def get_model(index, arguments, infer=False):
# return a model given index and arguments
if index == 1:
return SocialModel(arguments, infer)
elif index == 2:
return OLSTMModel(arguments, infer)
elif index == 3:
return VLSTMModel(arguments, infer)
else:
return SocialModel(arguments, infer)
def train(args):
origin = (0,0)
reference_point = (0,1)
validation_dataset_executed = False
prefix = '' # prefix = ''
f_prefix = args.data_dir
# if args.drive is True:
# prefix='drive/semester_project/social_lstm_final/'
# f_prefix = 'drive/semester_project/social_lstm_final'
print('data_dir:', args.data_dir)
# if not os.path.isdir("log/"):
# print("Directory creation script is running...")
# subprocess.call(['make_directories.sh'])
args.freq_validation = np.clip(args.freq_validation, 0, args.num_epochs)
validation_epoch_list = list(range(args.freq_validation, args.num_epochs+1, args.freq_validation))
validation_epoch_list[-1]-=1
# Create the data loader object. This object would preprocess the data in terms of
# batches each of size args.batch_size, of length args.seq_length
dataloader = DataLoader(f_prefix, args.batch_size, args.seq_length, args.num_validation, forcePreProcess=True)
model_name = "LSTM"
method_name = "SOCIALLSTM"
save_tar_name = method_name+"_lstm_model_"
if args.gru:
model_name = "GRU"
save_tar_name = method_name+"_gru_model_"
# Log directory
log_directory = os.path.join(prefix, 'log/')
plot_directory = os.path.join(prefix, 'plot/', method_name, model_name)
plot_train_file_directory = 'validation'
# Logging files
log_file_curve = open(os.path.join(log_directory, method_name, model_name,'log_curve.txt'), 'w+')
log_file = open(os.path.join(log_directory, method_name, model_name, 'val.txt'), 'w+')
# model directory
save_directory = os.path.join(prefix, 'model/')
# Save the arguments int the config file
import json
with open(os.path.join(save_directory, method_name, model_name,'config.pkl'), 'wb') as f:
args_dict = vars(args)
pickle.dump(args, f)
# Path to store the checkpoint file
def checkpoint_path(x):
return os.path.join(save_directory, method_name, model_name, save_tar_name+str(x)+'.tar')
# model creation
net = SocialModel(args)
if args.use_cuda:
net = net.cuda()
#optimizer = torch.optim.RMSprop(net.parameters(), lr=args.learning_rate)
optimizer = torch.optim.Adagrad(net.parameters(), weight_decay=args.lambda_param)
#optimizer = torch.optim.Adam(net.parameters(), weight_decay=args.lambda_param)
learning_rate = args.learning_rate
best_val_loss = 100
best_val_data_loss = 100
smallest_err_val = 100000
smallest_err_val_data = 100000
best_epoch_val = 0
best_epoch_val_data = 0
best_err_epoch_val = 0
best_err_epoch_val_data = 0
all_epoch_results = []
grids = []
num_batch = 0
dataset_pointer_ins_grid = -1
[grids.append([]) for dataset in range(dataloader.get_len_of_dataset())]
# Training
for epoch in range(args.num_epochs):
print('****************Training epoch beginning******************')
if dataloader.additional_validation and (epoch-1) in validation_epoch_list:
dataloader.switch_to_dataset_type(True)
dataloader.reset_batch_pointer(valid=False)
loss_epoch = 0
# For each batch
for batch in range(dataloader.num_batches):
start = time.time()
# Get batch data
x, y, d , numPedsList, PedsList ,target_ids= dataloader.next_batch()
loss_batch = 0
#if we are in a new dataset, zero the counter of batch
if dataset_pointer_ins_grid is not dataloader.dataset_pointer and epoch is not 0:
num_batch = 0
dataset_pointer_ins_grid = dataloader.dataset_pointer
# For each sequence
for sequence in range(dataloader.batch_size):
# Get the data corresponding to the current sequence
x_seq ,_ , d_seq, numPedsList_seq, PedsList_seq = x[sequence], y[sequence], d[sequence], numPedsList[sequence], PedsList[sequence]
target_id = target_ids[sequence]
#get processing file name and then get dimensions of file
folder_name = dataloader.get_directory_name_with_pointer(d_seq)
dataset_data = dataloader.get_dataset_dimension(folder_name)
#dense vector creation
x_seq, lookup_seq = dataloader.convert_proper_array(x_seq, numPedsList_seq, PedsList_seq)
target_id_values = x_seq[0][lookup_seq[target_id], 0:2]
#grid mask calculation and storage depending on grid parameter
if(args.grid):
if(epoch is 0):
grid_seq = getSequenceGridMask(x_seq, dataset_data, PedsList_seq,args.neighborhood_size, args.grid_size, args.use_cuda)
grids[dataloader.dataset_pointer].append(grid_seq)
else:
grid_seq = grids[dataloader.dataset_pointer][(num_batch*dataloader.batch_size)+sequence]
else:
grid_seq = getSequenceGridMask(x_seq, dataset_data, PedsList_seq,args.neighborhood_size, args.grid_size, args.use_cuda)
# vectorize trajectories in sequence
if args.use_cuda:
x_seq = x_seq.cuda()
x_seq, _ = vectorize_seq(x_seq, PedsList_seq, lookup_seq)
# <---------------------- Experimental block ----------------------->
# Main approach:
# 1) Translate all trajectories using first frame value of target trajectory so that target trajectory will start (0,0).
# 2) Get angle between first trajectory point of target ped and (0, 1) for turning.
# 3) Rotate all trajectories in the sequence using this angle.
# 4) Calculate grid mask for hidden layer pooling.
# 5) Vectorize all trajectories (substract first frame values of each trajectories from subsequent points in the trajectory).
