def __getitem__(self, idx):
ops = self.ops
af_dict = self.af_dict
# randomly choose emotion - wrt neutral probability
rndv = th_rand()
if rndv < ops.neutral_ratio:
# neutral
emo = 6
else:
# non neutral, choose again from 0~5
emo = th_randint(0, 6)
# emo = th_randint(0, 7)
# label vec
batch_emo = np.zeros(7)
batch_emo[emo] = 1
# randomly choose image sample and import
rndi = th_randint(len(af_dict[emo]))
sel_dict = af_dict[emo][rndi]
img = imread_to_rgb(ops.friends_path + sel_dict['img'])
batch_img = img
# returns
batch_img = torch.Tensor(batch_img.transpose(2, 0, 1).copy())
batch_emo = torch.Tensor(batch_emo.copy())
return batch_img, batch_emo
def __getitem__(self, idx):
ops = self.ops
af_dict = self.fr_dict
# randomly choose emotion - wrt neutral probability
rndv = th_rand()
if rndv < ops.neutral_ratio:
# neutral
emo = 6
else:
# non neutral, choose again from 0~5
emo = th_randint(0, 6)
# emo = th_randint(0, 7)
# label vec
batch_emo = np.zeros(7)
batch_emo[emo] = 1
# randomly choose image sample and import
rndi = th_randint(len(af_dict[emo]))
sel_dict = af_dict[emo][rndi]
if self.win_size == 1:
img = imread_to_rgb(ops.friends_path + sel_dict['img'][-1])
batch_img = img
batch_img = torch.Tensor(batch_img.transpose(2, 0, 1).copy())
batch_emo = torch.Tensor(batch_emo.copy())
else:
batch_imgs = []
batch_emos = []
for i in range(self.win_size):
img = imread_to_rgb(ops.friends_path + sel_dict['img'][i])
batch_imgs.append(img)
batch_emos.append(batch_emo)
batch_img = torch.Tensor(
np.array(batch_imgs).transpose(0, 3, 1, 2).copy())
batch_emo = torch.Tensor(np.array(batch_emos).copy())
# returns
return batch_img, batch_emo
def read_img_resize(self, imf):
img_orig = imread_to_rgb(imf)
h_orig, w_orig, _ = img_orig.shape
MAX_H, MAX_W = self.cfg.im_size
s_f = float(MAX_H) / float(h_orig)
if float(w_orig) * s_f > MAX_W:
s_f = float(MAX_W) / float(w_orig)
img_mod = cv2.resize(img_orig, (int(w_orig * s_f), int(h_orig * s_f)))
h_mod, w_mod, _ = img_mod.shape
img_zero = np.zeros([MAX_H, MAX_W, 3])
img_zero[:h_mod, :w_mod, :] = img_mod
return img_zero, s_f
def draw_vid_seq(self, seq_res, seq_name):
print '> make video seq...',
# preliminaries
seq_dict = self.track_dict[seq_name]
seq_path = seq_name if not seq_dict.has_key(
'path') else seq_dict['path']
if self.track_dbnm is not 'got10k':
seq_path = os.path.join(seq_path, 'img/')
seq_path = os.path.join(self.track_path, seq_path)
seq_len = len(seq_dict['img'])
# draw for all frames
im_slist = []
for i, imf in enumerate(seq_dict['img']):
# read img
im_frame = imread_to_rgb(os.path.join(seq_path, imf))
# draw bb = [xmin, ymin, width, height]
bb = seq_res[i].astype(int)
im_frame = cv2.rectangle(im_frame, (bb[0], bb[1]),
(bb[0] + bb[2], bb[1] + bb[3]), (1, 0, 0),
3)
# fnum text
fnum_str = str('%04d' % i)
im_frame = cv2.putText(im_frame, fnum_str, (0, im_frame.shape[0]),
cv2.FONT_HERSHEY_DUPLEX,
im_frame.shape[0] / 350., (1, 1, 0))
# save img
im_sname = os.path.join('.temp/',
seq_name + '_' + fnum_str + '.jpg')
im_slist.append(im_sname)
plt.imsave(im_sname, im_frame)
# encode video
vid_clip = mpe.ImageSequenceClip(im_slist, fps=30)
vid_clip.write_videofile('test.mp4', logger=None)
print 'done'
return
def __getitem__(self, idx):
ops = self.ops
if th_rand() < 0.2:
imf, gt, emo = self.get_raf()
else:
imf, gt, emo = self.get_af()
img = imread_to_rgb(imf)
xcen = gt[0] + gt[2]*0.5
ycen = gt[1] + gt[3]*0.5
margin = (gt[2]+gt[3])*ops.aug_marg
wh_mean = np.sqrt((gt[2]+margin)*(gt[2]+margin))
width = wh_mean
height = wh_mean
# data augmentation part ==
aug_prob = ops.aug_prob
aug_crop = ops.aug_crop
# random crops (always)
xmod = th_choice([1,-1])*th_rand_rng(0., width*aug_crop)
ymod = th_choice([1,-1])*th_rand_rng(0., height*aug_crop)
whmod = th_choice([1,-1])*th_rand_rng(0., aug_crop)
wmod = 1.0 + whmod
hmod = 1.0 + whmod
if img.shape[0]==0 or img.shape[1]==0:
print 'b: ',sel_dict['img'], gt
if wmod <= 0 or hmod <= 0:
print 'b: ',wmod,hmod
batch_img = crop_img(img, int(xcen+xmod), int(ycen+ymod), int(width*wmod), int(height*hmod), True, True)
batch_img = cv2.resize(batch_img, (ops.img_sz, ops.img_sz))
# flip horizontally
if th_rand() < 0.5:
batch_img = np.flip(batch_img, axis=1)
# add noise / change level
if th_rand() < aug_prob:
noise_std = th_rand()*0.05
if th_rand() < 0.5:
noise_img = np.repeat( (noise_std)*torch.randn(ops.img_sz, ops.img_sz, 1), 3, 2)
else:
noise_img = (noise_std)*torch.randn(ops.img_sz, ops.img_sz, 3)
batch_img += noise_img.contiguous().numpy()
if th_rand() < aug_prob:
batch_img += (th_rand()-0.5)/50.
# gaussian
if th_rand() < aug_prob:
k_sz = th_randint(1,5,2)*2 +1
batch_img = cv2.GaussianBlur(batch_img, (k_sz[0], k_sz[1]), 0)
# end of data augmentation ==
# reset range [0,1]
if batch_img.min() < 0.:
batch_img[ batch_img < 0. ] = 0.
if batch_img.max() > 1.:
batch_img[ batch_img > 1. ] = 1.
# returns
batch_img = torch.Tensor( batch_img.transpose(2,0,1).copy() )
batch_emo = torch.Tensor( emo.copy() )
#batch_var = torch.Tensor( [sel_dict['val'], sel_dict['aro']] )
return batch_img, batch_emo #, batch_var