def transform_screen_data(self, screen):
# Convert to float, rescale, convert to tensor
screen = np.ascontigousarray(screen, dtype=np.float32) / 255
screen = torch.from_numpy(screen)
# Use torchvision package to compose image transforms
resize = T.compose([T.ToPILImage(), T.Resize((40, 90)), T.Totensor()])
return resize(screen).unsqueeze(0).to(
self.device) # add a batch dimension
def image2Tensor(cv2im, resize_im=True):
mean = [0.485, 0.456, 0.406]
std = [0.229, 0.224, 0.225]
if resize_im:
cv2im = cv2.resize(cv2im, (224, 224))
im_as_arr = np.float32(cv2im)
im_as_arr = np.ascontigousarray(im_as_arr[..., ::-1])
im_as_arr = im_as_arr.transpose(2, 0, 1)
for channel, _ in enumerate(im_as_arr):
im_as_arr[channel] /= 255
im_as_arr[channel] -= mean[channel]
im_as_arr[channel] /= std[channel]
im_as_ten = torch.from_numpy(im_as_arr).float()
im_as_ten.unsqueeze_(0)
im_as_var = Variable(im_as_ten, requires_grad=True)
return im_as_var
def read_feat(self,
video_name,
f_init=None,
duration=None,
return_reshaped=True):
"""Read visual encoder features and stack them into memory.
Parameters
----------
video_name : str.
Video identifier.
f_init : int, optional.
Initial frame index. By default the feature is
sliced from frame 1.
duration : int, optional.
Duration in term of number of frames. By default
it is set till the last feature.
return_reshaped : bool.
Return stack of features reshaped after processing.
"""
if not self.fobj:
raise ValueError('The object instance is not open.')
T = self.t_size
s = self.t_delta
if f_init and duration:
frames_of_interest = range(f_init, f_init + duration - T + 1, s)
feat = self.fobj[video_name][self.feat_id][frames_of_interest, :]
elif f_init and (not duration):
feat = self.fobj[video_name][self.feat_id][f_init:-T + 1:s, :]
elif (not f_init) and duration:
feat = self.fobj[video_name][self.feat_id][:duration - T + 1:s, :]
else:
feat = self.fobj[video_name][self.feat_id][:-T + 1:s, :]
pooled_feat = self._feature_pooling(feat)
if not return_reshaped:
feat_dim = feat.shape[1]
pooled_feat = pooled_feat.reshape((-1, feat_dim))
if not pooled_feat.flags['C_CONTIGUOUS']:
return np.ascontigousarray(pooled_feat)
return pooled_feat
def read_feat(self, video_name, f_init=None, duration=None,
return_reshaped=True):
"""Read C3D features and stack them into memory.
Parameters
----------
video-name : str.
Video identifier.
f_init : int, optional.
Initial frame index. By default the feature is
sliced from frame 1.
duration : int, optional.
Duration in term of number of frames. By default
it is set till the last feature.
return_reshaped : bool.
Return stack of features reshaped when pooling is applied.
"""
if not self.fobj:
raise ValueError('The object instance is not open.')
s = self.t_stride
t_size = self.t_size
if f_init and duration:
frames_of_interest = range(f_init,
f_init + duration - t_size + 1, s)
feat = self.fobj[video_name][self.feat_id][frames_of_interest, :]
elif f_init and (not duration):
feat = self.fobj[video_name][self.feat_id][f_init:-t_size+1:s, :]
elif (not f_init) and duration:
feat = self.fobj[video_name][self.feat_id][:duration-t_size+1:s, :]
else:
feat = self.fobj[video_name][self.feat_id][:-t_size+1:s, :]
pooled_feat = self._feature_pooling(feat)
if not return_reshaped:
feat_dim = feat.shape[1]
pooled_feat = pooled_feat.reshape((-1, feat_dim))
if not pooled_feat.flags['C_CONTIGUOUS']:
return np.ascontigousarray(pooled_feat)
return pooled_feat