#
# Problem:
# Low accuracy
#
# Possible causes:
# *Each function has been already checked -> low possibility.
# *Logic errors or algorithm errors -> high possibility.
# *Wrong order of execution each step -> high possibility.
# <------------------------------------------------------------------------>
# x_seq = translate(x_seq, PedsList_seq, lookup_seq ,target_id_values)
# angle = angle_between(reference_point, (x_seq[1][lookup_seq[target_id], 0].data.numpy(), x_seq[1][lookup_seq[target_id], 1].data.numpy()))
# x_seq = rotate_traj_with_target_ped(x_seq, angle, PedsList_seq, lookup_seq)
# if(args.grid):
# if(epoch is 0):
# grid_seq = getSequenceGridMask(x_seq, dataset_data, PedsList_seq,args.neighborhood_size, args.grid_size, args.use_cuda)
# grids[dataloader.dataset_pointer].append(grid_seq)
# else:
# #grid_seq1 = getSequenceGridMask(x_seq, dataset_data, PedsList_seq,args.neighborhood_size, args.grid_size, args.use_cuda)
# grid_seq = grids[dataloader.dataset_pointer][(num_batch*dataloader.batch_size)+sequence]
# #print([ torch.equal(x.data, y.data) for (x,y) in zip(grid_seq1, grid_seq)])
# #if not (all([ torch.equal(x.data, y.data) for (x,y) in zip(grid_seq1, grid_seq)])):
# # print("not equal")
# # quit()
# else:
# grid_seq = getSequenceGridMask(x_seq, dataset_data, PedsList_seq,args.neighborhood_size, args.grid_size, args.use_cuda)
# x_seq, first_values_dict = vectorize_seq(x_seq, PedsList_seq, lookup_seq)
#print(grid_seq)
# Construct variables
#print("target id : ", target_id)
#print("look up : ", lookup_seq)
#print("pedlist_seq: ", PedsList_seq)
#print("before_xseq: ", x_seq)
#x_seq, target_id_values, first_values_dict = vectorize_seq_with_ped(x_seq, PedsList_seq, lookup_seq ,target_id)
#print("after_vectorize_seq: ", x_seq)
#print("angle: ", np.rad2deg(angle))
#print("after_xseq: ", x_seq)
#x_seq = rotate_traj_with_target_ped(x_seq, -angle, PedsList_seq, lookup_seq)
#x_seq = revert_seq(x_seq, PedsList_seq, lookup_seq, first_values_dict)
#number of peds in this sequence per frame
numNodes = len(lookup_seq)
hidden_states = Variable(torch.zeros(numNodes, args.rnn_size))
if args.use_cuda:
hidden_states = hidden_states.cuda()
cell_states = Variable(torch.zeros(numNodes, args.rnn_size))
if args.use_cuda:
cell_states = cell_states.cuda()
# Zero out gradients
net.zero_grad()
optimizer.zero_grad()
# Forward prop
outputs, _, _ = net(x_seq, grid_seq, hidden_states, cell_states, PedsList_seq,numPedsList_seq ,dataloader, lookup_seq)
# Compute loss
loss = Gaussian2DLikelihood(outputs, x_seq, PedsList_seq, lookup_seq)
loss_batch += loss.item()
# Compute gradients
loss.backward()
# Clip gradients
torch.nn.utils.clip_grad_norm_(net.parameters(), args.grad_clip)
# Update parameters
optimizer.step()
end = time.time()
loss_batch = loss_batch / dataloader.batch_size
loss_epoch += loss_batch
num_batch+=1
print('{}/{} (epoch {}), train_loss = {:.3f}, time/batch = {:.3f}'.format(epoch * dataloader.num_batches + batch,
args.num_epochs * dataloader.num_batches,
epoch,
loss_batch, end - start))
loss_epoch /= dataloader.num_batches
# Log loss values
log_file_curve.write("Training epoch: "+str(epoch)+" loss: "+str(loss_epoch)+'\n')
if dataloader.valid_num_batches > 0:
print('****************Validation epoch beginning******************')
# Validation
dataloader.reset_batch_pointer(valid=True)
loss_epoch = 0
err_epoch = 0
# For each batch
for batch in range(dataloader.valid_num_batches):
# Get batch data
x, y, d , numPedsList, PedsList ,target_ids= dataloader.next_valid_batch()
# Loss for this batch
loss_batch = 0
err_batch = 0
# For each sequence
for sequence in range(dataloader.batch_size):
# Get data corresponding to the current sequence
x_seq ,_ , d_seq, numPedsList_seq, PedsList_seq = x[sequence], y[sequence], d[sequence], numPedsList[sequence], PedsList[sequence]
target_id = target_ids[sequence]
#get processing file name and then get dimensions of file
folder_name = dataloader.get_directory_name_with_pointer(d_seq)
dataset_data = dataloader.get_dataset_dimension(folder_name)
#dense vector creation
x_seq, lookup_seq = dataloader.convert_proper_array(x_seq, numPedsList_seq, PedsList_seq)
target_id_values = x_seq[0][lookup_seq[target_id], 0:2]
#get grid mask
grid_seq = getSequenceGridMask(x_seq, dataset_data, PedsList_seq, args.neighborhood_size, args.grid_size, args.use_cuda)
if args.use_cuda:
x_seq = x_seq.cuda()
x_seq, first_values_dict = vectorize_seq(x_seq, PedsList_seq, lookup_seq)
# <---------------------- Experimental block ----------------------->
# x_seq = translate(x_seq, PedsList_seq, lookup_seq ,target_id_values)
# angle = angle_between(reference_point, (x_seq[1][lookup_seq[target_id], 0].data.numpy(), x_seq[1][lookup_seq[target_id], 1].data.numpy()))
# x_seq = rotate_traj_with_target_ped(x_seq, angle, PedsList_seq, lookup_seq)
# grid_seq = getSequenceGridMask(x_seq, dataset_data, PedsList_seq, args.neighborhood_size, args.grid_size, args.use_cuda)
# x_seq, first_values_dict = vectorize_seq(x_seq, PedsList_seq, lookup_seq)
#number of peds in this sequence per frame
numNodes = len(lookup_seq)
hidden_states = Variable(torch.zeros(numNodes, args.rnn_size))
if args.use_cuda:
hidden_states = hidden_states.cuda()
cell_states = Variable(torch.zeros(numNodes, args.rnn_size))
if args.use_cuda:
cell_states = cell_states.cuda()
# Forward prop
outputs, _, _ = net(x_seq[:-1], grid_seq[:-1], hidden_states, cell_states, PedsList_seq[:-1], numPedsList_seq , dataloader, lookup_seq)
# Compute loss
loss = Gaussian2DLikelihood(outputs, x_seq[1:], PedsList_seq[1:], lookup_seq)
# Extract the mean, std and corr of the bivariate Gaussian
mux, muy, sx, sy, corr = getCoef(outputs)
# Sample from the bivariate Gaussian
next_x, next_y = sample_gaussian_2d(mux.data, muy.data, sx.data, sy.data, corr.data, PedsList_seq[-1], lookup_seq)
next_vals = torch.FloatTensor(1,numNodes,2)
next_vals[:,:,0] = next_x
next_vals[:,:,1] = next_y
err = get_mean_error(next_vals, x_seq[-1].data[None, : ,:], [PedsList_seq[-1]], [PedsList_seq[-1]], args.use_cuda, lookup_seq)
loss_batch += loss.item()
err_batch += err
loss_batch = loss_batch / dataloader.batch_size
err_batch = err_batch / dataloader.batch_size
loss_epoch += loss_batch
err_epoch += err_batch
if dataloader.valid_num_batches != 0:
