def __test(self):
"""Testing function."""
self.net.eval()
data_list = list()
with torch.no_grad():
for i, data_tuple in enumerate(tqdm(self.data_loader,
desc="Test")):
"""
input, gt
"""
inputs = data_tuple[0].to(self.device)
tmp = "{};".format(len(data_list) + 1)
start = list()
end = list()
prediction_data = list()
s = StateMachine(threshold_window_S1=10, threshold_S3=25)
for idx, window in enumerate(inputs):
output = self.net(window.unsqueeze(0))
predicted = torch.argmax(output.detach(),
dim=-1).cpu().numpy()
prediction_data.append(predicted.squeeze())
retval = s.update(predicted.squeeze(), idx)
if retval:
start.append(s.get_start())
end.append(s.get_end())
pred = np.array(prediction_data)[start[-1]:end[-1] + 1]
pred = pred[pred != CLASS.index("NO_GESTURE")]
pred = np.bincount(pred.flatten()).argmax()
tmp += "{};{};{};".format(CLASS[pred], start[-1] + 2,
end[-1] + 2)
s.reset()
if s.active in [2, 3]:
if (len(inputs) - 1 - s.get_start()) > 3:
start.append(s.get_start())
end.append(len(inputs) - 1)
pred = np.array(prediction_data)[start[-1]:end[-1] + 1]
pred = pred[pred != CLASS.index("NO_GESTURE")]
pred = np.bincount(pred.flatten()).argmax()
tmp += "{};{};{};".format(CLASS[pred], start[-1] + 2,
end[-1] + 2)
s.reset()
tmp += "\n"
data_list.append(tmp)
if not os.path.exists("./outputs"):
os.makedirs("./outputs")
with open(
"./outputs/" + self.configer.get("checkpoints", "save_name") +
".txt", "wt") as outfile:
outfile.write("".join(data_list))
def state_machine(self, predicted, correct, start_end, data_list = None, data_list_gt = None):
s = StateMachine(threshold_window_S1=10, threshold_S3=25)
start_gt = list()
end_gt = list()
for idx, window in enumerate(correct):
retval = s.update(window, idx)
if retval:
start_gt.append(s.get_start())
end_gt.append(s.get_end())
s.reset()
if s.active in [2, 3]:
if (len(correct) - 1 - s.get_start()) > 3:
start_gt.append(s.get_start())
end_gt.append(len(correct) - 1)
s.reset()
start = list()
end = list()
for idx, window in enumerate(predicted):
retval = s.update(window, idx)
if retval:
start.append(s.get_start())
end.append(s.get_end())
s.reset()
if s.active in [2, 3]:
if (len(predicted) - 1 - s.get_start()) > 3:
start.append(s.get_start())
end.append(len(predicted) - 1)
TP = 0
FP = 0
couples = dict()
start_end = start_end[0]
for i in range(len(start)):
s = start[i]
e = end[i]
best_idx = ((np.abs((np.array(start_end[..., 0]) - s)) + np.abs((np.array(start_end[..., 1]) - e)) / 2)).argmin()
if best_idx not in couples:
couples[best_idx] = [i, ((np.abs((np.array(start_end[..., 0]) - s)) + np.abs((np.array(start_end[..., 1]) - e))) / 2).min()]
else:
if ((np.abs((np.array(start_end[..., 0]) - s)) + np.abs((np.array(start_end[..., 1]) - e))) / 2).min() < couples[best_idx][1]:
couples[best_idx] = [i, ((np.abs((np.array(start_end[..., 0]) - s)) + np.abs((np.array(start_end[..., 1]) - e))) / 2).min()]
FP += 1
FN = len(start_gt) - len(couples)
for k, v in couples.items():
self.frame_error_start.append(np.abs(start_end[k, 0] - start[v[0]]))
self.frame_error_end.append(np.abs(start_end[k, 1] - end[v[0]]))
pred = predicted[start[v[0]]:end[v[0]] + 1]
pred = pred[pred != CLASS.index("NO_GESTURE")]
try:
pred = np.bincount(pred.flatten()).argmax()
except ValueError:
print("OK")
gt = correct[start_end[k][0]:start_end[k][1] + 1]
gt = gt[gt != CLASS.index("NO_GESTURE")]
gt = np.bincount(gt.flatten()).argmax()
self.confusion[gt, pred] += 1
if pred == gt:
TP += 1
else:
FP += 1
if data_list is not None:
tmp = "{};".format(len(data_list) + 1)
for start, end in zip(start, end):
pred = predicted[start:end + 1]
pred = pred[pred != CLASS.index("NO_GESTURE")]
pred = np.bincount(pred.flatten()).argmax()
tmp += "{};{};{};".format(CLASS[pred], start+2, end+2)
tmp += "\n"
data_list.append(tmp)
if data_list_gt is not None:
tmp = "{};".format(len(data_list_gt) + 1)
for i, (s, e) in enumerate(zip(start_gt, end_gt)):
gt = correct[s:e + 1]
gt = gt[gt != CLASS.index("NO_GESTURE")]
gt = np.bincount(gt.flatten()).argmax()
tmp += "{};{};{};".format(CLASS[gt], start_end[i, 0], start_end[i, 1])
tmp += "\n"
data_list_gt.append(tmp)
self.AP.append((TP / (TP + FP)) if TP + FP != 0 else 0)
self.AR.append((TP / (TP + FN)) if TP + FN != 0 else 0)
self.FALSE_POS.append(FP)
self.CORR_CLASS.append(TP / len(couples))
self.MISLAB.append(1 - TP / len(couples))
