def execute(gpu, exp_batch, exp_alias):
# We set the visible cuda devices
os.environ["CUDA_VISIBLE_DEVICES"] = gpu
# At this point the log file with the correct naming is created.
merge_with_yaml(os.path.join(exp_batch, exp_alias+'.yaml'))
set_type_of_process('train')
sys.stdout = open(str(os.getpid()) + ".out", "a", buffering=1)
if monitorer.get_status(exp_batch, exp_alias, g_conf.PROCESS_NAME)[0] == "Finished":
# TODO: print some cool summary or not ?
return
#Define the dataset. This structure is has the __get_item__ redefined in a way
#that you can access the HDFILES positions from the root directory as a in a vector.
full_dataset = os.path.join(os.environ["COIL_DATASET_PATH"], g_conf.DATASET_NAME)
dataset = CoILDataset(full_dataset, transform=transforms.Compose([transforms.ToTensor()]))
# Creates the sampler, this part is responsible for managing the keys. It divides
# all keys depending on the measurements and produces a set of keys for each bach.
sampler = CoILSampler(splitter.control_steer_split(dataset.measurements, dataset.meta_data))
# The data loader is the multi threaded module from pytorch that release a number of
# workers to get all the data.
# TODO: batch size an number of workers go to some configuration file
data_loader = torch.utils.data.DataLoader(dataset, sampler=sampler, batch_size=120,
shuffle=False, num_workers=12, pin_memory=True)
# By instanciating the augmenter we get a callable that augment images and transform them
# into tensors.
augmenter = iag.Augmenter(g_conf.AUGMENTATION_SUITE)
# TODO: here there is clearly a posibility to make a cool "conditioning" system.
model = CoILModel(g_conf.MODEL_NAME)
model.cuda()
print(model)
criterion = Loss()
# TODO: DATASET SIZE SEEMS WEIRD
optimizer = optim.SGD(model.parameters(), lr=0.0001, momentum=0.9)
checkpoint_file = get_latest_saved_checkpoint()
if checkpoint_file != None:
checkpoint = torch.load(os.path.join('_logs', exp_batch, exp_alias,
'checkpoints', str(get_latest_saved_checkpoint())))
iteration = checkpoint['iteration']
else:
iteration = 0
# TODO: The checkpoint will continue, so the logs should restart ??? OR continue were it was
print (dataset.meta_data)
print (model)
for data in data_loader:
input_data, labels = data
#TODO we have to divide the input with other data.
#TODO, ADD ITERATION SCHEDULE
input_rgb_data = augmenter(0, input_data['rgb'])
# get the control commands from labels, size = [120,1]
controls = labels[:, 24, :]
# The output(branches) is a list of 5 branches results, each branch is with size [120,3]
model.zero_grad()
branches = model(input_rgb_data, labels[:, 10, :].cuda())
#print ("len ",len(branches))
# get the steer, gas and brake ground truth from labels
steer_gt = labels[:, 0, :]
gas_gt = labels[:, 1, :]
brake_gt = labels[:, 2, :]
speed_gt = labels[:, 10, :]
targets = torch.cat([steer_gt, gas_gt, brake_gt], 1)
loss = criterion.MSELoss(branches, targets.cuda(), controls.cuda(), speed_gt.cuda())
loss.backward()
optimizer.step()
# TODO: save also the optimizer state dictionary
if is_ready_to_save(iteration):
state = {
'iteration': iteration,
'state_dict': model.state_dict()
}
# TODO : maybe already summarize the best model ???
torch.save(state, os.path.join('_logs', exp_batch, exp_alias
, 'checkpoints', str(iteration) + '.pth'))
iteration += 1
def execute(gpu,
exp_batch,
exp_alias,
suppress_output=True,
number_of_workers=12,
encoder_params=None):
"""
The main training function. This functions loads the latest checkpoint
for a given, exp_batch (folder) and exp_alias (experiment configuration).
With this checkpoint it starts from the beginning or continue some training.
Args:
gpu: The GPU number
exp_batch: the folder with the experiments
exp_alias: the alias, experiment name
suppress_output: if the output are going to be saved on a file
number_of_workers: the number of threads used for data loading
Returns:
None
"""
try:
# We set the visible cuda devices to select the GPU
os.environ["CUDA_VISIBLE_DEVICES"] = gpu
g_conf.VARIABLE_WEIGHT = {}
# At this point the log file with the correct naming is created.
# You merge the yaml file with the global configuration structure.
merge_with_yaml(
os.path.join('configs', exp_batch, exp_alias + '.yaml'),
encoder_params)
set_type_of_process('train')
# Set the process into loading status.
coil_logger.add_message('Loading',
{'GPU': os.environ["CUDA_VISIBLE_DEVICES"]})
seed_everything(seed=g_conf.MAGICAL_SEED)
# Put the output to a separate file if it is the case
if suppress_output:
if not os.path.exists('_output_logs'):
os.mkdir('_output_logs')
sys.stdout = open(os.path.join(
'_output_logs', exp_alias + '_' + g_conf.PROCESS_NAME + '_' +
str(os.getpid()) + ".out"),
"a",
buffering=1)
sys.stderr = open(os.path.join(
'_output_logs', exp_alias + '_err_' + g_conf.PROCESS_NAME +
'_' + str(os.getpid()) + ".out"),
"a",
buffering=1)
if coil_logger.check_finish('train'):
coil_logger.add_message('Finished', {})
return
# Preload option
print(" GOING TO LOAD")
if g_conf.PRELOAD_MODEL_ALIAS is not None:
print(" LOADING A PRELOAD")
checkpoint = torch.load(
os.path.join('_logs', g_conf.PRELOAD_MODEL_BATCH,
g_conf.PRELOAD_MODEL_ALIAS, 'checkpoints',
str(g_conf.PRELOAD_MODEL_CHECKPOINT) + '.pth'))
else:
# Get the latest checkpoint to be loaded
# returns none if there are no checkpoints saved for this model
checkpoint_file = get_latest_saved_checkpoint()
if checkpoint_file is not None:
print('loading previous checkpoint ', checkpoint_file)
checkpoint = torch.load(
os.path.join('_logs', g_conf.EXPERIMENT_BATCH_NAME,
g_conf.EXPERIMENT_NAME, 'checkpoints',
str(get_latest_saved_checkpoint())))
iteration = checkpoint['iteration']
best_loss = checkpoint['best_loss']
best_loss_iter = checkpoint['best_loss_iter']
else:
iteration = 0
best_loss = 100000000.0
best_loss_iter = 0
# Define the dataset. This structure is has the __get_item__ redefined in a way
# that you can access the positions from the root directory as a in a vector.
#full_dataset = os.path.join(os.environ["COIL_DATASET_PATH"], g_conf.TRAIN_DATASET_NAME)
# By instantiating the augmenter we get a callable that augment images and transform them
# into tensors.
augmenter = Augmenter(g_conf.AUGMENTATION)
# We can save preload dataset depends on the json file name, then no need to load dataset for each time with the same dataset
if len(g_conf.EXPERIENCE_FILE) == 1:
json_file_name = str(
g_conf.EXPERIENCE_FILE[0]).split('/')[-1].split('.')[-2]
else:
json_file_name = str(g_conf.EXPERIENCE_FILE[0]).split(
'/')[-1].split('.')[-2] + '_' + str(
g_conf.EXPERIENCE_FILE[1]).split('/')[-1].split('.')[-2]
dataset = CoILDataset(transform=augmenter,
preload_name=g_conf.PROCESS_NAME + '_' +
json_file_name + '_' + g_conf.DATA_USED)
#dataset = CoILDataset(transform=augmenter, preload_name=str(g_conf.NUMBER_OF_HOURS)+ 'hours_' + g_conf.TRAIN_DATASET_NAME)
print("Loaded Training dataset")
data_loader = select_balancing_strategy(dataset, iteration,
number_of_workers)
if g_conf.MODEL_TYPE in ['separate-affordances']:
model = CoILModel(g_conf.MODEL_TYPE, g_conf.MODEL_CONFIGURATION,
g_conf.ENCODER_MODEL_CONFIGURATION)
model.cuda()
optimizer = optim.Adam(model.parameters(), lr=g_conf.LEARNING_RATE)
print(model)
# we use the pre-trained encoder model to extract bottleneck Z and train the E-t-E model
if g_conf.MODEL_TYPE in ['separate-affordances']:
encoder_model = EncoderModel(g_conf.ENCODER_MODEL_TYPE,
g_conf.ENCODER_MODEL_CONFIGURATION)
encoder_model.cuda()
encoder_model.eval()
# To freeze the pre-trained encoder model
if g_conf.FREEZE_ENCODER:
for param_ in encoder_model.parameters():
param_.requires_grad = False
if encoder_params is not None:
encoder_checkpoint = torch.load(
os.path.join(
'_logs', encoder_params['encoder_folder'],
encoder_params['encoder_exp'], 'checkpoints',
str(encoder_params['encoder_checkpoint']) + '.pth'))
print(
"Encoder model ",
str(encoder_params['encoder_checkpoint']), "loaded from ",
os.path.join('_logs', encoder_params['encoder_folder'],
encoder_params['encoder_exp'], 'checkpoints'))
encoder_model.load_state_dict(encoder_checkpoint['state_dict'])
if g_conf.FREEZE_ENCODER:
encoder_model.eval()
# To freeze the pre-trained encoder model
for param_ in encoder_model.parameters():
param_.requires_grad = False
else:
optimizer = optim.Adam(list(model.parameters()) +
list(encoder_model.parameters()),
lr=g_conf.LEARNING_RATE)
for name_encoder, param_encoder in encoder_model.named_parameters(
):
if param_encoder.requires_grad:
print(' Unfrozen layers', name_encoder)
else:
print(' Frozen layers', name_encoder)
if checkpoint_file is not None or g_conf.PRELOAD_MODEL_ALIAS is not None:
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
accumulated_time = checkpoint['total_time']
loss_window = coil_logger.recover_loss_window('train', iteration)
else: # We accumulate iteration time and keep the average speed
accumulated_time = 0
loss_window = []
for name, param in model.named_parameters():
if param.requires_grad:
print(' Unfrozen layers', name)
else:
print(' Frozen layers', name)
print("Before the loss")
# Loss time series window
for data in data_loader:
# Basically in this mode of execution, we validate every X Steps, if it goes up 3 times,
# add a stop on the _logs folder that is going to be read by this process
if g_conf.FINISH_ON_VALIDATION_STALE is not None and \
check_loss_validation_stopped(iteration, g_conf.FINISH_ON_VALIDATION_STALE):
break
"""
####################################
Main optimization loop
####################################
"""
if iteration % 1000 == 0:
adjust_learning_rate_auto(optimizer, loss_window)
model.zero_grad()
if not g_conf.FREEZE_ENCODER:
encoder_model.zero_grad()
if g_conf.LABELS_SUPERVISED:
inputs_data = torch.cat(
(data['rgb'], torch.zeros(g_conf.BATCH_SIZE, 1, 88, 200)),
dim=1).cuda()
else:
inputs_data = torch.squeeze(data['rgb'].cuda())
if g_conf.MODEL_TYPE in ['separate-affordances']:
#TODO: for this two encoder models training, we haven't put speed as input to train yet
if g_conf.ENCODER_MODEL_TYPE in [
'action_prediction', 'stdim', 'forward',
'one-step-affordances'
]:
e, inter = encoder_model.forward_encoder(
inputs_data,
dataset.extract_inputs(data).cuda(),
# We also add measurements and commands
torch.squeeze(dataset.extract_commands(data).cuda()))
elif g_conf.ENCODER_MODEL_TYPE in ['ETE']:
e, inter = encoder_model.forward_encoder(
inputs_data,
dataset.extract_inputs(data).cuda(),
torch.squeeze(dataset.extract_commands(data).cuda()))
loss_function_params = {
'classification_gt':