loss_epoch = loss_epoch / dataloader.valid_num_batches
err_epoch = err_epoch / dataloader.num_batches
# Update best validation loss until now
if loss_epoch < best_val_loss:
best_val_loss = loss_epoch
best_epoch_val = epoch
if err_epoch<smallest_err_val:
smallest_err_val = err_epoch
best_err_epoch_val = epoch
print('(epoch {}), valid_loss = {:.3f}, valid_err = {:.3f}'.format(epoch, loss_epoch, err_epoch))
print('Best epoch', best_epoch_val, 'Best validation loss', best_val_loss, 'Best error epoch',best_err_epoch_val, 'Best error', smallest_err_val)
log_file_curve.write("Validation epoch: "+str(epoch)+" loss: "+str(loss_epoch)+" err: "+str(err_epoch)+'\n')
# Validation dataset
if dataloader.additional_validation and (epoch) in validation_epoch_list:
dataloader.switch_to_dataset_type()
print('****************Validation with dataset epoch beginning******************')
dataloader.reset_batch_pointer(valid=False)
dataset_pointer_ins = dataloader.dataset_pointer
validation_dataset_executed = True
loss_epoch = 0
err_epoch = 0
f_err_epoch = 0
num_of_batch = 0
smallest_err = 100000
#results of one epoch for all validation datasets
epoch_result = []
#results of one validation dataset
results = []
# For each batch
for batch in range(dataloader.num_batches):
# Get batch data
x, y, d , numPedsList, PedsList ,target_ids = dataloader.next_batch()
if dataset_pointer_ins is not dataloader.dataset_pointer:
if dataloader.dataset_pointer is not 0:
print('Finished prosessed file : ', dataloader.get_file_name(-1),' Avarage error : ', err_epoch/num_of_batch)
num_of_batch = 0
epoch_result.append(results)
dataset_pointer_ins = dataloader.dataset_pointer
results = []
# Loss for this batch
loss_batch = 0
err_batch = 0
f_err_batch = 0
# For each sequence
for sequence in range(dataloader.batch_size):
# Get data corresponding to the current sequence
x_seq ,_ , d_seq, numPedsList_seq, PedsList_seq = x[sequence], y[sequence], d[sequence], numPedsList[sequence], PedsList[sequence]
target_id = target_ids[sequence]
#get processing file name and then get dimensions of file
folder_name = dataloader.get_directory_name_with_pointer(d_seq)
dataset_data = dataloader.get_dataset_dimension(folder_name)
#dense vector creation
x_seq, lookup_seq = dataloader.convert_proper_array(x_seq, numPedsList_seq, PedsList_seq)
#will be used for error calculation
orig_x_seq = x_seq.clone()
target_id_values = orig_x_seq[0][lookup_seq[target_id], 0:2]
#grid mask calculation
grid_seq = getSequenceGridMask(x_seq, dataset_data, PedsList_seq, args.neighborhood_size, args.grid_size, args.use_cuda)
#vectorize datapoints
if args.use_cuda:
x_seq = x_seq.cuda()
x_seq, first_values_dict = vectorize_seq(x_seq, PedsList_seq, lookup_seq)
# <---------------------- Experimental block ----------------------->
# x_seq = translate(x_seq, PedsList_seq, lookup_seq ,target_id_values)
# angle = angle_between(reference_point, (x_seq[1][lookup_seq[target_id], 0].data.numpy(), x_seq[1][lookup_seq[target_id], 1].data.numpy()))
# x_seq = rotate_traj_with_target_ped(x_seq, angle, PedsList_seq, lookup_seq)
# grid_seq = getSequenceGridMask(x_seq, dataset_data, PedsList_seq, args.neighborhood_size, args.grid_size, args.use_cuda)
# x_seq, first_values_dict = vectorize_seq(x_seq, PedsList_seq, lookup_seq)
if args.use_cuda:
x_seq = x_seq.cuda()
#sample predicted points from model
ret_x_seq, loss = sample_validation_data(x_seq, PedsList_seq, grid_seq, args, net, lookup_seq, numPedsList_seq, dataloader)
#revert the points back to original space
ret_x_seq = revert_seq(ret_x_seq, PedsList_seq, lookup_seq, first_values_dict)
# <---------------------- Experimental block revert----------------------->
# Revert the calculated coordinates back to original space:
# 1) Convert point from vectors to absolute coordinates
# 2) Rotate all trajectories in reverse angle
# 3) Translate all trajectories back to original space by adding the first frame value of target ped trajectory
# *It works without problems which mean that it reverts a trajectory back completely
# Possible problems:
# *Algoritmical errors caused by first experimental block -> High possiblity
# <------------------------------------------------------------------------>
# ret_x_seq = revert_seq(ret_x_seq, PedsList_seq, lookup_seq, first_values_dict)
# ret_x_seq = rotate_traj_with_target_ped(ret_x_seq, -angle, PedsList_seq, lookup_seq)
# ret_x_seq = translate(ret_x_seq, PedsList_seq, lookup_seq ,-target_id_values)
#get mean and final error
err = get_mean_error(ret_x_seq.data, orig_x_seq.data, PedsList_seq, PedsList_seq, args.use_cuda, lookup_seq)
f_err = get_final_error(ret_x_seq.data, orig_x_seq.data, PedsList_seq, PedsList_seq, args.use_cuda, lookup_seq)
loss_batch += loss.item()
err_batch += err
f_err_batch += f_err
print('Current file : ', dataloader.get_file_name(0),' Batch : ', batch+1, ' Sequence: ', sequence+1, ' Sequence mean error: ', err,' Sequence final error: ',f_err,' time: ', end - start)
results.append((orig_x_seq.data.cpu().numpy(), ret_x_seq.data.cpu().numpy(), PedsList_seq, lookup_seq, dataloader.get_frame_sequence(args.seq_length), target_id))
loss_batch = loss_batch / dataloader.batch_size
err_batch = err_batch / dataloader.batch_size
f_err_batch = f_err_batch / dataloader.batch_size
num_of_batch += 1
loss_epoch += loss_batch
err_epoch += err_batch
f_err_epoch += f_err_batch
epoch_result.append(results)
all_epoch_results.append(epoch_result)
if dataloader.num_batches != 0:
loss_epoch = loss_epoch / dataloader.num_batches
err_epoch = err_epoch / dataloader.num_batches
f_err_epoch = f_err_epoch / dataloader.num_batches
avarage_err = (err_epoch + f_err_epoch)/2
# Update best validation loss until now
if loss_epoch < best_val_data_loss:
best_val_data_loss = loss_epoch
best_epoch_val_data = epoch
if avarage_err<smallest_err_val_data:
smallest_err_val_data = avarage_err
best_err_epoch_val_data = epoch
print('(epoch {}), valid_loss = {:.3f}, valid_mean_err = {:.3f}, valid_final_err = {:.3f}'.format(epoch, loss_epoch, err_epoch, f_err_epoch))
print('Best epoch', best_epoch_val_data, 'Best validation loss', best_val_data_loss, 'Best error epoch',best_err_epoch_val_data, 'Best error', smallest_err_val_data)
log_file_curve.write("Validation dataset epoch: "+str(epoch)+" loss: "+str(loss_epoch)+" mean_err: "+str(err_epoch)+'final_err: '+str(f_err_epoch)+'\n')
