def __init__(self,
post_transform=None,
pre_transform=None,
p=0.5,
contrasts=None,
same_transforms=False,
colour_space='grey',
vision_type='trichromat',
mask_image=None,
grey_width=40,
avg_illuminant=0,
train_params=None,
repeat=False,
side_by_side=True):
self.p = p
self.grey_width = grey_width
self.contrasts = contrasts
self.same_transforms = same_transforms
self.colour_space = colour_space
self.vision_type = vision_type
self.mask_image = mask_image
self.post_transform = post_transform
self.pre_transform = pre_transform
self.avg_illuminant = avg_illuminant
if train_params is None:
self.train_params = train_params
else:
self.train_params = path_utils.read_pickle(train_params)
self.img_counter = 0
self.repeat = repeat
self.side_by_side = side_by_side
def match_results_to_input(result_file, geetup_info, model_in_size=(180, 320)):
model_preds = path_utils.read_pickle(result_file)
current_part_res = dict()
for j in range(geetup_info.__len__()):
f_path, f_gt = geetup_info.__getitem__(j)
f_path = f_path[-1]
f_gt = f_gt[-1]
splitted_parts = f_path.replace('//', '/').split('/')
part_folder = splitted_parts[-5]
if part_folder not in current_part_res:
current_part_res[part_folder] = {'1': dict(), '2': dict()}
folder_name = splitted_parts[-2]
image_name = splitted_parts[-1]
if '/segments/1/' in f_path:
seg = '1'
elif '/segments/2/' in f_path:
seg = '2'
else:
sys.exit('Ups unrecognised segment')
pred = model_preds[j]
pred = map_point_to_image_size(pred, (360, 640), model_in_size)
if folder_name not in current_part_res[part_folder][seg]:
current_part_res[part_folder][seg][folder_name] = []
euc_error = euclidean_distance(f_gt, pred)
current_part_res[part_folder][seg][folder_name].append(
[image_name, f_gt, pred, euc_error])
return current_part_res
def clip_visualise(db_path, pred_path, euc_path, out_dir,
video_clips_inds=None):
preds = path_utils.read_pickle(pred_path)
eucs = path_utils.read_pickle(euc_path)
geetup_info = geetup_db.GeetupDatasetInformative(db_path)
if video_clips_inds is None:
video_clips_inds = geetup_utils.get_video_clips_inds(geetup_info)
for clip_inds in video_clips_inds:
beg_ind, end_ind = clip_inds
print(
'Video [%d %d] %.2f' % (
beg_ind, end_ind, np.median(eucs[beg_ind:end_ind])
)
)
draw_circle_clips(preds, geetup_info, beg_ind, end_ind, out_dir)
def process_network(net_name):
all_layers_maxsf = []
for file_path in sorted(
glob.glob(os.path.join(activations_dir, net_name) + '/*.pickle'),
key=natural_keys
):
file_name = ntpath.basename(file_path)
layer_name = 'layer'
name_parts = file_name.split('_')
for pind, part in enumerate(name_parts):
if part == 'layer':
layer_name += name_parts[pind + 1].replace('.weight', '')
layer_name = layer_name.replace('conv', '')
break
png_name = os.path.join(
fig_out_dir, net_name, '%s_activation_%s.png' % (layer_name, 'max')
)
csv_name = os.path.join(
anl_out_dir, net_name, '%s_corrs_0.1.csv' % layer_name
)
print('reading', file_name, layer_name)
result_mat = path_utils.read_pickle(file_path)
contrast_activation, xvals = process_layer(result_mat)
# if not os.path.exists(png_name):
# plot_layer(contrast_activation, xvals, net_name, layer_name)
# if not os.path.exists(csv_name):
# corr_layer(contrast_activation, xvals, net_name, layer_name)
maxsf, header = maxsf_layer(
contrast_activation, xvals, net_name, layer_name
)
all_layers_maxsf.append(maxsf)
out_file = os.path.join(
anl_out_dir, 'peak_activations_avg', '%s_corrs.csv' % net_name
)
np.savetxt(
out_file, np.array(all_layers_maxsf), delimiter=',', header=header
)
return
def get_network_activation(net_name):
all_layers_maxsf = []
all_activations = []
for file_path in sorted(
glob.glob(os.path.join(activations_dir, net_name) + '/*.pickle'),
key=natural_keys
):
file_name = ntpath.basename(file_path)
layer_name = 'layer'
name_parts = file_name.split('_')
for pind, part in enumerate(name_parts):
if part == 'layer':
layer_name += name_parts[pind + 1].replace('.weight', '')
layer_name = layer_name.replace('conv', '')
break
print('reading', file_name, layer_name)
result_mat = path_utils.read_pickle(file_path)
contrast_activation, xvals = process_layer(result_mat)
all_activations.append(contrast_activation)
maxsf, header = maxsf_layer(
contrast_activation, xvals, net_name, layer_name
)
all_layers_maxsf.append(maxsf)
return all_layers_maxsf, all_activations, xvals