dataset.extract_affordances_targets(
data, 'classification').cuda(),
# harzard stop, red_light....
'class_weights':
g_conf.AFFORDANCES_CLASS_WEIGHT,
'regression_gt':
dataset.extract_affordances_targets(data,
'regression').cuda(),
'variable_weights':
g_conf.AFFORDANCES_VARIABLE_WEIGHT
}
loss = model(e, loss_function_params)
loss.backward()
optimizer.step()
else:
raise RuntimeError(
'Not implement yet, this branch is only work for g_conf.MODEL_TYPE in [separate-affordances]'
)
"""
####################################
Saving the model if necessary
####################################
"""
if is_ready_to_save(iteration):
state = {
'iteration': iteration,
'state_dict': model.state_dict(),
'best_loss': best_loss,
'total_time': accumulated_time,
'optimizer': optimizer.state_dict(),
'best_loss_iter': best_loss_iter
}
torch.save(
state,
os.path.join('_logs', g_conf.EXPERIMENT_BATCH_NAME,
g_conf.EXPERIMENT_NAME, 'checkpoints',
str(iteration) + '.pth'))
if not g_conf.FREEZE_ENCODER:
encoder_state = {
'iteration': iteration,
'state_dict': encoder_model.state_dict(),
'best_loss': best_loss,
'total_time': accumulated_time,
'optimizer': optimizer.state_dict(),
'best_loss_iter': best_loss_iter
}
torch.save(
encoder_state,
os.path.join('_logs', g_conf.EXPERIMENT_BATCH_NAME,
g_conf.EXPERIMENT_NAME, 'checkpoints',
str(iteration) + '_encoder.pth'))
iteration += 1
"""
################################################
Adding tensorboard logs.
Making calculations for logging purposes.
These logs are monitored by the printer module.
#################################################
"""
coil_logger.add_scalar('Loss', loss.data, iteration)
coil_logger.add_image('Image', torch.squeeze(data['rgb']),
iteration)
if loss.data < best_loss:
best_loss = loss.data.tolist()
best_loss_iter = iteration
if iteration % 100 == 0:
print('Train Iteration: {} [{}/{} ({:.0f}%)] \t Loss: {:.6f}'.
format(iteration, iteration, g_conf.NUMBER_ITERATIONS,
100. * iteration / g_conf.NUMBER_ITERATIONS,
loss.data))
coil_logger.add_message('Finished', {})
except KeyboardInterrupt:
coil_logger.add_message('Error', {'Message': 'Killed By User'})
except RuntimeError as e:
coil_logger.add_message('Error', {'Message': str(e)})
except:
traceback.print_exc()
coil_logger.add_message('Error', {'Message': 'Something Happened'})
def execute(gpu,
exp_batch,
exp_alias,
state_dict,
suppress_output=True,
number_of_workers=12):
"""
The main training function. This functions loads the latest checkpoint
for a given, exp_batch (folder) and exp_alias (experiment configuration).
With this checkpoint it starts from the beginning or continue some training.
Args:
gpu: The GPU number
exp_batch: the folder with the experiments
exp_alias: the alias, experiment name
suppress_output: if the output are going to be saved on a file
number_of_workers: the number of threads used for data loading
Returns:
None
"""
try:
# We set the visible cuda devices to select the GPU
os.environ["CUDA_VISIBLE_DEVICES"] = gpu
g_conf.VARIABLE_WEIGHT = {}
# At this point the log file with the correct naming is created.
# You merge the yaml file with the global configuration structure.
merge_with_yaml(os.path.join('configs', exp_batch,
exp_alias + '.yaml'))
set_type_of_process('train')
# Set the process into loading status.
coil_logger.add_message('Loading', {'GPU': gpu})
# Put the output to a separate file if it is the case
if suppress_output:
if not os.path.exists('_output_logs'):
os.mkdir('_output_logs')
sys.stdout = open(os.path.join(
'_output_logs', exp_alias + '_' + g_conf.PROCESS_NAME + '_' +
str(os.getpid()) + ".out"),
"a",
buffering=1)
sys.stderr = open(os.path.join(
'_output_logs', exp_alias + '_err_' + g_conf.PROCESS_NAME +
'_' + str(os.getpid()) + ".out"),
"a",
buffering=1)
if coil_logger.check_finish('train'):
coil_logger.add_message('Finished', {})
return
# Preload option
if g_conf.PRELOAD_MODEL_ALIAS is not None:
checkpoint = torch.load(
os.path.join('_logs', g_conf.PRELOAD_MODEL_BATCH,
g_conf.PRELOAD_MODEL_ALIAS, 'checkpoints',
str(g_conf.PRELOAD_MODEL_CHECKPOINT) + '.pth'))
# Get the latest checkpoint to be loaded
# returns none if there are no checkpoints saved for this model
checkpoint_file = get_latest_saved_checkpoint()
if checkpoint_file is not None:
checkpoint = torch.load(
os.path.join('_logs', exp_batch, exp_alias, 'checkpoints',
str(get_latest_saved_checkpoint())))
iteration = checkpoint['iteration']
best_loss = checkpoint['best_loss']
best_loss_iter = checkpoint['best_loss_iter']
else:
iteration = 0
best_loss = 10000.0
best_loss_iter = 0
# Define the dataset. This structure is has the __get_item__ redefined in a way
# that you can access the positions from the root directory as a in a vector.
full_dataset = os.path.join(os.environ["COIL_DATASET_PATH"],
g_conf.TRAIN_DATASET_NAME)
# By instantiating the augmenter we get a callable that augment images and transform them
# into tensors.
augmenter = Augmenter(g_conf.AUGMENTATION)
# Instantiate the class used to read a dataset. The coil dataset generator
# can be found
dataset = CoILDataset(full_dataset,
transform=augmenter,
preload_name=str(g_conf.NUMBER_OF_HOURS) +
'hours_' + g_conf.TRAIN_DATASET_NAME)
print("Loaded dataset")
data_loader = select_balancing_strategy(dataset, iteration,
number_of_workers)
model = CoILModel(g_conf.MODEL_TYPE, g_conf.MODEL_CONFIGURATION)
model.cuda()
if state_dict != '':
seg_model = ERFNet_Fast(2)
seg_model = load_my_state_dict(seg_model, torch.load(state_dict))
seg_model.cuda()
optimizer = optim.Adam(model.parameters(), lr=g_conf.LEARNING_RATE)
if checkpoint_file is not None or g_conf.PRELOAD_MODEL_ALIAS is not None:
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
accumulated_time = checkpoint['total_time']
loss_window = coil_logger.recover_loss_window('train', iteration)
else: # We accumulate iteration time and keep the average speed
accumulated_time = 0
loss_window = []
print("Before the loss")
criterion = Loss(g_conf.LOSS_FUNCTION)
color_transforms = Colorizes(2)
board = Dashboard(8097)
# Loss time series window
for data in data_loader:
# Basically in this mode of execution, we validate every X Steps, if it goes up 3 times,
# add a stop on the _logs folder that is going to be read by this process
if g_conf.FINISH_ON_VALIDATION_STALE is not None and \
check_loss_validation_stopped(iteration, g_conf.FINISH_ON_VALIDATION_STALE):
break
"""
####################################
Main optimization loop
####################################
"""
iteration += 1
if iteration % 1000 == 0:
adjust_learning_rate_auto(optimizer, loss_window)
# get the control commands from float_data, size = [120,1]
capture_time = time.time()
controls = data['directions']
# The output(branches) is a list of 5 branches results, each branch is with size [120,3]
model.zero_grad()
if state_dict != '':
with torch.no_grad():
repre = seg_model(torch.squeeze(data['rgb'].cuda()),
only_encode=False)
inputs = repre
imgs = color_transforms(inputs)
inputs = inputs.float().cuda()
else:
inputs = torch.squeeze(data['rgb'].cuda())
# vis
board.image(
torch.squeeze(data['rgb'])[0].cpu().data,
'(train) input iter: ' + str(iteration))
board.image(imgs[0].cpu().data,
'(train) output iter: ' + str(iteration))
branches = model(inputs, dataset.extract_inputs(data).cuda())
loss_function_params = {
'branches': branches,
'targets': dataset.extract_targets(data).cuda(),
'controls': controls.cuda(),
'inputs': dataset.extract_inputs(data).cuda(),
'branch_weights': g_conf.BRANCH_LOSS_WEIGHT,
'variable_weights': g_conf.VARIABLE_WEIGHT
}
loss, _ = criterion(loss_function_params)
loss.backward()
optimizer.step()
"""
####################################
Saving the model if necessary
####################################
"""
if is_ready_to_save(iteration):
state = {
'iteration': iteration,
'state_dict': model.state_dict(),
'best_loss': best_loss,
'total_time': accumulated_time,
'optimizer': optimizer.state_dict(),
'best_loss_iter': best_loss_iter
}
torch.save(
state,
os.path.join('_logs', exp_batch, exp_alias, 'checkpoints',
str(iteration) + '.pth'))
"""
################################################
Adding tensorboard logs.
