def read_frames(seq_path_i, as_dict=False):
if (as_dict):
return {
files.clean_str(path_j): cv2.imread(path_j, cv2.IMREAD_GRAYSCALE)
for path_j in files.top_files(seq_path_i)
}
return [
cv2.imread(path_j, cv2.IMREAD_GRAYSCALE)
for path_j in files.top_files(seq_path_i)
]
def read_frame_seqs(in_path, n_split=1):
frame_seqs = FrameSeqs()
for i, path_i in enumerate(files.top_files(in_path)):
name_i = files.Name(path_i.split('/')[-1]).clean()
if (len(name_i) == 0):
name_i = files.Name(str(i))
frames = [
read_frame(path_j, n_split) for path_j in files.top_files(path_i)
]
frame_seqs[name_i] = frames
return frame_seqs
def convert(in_path, out_path):
paths = files.top_files(in_path)
files.make_dir(out_path)
for path_i in paths:
out_i = out_path + '/' + path_i.split('/')[-1]
frames = from_binary(path_i)
to_imgs(frames, out_i)
def simple_exp(dataset):
if (os.path.isdir(dataset)):
dataset = files.top_files(dataset)
result = train_model(dataset)
result.report()
print(result.get_error())
print(result.get_acc())
def binary_feats(seq_path, feats_path=None):
feats_path = get_dirs(seq_path, "feats")
files.make_dir(feats_path)
for i, in_i in enumerate(files.top_files(seq_path)):
print(i)
out_i = feats_path + '/' + in_i.split('/')[-1]
feats.compute_feats(in_i, out_i)
def transform_action_img(in_path, out_path, fun):
files.make_dir(out_path)
for in_i in files.top_files(in_path):
out_i = "%s/%s" % (out_path, in_i.split('/')[-1])
img_i = cv2.imread(in_i, cv2.IMREAD_GRAYSCALE)
new_img_i = fun(img_i)
cv2.imwrite(out_i, new_img_i)
def read_actions(in_path):
actions = ActionImgs()
for path_i in files.top_files(in_path):
name_i = files.Name(path_i.split('/')[-1])
name_i = name_i.clean()
actions[name_i] = cv2.imread(path_i, cv2.IMREAD_GRAYSCALE)
return actions
def read_frames(seq_path_i):
if (seq_path_i.split(".")[-1] == "npy"):
return np.load(seq_path_i)
return [
cv2.imread(frame_path_j, cv2.IMREAD_GRAYSCALE)
for frame_path_j in files.top_files(seq_path_i)
]
def transform_template(fun, in_path, out_path, info=None, dir_ensemble=True):
if (not is_ens(in_path) and dir_ensemble):
return apply_fun(fun, in_path, out_path, info)
files.make_dir(out_path)
for in_i in files.top_files(in_path):
out_i = '%s/%s' % (out_path, in_i.split("/")[-1])
apply_fun(fun, in_i, out_i, info)
def extract_ensemble(img_path, dst_dir=None):
model_path, seq_path = get_dirs(img_path, ['models', 'seqs'], dst_dir)
files.make_dir(seq_path)
for i, model_i in enumerate(files.top_files(model_path)):
seq_i = "%s/nn%d" % (seq_path, i)
model_i = "%s/nn%d" % (model_path, i)
# resnet.extract_feats(seq_path,model_i,feat_i)
sim.frames.extract_feats(img_path, model_i, seq_i)
def ensemble_exp(in_path, out_name, n_epochs=1000, size=64):
input_paths = files.top_files("%s/seqs" % in_path)
out_path = "%s/%s" % (in_path, out_name)
ts_cnn = TS_CNN(dropout=0.5, activ='elu', optim_alg=deep.Nestrov())
train, extract = get_train(ts_cnn)
ensemble = ens.ts_ensemble(train, extract)
arg_dict = {'size': size, 'n_epochs': n_epochs}
ensemble(input_paths, out_path, arg_dict)
def read(in_path):
if (type(in_path) == list):
return [read_unified(in_path)]
if (not os.path.isdir(in_path)):
return [from_dict(read_single(in_path))]
return [
from_dict(read_single(path_i)) for path_i in files.top_files(in_path)
]
def check_model(in_path):
if (os.path.isdir(in_path)):
model_path = in_path + "/frame_models"
paths = files.top_files(model_path)
model = load_model(paths[0])
else:
model = load_model(in_path)
model.summary()
def read_local_feats(in_path):
paths = files.top_files(in_path)
feat_dict = {}
for path_i in paths:
name_i = path_i.split("/")[-1]
name_i = data.clean_name(path_i)
feat_dict[name_i] = load_seq(path_i)
return feat_dict
def ensemble1D(in_path, out_name, n_epochs=1000, size=64, atten=False):
input_paths = files.top_files("%s/seqs" % in_path)
out_path = "%s/%s" % (in_path, out_name)
files.make_dir(out_path)
train, extract = get_train(atten)
ensemble = ens.ts_ensemble(train, extract)
arg_dict = {'size': size, 'n_epochs': n_epochs}
ensemble(input_paths, out_path, arg_dict)
def convert_inert(in_path, out_path):
paths = files.top_files(in_path)
files.make_dir(out_path)
for path_i in paths:
out_i = out_path + '/' + path_i.split('/')[-1]
mat_i = scipy.io.loadmat(path_i)
