def load_data(type=1):
empty_extc = empty_extractor.PGEmptyFeatureExtractor()
empty_data = empty_extc.load_data()
empty_data = empty_data.tolist()
# print('empty_data: ', empty_data)
# 有目标的类别
data_extractor = extractor.FeatureExtractor() # 此处全使用默认的文件路径配置
car_data_list = data_extractor.load_data()
# car_data_list = random.shuffle(car_data_list)
input_data = []
# car_label = []
for item in car_data_list:
input_data.append(item[0]) # 强度信息
random.shuffle(input_data)
car_data = input_data[0:5490]
# 存储起来用来做验证的
pg_car_data = input_data[5490:]
np.save(
'D:\home\zeewei\projects\\77GRadar\model\cnn\\test_data_pg\\input_data_classification_pg_car_data.npy',
pg_car_data)
np.save(
'D:\home\zeewei\projects\\77GRadar\model\cnn\\test_data_pg\\input_data_classification_pg_empty_data.npy',
empty_data[300:])
if type != 1:
random.shuffle(car_data)
car_data = car_data[0:300]
road_extractor = classification_data_extractor.ClassificationExtractor(
)
road_car_data = road_extractor.load_car_data()
random.shuffle(road_car_data)
road_car_data = road_car_data[0:300]
for item in road_car_data:
car_data.append(item)
road_empty_data = road_extractor.load_empty_data()
print('road_empty_data: ', len(road_empty_data))
road_empty_data = road_empty_data
# random.shuffle(road_empty_data)
# empty_data = np.concatenate(empty_data, road_empty_data)
empty_data = empty_data[0:300]
for item in road_empty_data:
empty_data.append(item)
return car_data, empty_data
def load_playground_data():
if os.path.exists(PLAYGROUND_TRAIN_DATA_INPUT) \
and os.path.exists(PLAYGROUND_TRAIN_DATA_LABEL) \
and os.path.exists(PLAYGROUND_TEST_DATA_INPUT) \
and os.path.exists(PLAYGROUND_TEST_DATA_LABEL):
train_data_input = np.load(PLAYGROUND_TRAIN_DATA_INPUT)
train_data_label = np.load(PLAYGROUND_TRAIN_DATA_LABEL)
test_data_input = np.load(PLAYGROUND_TEST_DATA_INPUT)
test_data_label = np.load(PLAYGROUND_TEST_DATA_LABEL)
else:
train_data_path = os.path.join(config.processed_data_dir,
config.train_data_file_name)
if os.path.exists(train_data_path):
data_list = np.load(train_data_path)
else:
data_extractor = extractor.FeatureExtractor(
origin_data_dir=config.origin_train_data_dir,
processed_data_dir=config.processed_data_dir,
input_data_file_name=config.train_data_file_name)
data_list = data_extractor.load_data()
random.shuffle(data_list) # 随机打乱
train_num = int(10 * len(data_list) /
10) # 训练集与测试集7:3比例,此处如果测试集是按帧取数据而不是按录取的数据组划分则这么做
train_data = data_list[0:train_num]
test_data = data_list[train_num:]
train_data_input = []
train_data_label = []
for item in train_data:
train_data_input.append(item[0])
train_data_label.append(item[1])
test_data_input = []
test_data_label = []
for item in test_data:
test_data_input.append(item[0])
test_data_label.append(item[1])
np.save(PLAYGROUND_TRAIN_DATA_INPUT, train_data_input)
np.save(PLAYGROUND_TRAIN_DATA_LABEL, train_data_label)
np.save(PLAYGROUND_TEST_DATA_INPUT, test_data_input)
np.save(PLAYGROUND_TEST_DATA_LABEL, test_data_label)
return train_data_input, train_data_label, test_data_input, test_data_label
def load_val_data():
val_data_path = os.path.join(config.processed_data_dir,
config.val_data_file_name)
if os.path.exists(val_data_path):
val_data = np.load(val_data_path)
else:
data_extractor = extractor.FeatureExtractor(
origin_data_dir=config.origin_val_data_dir,
processed_data_dir=config.processed_data_dir,
input_data_file_name=config.val_data_file_name)
val_data = data_extractor.load_data()
random.shuffle(val_data)
val_data_input = []
val_data_label = []
for item in val_data:
val_data_input.append(item[0])
val_data_label.append(item[1])
return val_data_input, val_data_label
def load_all_data():
pg_data_extractor = extractor.FeatureExtractor() # 此处全使用默认的文件路径配置