Making calculations for logging purposes.
These logs are monitored by the printer module.
#################################################
"""
coil_logger.add_scalar('Loss', loss.data, iteration)
coil_logger.add_image('Image', torch.squeeze(data['rgb']),
iteration)
if loss.data < best_loss:
best_loss = loss.data.tolist()
best_loss_iter = iteration
# Log a random position
position = random.randint(0, len(data) - 1)
output = model.extract_branch(torch.stack(branches[0:4]), controls)
error = torch.abs(output - dataset.extract_targets(data).cuda())
accumulated_time += time.time() - capture_time
coil_logger.add_message(
'Iterating', {
'Iteration':
iteration,
'Loss':
loss.data.tolist(),
'Images/s':
(iteration * g_conf.BATCH_SIZE) / accumulated_time,
'BestLoss':
best_loss,
'BestLossIteration':
best_loss_iter,
'Output':
output[position].data.tolist(),
'GroundTruth':
dataset.extract_targets(data)[position].data.tolist(),
'Error':
error[position].data.tolist(),
'Inputs':
dataset.extract_inputs(data)[position].data.tolist()
}, iteration)
loss_window.append(loss.data.tolist())
coil_logger.write_on_error_csv('train', loss.data)
print("Iteration: %d Loss: %f" % (iteration, loss.data))
coil_logger.add_message('Finished', {})
except KeyboardInterrupt:
coil_logger.add_message('Error', {'Message': 'Killed By User'})
except RuntimeError as e:
coil_logger.add_message('Error', {'Message': str(e)})
except:
traceback.print_exc()
coil_logger.add_message('Error', {'Message': 'Something Happened'})
def execute(gpu,
exp_batch,
exp_alias,
suppress_output=True,
number_of_workers=12):
"""
The main training function. This functions loads the latest checkpoint
for a given, exp_batch (folder) and exp_alias (experiment configuration).
With this checkpoint it starts from the beginning or continue some training.
Args:
gpu: gpus ids for training
exp_batch: the folder with the experiments
exp_alias: the alias, experiment name
suppress_output: if the output are going to be saved on a file
number_of_workers: the number of threads used for data loading
Returns:
None
"""
try:
os.environ["CUDA_VISIBLE_DEVICES"] = ','.join(gpu)
g_conf.VARIABLE_WEIGHT = {}
# At this point the log file with the correct naming is created.
# You merge the yaml file with the global configuration structure.
merge_with_yaml(os.path.join('configs', exp_batch,
exp_alias + '.yaml'))
set_type_of_process('train')
# Set the process into loading status.
coil_logger.add_message('Loading', {'GPU': gpu})
# Put the output to a separate file if it is the case
if suppress_output:
if not os.path.exists('_output_logs'):
os.mkdir('_output_logs')
sys.stdout = open(os.path.join(
'_output_logs', exp_alias + '_' + g_conf.PROCESS_NAME + '_' +
str(os.getpid()) + ".out"),
"a",
buffering=1)
sys.stderr = open(os.path.join(
'_output_logs', exp_alias + '_err_' + g_conf.PROCESS_NAME +
'_' + str(os.getpid()) + ".out"),
"a",
buffering=1)
if coil_logger.check_finish('train'):
coil_logger.add_message('Finished', {})
return
# Preload option
if g_conf.PRELOAD_MODEL_ALIAS is not None:
checkpoint = torch.load(
os.path.join('_logs', g_conf.PRELOAD_MODEL_BATCH,
g_conf.PRELOAD_MODEL_ALIAS, 'checkpoints',
str(g_conf.PRELOAD_MODEL_CHECKPOINT) + '.pth'))
# Get the latest checkpoint to be loaded
# returns none if there are no checkpoints saved for this model
checkpoint_file = get_latest_saved_checkpoint()
if checkpoint_file is not None:
checkpoint = torch.load(
os.path.join('_logs', exp_batch, exp_alias, 'checkpoints',
str(get_latest_saved_checkpoint())))
iteration = checkpoint['iteration']
best_loss = checkpoint['best_loss']
best_loss_iter = checkpoint['best_loss_iter']
print('iteration: ', iteration, 'best_loss: ', best_loss)
else:
iteration = 0
best_loss = 10000.0
best_loss_iter = 0
# Define the dataset. This structure is has the __get_item__ redefined in a way
# that you can access the positions from the root directory as a in a vector.
full_dataset = os.path.join(os.environ["COIL_DATASET_PATH"],
g_conf.TRAIN_DATASET_NAME)
# By instantiating the augmenter we get a callable that augment images and transform them into tensors.
augmenter = Augmenter(g_conf.AUGMENTATION)
# Instantiate the class used to read the dataset
dataset = CoILDataset(full_dataset,
transform=augmenter,
preload_name=str(g_conf.NUMBER_OF_HOURS) +
'hours_' + g_conf.TRAIN_DATASET_NAME)
print("Loaded dataset")
# Creates the sampler, this part is responsible for managing the keys. It divides
# all keys depending on the measurements and produces a set of keys for each bach.
# define the sampling strategy for mini-batch, different samplers can be found in 'splitter.py'
data_loader = select_balancing_strategy(dataset, iteration,
number_of_workers)
# Instatiate the network architecture
model = CoILModel(g_conf.MODEL_TYPE, g_conf.MODEL_CONFIGURATION)
model.cuda()
optimizer = optim.Adam(model.parameters(), lr=g_conf.LEARNING_RATE
) # adabound and adamio can also be used here
if checkpoint_file is not None or g_conf.PRELOAD_MODEL_ALIAS is not None:
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
accumulated_time = checkpoint['total_time']
loss_window = coil_logger.recover_loss_window('train', iteration)
else:
# We accumulate iteration time and keep the average speed
accumulated_time = 0
loss_window = []
# freeze the perception module weights if required
# for m in model.perception.parameters():
# m.requires_grad = False
# total trainable parameters
model_parameters = filter(lambda p: p.requires_grad,
model.parameters())
total_params = sum([np.prod(p.size()) for p in model_parameters])
print('trainable parameters: ', total_params)
# multi-gpu
print('number of gpus: ', torch.cuda.device_count())
if torch.cuda.device_count() > 1:
model = nn.DataParallel(model)
criterion = Loss(g_conf.LOSS_FUNCTION)
print('Start Training')
st = time.time()
for data in data_loader:
# use this for early stopping if the validation loss is not coming down
if g_conf.FINISH_ON_VALIDATION_STALE is not None and \
check_loss_validation_stopped(iteration, g_conf.FINISH_ON_VALIDATION_STALE):
break
"""
####################################
Main optimization loop
####################################
"""
iteration += 1
if iteration % 1000 == 0:
adjust_learning_rate_auto(optimizer, loss_window)
# additional learning rate scheduler - cyclic cosine annealing (https://arxiv.org/pdf/1704.00109.pdf)
# adjust_learning_rate_cosine_annealing(optimizer, loss_window, iteration)
capture_time = time.time()
controls = data['directions']
model.zero_grad()
branches = model(torch.squeeze(data['rgb'].cuda()),
dataset.extract_inputs(data).cuda())
loss_function_params = {
'branches': branches,
'targets': dataset.extract_targets(data).cuda(),
'controls': controls.cuda(),
'inputs': dataset.extract_inputs(data).cuda(),
'branch_weights': g_conf.BRANCH_LOSS_WEIGHT,
'variable_weights': g_conf.VARIABLE_WEIGHT
}
loss, _ = criterion(loss_function_params)
loss.backward()
optimizer.step()
"""
####################################
Saving the model if necessary
####################################
"""
if is_ready_to_save(iteration):
if torch.cuda.device_count() > 1:
state_dict_save = model.module.state_dict()
else:
state_dict_save = model.state_dict()
state = {
'iteration': iteration,
'state_dict': state_dict_save,
'best_loss': best_loss,
'total_time': accumulated_time,
'optimizer': optimizer.state_dict(),
'best_loss_iter': best_loss_iter
}
torch.save(
state,
os.path.join('_logs', exp_batch, exp_alias, 'checkpoints',
str(iteration) + '.pth'))
"""
################################################
Adding tensorboard logs.
Making calculations for logging purposes.
These logs are monitored by the printer module.
