def test_model(model):
model.eval()
path_list = []
temp_list = []
pre_dict = {}
save_dict = {}
all_dict = {}
sun_set = roadDatasetInfer(TRAIN_DIR)
test_json = '/data4/mjx/gd/raw_data/amap_traffic_annotations_test.json'
train_json = '/data4/mjx/gd/raw_data/amap_traffic_annotations_train.json'
data_loaders = DataLoader(sun_set,
batch_size=TRAIN_BATCH_SIZE,
shuffle=False,
num_workers=4)
for data in data_loaders:
inputs, paths = data
inputs = inputs.cuda()
output = model(inputs)
output = output.data.cpu().detach().numpy().tolist()
for idx in range(len(output)):
# temp[idx] += output[idx]
# output[idx].insert(0, paths[idx].split('/')[-1])
pre_dict[paths[idx].split('/')[-1]] = output[idx]
# temp_list += output
with open(train_json) as f:
submit = json.load(f)
submit_annos = submit['annotations']
for i in range(len(submit_annos)):
submit_anno = submit_annos[i]
imgId = submit_anno['id']
key_frame = submit_anno['key_frame']
all_res = []
for name in submit_anno['frames']:
try:
name_pre = pre_dict[imgId + '_' + name['frame_name']]
all_res.append(name_pre)
except:
continue
if all_res != []:
all_dict[imgId] = np.array(all_res).mean(axis=0)
try:
save_dict[imgId] = pre_dict[imgId + '_' + key_frame]
except:
continue
df = pd.DataFrame.from_dict(save_dict,
orient='index',
columns=['P1', 'P2', 'P3'])
df = df.reset_index().rename(columns={'index': 'map_id'})
df.to_csv('feature/train_key_feature_seg_w.csv', index=False)
all_df = pd.DataFrame.from_dict(all_dict,
orient='index',
columns=['ave_P1', 'ave_P2', 'ave_P3'])
all_df = all_df.reset_index().rename(columns={'index': 'map_id'})
all_df.to_csv('feature/train_all_feature_seg_w.csv', index=False)
def train_infer(self, model, epoch):
rawLabelDir = '/data4/mjx/gd/raw_data/amap_traffic_annotations_test.json'
image_datasets = roadDatasetInfer(self.test_dir)
dataset_loaders = torch.utils.data.DataLoader(
image_datasets,
batch_size=self.test_batch_size,
shuffle=False,
num_workers=self.num_workers)
model.eval()
pre_result = []
pre_name = []
pre_dict = {}
for data in dataset_loaders:
inputs, paths = data
inputs = inputs.cuda()
outputs = model(inputs)
_, preds = torch.max(outputs.data, 1)
pre_result += preds.cpu().numpy().tolist()
for frame in paths:
pre_name.append(frame.split('/')[-1])
assert len(pre_name) == len(pre_result)
for idx in range(len(pre_result)):
pre_dict[pre_name[idx]] = pre_result[idx]
count_result = {'畅通': 0, '缓行': 0, '拥堵': 0}
with open(rawLabelDir) as f:
submit = json.load(f)
submit_annos = submit['annotations']
submit_result = []
for i in range(len(submit_annos)):
submit_anno = submit_annos[i]
imgId = submit_anno['id']
frame_name = [
imgId + '_' + i['frame_name'] for i in submit_anno['frames']
]
status_all = [pre_dict[i] for i in frame_name]
status = max(status_all, key=status_all.count)
submit['annotations'][i]['status'] = status
submit_json = '{}/{}/{}_{}.json'.format(self.sub_dir, self.model_name,
self.model_name, epoch)
json_data = json.dumps(submit)
with open(submit_json, 'w') as w:
w.write(json_data)
f_class, score, P, R, real_f1 = compare(submit_json, self.real_json)
count_result = count(submit_json)
count_result_real = count(self.real_json)
self.logger.info("{} 第{} epoch 预测结果:{}".format(self.model_name, epoch,
count_result))
self.logger.info("{} 预测结果:{}".format(self.real_json,
count_result_real))
self.logger.info("{} 和 {} 的 f1:{} 加权f1:{} ".format(
self.model_name, self.real_json, f_class, real_f1))
self.logger.info("{} 和 {} 的 Acc:{} Precision:{} Recall:{}".format(
self.model_name, self.real_json, score, P, R))
def test_model(model):
model.eval()
path_list = []
temp_list = []
temp_dict = {}
save_dict = {}
all_dict = {}
sun_set = roadDatasetInfer(TRAIN_DIR)
train_json = '/data4/mjx/GD-B/amap_traffic_final_train_0906.json'
data_loaders = DataLoader(sun_set,
batch_size=TRAIN_BATCH_SIZE,
shuffle=False,
num_workers=4)
for data in data_loaders:
inputs, paths = data
inputs = inputs.cuda()
temp, output = model(inputs)
output = output.data.cpu().detach().numpy().tolist()
temp = temp.cpu().detach().numpy().tolist()
# print(paths[idx].split('/')[-1])
# assert False
for idx in range(len(output)):
temp_dict[paths[idx].split('/')[-1]] = temp[idx]
# temp_list += output
with open(train_json) as f:
submit = json.load(f)
submit_annos = submit['annotations']
for i in range(len(submit_annos)):
submit_anno = submit_annos[i]
imgId = submit_anno['id']
key_frame = submit_anno['key_frame']
all_res = []
for name in submit_anno['frames']:
name_pre = temp_dict[imgId + '_' + name['frame_name']]
all_res.append(name_pre)
all_dict[imgId] = np.array(all_res).mean(axis=0)
save_dict[imgId] = temp_dict[imgId + '_' + key_frame]
all_df = pd.DataFrame.from_dict(
all_dict,
orient='index',
columns=['F{}'.format(i) for i in range(2048)])
all_df = all_df.reset_index().rename(columns={'index': 'map_id'})
all_df.to_csv('to_dw.csv', index=False)
def test_model(model):
model.eval()
path_list = []
temp_list = []
res_temp = []
sun_set = roadDatasetInfer(TRAIN_DIR)
data_loaders = DataLoader(sun_set,
batch_size=TRAIN_BATCH_SIZE,
shuffle=False,
num_workers=4)
for data in data_loaders:
inputs, paths = data
inputs = inputs.cuda()
temp, output = model(inputs)
temp = temp.cpu().detach().numpy().tolist()
output = output.data.cpu().detach().numpy().tolist()
for idx in range(len(temp)):
temp[idx] += output[idx]
temp[idx].insert(0, paths[idx].split('/')[-1])
temp_list += temp
pd.DataFrame(temp_list).to_csv('feature/train_feature_merge.csv',
index=False,
header=None)