def train_from_cfg(cfg: DictConfig) -> Type[nn.Module]:
rundir = os.getcwd() # done by hydra
device = torch.device(
"cuda:" +
str(cfg.compute.gpu_id) if torch.cuda.is_available() else "cpu")
torch.cuda.set_device(device)
log.info('Training sequence model...')
dataloaders = get_dataloaders_from_cfg(cfg, model_type='sequence')
utils.save_dict_to_yaml(dataloaders['split'],
os.path.join(rundir, 'split.yaml'))
log.debug('Num training batches {}, num val: {}'.format(
len(dataloaders['train']), len(dataloaders['val'])))
model = build_model_from_cfg(cfg,
dataloaders['num_features'],
dataloaders['num_classes'],
pos=dataloaders['pos'],
neg=dataloaders['neg'])
weights = projects.get_weightfile_from_cfg(cfg, model_type='sequence')
if weights is not None:
model = utils.load_weights(model, weights)
model = model.to(device)
log.info('Total trainable params: {:,}'.format(
utils.get_num_parameters(model)))
optimizer = optim.Adam(filter(lambda p: p.requires_grad,
model.parameters()),
lr=cfg.train.lr)
torch.save(model, os.path.join(rundir,
cfg.sequence.arch + '_definition.pt'))
stopper = get_stopper(cfg)
scheduler = initialize_scheduler(
optimizer,
cfg,
mode='max',
reduction_factor=cfg.train.reduction_factor)
metrics = get_metrics(rundir,
num_classes=len(cfg.project.class_names),
num_parameters=utils.get_num_parameters(model),
key_metric='f1')
criterion = get_criterion(cfg.feature_extractor.final_activation,
dataloaders, device)
steps_per_epoch = dict(cfg.train.steps_per_epoch)
model = train(model,
dataloaders,
criterion,
optimizer,
metrics,
scheduler,
rundir,
stopper,
device,
steps_per_epoch,
final_activation=cfg.feature_extractor.final_activation,
sequence=True,
normalizer=None)
def get_cnn(model_name: str,
in_channels: int = 3,
reload_imagenet: bool = True,
num_classes: int = 1000,
freeze: bool = False,
pos: np.ndarray = None,
neg: np.ndarray = None,
**kwargs):
""" Initializes a pretrained CNN from Torchvision.
Currently supported models:
AlexNet, DenseNet, Inception, VGGXX, ResNets, SqueezeNets, and Resnet3Ds (not torchvision)
Args:
model_name (str):
in_channels (int): number of input channels. If not 3, the per-channel weights will be averaged and replicated
in_channels times
reload_imagenet (bool): if True, reload imagenet weights from Torchvision
num_classes (int): number of output classes (neurons in final FC layer)
freeze (bool): if true, model weights will be freezed
pos (np.ndarray): number of positive examples in training set. Used for custom bias initialization in
final layer
neg (np.ndarray): number of negative examples in training set. Used for custom bias initialization in
final layer
**kwargs (): passed to model initialization function
Returns:
model: a pytorch CNN
"""
model = models[model_name](pretrained=reload_imagenet,
in_channels=in_channels,
**kwargs)
if freeze:
log.info('Before freezing: {:,}'.format(
utils.get_num_parameters(model)))
for param in model.parameters():
param.requires_grad = False
log.info('After freezing: {:,}'.format(
utils.get_num_parameters(model)))
# we have to use the pop function because the final layer in these models has different names
model, num_features, final_layer = pop(model, model_name, 1)
linear_layer = nn.Linear(num_features, num_classes)
# initialize bias to roughly approximate the probability of positive examples in the training set
# https://www.tensorflow.org/tutorials/structured_data/imbalanced_data#optional_set_the_correct_initial_bias
if pos is not None and neg is not None:
with torch.no_grad():
bias = np.nan_to_num(np.log(pos / neg), neginf=0.0)
bias = torch.nn.Parameter(torch.from_numpy(bias).float())
linear_layer.bias = bias
log.info('Custom bias: {}'.format(linear_layer.bias))
model = nn.Sequential(model, linear_layer)
return model
def flow_generator_train(cfg: DictConfig) -> nn.Module:
"""Trains flow generator models from a configuration.