data_list = pg_data_extractor.load_data() # 获取操场数据
data_list = data_list.tolist()
random.shuffle(data_list)
data_list = data_list[0:3000]
input_data, label_data, test_data, test_label_data = load_road_data(
) # 获取道路数据
input_data = input_data.tolist()
label_data = label_data.tolist()
for i in range(len(input_data)):
a_group_data = []
a_group_data.append(input_data[i])
a_group_data.append(label_data[i])
data_list.append(a_group_data)
for i in range(len(test_data)):
a_group_data = []
a_group_data.append(test_data[i])
a_group_data.append(test_label_data[i])
data_list.append(a_group_data)
return data_list
def load_pg_data_by_range(start=0, end=64):
'''
start,开始的距离单元
end，结束的距离单元位置
'''
pg_train_data_input_range = os.path.join(
DATA_DIR,
str(start) + '_' + str(end) + '_pg_train_data_input.npy')
pg_train_data_label_range = os.path.join(
DATA_DIR,
str(start) + '_' + str(end) + '_pg_train_data_label.npy')
pg_test_data_input_range = os.path.join(
DATA_DIR,
str(start) + '_' + str(end) + '_pg_test_data_input.npy')
pg_test_data_label_range = os.path.join(
DATA_DIR,
str(start) + '_' + str(end) + '_pg_test_data_label.npy')
if os.path.exists(pg_train_data_input_range) and os.path.exists(
pg_train_data_label_range) and os.path.exists(
pg_test_data_input_range) and os.path.exists(
pg_test_data_label_range):
train_data_input = np.load(pg_train_data_input_range)
train_data_label = np.load(pg_train_data_label_range)
test_data_input = np.load(pg_test_data_input_range)
test_data_label = np.load(pg_test_data_label_range)
return train_data_input, train_data_label, test_data_input, test_data_label
data_extractor = extractor.FeatureExtractor() # 此处全使用默认的文件路径配置
data_list = data_extractor.load_data()
# print('划分数据集')
# data_list = np.load('D:\home\zeewei\projects\\77GRadar\data\\all\\all_train_data.npy')
remain_list = []
'''
对数据进行范围过滤
'''
for item in data_list:
max_index = item[1].argmax(axis=0) # 求每一行的最大下标，即目标的位置
if max_index >= start and max_index <= end:
remain_list.append(item)
data_list = remain_list
random.shuffle(data_list) # 随机打乱
train_num = int(7 * len(data_list) / 10) # 训练集与测试集7:3比例
train_data = data_list[0:train_num]
test_data = data_list[train_num:]
train_data_input = []
train_data_label = []
for item in train_data:
train_data_input.append(item[0])
train_data_label.append(item[1])
test_data_input = []
test_data_label = []
for item in test_data:
test_data_input.append(item[0])
test_data_label.append(item[1])
np.save(pg_train_data_input_range, train_data_input)
np.save(pg_train_data_label_range, train_data_label)
np.save(pg_test_data_input_range, test_data_input)
np.save(pg_test_data_label_range, test_data_label)
return train_data_input, train_data_label, test_data_input, test_data_label
def train():
model = BP_Net().cuda(0)
optimizer = torch.optim.Adam(model.parameters(), lr=5e-5)
# data_extractor = feature_extractor.FeatureExtractor() # 此处全使用默认的文件路径配置
data_extractor = extractor.FeatureExtractor() # 此处全使用默认的文件路径配置
data_list = data_extractor.load_data()
# data_list = load_all_data() # 道路操场数据混合
print('the length of data_list : ', len(data_list))
random.shuffle(data_list)
batch_num = int(7 * len(data_list) / 10)
train_data = data_list[0:batch_num]
test_data = data_list[batch_num:]
train_data_input = []
train_data_label = []
for item in train_data:
train_data_input.append(item[0])
train_data_label.append(item[1])
test_data_input = []
test_data_label = []
for item in test_data:
test_data_input.append(item[0])
test_data_label.append(item[1])
'''将距离分辨率改为6m'''
train_data_input_n = []
train_data_label_n = []
for i in range(len(train_data_label)):
input = decrease_range_resolution(train_data_input[i])
label = decrease_range_resolution(train_data_label[i])