optimizer = time_lr_scheduler(optimizer, epoch, lr_decay_epoch = args.freq_optimizer)
# Save the model after each epoch
print('Saving model')
torch.save({
'epoch': epoch,
'state_dict': net.state_dict(),
'optimizer_state_dict': optimizer.state_dict()
}, checkpoint_path(epoch))
if dataloader.valid_num_batches != 0:
print('Best epoch', best_epoch_val, 'Best validation Loss', best_val_loss, 'Best error epoch',best_err_epoch_val, 'Best error', smallest_err_val)
# Log the best epoch and best validation loss
log_file.write('Validation Best epoch:'+str(best_epoch_val)+','+' Best validation Loss: '+str(best_val_loss))
if dataloader.additional_validation:
print('Best epoch acording to validation dataset', best_epoch_val_data, 'Best validation Loss', best_val_data_loss, 'Best error epoch',best_err_epoch_val_data, 'Best error', smallest_err_val_data)
log_file.write("Validation dataset Best epoch: "+str(best_epoch_val_data)+','+' Best validation Loss: '+str(best_val_data_loss)+'\n')
#dataloader.write_to_plot_file(all_epoch_results[best_epoch_val_data], plot_directory)
#elif dataloader.valid_num_batches != 0:
# dataloader.write_to_plot_file(all_epoch_results[best_epoch_val], plot_directory)
#else:
if validation_dataset_executed:
dataloader.switch_to_dataset_type(load_data=False)
create_directories(plot_directory, [plot_train_file_directory])
dataloader.write_to_plot_file(all_epoch_results[len(all_epoch_results)-1], os.path.join(plot_directory, plot_train_file_directory))
# Close logging files
log_file.close()
log_file_curve.close()
def main():
# Parse the arguments received from command line
parser = argparse.ArgumentParser(description="Train a social LSTM")
parser.add_argument(
"modelParams",
type=str,
help=
"Path to the file or folder with the parameters of the experiments",
)
parser.add_argument(
"-l",
"--logLevel",
help="logging level of the logger. Default is INFO",
metavar="level",
type=str,
)
parser.add_argument(
"-f",
"--logFolder",
help=
"path to the folder where to save the logs. If None, logs are only printed in stderr",
type=str,
metavar="path",
)
args = parser.parse_args()
if os.path.isdir(args.modelParams):
names_experiments = os.listdir(args.modelParams)
experiments = [
os.path.join(args.modelParams, experiment)
for experiment in names_experiments
]
else:
experiments = [args.modelParams]
for experiment in experiments:
# Load the parameters
hparams = utils.YParams(experiment)
# Define the logger
setLogger(hparams, args, PHASE)
remainSpaces = 29 - len(hparams.name)
logging.info(
"\n" +
"--------------------------------------------------------------------------------\n"
+ "| Training experiment: " +
hparams.name + " " * remainSpaces + "|\n" +
"--------------------------------------------------------------------------------\n"
)
trajectory_size = hparams.obsLen + hparams.predLen
logging.info("Loading the training datasets...")
train_loader = utils.DataLoader(
hparams.dataPath,
hparams.trainDatasets,
hparams.trainMaps,
hparams.semanticMaps,
hparams.trainMapping,
hparams.homography,
num_labels=hparams.numLabels,
delimiter=hparams.delimiter,
skip=hparams.skip,
max_num_ped=hparams.maxNumPed,
trajectory_size=trajectory_size,
neighborood_size=hparams.neighborhoodSize,
)
logging.info("Loading the validation datasets...")
val_loader = utils.DataLoader(
hparams.dataPath,
hparams.validationDatasets,
hparams.validationMaps,
hparams.semanticMaps,
hparams.validationMapping,
hparams.homography,
num_labels=hparams.numLabels,
delimiter=hparams.delimiter,
skip=hparams.skip,
max_num_ped=hparams.maxNumPed,
trajectory_size=trajectory_size,
neighborood_size=hparams.neighborhoodSize,
)
logging.info(
"Creating the training and validation dataset pipeline...")
dataset = utils.TrajectoriesDataset(
train_loader,
val_loader=val_loader,
batch=False,
shuffle=hparams.shuffle,
prefetch_size=hparams.prefetchSize,
)
hparams.add_hparam("learningRateSteps", train_loader.num_sequences)
logging.info("Creating the model...")
start = time.time()
model = SocialModel(dataset, hparams, phase=PHASE)
end = time.time() - start
logging.debug("Model created in {:.2f}s".format(end))
# Define the path to where save the model and the checkpoints
if hparams.modelFolder:
save_model = True
model_folder = os.path.join(hparams.modelFolder, hparams.name)
if not os.path.exists(model_folder):
os.makedirs(model_folder)
os.makedirs(os.path.join(model_folder, "checkpoints"))
model_path = os.path.join(model_folder, hparams.name)
checkpoints_path = os.path.join(model_folder, "checkpoints",
hparams.name)
# Create the saver
saver = tf.train.Saver()
# Zero padding
padding = len(str(train_loader.num_sequences))
# ============================ START TRAINING ============================
with tf.Session() as sess:
logging.info(
"\n" +
"--------------------------------------------------------------------------------\n"
+
"| Start training |\n"
+
"--------------------------------------------------------------------------------\n"
)
# Initialize all the variables in the graph
sess.run(tf.global_variables_initializer())
for epoch in range(hparams.epochs):
logging.info("Starting epoch {}".format(epoch + 1))
# ==================== TRAINING PHASE ====================
# Initialize the iterator of the training dataset
sess.run(dataset.init_train)
for sequence in range(train_loader.num_sequences):
start = time.time()
loss, _ = sess.run([model.loss, model.train_optimizer])
end = time.time() - start
logging.info(
"{:{width}d}/{} epoch: {} time/Batch = {:.2f}s. Loss = {:.4f}"
.format(
sequence + 1,
train_loader.num_sequences,
epoch + 1,
end,
loss,
width=padding,
))
# ==================== VALIDATION PHASE ====================
logging.info(" ========== Validation ==========")
# Initialize the iterator of the validation dataset
sess.run(dataset.init_val)
loss_val = 0
for _ in range(val_loader.num_sequences):
loss = sess.run(model.loss)
loss_val += loss
mean_val = loss_val / val_loader.num_sequences
logging.info("Epoch: {}. Validation loss = {:.4f}".format(
epoch + 1, mean_val))
# Save the model
if save_model:
logging.info("Saving model...")
saver.save(
sess,
checkpoints_path,
global_step=epoch + 1,
write_meta_graph=False,
)
logging.info("Model saved...")