#################################################
"""
coil_logger.add_scalar('Loss', loss.data, iteration)
coil_logger.add_image('Image', torch.squeeze(data['rgb']),
iteration)
if loss.data < best_loss:
best_loss = loss.data.tolist()
best_loss_iter = iteration
# Log a random position
position = random.randint(0, len(data) - 1)
if torch.cuda.device_count() > 1:
output = model.module.extract_branch(
torch.stack(branches[0:4]), controls)
else:
output = model.extract_branch(torch.stack(branches[0:4]),
controls)
error = torch.abs(output - dataset.extract_targets(data).cuda())
accumulated_time += time.time() - capture_time
coil_logger.add_message(
'Iterating', {
'Iteration':
iteration,
'Loss':
loss.data.tolist(),
'Images/s':
(iteration * g_conf.BATCH_SIZE) / accumulated_time,
'BestLoss':
best_loss,
'BestLossIteration':
best_loss_iter,
'Output':
output[position].data.tolist(),
'GroundTruth':
dataset.extract_targets(data)[position].data.tolist(),
'Error':
error[position].data.tolist(),
'Inputs':
dataset.extract_inputs(data)[position].data.tolist()
}, iteration)
loss_window.append(loss.data.tolist())
coil_logger.write_on_error_csv('train', loss.data)
print("Iteration: %d Loss: %f" % (iteration, loss.data))
st = time.time()
coil_logger.add_message('Finished', {})
except KeyboardInterrupt:
coil_logger.add_message('Error', {'Message': 'Killed By User'})
except RuntimeError as e:
coil_logger.add_message('Error', {'Message': str(e)})
except:
traceback.print_exc()
coil_logger.add_message('Error', {'Message': 'Something Happened'})
def execute(gpu, exp_batch, exp_alias, dataset_name):
# We set the visible cuda devices
os.environ["CUDA_VISIBLE_DEVICES"] = '0'
# At this point the log file with the correct naming is created.
merge_with_yaml(os.path.join('configs', exp_batch, exp_alias + '.yaml'))
set_type_of_process('validation', dataset_name)
if not os.path.exists('_output_logs'):
os.mkdir('_output_logs')
sys.stdout = open(os.path.join(
'_output_logs', g_conf.PROCESS_NAME + '_' + str(os.getpid()) + ".out"),
"a",
buffering=1)
if monitorer.get_status(exp_batch, exp_alias + '.yaml',
g_conf.PROCESS_NAME)[0] == "Finished":
# TODO: print some cool summary or not ?
return
#Define the dataset. This structure is has the __get_item__ redefined in a way
#that you can access the HDFILES positions from the root directory as a in a vector.
full_dataset = os.path.join(os.environ["COIL_DATASET_PATH"], dataset_name)
print(full_dataset)
dataset = CoILDataset(full_dataset,
transform=transforms.Compose([transforms.ToTensor()
]))
# Creates the sampler, this part is responsible for managing the keys. It divides
# all keys depending on the measurements and produces a set of keys for each bach.
# The data loader is the multi threaded module from pytorch that release a number of
# workers to get all the data.
# TODO: batch size an number of workers go to some configuration file
data_loader = torch.utils.data.DataLoader(dataset,
batch_size=120,
shuffle=False,
num_workers=12,
pin_memory=True)
# TODO: here there is clearly a posibility to make a cool "conditioning" system.
model = CoILModel(g_conf.MODEL_NAME)
model.cuda()
model.eval()
criterion = Loss()
latest = get_latest_evaluated_checkpoint()
if latest is None: # When nothing was tested, get latest returns none, we fix that.
latest = 0
latest = 200000
best_loss = 1000.0
best_error = 1000.0
best_loss_iter = 0
best_error_iter = 0
print(dataset.meta_data[0][0])
for k in dataset.meta_data:
k[0] = str(k[0], 'utf-8')
print(dataset.meta_data[0][0])
cpts = glob.glob(
'/home-local/rohitrishabh/coil_20-06/_logs/eccv/experiment_1/checkpoints/*.pth'
)
# while not maximun_checkpoint_reach(latest, g_conf.TEST_SCHEDULE):
for ckpt in cpts:
# if is_next_checkpoint_ready(g_conf.TEST_SCHEDULE):
# latest = get_next_checkpoint(g_conf.TEST_SCHEDULE)
latest = int(ckpt[-10:-4])
# checkpoint = torch.load(os.path.join('_logs', exp_batch, exp_alias
# , 'checkpoints', str(latest) + '.pth'))
checkpoint = torch.load(ckpt)
checkpoint_iteration = checkpoint['iteration']
print("Validation loaded ", checkpoint_iteration)
accumulated_loss = 0.0
accumulated_error = 0.0
iteration_on_checkpoint = 0
for data in data_loader:
input_data, float_data = data
control_position = np.where(
dataset.meta_data[:, 0] == 'control')[0][0]
speed_position = np.where(
dataset.meta_data[:, 0] == 'speed_module')[0][0]
# print (torch.squeeze(input_data['rgb']).shape)
# print (control_position)
# print (speed_position)
# Obs : Maybe we could also check for other branches ??
output = model.forward_branch(
torch.squeeze(input_data['rgb']).cuda(),
float_data[:, speed_position, :].cuda(),
float_data[:, control_position, :].cuda())
for i in range(input_data['rgb'].shape[0]):
coil_logger.write_on_csv(
checkpoint_iteration,
[output[i][0], output[i][1], output[i][2]])
# TODO: Change this a functional standard using the loss functions.
loss = torch.mean(
(output -
dataset.extract_targets(float_data).cuda())**2).data.tolist()
mean_error = torch.mean(
torch.abs(
output -
dataset.extract_targets(float_data).cuda())).data.tolist()
accumulated_error += mean_error
accumulated_loss += loss
error = torch.abs(output -
dataset.extract_targets(float_data).cuda())
# Log a random position
position = random.randint(0, len(float_data) - 1)
#print (output[position].data.tolist())
coil_logger.add_message(
'Iterating in Validation', {
'Checkpoint':
latest,
'Iteration': (str(iteration_on_checkpoint * 120) + '/' +
str(len(dataset))),
'MeanError':
mean_error,
'Loss':
loss,
'Output':
output[position].data.tolist(),
'GroundTruth':
dataset.extract_targets(float_data)
[position].data.tolist(),
'Error':
error[position].data.tolist(),
'Inputs':
dataset.extract_inputs(float_data)[position].data.tolist()
}, latest)
iteration_on_checkpoint += 1
checkpoint_average_loss = accumulated_loss / len(dataset)
checkpoint_average_error = accumulated_error / len(dataset)
coil_logger.add_scalar('Loss', checkpoint_average_loss, latest)
coil_logger.add_scalar('Error', checkpoint_average_error, latest)
print('Loss: ', checkpoint_average_loss, "----Error: ",
checkpoint_average_error)
if checkpoint_average_loss < best_loss:
best_loss = checkpoint_average_loss
best_loss_iter = latest
state = {
'state_dict': model.state_dict(),
'best_loss': best_loss,
'best_loss_iter': best_loss_iter
}
# TODO : maybe already summarize the best model ???
torch.save(
state,
os.path.join('_logs', exp_batch, exp_alias,
'best_model_l2' + '.pth'))
if checkpoint_average_error < best_error:
best_error = checkpoint_average_error
best_error_iter = latest
state = {
'state_dict': model.state_dict(),
'best_error': best_error,
'best_error_iter': best_error_iter
}
# TODO : maybe already summarize the best model ???
torch.save(
state,
os.path.join('_logs', exp_batch, exp_alias,
'best_model_l1' + '.pth'))
print('Best Loss: ', best_loss, "Checkpoint", best_loss_iter)
print('Best Error: ', best_error, "Checkpoint", best_error_iter)
coil_logger.add_message(
'Iterating in Validation', {
'Summary': {
'Error': checkpoint_average_error,
'Loss': checkpoint_average_loss,
'BestError': best_error,
'BestLoss': best_loss,
'BestLossCheckpoint': best_loss_iter,
'BestErrorCheckpoint': best_error_iter
},
'Checkpoint': latest
})
def execute(gpu, exp_batch, exp_alias, suppress_output=True, number_of_workers=12):
"""
The main training function. This functions loads the latest checkpoint
for a given, exp_batch (folder) and exp_alias (experiment configuration).
With this checkpoint it starts from the beginning or continue some training.
Args:
gpu: The GPU number
exp_batch: the folder with the experiments
exp_alias: the alias, experiment name
suppress_output: if the output are going to be saved on a file
number_of_workers: the number of threads used for data loading
Returns:
None
"""
try:
# We set the visible cuda devices to select the GPU
os.environ["CUDA_VISIBLE_DEVICES"] = gpu
g_conf.VARIABLE_WEIGHT = {}
# At this point the log file with the correct naming is created.
# You merge the yaml file with the global configuration structure.
merge_with_yaml(os.path.join('configs', exp_batch, exp_alias + '.yaml'))
set_type_of_process('train')
# Set the process into loading status.
coil_logger.add_message('Loading', {'GPU': gpu})
# Put the output to a separate file if it is the case
if suppress_output:
if not os.path.exists('_output_logs'):
os.mkdir('_output_logs')
sys.stdout = open(os.path.join('_output_logs', exp_alias + '_' +
g_conf.PROCESS_NAME + '_' + str(os.getpid()) + ".out"), "a",
buffering=1)
sys.stderr = open(os.path.join('_output_logs',
exp_alias + '_err_'+g_conf.PROCESS_NAME + '_'
+ str(os.getpid()) + ".out"),
"a", buffering=1)
if coil_logger.check_finish('train'):
coil_logger.add_message('Finished', {})
return
# Preload option
if g_conf.PRELOAD_MODEL_ALIAS is not None:
checkpoint = torch.load(os.path.join('_logs', g_conf.PRELOAD_MODEL_BATCH,
g_conf.PRELOAD_MODEL_ALIAS,
'checkpoints',
str(g_conf.PRELOAD_MODEL_CHECKPOINT)+'.pth'))
# Get the latest checkpoint to be loaded
# returns none if there are no checkpoints saved for this model
checkpoint_file = get_latest_saved_checkpoint()
if checkpoint_file is not None:
checkpoint = torch.load(os.path.join('_logs', exp_batch, exp_alias,
'checkpoints', str(get_latest_saved_checkpoint())))
iteration = checkpoint['iteration']
best_loss = checkpoint['best_loss']
best_loss_iter = checkpoint['best_loss_iter']
else:
iteration = 0
best_loss = 10000.0
best_loss_iter = 0
# Define the dataset. This structure is has the __get_item__ redefined in a way
# that you can access the positions from the root directory as a in a vector.
full_dataset = os.path.join(os.environ["COIL_DATASET_PATH"], g_conf.TRAIN_DATASET_NAME)
