def train_social_model(out_dir: str, config: ModelConfig) -> None:
# load data
train_data, test_data = provide_train_test(config)
# prepare train data
obs_len_train, pred_len_train = obs_pred_split(config.obs_len,
config.pred_len,
*train_data)
x_obs_len_train, _, grid_obs_len_train, zeros_obs_len_train = obs_len_train
_, y_pred_len_train, _, _ = pred_len_train
# prepare test data
obs_len_test, pred_len_test = obs_pred_split(config.obs_len,
config.pred_len, *test_data)
x_obs_len_test, _, grid_obs_len_test, zeros_obs_len_test = obs_len_test
_, y_pred_len_test, _, _ = pred_len_test
os.makedirs(out_dir, exist_ok=True)
# training
my_model = MySocialModel(config)
print(config.steps_per_epoch)
# history = my_model.train_model.fit_generator(
# train_generator([x_obs_len_train, grid_obs_len_train, zeros_obs_len_train, y_pred_len_train], config.batch_size),
# # batch_size=config.batch_size,
# epochs=config.n_epochs,
# verbose=1,
# steps_per_epoch=config.steps_per_epoch,
# validation_data=(
# [x_obs_len_test, grid_obs_len_test, zeros_obs_len_test],
# y_pred_len_test
# )
# )
history = my_model.train_model.fit(
[x_obs_len_train, grid_obs_len_train, zeros_obs_len_train],
y_pred_len_train,
batch_size=config.batch_size,
epochs=config.n_epochs,
verbose=1,
# steps_per_epoch=config.steps_per_epoch,
validation_data=([
x_obs_len_test, grid_obs_len_test, zeros_obs_len_test
], y_pred_len_test))
# save the trained model weights
weights_file = os.path.join(out_dir,
_make_weights_file_name(config.n_epochs))
my_model.train_model.save_weights(weights_file)
history_file = os.path.join(out_dir, "history.json")
dump_json_file(history.history, history_file)
# save loss plot
plt_save_loss(config, out_dir, history.history)
def train_social_model(out_dir: str, config: ModelConfig) -> None:
# load data
train_data, test_data = provide_train_test(config)
# prepare train data
obs_len_train, pred_len_train = obs_pred_split(config.obs_len,
config.pred_len,
*train_data)
x_obs_len_train, _, grid_obs_len_train, zeros_obs_len_train = obs_len_train
_, y_pred_len_train, _, _ = pred_len_train
# prepare test data
obs_len_test, pred_len_test = obs_pred_split(config.obs_len,
config.pred_len, *test_data)
x_obs_len_test, _, grid_obs_len_test, zeros_obs_len_test = obs_len_test
_, y_pred_len_test, _, _ = pred_len_test
os.makedirs(out_dir, exist_ok=True)
# training
my_model = MySocialModel(config)
history = my_model.train_model.fit(
[x_obs_len_train, grid_obs_len_train, zeros_obs_len_train],
y_pred_len_train,
batch_size=config.batch_size,
epochs=config.n_epochs,
verbose=1,
validation_data=([
x_obs_len_test, grid_obs_len_test, zeros_obs_len_test
], y_pred_len_test))
# save loss plot
plt.plot(history.history["loss"])
plt.plot(history.history["val_loss"])
plt.title("social model loss")
plt.ylabel("loss")
plt.xlabel("epoch")
plt.legend(["train", "test"], loc="upper right")
plt.savefig(
os.path.join(out_dir,
"test={}_loss.png".format(config.test_dataset_kind)))
history_file = os.path.join(out_dir, "history.json")
dump_json_file(history.history, history_file)
# save the trained model weights
weights_file = os.path.join(out_dir,
_make_weights_file_name(config.n_epochs))
my_model.train_model.save_weights(weights_file)
def main() -> None:
args = _load_eval_args()
config = load_model_config(args.trained_model_config)
out_dir = os.path.join(os.path.dirname(args.trained_model_file), "eval")
os.makedirs(out_dir, exist_ok=True)
# load data
_, test_data = provide_train_test(config)
obs_len_test, pred_len_test = obs_pred_split(config.obs_len,
config.pred_len, *test_data)
# load trained model weights
my_model = MySocialModel(config)
my_model.train_model.load_weights(args.trained_model_file)
# first, predict `pred_len` sequences following observation sequence
x_obs_len_test, _, grid_obs_len_test, zeros_obs_len_test = obs_len_test
x_pred_len_test, *_ = pred_len_test
x_pred_len_model = my_model.sample_model.predict(
[x_obs_len_test, grid_obs_len_test, zeros_obs_len_test],
batch_size=config.batch_size,
verbose=1)
# --------------------------------------------------------------------------
# visualization
# --------------------------------------------------------------------------
x_concat_test = np.concatenate([x_obs_len_test, x_pred_len_test], axis=1)
x_concat_model = np.concatenate([x_obs_len_test, x_pred_len_model], axis=1)