mat_i = mat_i['d_iner']
np.savetxt(out_i, mat_i, delimiter=',')
def read_seqs(in_path):
seqs = {}
for seq_path_i in files.top_files(in_path):
frames = read_frames(seq_path_i)
name_i = seq_path_i.split('/')[-1]
print(name_i)
seqs[name_i] = frames
return seqs
def read_seqs(in_path):
seqs=Seqs()
for path_i in files.top_files(in_path):
data_i=np.loadtxt(path_i, delimiter=',')
name_i=path_i.split('/')[-1]
name_i=files.Name(name_i).clean()#clean(name_i)
seqs[name_i]=data_i
return seqs
def ensemble_agum(in_path, out_path, n_epochs=1000, size=128):
model = ts_cnn.TS_CNN(dropout=0.5, optim_alg=deep.Nestrov())
train, extract = ts_cnn.get_train(model)
ensemble = ens.ts_ensemble(train, extract, preproc=norm_partial)
input_paths = files.top_files(in_path)
arg_dict = {"n_epochs": n_epochs, "size": size}
files.make_dir(out_path)
ensemble(input_paths, out_path, arg_dict)
def check_feature(in_path):
all_ts = []
for path_i in files.top_files(in_path):
seq_dict_i = data.seqs.read_seqs(path_i)
feat_dict_i = seq_dict_i.to_feats(scipy.stats.kurtosis, single=True)
X, y = feat_dict_i.to_dataset()
all_ts.append(X.flatten())
all_ts = np.array(all_ts).flatten()
print(basic_stats(all_ts))
def img_dataset(in_path):
img_dict={}
for path_i in files.top_files(in_path):
name_i=files.clean_str(path_i.split("/")[-1])
img_i=cv2.imread(path_i, cv2.IMREAD_GRAYSCALE)
img_i=np.expand_dims(img_i,-1)
img_dict[name_i]=img_i
print(img_dict.keys())
return img_dict
def extract_template(fun, frame_path, model_path, out_path):
if (not is_ens(frame_path)):
return fun(frame_path, model_path, out_path)
files.make_dir(out_path)
for in_i in files.top_files(frame_path):
model_i = '%s/%s' % (model_path, in_i.split("/")[-1])
out_i = '%s/%s' % (out_path, in_i.split("/")[-1])
print(in_i)
fun(in_i, model_i, out_i)
def action_img(in_path, out_path, action_fun):
if (type(action_fun) == list):
action_fun = Pipeline(action_fun)
files.make_dir(out_path)
for in_i in files.top_files(in_path):
frames = read_frames(in_i)
action_img_i = action_fun(frames)
out_i = out_path + '/' + in_i.split('/')[-1] + ".png"
cv2.imwrite(out_i, action_img_i)
def read_seqs(in_path):
paths = files.top_files(in_path)
seq_dict = Seqs()
for path_i in paths:
data_i = np.load(path_i)
name_i = path_i.split('/')[-1]
name_i = files.Name(name_i).clean()
seq_dict[name_i] = data_i
return seq_dict
def binary_extract(frame_path,model_path,seq_path):
files.make_dir(seq_path)
paths=model_path if(type(model_path)==list) else files.top_files(model_path)
frame_dict=imgs.read_seqs(frame_path)
for i,in_i in enumerate(paths):
print(i)
out_i= seq_path+'/'+in_i.split('/')[-1]
extract.extract_features(frame_dict,in_i,out_i)
gc.collect()
def proj_template(in_path,out_path,proj_funcs):
files.make_dir(out_path)
paths=files.top_files(in_path)
for seq_path_i in paths:
print(seq_path_i)
seq_i=imgs.read_frames(seq_path_i,as_dict=False)
proj_seq_i=[proj_i(seq_i) for proj_i in proj_funcs]
new_imgs=np.concatenate(proj_seq_i,axis=1)
out_i="%s/%s" % (out_path,seq_path_i.split("/")[-1])
imgs.save_frames(out_i,new_imgs)
def read_seqs(in_path):
seq_dict = {}
for path_i in files.top_files(in_path):
postfix = in_path.split(".")[-1]
if (postfix == "npy"):
seq_i = np.loadtxt(path_i, delimiter=',')
else:
seq_i = np.load(path_i)
seq_dict[path_i.split('/')[-1]] = seq_i
return seq_dict
def __call__(self, common_path, binary_path, info):
lines = []
for binary_i in files.top_files(binary_path):
id_i = binary_i.split('/')[-1]
path_i = "%s/feats" % binary_i
desc_i = self.fun(common_path, path_i)
desc_i = "%s,%s,%s" % (info, id_i, desc_i)
print(desc_i)
lines.append(str(desc_i))
return lines
def make_plot(in_path, out_path):
if (not os.path.isdir(in_path)):
plot_i = tsne_plot(in_path, show=False)
plot_i.savefig(out_path)
else:
files.make_dir(out_path)
for in_i in files.top_files(in_path):
out_i = "%s/%s" % (out_path, in_i.split('/')[-1])
plot_i = tsne_plot(in_i, show=False)
plot_i.savefig(out_i)
def reformat(in_path,out_path,in_dict="rank_rev"):
pattern="%s/%i/image2/"+in_dict
files.make_dir(out_path)
for i in range(1,21):
in_i=pattern % (in_path,i)
for in_ij in files.top_files(in_i):
print(in_ij)
name_ij=in_ij.split('/')[-1]
out_ij="%s/%d_%s" %(out_path,i,name_ij)
print(out_ij)
shutil.copyfile(in_ij, out_ij)