# By instantiating the augmenter we get a callable that augment images and transform them
# into tensors.
augmenter = Augmenter(g_conf.AUGMENTATION)
# Instantiate the class used to read a dataset. The coil dataset generator
# can be found
dataset = CoILDataset(full_dataset, transform=augmenter,
preload_name=str(g_conf.NUMBER_OF_HOURS)
+ 'hours_' + g_conf.TRAIN_DATASET_NAME)
print ("Loaded dataset")
data_loader = select_balancing_strategy(dataset, iteration, number_of_workers)
model = CoILModel(g_conf.MODEL_TYPE, g_conf.MODEL_CONFIGURATION)
model.cuda()
optimizer = optim.Adam(model.parameters(), lr=g_conf.LEARNING_RATE)
# Set ERFnet for segmentation
model_erf = ERFNet(20)
model_erf = torch.nn.DataParallel(model_erf)
model_erf = model_erf.cuda()
print("LOAD ERFNet")
def load_my_state_dict(model, state_dict): #custom function to load model when not all dict elements
own_state = model.state_dict()
for name, param in state_dict.items():
if name not in own_state:
continue
own_state[name].copy_(param)
return model
model_erf = load_my_state_dict(model_erf, torch.load(os.path.join('trained_models/erfnet_pretrained.pth')))
model_erf.eval()
print ("ERFNet and weights LOADED successfully")
if checkpoint_file is not None or g_conf.PRELOAD_MODEL_ALIAS is not None:
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
accumulated_time = checkpoint['total_time']
loss_window = coil_logger.recover_loss_window('train', iteration)
else: # We accumulate iteration time and keep the average speed
accumulated_time = 0
loss_window = []
print ("Before the loss")
criterion = Loss(g_conf.LOSS_FUNCTION)
# Loss time series window
for data in data_loader:
# Basically in this mode of execution, we validate every X Steps, if it goes up 3 times,
# add a stop on the _logs folder that is going to be read by this process
if g_conf.FINISH_ON_VALIDATION_STALE is not None and \
check_loss_validation_stopped(iteration, g_conf.FINISH_ON_VALIDATION_STALE):
break
"""
####################################
Main optimization loop
####################################
"""
iteration += 1
if iteration % 1000 == 0:
adjust_learning_rate_auto(optimizer, loss_window)
# get the control commands from float_data, size = [120,1]
capture_time = time.time()
controls = data['directions']
# The output(branches) is a list of 5 branches results, each branch is with size [120,3]
model.zero_grad()
# print("Segmentation")
# use ERFNet to convert RGB to Segmentation
rgbs = data['rgb']
filenames = data['rgb_name']
# # seg one by one
# seg_road = []
# seg_not_road = []
# i = 0
# for inputs in rgbs:
# inputs = inputs.unsqueeze(0)
# # print("inputs ",inputs.shape)
# with torch.no_grad():
# outputs = model_erf(inputs)
# label = outputs[0].max(0)[1].byte().cpu().data
# road = (label == 0)
# not_road = (label != 0)
# seg_road.append(road)
# seg_not_road.append(not_road)
# # # print("label ",label.shape)
# # label_color = Colorize()(label.unsqueeze(0))
# # filename = filenames[i]
# # filenameSave = "./save_color/" + filename.split("CoILTrain/")[1]
# # os.makedirs(os.path.dirname(filenameSave), exist_ok=True)
# # label_save = ToPILImage()(label_color)
# # label_save.save(filenameSave)
# # # print (i, filenameSave)
# # i += 1
# seg_road = torch.stack(seg_road)
# seg_not_road = torch.stack(seg_not_road)
# seg = torch.stack([seg_road,seg_not_road]).transpose(0,1).float()
# # print(seg.shape)
# seg batch
with torch.no_grad():
outputs = model_erf(rgbs)
# print("outputs.shape ",outputs.shape)
labels = outputs.max(1)[1].byte().cpu().data
# print("labels.shape",labels.shape)
# print(np.unique(labels[0]))
seg_road = (labels==0)
seg_not_road = (labels!=0)
seg = torch.stack((seg_road,seg_not_road),1).float()
# save 1st batch's segmentation results
if iteration == 1:
for i in range(120):
label = seg[i,0,:,:]
label_color = Colorize()(label.unsqueeze(0))
filenameSave = "./save_color/batch_road_mask/%d.png"%(i)
os.makedirs(os.path.dirname(filenameSave), exist_ok=True)
label_save = ToPILImage()(label_color)
label_save.save(filenameSave)
label = labels[i,:,:]
label_color = Colorize()(label.unsqueeze(0))
filenameSave = "./save_color/batch_road/%d.png"%(i)
os.makedirs(os.path.dirname(filenameSave), exist_ok=True)
label_save = ToPILImage()(label_color)
label_save.save(filenameSave)
branches = model(torch.squeeze(seg).cuda(),
dataset.extract_inputs(data).cuda())
# branches = model(torch.squeeze(rgbs.cuda()),
# dataset.extract_input(data).cuda())
loss_function_params = {
'branches': branches,
'targets': dataset.extract_targets(data).cuda(),
'controls': controls.cuda(),
'inputs': dataset.extract_inputs(data).cuda(),
'branch_weights': g_conf.BRANCH_LOSS_WEIGHT,
'variable_weights': g_conf.VARIABLE_WEIGHT
}
loss, _ = criterion(loss_function_params)
loss.backward()
optimizer.step()
"""
####################################
Saving the model if necessary
####################################
"""
if is_ready_to_save(iteration):
state = {
'iteration': iteration,
'state_dict': model.state_dict(),
'best_loss': best_loss,
'total_time': accumulated_time,
'optimizer': optimizer.state_dict(),
'best_loss_iter': best_loss_iter
}
torch.save(state, os.path.join('_logs', exp_batch, exp_alias
, 'checkpoints', str(iteration) + '.pth'))
"""
################################################
Adding tensorboard logs.
Making calculations for logging purposes.
These logs are monitored by the printer module.
#################################################
"""
coil_logger.add_scalar('Loss', loss.data, iteration)
coil_logger.add_image('Image', torch.squeeze(data['rgb']), iteration)
if loss.data < best_loss:
best_loss = loss.data.tolist()
best_loss_iter = iteration
# Log a random position
position = random.randint(0, len(data) - 1)
output = model.extract_branch(torch.stack(branches[0:4]), controls)
error = torch.abs(output - dataset.extract_targets(data).cuda())
accumulated_time += time.time() - capture_time
coil_logger.add_message('Iterating',
{'Iteration': iteration,
'Loss': loss.data.tolist(),
'Images/s': (iteration * g_conf.BATCH_SIZE) / accumulated_time,
'BestLoss': best_loss, 'BestLossIteration': best_loss_iter,
'Output': output[position].data.tolist(),
'GroundTruth': dataset.extract_targets(data)[
position].data.tolist(),
'Error': error[position].data.tolist(),
'Inputs': dataset.extract_inputs(data)[
position].data.tolist()},
iteration)
loss_window.append(loss.data.tolist())
coil_logger.write_on_error_csv('train', loss.data)
print("Iteration: %d Loss: %f" % (iteration, loss.data))
coil_logger.add_message('Finished', {})
except KeyboardInterrupt:
coil_logger.add_message('Error', {'Message': 'Killed By User'})
except RuntimeError as e:
coil_logger.add_message('Error', {'Message': str(e)})
except:
traceback.print_exc()
coil_logger.add_message('Error', {'Message': 'Something Happened'})
def execute(gpu,
exp_batch,
exp_alias,
suppress_output=True,
number_of_workers=12):
"""
The main training function. This functions loads the latest checkpoint
for a given, exp_batch (folder) and exp_alias (experiment configuration).
With this checkpoint it starts from the beginning or continue some training.
Args:
gpu: The GPU number
exp_batch: the folder with the experiments
exp_alias: the alias, experiment name
suppress_output: if the output are going to be saved on a file
number_of_workers: the number of threads used for data loading
Returns:
None
"""
try:
# We set the visible cuda devices to select the GPU
os.environ["CUDA_VISIBLE_DEVICES"] = gpu
g_conf.VARIABLE_WEIGHT = {}
# At this point the log file with the correct naming is created.
# You merge the yaml file with the global configuration structure.
merge_with_yaml(os.path.join('configs', exp_batch,
exp_alias + '.yaml'))
set_type_of_process('train')
# Set the process into loading status.
coil_logger.add_message('Loading', {'GPU': gpu})
# Seed RNGs
torch.manual_seed(g_conf.MAGICAL_SEED)
random.seed(g_conf.MAGICAL_SEED)
# Put the output to a separate file if it is the case
if suppress_output:
if not os.path.exists('_output_logs'):
os.mkdir('_output_logs')
sys.stdout = open(os.path.join(
'_output_logs', exp_alias + '_' + g_conf.PROCESS_NAME + '_' +
str(os.getpid()) + ".out"),
"a",
buffering=1)
sys.stderr = open(os.path.join(
'_output_logs', exp_alias + '_err_' + g_conf.PROCESS_NAME +
'_' + str(os.getpid()) + ".out"),
"a",
buffering=1)
if coil_logger.check_finish('train'):
coil_logger.add_message('Finished', {})
return
# Preload option
if g_conf.PRELOAD_MODEL_ALIAS is not None:
checkpoint = torch.load(
os.path.join('_logs', g_conf.PRELOAD_MODEL_BATCH,
g_conf.PRELOAD_MODEL_ALIAS, 'checkpoints',
str(g_conf.PRELOAD_MODEL_CHECKPOINT) + '.pth'))
# Get the latest checkpoint to be loaded
# returns none if there are no checkpoints saved for this model
checkpoint_file = get_latest_saved_checkpoint()
if checkpoint_file is not None:
checkpoint = torch.load(
os.path.join('_logs', exp_batch, exp_alias, 'checkpoints',
str(get_latest_saved_checkpoint())))
iteration = checkpoint['iteration']
best_loss = checkpoint['best_loss']
best_loss_iter = checkpoint['best_loss_iter']
else:
iteration = 0
best_loss = 10000.0
best_loss_iter = 0
# Define the dataset.
# Can specify a list of training datasets or just a single training dataset
if len(g_conf.TRAIN_DATASET_NAMES) == 0:
train_dataset_list = [g_conf.TRAIN_DATASET_NAME]
else:
train_dataset_list = g_conf.TRAIN_DATASET_NAMES
full_dataset = [
os.path.join(os.environ["COIL_DATASET_PATH"], dataset_name)
for dataset_name in train_dataset_list
]
# By instantiating the augmenter we get a callable that augment images and transform them
# into tensors.
augmenter = Augmenter(g_conf.AUGMENTATION)
# Instantiate the class used to read a dataset. The coil dataset generator
# can be found
dataset = CoILDataset(full_dataset,
transform=augmenter,
preload_names=[
str(g_conf.NUMBER_OF_HOURS) + 'hours_' +
dataset_name
for dataset_name in train_dataset_list
],
train_dataset=True)
print("Loaded dataset")
# Create dataloader, model, and optimizer
data_loader = select_balancing_strategy(dataset, iteration,
number_of_workers)
model = CoILModel(g_conf.MODEL_TYPE, g_conf.MODEL_CONFIGURATION)
model.cuda()
optimizer = optim.Adam(model.parameters(), lr=g_conf.LEARNING_RATE)
# If we have a previous checkpoint, load model, optimizer, and record of previous
# train loss values (used for the learning rate schedule)
if checkpoint_file is not None or g_conf.PRELOAD_MODEL_ALIAS is not None:
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
accumulated_time = checkpoint['total_time']
loss_window = coil_logger.recover_loss_window('train', iteration)
else: # We accumulate iteration time and keep the average speed
accumulated_time = 0
loss_window = []
print("Before the loss")
# Define control loss function
criterion = Loss(g_conf.LOSS_FUNCTION)
if iteration == 0 and is_ready_to_save(iteration):
state = {
'iteration': iteration,
'state_dict': model.state_dict(),
'best_loss': best_loss,
'total_time': accumulated_time,
'optimizer': optimizer.state_dict(),
'best_loss_iter': best_loss_iter
}
torch.save(
state,
os.path.join('_logs', exp_batch, exp_alias, 'checkpoints',
str(iteration) + '.pth'))
# Training loop
for data in data_loader:
# Basically in this mode of execution, we validate every X Steps, if it goes up 3 times,
# add a stop on the _logs folder that is going to be read by this process
if g_conf.FINISH_ON_VALIDATION_STALE is not None and \
check_loss_validation_stopped(iteration, g_conf.FINISH_ON_VALIDATION_STALE):
break
"""
####################################
Main optimization loop
####################################
"""
iteration += 1
# Adjust learning rate based on training loss
if iteration % 1000 == 0:
adjust_learning_rate_auto(optimizer, loss_window)
capture_time = time.time()
model.zero_grad()
controls = data['directions']
# Run model forward and get outputs
# First case corresponds to training squeeze network, second case corresponds to training driving model without
# mimicking losses, last case corresponds to training mimic network
if "seg" in g_conf.SENSORS.keys():
branches = model(data,
dataset.extract_inputs(data).cuda(),
dataset.extract_intentions(data).cuda())
elif not g_conf.USE_REPRESENTATION_LOSS:
branches = model(data, dataset.extract_inputs(data).cuda())
else:
branches, intermediate_reps = model(
data,
dataset.extract_inputs(data).cuda())
# Compute control loss
targets_to_use = dataset.extract_targets(data)
loss_function_params = {
'branches': branches,
'targets': targets_to_use.cuda(),
'controls': controls.cuda(),
'inputs': dataset.extract_inputs(data).cuda(),
'branch_weights': g_conf.BRANCH_LOSS_WEIGHT,
'variable_weights': g_conf.VARIABLE_WEIGHT
}
loss, _ = criterion(loss_function_params)
# Compute mimicking loss
if g_conf.USE_REPRESENTATION_LOSS:
expert_reps = dataset.extract_representations(data)
# Seg mask mimicking loss
if g_conf.USE_PERCEPTION_REP_LOSS:
perception_rep_loss_elementwise = (
intermediate_reps[0] - expert_reps[0].cuda())**2
perception_rep_loss = g_conf.PERCEPTION_REP_WEIGHT * torch.sum(
perception_rep_loss_elementwise) / branches[0].shape[0]
else:
perception_rep_loss = torch.tensor(0.).cuda()
# Speed mimicking loss
if g_conf.USE_SPEED_REP_LOSS:
speed_rep_loss_elementwise = (intermediate_reps[1] -
expert_reps[1].cuda())**2
speed_rep_loss = g_conf.SPEED_REP_WEIGHT * torch.sum(
speed_rep_loss_elementwise) / branches[0].shape[0]
else:
speed_rep_loss = torch.tensor(0.).cuda()
# Stop intentions mimicking loss
if g_conf.USE_INTENTION_REP_LOSS:
intentions_rep_loss_elementwise = (
intermediate_reps[2] - expert_reps[2].cuda())**2
intentions_rep_loss = g_conf.INTENTIONS_REP_WEIGHT * torch.sum(
intentions_rep_loss_elementwise) / branches[0].shape[0]
else:
intentions_rep_loss = torch.tensor(0.).cuda()
rep_loss = g_conf.REP_LOSS_WEIGHT * (
perception_rep_loss + speed_rep_loss + intentions_rep_loss)
overall_loss = loss + rep_loss
else:
overall_loss = loss
overall_loss.backward()
optimizer.step()
"""
####################################
Saving the model if necessary
####################################
"""
if is_ready_to_save(iteration):
state = {
'iteration': iteration,
'state_dict': model.state_dict(),
'best_loss': best_loss,
'total_time': accumulated_time,
'optimizer': optimizer.state_dict(),
'best_loss_iter': best_loss_iter
}
torch.save(
state,
os.path.join('_logs', exp_batch, exp_alias, 'checkpoints',
str(iteration) + '.pth'))
"""
################################################
Adding tensorboard logs.
Making calculations for logging purposes.
These logs are monitored by the printer module.
#################################################
"""
coil_logger.add_scalar('Loss', loss.data, iteration)
if g_conf.USE_REPRESENTATION_LOSS:
coil_logger.add_scalar('Perception Rep Loss',
perception_rep_loss.data, iteration)
coil_logger.add_scalar('Speed Rep Loss', speed_rep_loss.data,
iteration)
coil_logger.add_scalar('Intentions Rep Loss',
intentions_rep_loss.data, iteration)
coil_logger.add_scalar('Overall Rep Loss', rep_loss.data,
iteration)
coil_logger.add_scalar('Total Loss', overall_loss.data,
iteration)
if 'rgb' in data:
coil_logger.add_image('Image', torch.squeeze(data['rgb']),
iteration)
if overall_loss.data < best_loss:
best_loss = overall_loss.data.tolist()
best_loss_iter = iteration
# Log a random position
position = random.randint(0, len(data) - 1)
output = model.extract_branch(torch.stack(branches[0:4]), controls)
error = torch.abs(output - targets_to_use.cuda())
accumulated_time += time.time() - capture_time
# Log to terminal and log file
if g_conf.USE_REPRESENTATION_LOSS:
coil_logger.add_message(
'Iterating', {
'Iteration':
iteration,
'Loss':
overall_loss.data.tolist(),
'Control Loss':
loss.data.tolist(),
'Rep Loss':
rep_loss.data.tolist(),
'Images/s':
(iteration * g_conf.BATCH_SIZE) / accumulated_time,
'BestLoss':
best_loss,
'BestLossIteration':
best_loss_iter,
'Output':
output[position].data.tolist(),
'GroundTruth':
targets_to_use[position].data.tolist(),
'Error':
error[position].data.tolist(),
'Inputs':
dataset.extract_inputs(data)[position].data.tolist()
}, iteration)
else:
coil_logger.add_message(
'Iterating', {
'Iteration':
iteration,
'Loss':
loss.data.tolist(),
'Images/s':
(iteration * g_conf.BATCH_SIZE) / accumulated_time,
'BestLoss':
best_loss,
'BestLossIteration':
best_loss_iter,
'Output':
output[position].data.tolist(),
'GroundTruth':
targets_to_use[position].data.tolist(),
'Error':
error[position].data.tolist(),
'Inputs':
dataset.extract_inputs(data)[position].data.tolist()
}, iteration)
# Save training loss history (useful for restoring training runs since learning rate is adjusted
# based on training loss)
loss_window.append(overall_loss.data.tolist())
coil_logger.write_on_error_csv('train', overall_loss.data)
print("Iteration: %d Loss: %f" % (iteration, overall_loss.data))
coil_logger.add_message('Finished', {})
except KeyboardInterrupt:
coil_logger.add_message('Error', {'Message': 'Killed By User'})
except RuntimeError as e:
coil_logger.add_message('Error', {'Message': str(e)})
except:
traceback.print_exc()
coil_logger.add_message('Error', {'Message': 'Something Happened'})
def execute(gpu, exp_batch, exp_alias):
# We set the visible cuda devices
try:
os.environ["CUDA_VISIBLE_DEVICES"] = gpu
# At this point the log file with the correct naming is created.
merge_with_yaml(os.path.join('configs', exp_batch,
exp_alias + '.yaml'))
set_type_of_process('train')
coil_logger.add_message('Loading', {'GPU': gpu})
if not os.path.exists('_output_logs'):
os.mkdir('_output_logs')
sys.stdout = open(os.path.join(
'_output_logs',
g_conf.PROCESS_NAME + '_' + str(os.getpid()) + ".out"),
"a",
buffering=1)
if monitorer.get_status(exp_batch, exp_alias + '.yaml',
g_conf.PROCESS_NAME)[0] == "Finished":
# TODO: print some cool summary or not ?
return
#Define the dataset. This structure is has the __get_item__ redefined in a way
#that you can access the HDFILES positions from the root directory as a in a vector.