train_data_input_n.append(input)
train_data_label_n.append(label)
test_data_input_n = []
test_data_label_n = []
for i in range(len(test_data_label)):
input = decrease_range_resolution(test_data_input[i])
label = decrease_range_resolution(test_data_label[i])
test_data_input_n.append(input)
test_data_label_n.append(label)
input_data_tensor = torch.FloatTensor(train_data_input_n).cuda(0)
label_data_tensor = torch.FloatTensor(train_data_label_n).cuda(0)
# test_data, test_label_data = data_extractor.load_test_data()
test_data_tensor = torch.FloatTensor(test_data_input_n).cuda(0)
test_label_tensor = torch.FloatTensor(test_data_label_n).cuda(0)
np.save('D:\home\zeewei\projects\\77GRadar\model\\bp\\test_data\\bp3.npy',
test_data_input_n)
np.save('D:\home\zeewei\projects\\77GRadar\model\\bp\\test_data\\bp3.npy',
test_data_label_n)
min_loss = 200
# L = len(input_data)
for i in range(300000):
# loss_sum = 0
prediction = model(input_data_tensor)
loss = loss_fn(prediction, label_data_tensor)
loss_sum = loss.data.cpu().numpy()
loss.backward()
optimizer.step()
test_prediction = model(test_data_tensor)
test_loss = loss_fn(test_prediction, test_label_tensor)
test_loss = test_loss.data.cpu().numpy()
if i % 50 == 0:
if test_loss < min_loss:
min_loss = test_loss
if test_loss < 0.711:
torch.save(model, MODEL_SAVE_DIR + 'bp2_' + str(i) + '.pkl')
print(i, ' train_mean_loss: ', loss_sum, ' test_loss: ', test_loss,
' min_test_loss: ', min_loss)
print('test_min_loss: ', min_loss)
torch.save(model, 'BP2_data_with_empty_loss.pkl')
def train_playground():
model = Net().cuda(0)
optimizer = torch.optim.Adam(model.parameters(), lr=LR)
data_extractor = extractor.FeatureExtractor() # 此处全使用默认的文件路径配置
data_list = data_extractor.load_data()
# data_list = load_all_data() # 道路操场数据混合
# print('the length of data_list : ', len(data_list))
random.shuffle(data_list)
batch_num = int(7 * len(data_list) / 10)
train_data = data_list[0:batch_num]
test_data = data_list[batch_num:]
train_data_input = []
train_data_label = []
for item in train_data:
train_data_input.append(item[0])
train_data_label.append(item[1])
test_data_input = []
test_data_label = []
for item in test_data:
test_data_input.append(item[0])
test_data_label.append(item[1])
np.save('D:\home\zeewei\projects\\77GRadar\model\cnn\\test_data_pg\\input_data_bp2.npy', test_data_input)
np.save('D:\home\zeewei\projects\\77GRadar\model\cnn\\test_data_pg\\label_data_bp2.npy', test_data_label)
# test_batch_num = len(test_data_input)
# test_data_input = np.array(test_data_input).reshape(test_batch_num, 1, 64)
# test_data_label = np.array(test_data_label).reshape(test_batch_num, 1, 64)
test_data_tensor = torch.FloatTensor(test_data_input).cuda(0)
test_label_tensor = torch.FloatTensor(test_data_label).cuda(0)
# train_data_input = np.array(train_data_input).reshape(batch_num, 1, 64)
# train_data_label = np.array(train_data_label).reshape(batch_num, 1, 64)
train_data_tensor = torch.FloatTensor(train_data_input).cuda(0)
train_label_tensor = torch.FloatTensor(train_data_label).cuda(0)
min_loss = 2
for i in range(3000000):
optimizer.zero_grad()
prediction = model(train_data_tensor)
loss = loss_fn(prediction, train_label_tensor)
loss_val = loss.data.cpu().numpy()
loss.backward()
optimizer.step()
test_prediction = model(test_data_tensor)
test_loss = loss_fn(test_prediction, test_label_tensor)
test_loss = test_loss.data.cpu().numpy()
if i % 50 == 0:
if test_loss < min_loss:
min_loss = test_loss
if test_loss < 0.6993:
torch.save(model, MODEL_SAVE_DIR + 'bp2_pg_' + str(i) + '.pkl')
print(i, ' train_mean_loss: ', loss_val,
' test_loss: ', test_loss, 'min_loss: ', min_loss)
print('test_min_loss: ', min_loss)