# Save the final model
if save_model:
saver.save(sess, model_path)
tf.reset_default_graph()
def train(args):
origin = (0, 0)
reference_point = (0, 1)
validation_dataset_executed = False
# 这一片是否可以去掉
prefix = ''
f_prefix = '.'
if args.drive is True:
prefix = 'drive/semester_project/social_lstm_final/'
f_prefix = 'drive/semester_project/social_lstm_final'
# 用于从云端读取数据
if not os.path.isdir("log/"):
print("Directory creation script is running...")
subprocess.call([f_prefix + '/make_directories.sh'])
args.freq_validation = np.clip(args.freq_validation, 0, args.num_epochs)
validation_epoch_list = list(
range(args.freq_validation, args.num_epochs + 1, args.freq_validation))
validation_epoch_list[-1] -= 1
# Create the data loader object. This object would preprocess the data in terms of
# batches each of size args.batch_size, of length args.seq_length
# 读取数据
dataloader = DataLoader(f_prefix,
args.batch_size,
args.seq_length,
args.num_validation,
forcePreProcess=True)
model_name = "LSTM"
method_name = "SOCIALLSTM"
save_tar_name = method_name + "_lstm_model_"
if args.gru:
model_name = "GRU"
save_tar_name = method_name + "_gru_model_"
# Log directory
log_directory = os.path.join(prefix, 'log/')
plot_directory = os.path.join(prefix, 'plot/', method_name, model_name)
# print(plot_directory)
plot_train_file_directory = 'validation'
# Logging files
log_file_curve = open(
os.path.join(log_directory, method_name, model_name, 'log_curve.txt'),
'w+')
log_file = open(
os.path.join(log_directory, method_name, model_name, 'val.txt'), 'w+')
# model directory
save_directory = os.path.join(prefix, 'model/')
# Save the arguments int the config file
# 将参数保存在配置文件中
with open(
os.path.join(save_directory, method_name, model_name,
'config.pkl'), 'wb') as f:
pickle.dump(args, f)
# Path to store the checkpoint file
# 存储检查点文件的路径
def checkpoint_path(x):
return os.path.join(save_directory, method_name, model_name,
save_tar_name + str(x) + '.tar')
# model creation
net = SocialModel(args)
if args.use_cuda:
net = net.cuda()
optimizer = torch.optim.Adagrad(net.parameters(),
weight_decay=args.lambda_param)
learning_rate = args.learning_rate
best_val_loss = 100
best_val_data_loss = 100
smallest_err_val = 100000
smallest_err_val_data = 100000
best_epoch_val = 0
best_epoch_val_data = 0
best_err_epoch_val = 0
best_err_epoch_val_data = 0
all_epoch_results = []
grids = []
num_batch = 0
dataset_pointer_ins_grid = -1
[grids.append([]) for dataset in range(dataloader.get_len_of_dataset())]
# Training
for epoch in range(args.num_epochs):
print('****************Training epoch beginning******************')
if dataloader.additional_validation and (epoch -
1) in validation_epoch_list:
dataloader.switch_to_dataset_type(True)
dataloader.reset_batch_pointer(valid=False)
loss_epoch = 0
# For each batch
for batch in range(dataloader.num_batches):
start = time.time()
# Get batch data
x, y, d, numPedsList, PedsList, target_ids = dataloader.next_batch(
)
loss_batch = 0
# 如果我们在新数据集中，则将批处理计数器清零
if dataset_pointer_ins_grid is not dataloader.dataset_pointer and epoch is not 0:
num_batch = 0
dataset_pointer_ins_grid = dataloader.dataset_pointer
# For each sequence
for sequence in range(dataloader.batch_size):
# 获取与当前序列相对应的数据
x_seq, _, d_seq, numPedsList_seq, PedsList_seq = x[
sequence], y[sequence], d[sequence], numPedsList[
sequence], PedsList[sequence]
target_id = target_ids[sequence]
# 获取处理文件名，然后获取文件尺寸
folder_name = dataloader.get_directory_name_with_pointer(d_seq)
dataset_data = dataloader.get_dataset_dimension(folder_name)
# dense vector creation
# 密集矢量创建
x_seq, lookup_seq = dataloader.convert_proper_array(
x_seq, numPedsList_seq, PedsList_seq)
target_id_values = x_seq[0][lookup_seq[target_id], 0:2]
# grid mask calculation and storage depending on grid parameter
# 网格掩码的计算和存储取决于网格参数
# 应该是用于判断是否有social性
if (args.grid):
if (epoch is 0):
grid_seq = getSequenceGridMask(x_seq, dataset_data,
PedsList_seq,
args.neighborhood_size,
args.grid_size,
args.use_cuda)
grids[dataloader.dataset_pointer].append(grid_seq)
else:
temp = (num_batch * dataloader.batch_size) + sequence
if temp > 128:
temp = 128
grid_seq = grids[dataloader.dataset_pointer][temp]
else:
grid_seq = getSequenceGridMask(x_seq, dataset_data,
PedsList_seq,
args.neighborhood_size,
args.grid_size,
args.use_cuda)
# 按顺序矢量化轨迹
x_seq, _ = vectorize_seq(x_seq, PedsList_seq, lookup_seq)
if args.use_cuda:
x_seq = x_seq.cuda()
# number of peds in this sequence per frame
# 每帧此序列中的ped数
numNodes = len(lookup_seq)
hidden_states = Variable(torch.zeros(numNodes, args.rnn_size))
if args.use_cuda:
hidden_states = hidden_states.cuda()
cell_states = Variable(torch.zeros(numNodes, args.rnn_size))
if args.use_cuda:
cell_states = cell_states.cuda()
# 零梯度
net.zero_grad()
optimizer.zero_grad()
# Forward prop
outputs, _, _ = net(x_seq, grid_seq, hidden_states,
cell_states, PedsList_seq, numPedsList_seq,
dataloader, lookup_seq)
# 计算损失loss
loss = Gaussian2DLikelihood(outputs, x_seq, PedsList_seq,
lookup_seq)
loss_batch += loss.item()
# 计算梯度
loss.backward()
# 裁剪梯度
torch.nn.utils.clip_grad_norm_(net.parameters(),
args.grad_clip)
# 更新梯度
optimizer.step()
end = time.time()
loss_batch = loss_batch / dataloader.batch_size
loss_epoch += loss_batch
num_batch += 1
print('{}/{} (epoch {}), train_loss = {:.3f}, time/batch = {:.3f}'.