full_dataset = os.path.join(os.environ["COIL_DATASET_PATH"],
g_conf.TRAIN_DATASET_NAME)
#augmenter_cpu = iag.AugmenterCPU(g_conf.AUGMENTATION_SUITE_CPU)
dataset = CoILDataset(full_dataset,
transform=transforms.Compose(
[transforms.ToTensor()]))
# Creates the sampler, this part is responsible for managing the keys. It divides
# all keys depending on the measurements and produces a set of keys for each bach.
sampler = BatchSequenceSampler(
splitter.control_steer_split(dataset.measurements,
dataset.meta_data), g_conf.BATCH_SIZE,
g_conf.NUMBER_IMAGES_SEQUENCE, g_conf.SEQUENCE_STRIDE)
# The data loader is the multi threaded module from pytorch that release a number of
# workers to get all the data.
# TODO: batch size an number of workers go to some configuration file
data_loader = torch.utils.data.DataLoader(dataset,
batch_sampler=sampler,
shuffle=False,
num_workers=12,
pin_memory=False)
# By instanciating the augmenter we get a callable that augment images and transform them
# into tensors.
st = lambda aug: iag.Sometimes(aug, 0.4)
oc = lambda aug: iag.Sometimes(aug, 0.3)
rl = lambda aug: iag.Sometimes(aug, 0.09)
augmenter = iag.Augmenter([iag.ToGPU()] + [
rl(iag.GaussianBlur(
(0, 1.5))), # blur images with a sigma between 0 and 1.5
rl(
iag.AdditiveGaussianNoise(
loc=0, scale=(0.0, 0.05),
per_channel=0.5)), # add gaussian noise to images
oc(iag.Dropout((0.0, 0.10), per_channel=0.5)
), # randomly remove up to X% of the pixels
oc(
iag.CoarseDropout(
(0.0, 0.10), size_percent=(0.08, 0.2), per_channel=0.5)
), # randomly remove up to X% of the pixels
oc(iag.Add((-40, 40), per_channel=0.5)
), # change brightness of images (by -X to Y of original value)
st(iag.Multiply((0.10, 2), per_channel=0.2)
), # change brightness of images (X-Y% of original value)
rl(iag.ContrastNormalization((
0.5, 1.5), per_channel=0.5)), # improve or worsen the contrast
rl(iag.Grayscale((0.0, 1))), # put grayscale
] # do all of the above in random order
)
# augmenter = iag.Augmenter(g_conf.AUGMENTATION_SUITE)
# TODO: here there is clearly a posibility to make a cool "conditioning" system.
model = CoILModel(g_conf.MODEL_NAME)
model.cuda()
print(model)
criterion = Loss()
# TODO: DATASET SIZE SEEMS WEIRD
optimizer = optim.Adam(model.parameters(), lr=0.0002)
checkpoint_file = get_latest_saved_checkpoint()
if checkpoint_file != None:
checkpoint = torch.load(
os.path.join('_logs', exp_batch, exp_alias, 'checkpoints',
str(get_latest_saved_checkpoint())))
iteration = checkpoint['iteration']
accumulated_time = checkpoint['total_time']
best_loss = checkpoint['best_loss']
best_loss_iter = checkpoint['best_loss_iter']
else:
iteration = 0
best_loss = 10000.0
accumulated_time = 0 # We accumulate iteration time and keep the average speed
best_loss_iter = 0
# TODO: The checkpoint will continue, so it should erase everything up to the iteration
best_loss_save = 10000.0
best_loss_save_iter = 0
curr_loss_save = 0.0
print(dataset.meta_data)
print(model)
capture_time = time.time()
model.train()
for data in data_loader:
input_data, float_data = data
#TODO, ADD ITERATION SCHEDULE
input_rgb_data = augmenter(0, input_data['rgb'])
augment_for_controls = 1
adjustlr = 1
if augment_for_controls: #and self._config.targets_names[j] == "Steer":
camera_angle = float_data[:, 26, :]
camera_angle = camera_angle.cuda(
) #self._config.variable_names.index('Angle'),i]
print("Camera angle", camera_angle[0])
steer = float_data[:, 0, :]
# print("Original", steer[0])
steer = steer.cuda()
speed = float_data[:, 10, :]
speed = speed.cuda()
# print (steer)
time_use = 1.0
car_length = 3.0
extra_factor = 2.5
threshold = 1.0
pos = camera_angle > 0.0
pos = pos.type(torch.FloatTensor)
neg = camera_angle <= 0.0
neg = neg.type(torch.FloatTensor)
pos = pos.cuda()
neg = neg.cuda()
rad_camera_angle = math.pi * (torch.abs(camera_angle)) / 180.0
val = extra_factor * (torch.atan(
(rad_camera_angle * car_length) /
(time_use * speed + 0.05))) / 3.1415
# print(val)
steer -= pos * torch.min(val, torch.tensor([0.6]).cuda())
steer += neg * torch.min(val, torch.tensor([0.6]).cuda())
print("val", val[0])
print("speed", speed[0])
steer = steer.cpu()
float_data[:, 0, :] = steer
float_data[:, 0, :][float_data[:, 0, :] > 1.0] = 1.0
float_data[:, 0, :][float_data[:, 0, :] < -1.0] = -1.0
#coil_logger.add_images(input_rgb_data)
# get the control commands from float_data, size = [120,1]
controls = float_data[:, dataset.controls_position(), :]
# print(" CONTROLS ", controls.shape)
# The output(branches) is a list of 5 branches results, each branch is with size [120,3]
model.zero_grad()
# print ( 'INPUTS', dataset.extract_inputs(float_data).shape )
branches = model(input_rgb_data,
dataset.extract_inputs(float_data).cuda())
#print ("len ",len(branches))
#targets = torch.cat([steer_gt, gas_gt, brake_gt], 1)
# print ("Extracted targets ", dataset.extract_targets(float_data).shape[0])
loss = criterion.MSELoss(
branches,
dataset.extract_targets(float_data).cuda(), controls.cuda(),
dataset.extract_inputs(float_data).cuda())
# TODO: All these logging things could go out to clean up the main
if loss.data < best_loss:
best_loss = loss.data.tolist()
best_loss_iter = iteration
curr_loss_save += loss.data
# Log a random position
position = random.randint(0, len(float_data) - 1)
output = model.extract_branch(torch.stack(branches[0:4]), controls)
error = torch.abs(output -
dataset.extract_targets(float_data).cuda())
# TODO: For now we are computing the error for just the correct branch, it could be multi- branch,
coil_logger.add_scalar('Loss', loss.data, iteration)
loss.backward()
optimizer.step()
accumulated_time += time.time() - capture_time
capture_time = time.time()
# TODO: Get only the float_data that are actually generating output
# TODO: itearation is repeating , and that is dumb
coil_logger.add_message(
'Iterating', {
'Iteration':
iteration,
'Loss':
loss.data.tolist(),
'Images/s':
(iteration * g_conf.BATCH_SIZE) / accumulated_time,
'BestLoss':
best_loss,
'BestLossIteration':
best_loss_iter,
'BestLossSave':
best_loss_save,
'Output':
output[position].data.tolist(),
'GroundTruth':
dataset.extract_targets(
float_data)[position].data.tolist(),
'Error':
error[position].data.tolist(),
'Inputs':
dataset.extract_inputs(float_data)[position].data.tolist()
}, iteration)
# TODO: For now we are computing the error for just the correct branch, it could be multi-branch,
# TODO: save also the optimizer state dictionary
if is_ready_to_save(iteration):
state = {
'iteration': iteration,
'state_dict': model.state_dict(),
'best_loss': best_loss,
'total_time': accumulated_time,
'best_loss_iter': best_loss_iter
}
# TODO : maybe already summarize the best model ???
torch.save(
state,
os.path.join('_logs', exp_batch, exp_alias, 'checkpoints',
str(iteration) + '.pth'))
print("before best save")
if iteration % 5 == 0 and iteration > 4:
curr_loss_save /= 5000.0
if curr_loss_save < best_loss_save:
best_loss_save = curr_loss_save
curr_loss_save = 0
state = {
'iteration': iteration,
'state_dict': model.state_dict(),
'best_loss': best_loss_save,
'total_time': accumulated_time,
'best_loss_iter': best_loss_save_iter
}
# TODO : maybe already summarize the best model ???
torch.save(
state,
os.path.join('_logs', exp_batch, exp_alias,
'best_loss_save' + '.pth'))
print("after best save")
if iteration == best_loss_iter:
state = {
'iteration': iteration,
'state_dict': model.state_dict(),
'best_loss': best_loss,
'total_time': accumulated_time,
'best_loss_iter': best_loss_iter
}
# TODO : maybe already summarize the best model ???
torch.save(
state,
os.path.join('_logs', exp_batch, exp_alias,
'best_loss' + '.pth'))
iteration += 1
if adjustlr and iteration % 1000:
adjust_learning_rate(optimizer, iteration)
except KeyboardInterrupt:
coil_logger.add_message('Error', {'Message': 'Killed By User'})
except:
traceback.print_exc()
coil_logger.add_message('Error', {'Message': 'Something Happened'})
def execute(gpu, exp_batch, exp_alias):
# We set the visible cuda devices
try:
os.environ["CUDA_VISIBLE_DEVICES"] = gpu
# At this point the log file with the correct naming is created.
merge_with_yaml(os.path.join('configs', exp_batch,
exp_alias + '.yaml'))
set_type_of_process('train')
coil_logger.add_message('Loading', {'GPU': gpu})
if not os.path.exists('_output_logs'):
os.mkdir('_output_logs')
sys.stdout = open(os.path.join(
'_output_logs',
g_conf.PROCESS_NAME + '_' + str(os.getpid()) + ".out"),
"a",
buffering=1)
if monitorer.get_status(exp_batch, exp_alias + '.yaml',
g_conf.PROCESS_NAME)[0] == "Finished":
# TODO: print some cool summary or not ?
return
#Define the dataset. This structure is has the __get_item__ redefined in a way
#that you can access the HDFILES positions from the root directory as a in a vector.