Parameters
----------
cfg : DictConfig
Configuration, e.g. that returned by deepethogram.configration.make_flow_generator_train_cfg
Returns
-------
nn.Module
Trained flow generator
"""
cfg = projects.setup_run(cfg)
log.info('args: {}'.format(' '.join(sys.argv)))
# only two custom overwrites of the configuration file
# allow for editing
OmegaConf.set_struct(cfg, False)
# second, use the model directory to find the most recent run of each model type
# cfg = projects.overwrite_cfg_with_latest_weights(cfg, cfg.project.model_path, model_type='flow_generator')
# SHOULD NEVER MODIFY / MAKE ASSIGNMENTS TO THE CFG OBJECT AFTER RIGHT HERE!
log.info('configuration used ~~~~~')
log.info(OmegaConf.to_yaml(cfg))
datasets, data_info = get_datasets_from_cfg(
cfg, 'flow_generator', input_images=cfg.flow_generator.n_rgb)
flow_generator = build_model_from_cfg(cfg)
log.info('Total trainable params: {:,}'.format(
utils.get_num_parameters(flow_generator)))
utils.save_dict_to_yaml(data_info['split'],
os.path.join(os.getcwd(), 'split.yaml'))
flow_weights = deepethogram.projects.get_weightfile_from_cfg(
cfg, 'flow_generator')
if flow_weights is not None:
print('reloading weights...')
flow_generator = utils.load_weights(flow_generator,
flow_weights,
device='cpu')
stopper = get_stopper(cfg)
metrics = get_metrics(cfg, os.getcwd(),
utils.get_num_parameters(flow_generator))
lightning_module = OpticalFlowLightning(flow_generator, cfg, datasets,
metrics,
viz.visualize_logger_optical_flow)
trainer = get_trainer_from_cfg(cfg, lightning_module, stopper)
trainer.fit(lightning_module)
return flow_generator
def sequence_train(cfg: DictConfig) -> nn.Module:
"""Trains sequence models from a configuration.
Parameters
----------
cfg : DictConfig
Configuration, e.g. that returned by deepethogram.configration.make_sequence_train_cfg
Returns
-------
nn.Module
Trained sequence model
"""
cfg = projects.setup_run(cfg)
log.info('args: {}'.format(' '.join(sys.argv)))
if cfg.sequence.latent_name is None:
cfg.sequence.latent_name = cfg.feature_extractor.arch
# allow for editing
OmegaConf.set_struct(cfg, False)
log.info('Configuration used: ')
log.info(OmegaConf.to_yaml(cfg))
datasets, data_info = get_datasets_from_cfg(cfg, 'sequence')
utils.save_dict_to_yaml(data_info['split'], os.path.join(os.getcwd(), 'split.yaml'))
model = build_model_from_cfg(cfg,
data_info['num_features'],
data_info['num_classes'],
pos=data_info['pos'],
neg=data_info['neg'])
weights = projects.get_weightfile_from_cfg(cfg, model_type='sequence')
if weights is not None:
model = utils.load_weights(model, weights)
log.debug('model arch: {}'.format(model))
log.info('Total trainable params: {:,}'.format(utils.get_num_parameters(model)))
stopper = get_stopper(cfg)
metrics = get_metrics(os.getcwd(),
data_info['num_classes'],
num_parameters=utils.get_num_parameters(model),
key_metric='f1_class_mean',
num_workers=cfg.compute.metrics_workers)
criterion = get_criterion(cfg, model, data_info)
lightning_module = SequenceLightning(model, cfg, datasets, metrics, criterion)
# change auto batch size parameters because large sequences can overflow RAM
trainer = get_trainer_from_cfg(cfg, lightning_module, stopper)
trainer.fit(lightning_module)
return model
def train_from_cfg(cfg: DictConfig) -> Type[nn.Module]:
device = torch.device(
"cuda:" +
str(cfg.compute.gpu_id) if torch.cuda.is_available() else "cpu")
if device != 'cpu': torch.cuda.set_device(device)
log.info('Training flow generator....')
dataloaders = get_dataloaders_from_cfg(
cfg,
model_type='flow_generator',
input_images=cfg.flow_generator.n_rgb)
# print(dataloaders)
log.info('Num training batches {}, num val: {}'.format(
len(dataloaders['train']), len(dataloaders['val'])))
flow_generator = build_model_from_cfg(cfg)
flow_generator = flow_generator.to(device)
log.info('Total trainable params: {:,}'.format(
utils.get_num_parameters(flow_generator)))
rundir = os.getcwd() # this is configured by hydra
# save model definition
torch.save(
flow_generator,
os.path.join(rundir, cfg.flow_generator.arch + '_definition.pt'))
utils.save_dict_to_yaml(dataloaders['split'],
os.path.join(rundir, 'split.yaml'))
optimizer = optim.Adam(filter(lambda p: p.requires_grad,
flow_generator.parameters()),
lr=cfg.train.lr)
flow_weights = deepethogram.projects.get_weightfile_from_cfg(
cfg, 'flow_generator')
if flow_weights is not None:
print('reloading weights...')