format(epoch * dataloader.num_batches + batch,
args.num_epochs * dataloader.num_batches, epoch,
loss_batch, end - start))
loss_epoch /= dataloader.num_batches
# 记录loss
log_file_curve.write("Training epoch: " + str(epoch) + " loss: " +
str(loss_epoch) + '\n')
if dataloader.valid_num_batches > 0:
print(
'****************Validation epoch beginning******************')
# Validation
dataloader.reset_batch_pointer(valid=True)
loss_epoch = 0
err_epoch = 0
# 每一个batch
for batch in range(dataloader.valid_num_batches):
# 获取batch数据
x, y, d, numPedsList, PedsList, target_ids = dataloader.next_valid_batch(
)
# batch的损失loss
loss_batch = 0
err_batch = 0
# 对于每个序列
for sequence in range(dataloader.batch_size):
# 获取与当前序列相对应的数据
x_seq, _, d_seq, numPedsList_seq, PedsList_seq = x[
sequence], y[sequence], d[sequence], numPedsList[
sequence], PedsList[sequence]
target_id = target_ids[sequence]
# 获取处理文件名，然后获取文件尺寸
folder_name = dataloader.get_directory_name_with_pointer(
d_seq)
dataset_data = dataloader.get_dataset_dimension(
folder_name)
# 密集矢量创建
x_seq, lookup_seq = dataloader.convert_proper_array(
x_seq, numPedsList_seq, PedsList_seq)
target_id_values = x_seq[0][lookup_seq[target_id], 0:2]
# get grid mask
# 应该是用于判断是否有social
grid_seq = getSequenceGridMask(x_seq, dataset_data,
PedsList_seq,
args.neighborhood_size,
args.grid_size,
args.use_cuda)
x_seq, first_values_dict = vectorize_seq(
x_seq, PedsList_seq, lookup_seq)
if args.use_cuda:
x_seq = x_seq.cuda()
# number of peds in this sequence per frame
numNodes = len(lookup_seq)
hidden_states = Variable(
torch.zeros(numNodes, args.rnn_size))
if args.use_cuda:
hidden_states = hidden_states.cuda()
cell_states = Variable(torch.zeros(numNodes,
args.rnn_size))
if args.use_cuda:
cell_states = cell_states.cuda()
# Forward prop
outputs, _, _ = net(x_seq[:-1], grid_seq[:-1],
hidden_states, cell_states,
PedsList_seq[:-1], numPedsList_seq,
dataloader, lookup_seq)
# 计算损失loss
loss = Gaussian2DLikelihood(outputs, x_seq[1:],
PedsList_seq[1:], lookup_seq)
# 提取二元高斯的均值mean，std标准差和corr相关性
mux, muy, sx, sy, corr = getCoef(outputs)
# 来自二元高斯的样本
next_x, next_y = sample_gaussian_2d(
mux.data, muy.data, sx.data, sy.data, corr.data,
PedsList_seq[-1], lookup_seq)
next_vals = torch.FloatTensor(1, numNodes, 2)
next_vals[:, :, 0] = next_x
next_vals[:, :, 1] = next_y
err = get_mean_error(next_vals, x_seq[-1].data[None, :, :],
[PedsList_seq[-1]],
[PedsList_seq[-1]], args.use_cuda,
lookup_seq)
loss_batch += loss.item()
err_batch += err
loss_batch = loss_batch / dataloader.batch_size
err_batch = err_batch / dataloader.batch_size
loss_epoch += loss_batch
err_epoch += err_batch
if dataloader.valid_num_batches != 0:
loss_epoch = loss_epoch / dataloader.valid_num_batches
err_epoch = err_epoch / dataloader.num_batches
# 到目前为止更新最佳验证损失loss
if loss_epoch < best_val_loss:
best_val_loss = loss_epoch
best_epoch_val = epoch
if err_epoch < smallest_err_val:
smallest_err_val = err_epoch
best_err_epoch_val = epoch
print('(epoch {}), valid_loss = {:.3f}, valid_err = {:.3f}'.
format(epoch, loss_epoch, err_epoch))
print('Best epoch', best_epoch_val, 'Best validation loss',
best_val_loss, 'Best error epoch', best_err_epoch_val,
'Best error', smallest_err_val)
log_file_curve.write("Validation epoch: " + str(epoch) +
" loss: " + str(loss_epoch) + " err: " +
str(err_epoch) + '\n')
# Validation验证数据集
if dataloader.additional_validation and (
epoch) in validation_epoch_list:
dataloader.switch_to_dataset_type()
print(
'****************Validation with dataset epoch beginning******************'
)
dataloader.reset_batch_pointer(valid=False)
dataset_pointer_ins = dataloader.dataset_pointer
validation_dataset_executed = True
loss_epoch = 0
err_epoch = 0
f_err_epoch = 0
num_of_batch = 0
smallest_err = 100000
# results of one epoch for all validation datasets
# 所有验证数据集的一个时期的结果
epoch_result = []
# results of one validation dataset
# 一个验证数据集的结果
results = []
# For each batch
for batch in range(dataloader.num_batches):
# Get batch data
x, y, d, numPedsList, PedsList, target_ids = dataloader.next_batch(
)
if dataset_pointer_ins is not dataloader.dataset_pointer:
if dataloader.dataset_pointer is not 0:
print('Finished prosessed file : ',
dataloader.get_file_name(-1),
' Avarage error : ', err_epoch / num_of_batch)
num_of_batch = 0
epoch_result.append(results)
dataset_pointer_ins = dataloader.dataset_pointer
results = []
# Loss for this batch
loss_batch = 0
err_batch = 0
f_err_batch = 0
# For each sequence
for sequence in range(dataloader.batch_size):
# Get data corresponding to the current sequence
x_seq, _, d_seq, numPedsList_seq, PedsList_seq = x[
sequence], y[sequence], d[sequence], numPedsList[
sequence], PedsList[sequence]
target_id = target_ids[sequence]
# get processing file name and then get dimensions of file
folder_name = dataloader.get_directory_name_with_pointer(
d_seq)
dataset_data = dataloader.get_dataset_dimension(
folder_name)
# dense vector creation
x_seq, lookup_seq = dataloader.convert_proper_array(
x_seq, numPedsList_seq, PedsList_seq)
# will be used for error calculation
orig_x_seq = x_seq.clone()
target_id_values = orig_x_seq[0][lookup_seq[target_id],
0:2]
# grid mask calculation
grid_seq = getSequenceGridMask(x_seq, dataset_data,
PedsList_seq,
args.neighborhood_size,
args.grid_size,
args.use_cuda)
if args.use_cuda:
x_seq = x_seq.cuda()
orig_x_seq = orig_x_seq.cuda()
# 向量化数据点
x_seq, first_values_dict = vectorize_seq(
x_seq, PedsList_seq, lookup_seq)
# 从模型中抽取预测点
ret_x_seq, loss = sample_validation_data(
x_seq, PedsList_seq, grid_seq, args, net, lookup_seq,
numPedsList_seq, dataloader)
# 将点还原回原始空间
ret_x_seq = revert_seq(ret_x_seq, PedsList_seq, lookup_seq,
first_values_dict)
# get mean and final error
err = get_mean_error(ret_x_seq.data, orig_x_seq.data,
PedsList_seq, PedsList_seq,
args.use_cuda, lookup_seq)
f_err = get_final_error(ret_x_seq.data, orig_x_seq.data,
PedsList_seq, PedsList_seq,
lookup_seq)
loss_batch += loss.item()
err_batch += err
f_err_batch += f_err
print('Current file : ', dataloader.get_file_name(0),
' Batch : ', batch + 1, ' Sequence: ', sequence + 1,
' Sequence mean error: ', err,
' Sequence final error: ', f_err, ' time: ',
end - start)
results.append(
(orig_x_seq.data.cpu().numpy(),
ret_x_seq.data.cpu().numpy(), PedsList_seq,
lookup_seq,
dataloader.get_frame_sequence(args.seq_length),
target_id))
loss_batch = loss_batch / dataloader.batch_size
err_batch = err_batch / dataloader.batch_size
f_err_batch = f_err_batch / dataloader.batch_size
num_of_batch += 1
loss_epoch += loss_batch