full_dataset = os.path.join(os.environ["COIL_DATASET_PATH"],
g_conf.TRAIN_DATASET_NAME)
#augmenter_cpu = iag.AugmenterCPU(g_conf.AUGMENTATION_SUITE_CPU)
dataset = CoILDataset(full_dataset,
transform=transforms.Compose(
[transforms.ToTensor()]))
# Creates the sampler, this part is responsible for managing the keys. It divides
# all keys depending on the measurements and produces a set of keys for each bach.
sampler = BatchSequenceSampler(
splitter.control_steer_split(dataset.measurements,
dataset.meta_data), g_conf.BATCH_SIZE,
g_conf.NUMBER_IMAGES_SEQUENCE, g_conf.SEQUENCE_STRIDE)
# The data loader is the multi threaded module from pytorch that release a number of
# workers to get all the data.
# TODO: batch size an number of workers go to some configuration file
data_loader = torch.utils.data.DataLoader(dataset,
batch_sampler=sampler,
shuffle=False,
num_workers=12,
pin_memory=True)
# By instanciating the augmenter we get a callable that augment images and transform them
# into tensors.
augmenter = iag.Augmenter(g_conf.AUGMENTATION_SUITE)
# TODO: here there is clearly a posibility to make a cool "conditioning" system.
model = CoILModel(g_conf.MODEL_NAME)
model.cuda()
exit()
print(model)
criterion = Loss()
# TODO: DATASET SIZE SEEMS WEIRD
optimizer = optim.SGD(model.parameters(), lr=0.0001, momentum=0.9)
checkpoint_file = get_latest_saved_checkpoint()
if checkpoint_file != None:
checkpoint = torch.load(
os.path.join('_logs', exp_batch, exp_alias, 'checkpoints',
str(get_latest_saved_checkpoint())))
iteration = checkpoint['iteration']
accumulated_time = checkpoint['total_time']
best_loss = checkpoint['best_loss']
best_loss_iter = checkpoint['best_loss_iter']
else:
iteration = 0
best_loss = 10000.0
accumulated_time = 0 # We accumulate iteration time and keep the average speed
best_loss_iter = 0
# TODO: The checkpoint will continue, so it should erase everything up to the iteration
print(dataset.meta_data)
print(model)
capture_time = time.time()
for data in data_loader:
input_data, float_data = data
#TODO, ADD ITERATION SCHEDULE
input_rgb_data = augmenter(0, input_data['rgb'])
#coil_logger.add_images(input_rgb_data)
# get the control commands from float_data, size = [120,1]
controls = float_data[:, dataset.controls_position(), :]
print(" CONTROLS ", controls.shape)
# The output(branches) is a list of 5 branches results, each branch is with size [120,3]
model.zero_grad()
print('INPUTS', dataset.extract_inputs(float_data).shape)
branches = model(input_rgb_data,
dataset.extract_inputs(float_data).cuda())
#print ("len ",len(branches))
#targets = torch.cat([steer_gt, gas_gt, brake_gt], 1)
print("Extracted targets ",
dataset.extract_targets(float_data).shape[0])
loss = criterion.MSELoss(
branches,
dataset.extract_targets(float_data).cuda(), controls.cuda(),
dataset.extract_inputs(float_data).cuda())
# TODO: All these logging things could go out to clean up the main
if loss.data < best_loss:
best_loss = loss.data.tolist()
best_loss_iter = iteration
# Log a random position
position = random.randint(0, len(float_data) - 1)
output = model.extract_branch(torch.stack(branches[0:4]), controls)
error = torch.abs(output -
dataset.extract_targets(float_data).cuda())
# TODO: For now we are computing the error for just the correct branch, it could be multi- branch,
coil_logger.add_scalar('Loss', loss.data, iteration)
loss.backward()
optimizer.step()
accumulated_time += time.time() - capture_time
capture_time = time.time()
# TODO: Get only the float_data that are actually generating output
# TODO: itearation is repeating , and that is dumb
coil_logger.add_message(
'Iterating', {
'Iteration':
iteration,
'Loss':
loss.data.tolist(),
'Images/s':
(iteration * g_conf.BATCH_SIZE) / accumulated_time,
'BestLoss':
best_loss,
'BestLossIteration':
best_loss_iter,
'Output':
output[position].data.tolist(),
'GroundTruth':
dataset.extract_targets(
float_data)[position].data.tolist(),
'Error':
error[position].data.tolist(),
'Inputs':
dataset.extract_inputs(float_data)[position].data.tolist()
}, iteration)
# TODO: For now we are computing the error for just the correct branch, it could be multi-branch,
# TODO: save also the optimizer state dictionary
if is_ready_to_save(iteration):
state = {
'iteration': iteration,
'state_dict': model.state_dict(),
'best_loss': best_loss,
'total_time': accumulated_time,
'best_loss_iter': best_loss_iter
}
# TODO : maybe already summarize the best model ???
torch.save(
state,
os.path.join('_logs', exp_batch, exp_alias, 'checkpoints',
str(iteration) + '.pth'))
iteration += 1
except KeyboardInterrupt:
coil_logger.add_message('Error', {'Message': 'Killed By User'})
except:
traceback.print_exc()
coil_logger.add_message('Error', {'Message': 'Something Happened'})
def execute(gpu, exp_batch, exp_alias, suppress_output=True):
# We set the visible cuda devices
# TODO: probable race condition, the train has to be started before.
try:
os.environ["CUDA_VISIBLE_DEVICES"] = gpu
# At this point the log file with the correct naming is created.
merge_with_yaml(os.path.join('configs', exp_batch,
exp_alias + '.yaml'))
set_type_of_process('train')
coil_logger.add_message('Loading', {'GPU': gpu})
if not os.path.exists('_output_logs'):
os.mkdir('_output_logs')
# Put the output to a separate file
if suppress_output:
sys.stdout = open(os.path.join(
'_output_logs',
g_conf.PROCESS_NAME + '_' + str(os.getpid()) + ".out"),
"a",
buffering=1)
checkpoint_file = get_latest_saved_checkpoint()
if checkpoint_file is not None:
checkpoint = torch.load(
os.path.join('_logs', exp_batch, exp_alias, 'checkpoints',
str(get_latest_saved_checkpoint())))
iteration = checkpoint['iteration']
best_loss = checkpoint['best_loss']
best_loss_iter = checkpoint['best_loss_iter']
else:
iteration = 0
best_loss = 10000.0
best_loss_iter = 0
# TODO: The checkpoint will continue, so it should erase everything up to the iteration on tensorboard
# Define the dataset. This structure is has the __get_item__ redefined in a way
# that you can access the HD_FILES positions from the root directory as a in a vector.
full_dataset = os.path.join(os.environ["COIL_DATASET_PATH"],
g_conf.TRAIN_DATASET_NAME)
# augmenter_cpu = iag.AugmenterCPU(g_conf.AUGMENTATION_SUITE_CPU)
# By instanciating the augmenter we get a callable that augment images and transform them
# into tensors.
augmenter = Augmenter(g_conf.AUGMENTATION)
dataset = CoILDataset(full_dataset, transform=augmenter)
data_loader = select_balancing_strategy(dataset, iteration)
model = CoILModel(g_conf.MODEL_TYPE, g_conf.MODEL_CONFIGURATION)
model.cuda()
if checkpoint_file is not None:
model.load_state_dict(checkpoint['state_dict'])
print(model)
criterion = Loss(g_conf.LOSS_FUNCTION)
optimizer = optim.Adam(model.parameters(), lr=g_conf.LEARNING_RATE)
print(dataset.meta_data)
print(model)
if checkpoint_file is not None:
accumulated_time = checkpoint['total_time']
else:
accumulated_time = 0 # We accumulate iteration time and keep the average speed
#TODO: test experiment continuation. Is the data sampler going to continue were it started.. ?
capture_time = time.time()
for data in data_loader:
input_data, float_data = data
# get the control commands from float_data, size = [120,1]
controls = float_data[:, dataset.controls_position(), :]
# The output(branches) is a list of 5 branches results, each branch is with size [120,3]
model.zero_grad()
branches = model(torch.squeeze(input_data['rgb'].cuda()),
dataset.extract_inputs(float_data).cuda())
loss = criterion(branches,
dataset.extract_targets(float_data).cuda(),
controls.cuda(),
dataset.extract_inputs(float_data).cuda(),
branch_weights=g_conf.BRANCH_LOSS_WEIGHT,
variable_weights=g_conf.VARIABLE_WEIGHT)
# TODO: All these logging things could go out to clean up the main
if loss.data < best_loss:
best_loss = loss.data.tolist()
best_loss_iter = iteration
# Log a random position
position = random.randint(0, len(float_data) - 1)
output = model.extract_branch(torch.stack(branches[0:4]), controls)
error = torch.abs(output -
dataset.extract_targets(float_data).cuda())
# TODO: For now we are computing the error for just the correct branch, it could be multi- branch,
coil_logger.add_scalar('Loss', loss.data, iteration)
coil_logger.add_image('Image', torch.squeeze(input_data['rgb']),
iteration)
loss.backward()
optimizer.step()
accumulated_time += time.time() - capture_time
capture_time = time.time()
# TODO: Get only the float_data that are actually generating output
# TODO: itearation is repeating , and that is dumb
coil_logger.add_message(
'Iterating', {
'Iteration':
iteration,
'Loss':
loss.data.tolist(),
'Images/s':
(iteration * g_conf.BATCH_SIZE) / accumulated_time,
'BestLoss':
best_loss,
'BestLossIteration':
best_loss_iter,
'Output':
output[position].data.tolist(),
'GroundTruth':
dataset.extract_targets(
float_data)[position].data.tolist(),
'Error':
error[position].data.tolist(),
'Inputs':
dataset.extract_inputs(float_data)[position].data.tolist()
}, iteration)
# TODO: For now we are computing the error for just the correct branch, it could be multi-branch,
# TODO: save also the optimizer state dictionary
if is_ready_to_save(iteration):
state = {
'iteration': iteration,
'state_dict': model.state_dict(),
'best_loss': best_loss,
'total_time': accumulated_time,
'best_loss_iter': best_loss_iter
}
# TODO : maybe already summarize the best model ???
torch.save(
state,
os.path.join('_logs', exp_batch, exp_alias, 'checkpoints',
str(iteration) + '.pth'))
iteration += 1
print(iteration)
if iteration % 1000 == 0:
adjust_learning_rate(optimizer, iteration)
del data
coil_logger.add_message('Finished', {})
except KeyboardInterrupt:
coil_logger.add_message('Error', {'Message': 'Killed By User'})
except:
traceback.print_exc()
coil_logger.add_message('Error', {'Message': 'Something Happened'})