def subsample(self, train):
all_names = list(train.keys())
if (self.s_name is None):
self.s_name = [
name_i for name_i in all_names
if (np.random.uniform() < self.p)
]
s_train = {name_i: train[name_i] for name_i in self.s_name}
return feats.Feats(s_train)
def cross_acc(datasets):
datasets = [data_i.split()[0] for data_i in datasets]
acc = []
for data_i in datasets:
new_data_i = feats.Feats()
for j, name_j in enumerate(data_i.keys()):
new_name = "%s_%d" % (name_j.split('_')[0], j)
new_data_i[feats.Name(new_name)] = data_i[name_j]
result_i = train.train_model(new_data_i, binary=True)
acc.append(result_i.get_acc())
return np.array(acc)
def to_feats(in_path, out_path):
pairs = read(in_path)
dtw_feats = feats.Feats()
train, test = feats.split(pairs, names_only=True)
# raise Exception(train)
for name_i in pairs.keys():
dtw_feats[name_i] = np.array(
[pairs[name_i][name_j] for name_j in train])
# raise Exception(dtw_feats[name_i].shape)
print(dtw_feats[name_i].shape)
dtw_feats.save(out_path)
def __call__(self, datasets, clf="LR"):
results = []
new_datasets = []
for data_i in datasets:
train_i, test_i = data_i.split()
s_train_i = self.subsample(train_i)
s_data_i = {**s_train_i, **test_i}
s_data_i = feats.Feats(s_data_i)
new_datasets.append(s_data_i)
result_i = learn.train_model(s_data_i, binary=False, clf_type=clf)
results.append(result_i)
return new_datasets, results
def reduce(data_i,n=100):
if( not n or n>data_i.dim()[0]):
return data_i
print("Old dim:" + str(data_i.dim()))
X,y,names=data_i.as_dataset()
train_i=data_i.split()[0]
new_X=recursive(train_i,data_i,n)
new_data_i=feats.Feats()
for j,name_j in enumerate(names):
new_data_i[name_j]=new_X[j]
print("New dim:" + str(new_data_i.dim()))
return new_data_i
def random_split(train):
sss = StratifiedShuffleSplit(n_splits=1, test_size=0.5, random_state=0)
X, y, names = train.as_dataset()
a = sss.split(X, y)
new_dict = feats.Feats()
for train_index, test_index in a:
for i in test_index:
name_i = "%d_0_%d" % (names[i].get_cat(), i)
new_dict[names[i]] = train[names[i]]
for i in train_index:
name_i = "%d_1_%d" % (names[i].get_cat(), i)
new_dict[names[i]] = train[names[i]]
return new_dict
def one_shot(main,add,out_path): dtw_pairs=read(main) names=dtw_pairs.selection(person_id,center_selector) full=dtw_pairs.with_test(names) pairs=[dtw_pairs,read(add)] if(add) else [dtw_pairs] s_feats=feats.Feats() for name_i in full: vectors=[ pairs_i.get_vector(name_i,names) for pairs_i in pairs] # v0=dtw_pairs.get_vector(name_i,names) # v1=add_pairs.get_vector(name_i,names) s_feats[name_i]=np.concatenate(vectors,axis=0) s_feats.save(out_path)
def helper(data_i):
feat_i = feats.Feats()
for name_i, rename_i in rename.items():
print((rename_i, name_i))
feat_i[rename_i] = data_i[name_i]
return feat_i
def compute_stats(in_path,out_path):
seqs=files.get_seqs(in_path)
feat_dict=feats.Feats()
for name_i,seq_i in seqs.items():
feat_dict[name_i]=EBTF(seq_i)
feat_dict.save(out_path)
def to_feats(train,full,pairs,feat_path): dtw_feats=feats.Feats() for name_i in full: dtw_feats[name_i]=pairs.features(name_i,train) dtw_feats.save(feat_path)