err_epoch += err_batch
f_err_epoch += f_err_batch
epoch_result.append(results)
all_epoch_results.append(epoch_result)
if dataloader.num_batches != 0:
loss_epoch = loss_epoch / dataloader.num_batches
err_epoch = err_epoch / dataloader.num_batches
f_err_epoch = f_err_epoch / dataloader.num_batches
avarage_err = (err_epoch + f_err_epoch) / 2
# Update best validation loss until now
if loss_epoch < best_val_data_loss:
best_val_data_loss = loss_epoch
best_epoch_val_data = epoch
if avarage_err < smallest_err_val_data:
smallest_err_val_data = avarage_err
best_err_epoch_val_data = epoch
print('(epoch {}), valid_loss = {:.3f}, '
'valid_mean_err = {:.3f}, '
'valid_final_err = {:.3f}'.format(
epoch, loss_epoch, err_epoch, f_err_epoch))
print('Best epoch', best_epoch_val_data,
'Best validation loss', best_val_data_loss,
'Best error epoch', best_err_epoch_val_data,
'Best error', smallest_err_val_data)
log_file_curve.write("Validation dataset epoch: " +
str(epoch) + " loss: " + str(loss_epoch) +
" mean_err: " + str(err_epoch) +
'final_err: ' + str(f_err_epoch) + '\n')
optimizer = time_lr_scheduler(optimizer,
epoch,
lr_decay_epoch=args.freq_optimizer)
# Save the model after each epoch
print('Saving model')
torch.save(
{
'epoch': epoch,
'state_dict': net.state_dict(),
'optimizer_state_dict': optimizer.state_dict()
}, checkpoint_path(epoch))
if dataloader.valid_num_batches != 0:
print('Best epoch', best_epoch_val, 'Best validation Loss',
best_val_loss, 'Best error epoch', best_err_epoch_val,
'Best error', smallest_err_val)
# Log the best epoch and best validation loss
log_file.write('Validation Best epoch:' + str(best_epoch_val) + ',' +
' Best validation Loss: ' + str(best_val_loss))
if dataloader.additional_validation:
print('Best epoch acording to validation dataset', best_epoch_val_data,
'Best validation Loss', best_val_data_loss, 'Best error epoch',
best_err_epoch_val_data, 'Best error', smallest_err_val_data)
log_file.write("Validation dataset Best epoch: " +
str(best_epoch_val_data) + ',' +
' Best validation Loss: ' + str(best_val_data_loss) +
'\n')
# FileNotFoundError: [Errno 2] No such file or directory: 'plot/SOCIALLSTM\\LSTM\\validation\\biwi\\biwi_hotel_4.pkl'
# validation_dataset_executed = True
if validation_dataset_executed:
# print("用于绘图的文件开始保存了")
dataloader.switch_to_dataset_type(load_data=False)
create_directories(plot_directory, [plot_train_file_directory])
# 找不到这个文件，是我手动添加的
# print(all_epoch_results)
# print(len(all_epoch_results) - 1)
dataloader.write_to_plot_file(
all_epoch_results[len(all_epoch_results) - 1],
os.path.join(plot_directory, plot_train_file_directory))
# Close logging files
log_file.close()
log_file_curve.close()
def train(args):
datasets = list(range(5))
#从数据集中删除leaveDataset
datasets.remove(args.leaveDataset)
# Create the SocialDataLoader object
data_loader = SocialDataLoader(args.batch_size,
args.seq_length,
args.maxNumPeds,
datasets,
forcePreProcess=True,
infer=False)
# 日志目录
log_directory = 'log/'
log_directory += str(args.leaveDataset) + '/'
# Logging files
log_file_curve = open(os.path.join(log_directory, 'log_curve.txt'), 'w')
log_file = open(os.path.join(log_directory, 'val.txt'), 'w')
#存储目录
save_directory = 'save/'
save_directory += str(args.leaveDataset) + '/'
with open(os.path.join(save_directory, 'social_config.pkl'), 'wb') as f:
pickle.dump(args, f)
# Create a SocialModel object with the arguments
model = SocialModel(args)
with tf.Session() as sess:
# Initialize all variables in the graph
sess.run(tf.global_variables_initializer())
# Initialize a saver that saves all the variables in the graph
saver = tf.train.Saver(tf.global_variables(), max_to_keep=None)
# summary_writer = tf.train.SummaryWriter('/tmp/lstm/logs', graph_def=sess.graph_def)
print('Training begin')
best_val_loss = 100
best_epoch = 0
# For each epoch
for e in range(args.num_epochs):
# Assign the learning rate value for this epoch
sess.run(
tf.assign(model.lr, args.learning_rate * (args.decay_rate**e)))
# Reset the data pointers in the data_loader
data_loader.reset_batch_pointer(valid=False)
loss_epoch = 0
# For each batch
for b in range(data_loader.num_batches):
start = time.time()
# Get the source, target and dataset data for the next batch
# 获取下一批的源，目标和数据集数据
# x, y are input and target data which are lists containing numpy arrays of size seq_length x maxNumPeds x 3
# x，y是输入和目标数据，它们是包含大小为seq_length x maxNumPeds x 3的numpy数组的列表
# d is the list of dataset indices from which each batch is generated (used to differentiate between datasets)
# d是生成每个批次的数据集索引列表（用于区分数据集）
x, y, d = data_loader.next_batch()
# variable to store the loss for this batch
# 变量来存储此批次的损失
loss_batch = 0
# For each sequence in the batch
for batch in range(data_loader.batch_size):
# x_batch, y_batch and d_batch contains the source, target and dataset index data for
# x_batch，y_batch和d_batch包含源，目标和数据集索引数据
# seq_length long consecutive frames in the dataset
# seq_length数据集中的长连续帧
# x_batch, y_batch would be numpy arrays of size seq_length x maxNumPeds x 3
# d_batch would be a scalar identifying the dataset from which this sequence is extracted
# d_batch将是一个标量，用于标识从中提取此序列的数据集
x_batch, y_batch, d_batch = x[batch], y[batch], d[batch]
if d_batch == 0 and datasets[0] == 0:
dataset_data = [640, 480]
else:
dataset_data = [720, 576]
grid_batch = getSequenceGridMask(x_batch, dataset_data,
args.neighborhood_size,
args.grid_size)
# Feed the source, target data
# 提供源，目标数据
feed = {
model.input_data: x_batch,
model.target_data: y_batch,
model.grid_data: grid_batch
}
# 运行模型，跑出来一个结果
train_loss, _ = sess.run([model.cost, model.train_op],
feed)
loss_batch += train_loss
end = time.time()
loss_batch = loss_batch / data_loader.batch_size
loss_epoch += loss_batch
print(
"{}/{} (epoch {}), train_loss = {:.3f}, time/batch = {:.3f}"
.format(e * data_loader.num_batches + b,
args.num_epochs * data_loader.num_batches, e,
loss_batch, end - start))
loss_epoch /= data_loader.num_batches
log_file_curve.write(str(e) + ',' + str(loss_epoch) + ',')
print('*****************')
# 验证模型
data_loader.reset_batch_pointer(valid=True)
loss_epoch = 0
for b in range(data_loader.num_batches):
# Get the source, target and dataset data for the next batch
# x, y are input and target data which are lists containing numpy arrays of size seq_length x maxNumPeds x 3
# d is the list of dataset indices from which each batch is generated (used to differentiate between datasets)
x, y, d = data_loader.next_valid_batch()
# variable to store the loss for this batch
loss_batch = 0