# visualize true and predicted trajectories, and save as png files
out_fig_dir = os.path.join(out_dir, "figs")
os.makedirs(out_fig_dir, exist_ok=True)
for s in range(len(x_concat_test)):
fig = visualize_trajectories(x_concat_test[s], x_concat_model[s],
config.obs_len, config.pred_len)
fig_file = os.path.join(out_fig_dir, "{0:04d}.png".format(s))
fig.savefig(fig_file)
plt.close(fig)
# --------------------------------------------------------------------------
# evaluation
# --------------------------------------------------------------------------
# evaluation
ade = compute_abe(x_pred_len_test, x_pred_len_model)
fde = compute_fde(x_pred_len_test, x_pred_len_model)
report = {"ade": float(ade), "fde": float(fde)}
# write to a json file
report_file = os.path.join(out_dir, "report.json")
with open(report_file, "w") as f:
json.dump(report, f)
print("Average displacement error: {}".format(ade))
print("Final displacement error: {}".format(fde))
def train_social_model(out_dir: str, config: ModelConfig) -> None:
# 加载数据
train_data, test_data = provide_train_test(config)
# 训练数据
obs_len_train, pred_len_train = obs_pred_split(config.obs_len,
config.pred_len,
*train_data)
x_obs_len_train, _, grid_obs_len_train, zeros_obs_len_train = obs_len_train
_, y_pred_len_train, _, _ = pred_len_train
# 测试数据
obs_len_test, pred_len_test = obs_pred_split(config.obs_len,
config.pred_len, *test_data)
x_obs_len_test, _, grid_obs_len_test, zeros_obs_len_test = obs_len_test
_, y_pred_len_test, _, _ = pred_len_test
os.makedirs(out_dir, exist_ok=True)
# 训练部分，加载历史运动轨迹，batch_size,训练批次
my_model = MySocialModel(config)
print(config.steps_per_epoch)
history = my_model.train_model.fit(
[x_obs_len_train, grid_obs_len_train, zeros_obs_len_train],
y_pred_len_train,
batch_size=config.batch_size,
epochs=config.n_epochs,
verbose=1,
validation_data=([
x_obs_len_test, grid_obs_len_test, zeros_obs_len_test
], y_pred_len_test))
# 保存训练权重
weights_file = os.path.join(out_dir,
_make_weights_file_name(config.n_epochs))
my_model.train_model.save_weights(weights_file)
history_file = os.path.join(out_dir, "history.json")
dump_json_file(history.history, history_file)
# 保存训练过程loss曲线图
plt_save_loss(config, out_dir, history.history)
def main() -> None:
# 'circle', 'rect
grid_type = 'rect'
is_test = True
args = _load_eval_args()
config = load_model_config(args.trained_model_config)
config.grid_type = grid_type
config.test_dataset_kind = 'students003'
out_dir = os.path.join(os.path.dirname(args.trained_model_file), "eval")
os.makedirs(out_dir, exist_ok=True)
# 加载数据
_, test_data = provide_train_test(config, is_test=is_test)
obs_len_test, pred_len_test = obs_pred_split(
config.obs_len, config.pred_len, *test_data)
# 加载训练权重
my_model = MySocialModel(config)
my_model.train_model.load_weights(args.trained_model_file)
# obs_len代表历史运动序列；`pred_len`代表预测序列
x_obs_len_test, _, grid_obs_len_test, zeros_obs_len_test = obs_len_test
x_pred_len_test, *_ = pred_len_test
x_pred_len_model = my_model.sample_model.predict(
[x_obs_len_test, grid_obs_len_test, zeros_obs_len_test],
batch_size=config.batch_size, verbose=1)
# --------------------------------------------------------------------------
# --------------------------------------------------------------------------
x_concat_test = np.concatenate([x_obs_len_test, x_pred_len_test], axis=1)
x_concat_model = np.concatenate([x_obs_len_test, x_pred_len_model], axis=1)
# 将预测坐标保存至person_n.txt文件
out_fig_dir = os.path.join(out_dir, "figs")
os.makedirs(out_fig_dir, exist_ok=True)
for s in range(len(x_concat_test)):
fig = visualize_trajectories(x_concat_test[s], x_concat_model[s],
config.obs_len, config.pred_len)
fig_file = os.path.join(out_fig_dir, "{0:04d}.person_n.txt".format(s))
fig.savefig(fig_file)
plt.close(fig)
# --------------------------------------------------------------------------
# 评估指标
# --------------------------------------------------------------------------
ade = compute_abe(x_pred_len_test, x_pred_len_model)
fde = compute_fde(x_pred_len_test, x_pred_len_model)
report = {"ade": float(ade), "fde": float(fde)}
# 将定性测试结果保存为一个json文件，包含平均位移误差和最终位移误差
report_file = os.path.join(out_dir, "report.json")
with open(report_file, "w") as f:
json.dump(report, f)
print("Average displacement error: {}".format(ade))
print("Final displacement error: {}".format(fde))