flow_generator = utils.load_weights(flow_generator,
flow_weights,
device=device)
# stopper, early_stopping_begins = get_stopper(cfg)
stopper = get_stopper(cfg)
scheduler = initialize_scheduler(optimizer, cfg, mode='min')
if cfg.flow_generator.loss == 'MotionNet':
criterion = MotionNetLoss(
flow_sparsity=cfg.flow_generator.flow_sparsity,
sparsity_weight=cfg.flow_generator.sparsity_weight,
smooth_weight_multiplier=cfg.flow_generator.
smooth_weight_multiplier)
else:
raise NotImplementedError
metrics = get_metrics(cfg, rundir,
utils.get_num_parameters(flow_generator))
reconstructor = Reconstructor(cfg)
steps_per_epoch = cfg.train.steps_per_epoch
if cfg.compute.fp16:
assert torch_amp, 'must install torch 1.6 or greater to use FP16 training'
flow_generator = train(flow_generator,
dataloaders,
criterion,
optimizer,
metrics,
scheduler,
reconstructor,
rundir,
stopper,
device,
num_epochs=cfg.train.num_epochs,
steps_per_epoch=steps_per_epoch['train'],
steps_per_validation_epoch=steps_per_epoch['val'],
steps_per_test_epoch=steps_per_epoch['test'],
max_flow=cfg.flow_generator.max,
dali=cfg.compute.dali,
fp16=cfg.compute.fp16)
return flow_generator
def train_from_cfg(cfg: DictConfig) -> Type[nn.Module]:
""" train DeepEthogram feature extractors from a configuration object.
Args:
cfg (DictConfig): configuration object generated by Hydra
Returns:
trained feature extractor
"""
rundir = os.getcwd() # done by hydra
device = torch.device(
"cuda:" +
str(cfg.compute.gpu_id) if torch.cuda.is_available() else "cpu")
if device != 'cpu': torch.cuda.set_device(device)
flow_generator = build_flow_generator(cfg)
flow_weights = get_weightfile_from_cfg(cfg, 'flow_generator')
assert flow_weights is not None, (
'Must have a valid weightfile for flow generator. Use '
'deepethogram.flow_generator.train or cfg.reload.latest')
log.info('loading flow generator from file {}'.format(flow_weights))
flow_generator = utils.load_weights(flow_generator,
flow_weights,
device=device)
flow_generator = flow_generator.to(device)
dataloaders = get_dataloaders_from_cfg(
cfg,
model_type='feature_extractor',
input_images=cfg.feature_extractor.n_flows + 1)
spatial_classifier, flow_classifier = build_model_from_cfg(
cfg,
return_components=True,
pos=dataloaders['pos'],
neg=dataloaders['neg'])
spatial_classifier = spatial_classifier.to(device)
flow_classifier = flow_classifier.to(device)
num_classes = len(cfg.project.class_names)
utils.save_dict_to_yaml(dataloaders['split'],
os.path.join(rundir, 'split.yaml'))
criterion = get_criterion(cfg.feature_extractor.final_activation,
dataloaders, device)
steps_per_epoch = dict(cfg.train.steps_per_epoch)
metrics = get_metrics(
rundir,
num_classes=num_classes,
num_parameters=utils.get_num_parameters(spatial_classifier))
dali = cfg.compute.dali
# training in a curriculum goes as follows:
# first, we train the spatial classifier, which takes still images as input
# second, we train the flow classifier, which generates optic flow with the flow_generator model and then classifies
# it. Thirdly, we will train the whole thing end to end
# Without the curriculum we just train end to end from the start
if cfg.feature_extractor.curriculum:
del dataloaders
# train spatial model, then flow model, then both end-to-end
dataloaders = get_dataloaders_from_cfg(
cfg,
model_type='feature_extractor',
input_images=cfg.feature_extractor.n_rgb)
log.info('Num training batches {}, num val: {}'.format(
len(dataloaders['train']), len(dataloaders['val'])))
# we'll use this to visualize our data, because it is loaded z-scored. we want it to be in the range [0-1] or
# [0-255] for visualization, and for that we need to know mean and std
normalizer = get_normalizer(cfg,
input_images=cfg.feature_extractor.n_rgb)
optimizer = optim.Adam(filter(lambda p: p.requires_grad,
spatial_classifier.parameters()),
lr=cfg.train.lr,
weight_decay=cfg.feature_extractor.weight_decay)
spatialdir = os.path.join(rundir, 'spatial')
if not os.path.isdir(spatialdir):
os.makedirs(spatialdir)
stopper = get_stopper(cfg)
# we're using validation loss as our key metric
scheduler = initialize_scheduler(
optimizer,
cfg,
mode='min',
reduction_factor=cfg.train.reduction_factor)
log.info('key metric: {}'.format(metrics.key_metric))
spatial_classifier = train(
spatial_classifier,
dataloaders,
criterion,
optimizer,
metrics,
scheduler,
spatialdir,
stopper,
device,
steps_per_epoch,
final_activation=cfg.feature_extractor.final_activation,
sequence=False,
normalizer=normalizer,
dali=dali)
log.info('Training flow stream....')