# For each sequence in the batch
for batch in range(data_loader.batch_size):
# x_batch, y_batch and d_batch contains the source, target and dataset index data for
# seq_length long consecutive frames in the dataset
# x_batch, y_batch would be numpy arrays of size seq_length x maxNumPeds x 3
# d_batch would be a scalar identifying the dataset from which this sequence is extracted
x_batch, y_batch, d_batch = x[batch], y[batch], d[batch]
if d_batch == 0 and datasets[0] == 0:
dataset_data = [640, 480]
else:
dataset_data = [720, 576]
grid_batch = getSequenceGridMask(x_batch, dataset_data,
args.neighborhood_size,
args.grid_size)
# Feed the source, target data
feed = {
model.input_data: x_batch,
model.target_data: y_batch,
model.grid_data: grid_batch
}
train_loss = sess.run(model.cost, feed)
loss_batch += train_loss
loss_batch = loss_batch / data_loader.batch_size
loss_epoch += loss_batch
loss_epoch /= data_loader.valid_num_batches
# Update best validation loss until now
if loss_epoch < best_val_loss:
best_val_loss = loss_epoch
best_epoch = e
print('(epoch {}), valid_loss = {:.3f}'.format(e, loss_epoch))
print('Best epoch', best_epoch, 'Best validation loss',
best_val_loss)
log_file_curve.write(str(loss_epoch) + '\n')
print('*****************')
# Save the model after each epoch
print('Saving model')
checkpoint_path = os.path.join(save_directory, 'social_model.ckpt')
saver.save(sess, checkpoint_path, global_step=e)
print("model saved to {}".format(checkpoint_path))
print('Best epoch', best_epoch, 'Best validation loss', best_val_loss)
log_file.write(str(best_epoch) + ',' + str(best_val_loss))
# CLose logging files
log_file.close()
log_file_curve.close()
def main():
parser = argparse.ArgumentParser(
description="Sample new trajectories with a social LSTM")
parser.add_argument(
"modelParams",
type=str,
help=
"Path to the file or folder with the parameters of the experiments",
)
parser.add_argument(
"-l",
"--logLevel",
help="logging level of the logger. Default is INFO",
metavar="level",
type=str,
)
parser.add_argument(
"-f",
"--logFolder",
help=
"path to the folder where to save the logs. If None, logs are only printed in stderr",
metavar="path",
type=str,
)
parser.add_argument(
"-ns",
"--noSaveCoordinates",
help="Flag to not save the predicted and ground truth coordinates",
action="store_true",
)
args = parser.parse_args()
if os.path.isdir(args.modelParams):
names_experiments = os.listdir(args.modelParams)
experiments = [
os.path.join(args.modelParams, experiment)
for experiment in names_experiments
]
else:
experiments = [args.modelParams]
# Table will show the metrics of each experiment
results = BeautifulTable()
results.column_headers = ["Name experiment", "ADE", "FDE"]
for experiment in experiments:
# Load the parameters
hparams = utils.YParams(experiment)
# Define the logger
setLogger(hparams, args, PHASE)
remainSpaces = 29 - len(hparams.name)
logging.info(
"\n" +
"--------------------------------------------------------------------------------\n"
+ "| Sampling experiment: " +
hparams.name + " " * remainSpaces + "|\n" +
"--------------------------------------------------------------------------------\n"
)
trajectory_size = hparams.obsLen + hparams.predLen
saveCoordinates = False
if args.noSaveCoordinates is True:
saveCoordinates = False
elif hparams.saveCoordinates:
saveCoordinates = hparams.saveCoordinates
if saveCoordinates:
coordinates_path = os.path.join("coordinates", hparams.name)
if not os.path.exists("coordinates"):
os.makedirs("coordinates")
logging.info("Loading the test datasets...")
test_loader = utils.DataLoader(
hparams.dataPath,
hparams.testDatasets,
hparams.testMaps,
hparams.semanticMaps,
hparams.testMapping,
hparams.homography,
num_labels=hparams.numLabels,
delimiter=hparams.delimiter,
skip=hparams.skip,
max_num_ped=hparams.maxNumPed,
trajectory_size=trajectory_size,
neighborood_size=hparams.neighborhoodSize,
)
logging.info("Creating the test dataset pipeline...")
dataset = utils.TrajectoriesDataset(
test_loader,
val_loader=None,
batch=False,
shuffle=hparams.shuffle,
prefetch_size=hparams.prefetchSize,
)
logging.info("Creating the model...")
start = time.time()
model = SocialModel(dataset, hparams, phase=PHASE)
end = time.time() - start
logging.debug("Model created in {:.2f}s".format(end))
# Define the path to the file that contains the variables of the model
model_folder = os.path.join(hparams.modelFolder, hparams.name)
model_path = os.path.join(model_folder, hparams.name)
# Create a saver
saver = tf.train.Saver()
# Add to the computation graph the evaluation functions
ade_sequence = utils.average_displacement_error(
model.new_pedestrians_coordinates[-hparams.predLen:],
model.pedestrians_coordinates[-hparams.predLen:],
model.num_peds_frame,
)
fde_sequence = utils.final_displacement_error(
model.new_pedestrians_coordinates[-1],
model.pedestrians_coordinates[-1],
model.num_peds_frame,
)
ade = 0
fde = 0
coordinates_predicted = []
coordinates_gt = []
peds_in_sequence = []
# Zero padding
padding = len(str(test_loader.num_sequences))
# ============================ START SAMPLING ============================
with tf.Session() as sess:
# Restore the model trained
saver.restore(sess, model_path)
# Initialize the iterator of the sample dataset
sess.run(dataset.init_train)
logging.info(
"\n" +
"--------------------------------------------------------------------------------\n"
+
"| Start sampling |\n"
+
"--------------------------------------------------------------------------------\n"
)
for seq in range(test_loader.num_sequences):
logging.info("Sample trajectory number {:{width}d}/{}".format(
seq + 1, test_loader.num_sequences, width=padding))
ade_value, fde_value, coordinates_pred_value, coordinates_gt_value, num_peds = sess.run(
[
ade_sequence,
fde_sequence,
model.new_pedestrians_coordinates,
model.pedestrians_coordinates,
model.num_peds_frame,
])
ade += ade_value
fde += fde_value
coordinates_predicted.append(coordinates_pred_value)
coordinates_gt.append(coordinates_gt_value)
peds_in_sequence.append(num_peds)
ade = ade / test_loader.num_sequences
fde = fde / test_loader.num_sequences
logging.info("Sampling finished. ADE: {:.4f} FDE: {:.4f}".format(
ade, fde))
results.append_row([hparams.name, ade, fde])
if saveCoordinates:
coordinates_predicted = np.array(coordinates_predicted)
coordinates_gt = np.array(coordinates_gt)
saveCoords(
coordinates_predicted,
coordinates_gt,
peds_in_sequence,
hparams.predLen,
coordinates_path,
)
tf.reset_default_graph()
logging.info("\n{}".format(results))