples = self.n_samples - (self.n_samples % batch_size)
features = features[:self.n_samples]
targets = targets[:self.n_samples]
'''
#view_converter = dense_design_matrix.DefaultViewConverter((sequence_length, 96, 96, 3), axes = ('b', 't', 0, 1, 'c'))
super(AFEW2FaceTubes, self).__init__(X = features, y = targets, axes = ('b', 'c', 0, 1))
def run_model(which_set,
classifier,
batch_size,
model_axes,
data_axes=('b', 0, 1, 'c')):
dset = SeqDataset(which_set=which_set)
#indices = dset.get_filtered_indices(perturbations=['0'], flips=[False])
indices = range(dset.numSamples)
#indices = [2,3,4]
#print indices
targets_ints = []
stats = []
fileName = []
for n in indices:
features, targets, fname = dset.get_clip(n)
misclass = []
frame_misclass = []
feature = features.reshape(len(features), 48, 48, 1)
target = None
if targets is not None:
target = basic_7emotion_names.index(targets.lower().replace(
"angry", "anger"))
#feature = feature / 255.
#feature = feature.astype('float32')
if data_axes != model_axes:
feature = feature.transpose(
*[data_axes.index(axis) for axis in model_axes])
num_samples = feature.shape[3]
predictions = []
for i in range(num_samples / batch_size):
# TODO FIX ME, after grayscale
predictions.append(
classifier(feature[0, :, :,
i * batch_size:(i + 1) * batch_size]
[numpy.newaxis, :, :, :])) #XXX:numpy.newaxis
# for modulo we pad with garbage
if batch_size > num_samples:
modulo = batch_size - num_samples
else:
modulo = num_samples % batch_size
if modulo != 0:
# TODO FIX ME, after grayscale
shape = [1, feature.shape[1], feature.shape[2], modulo]
padding = numpy.ones((shape)).astype('float32')
# TODO FIX ME, after grayscale
feature = numpy.concatenate(
(feature[0, :, :, (num_samples / batch_size) *
batch_size:][numpy.newaxis, :, :, :], padding),
axis=3) #XXX:axis,numpy.newaxis
predictions.append(classifier(feature)[:batch_size - modulo])
targets_ints.append(target)
predictions = numpy.concatenate(predictions, axis=0)
misclass.append(accuracy(predictions, target))
frame_misclass.append(framewise_accuracy(predictions, target))
stats.append(get_stats(predictions, target))
fileName.append(fname)
# error = numpy.sum(misclass) / float(len(features))
# print "clip wise: ", error, 1-error
# frame_misclass = numpy.concatenate(frame_misclass)
# error = frame_misclass.sum() / float(len(frame_misclass))
# print "frame wise: ", error, 1-error
return numpy.vstack(stats), targets_ints, fileName