input_images = cfg.feature_extractor.n_flows + 1
del dataloaders
dataloaders = get_dataloaders_from_cfg(cfg,
model_type='feature_extractor',
input_images=input_images)
normalizer = get_normalizer(cfg, input_images=input_images)
log.info('Num training batches {}, num val: {}'.format(
len(dataloaders['train']), len(dataloaders['val'])))
flowdir = os.path.join(rundir, 'flow')
if not os.path.isdir(flowdir):
os.makedirs(flowdir)
flow_generator_and_classifier = FlowOnlyClassifier(
flow_generator, flow_classifier).to(device)
optimizer = optim.Adam(filter(lambda p: p.requires_grad,
flow_classifier.parameters()),
lr=cfg.train.lr,
weight_decay=cfg.feature_extractor.weight_decay)
stopper = get_stopper(cfg)
# we're using validation loss as our key metric
scheduler = initialize_scheduler(
optimizer,
cfg,
mode='min',
reduction_factor=cfg.train.reduction_factor)
flow_generator_and_classifier = train(
flow_generator_and_classifier,
dataloaders,
criterion,
optimizer,
metrics,
scheduler,
flowdir,
stopper,
device,
steps_per_epoch,
final_activation=cfg.feature_extractor.final_activation,
sequence=False,
normalizer=normalizer,
dali=dali)
flow_classifier = flow_generator_and_classifier.flow_classifier
# overwrite checkpoint
utils.checkpoint(flow_classifier, flowdir, stopper.epoch_counter)
model = HiddenTwoStream(flow_generator,
spatial_classifier,
flow_classifier,
cfg.feature_extractor.arch,
fusion_style=cfg.feature_extractor.fusion,
num_classes=num_classes).to(device)
# setting the mode to end-to-end would allow to backprop gradients into the flow generator itself
# the paper does this, but I don't expect that users would have enough data for this to make sense
model.set_mode('classifier')
log.info('Training end to end...')
input_images = cfg.feature_extractor.n_flows + 1
dataloaders = get_dataloaders_from_cfg(cfg,
model_type='feature_extractor',
input_images=input_images)
normalizer = get_normalizer(cfg, input_images=input_images)
optimizer = optim.Adam(filter(lambda p: p.requires_grad,
model.parameters()),
lr=cfg.train.lr,
weight_decay=cfg.feature_extractor.weight_decay)
stopper = get_stopper(cfg)
# we're using validation loss as our key metric
scheduler = initialize_scheduler(
optimizer,
cfg,
mode='min',
reduction_factor=cfg.train.reduction_factor)
log.info('Total trainable params: {:,}'.format(
utils.get_num_parameters(model)))
model = train(model,
dataloaders,
criterion,
optimizer,
metrics,
scheduler,
rundir,
stopper,
device,
steps_per_epoch,
final_activation=cfg.feature_extractor.final_activation,
sequence=False,
normalizer=normalizer,
dali=dali)
utils.save_hidden_two_stream(model, rundir, dict(cfg),
stopper.epoch_counter)
return model
def feature_extractor_train(cfg: DictConfig) -> nn.Module:
"""Trains feature extractor models from a configuration.
Parameters
----------
cfg : DictConfig
Configuration, e.g. that returned by deepethogram.configration.make_feature_extractor_train_cfg
Returns
-------
nn.Module
Trained feature extractor
"""
# rundir = os.getcwd()
cfg = projects.setup_run(cfg)
log.info('args: {}'.format(' '.join(sys.argv)))
# change the project paths from relative to absolute
# allow for editing
OmegaConf.set_struct(cfg, False)
# SHOULD NEVER MODIFY / MAKE ASSIGNMENTS TO THE CFG OBJECT AFTER RIGHT HERE!
log.info('configuration used ~~~~~')
log.info(OmegaConf.to_yaml(cfg))
# we build flow generator independently because you might want to load it from a different location
flow_generator = build_flow_generator(cfg)
flow_weights = projects.get_weightfile_from_cfg(cfg, 'flow_generator')
assert flow_weights is not None, (
'Must have a valid weightfile for flow generator. Use '
'deepethogram.flow_generator.train or cfg.reload.latest')
log.info('loading flow generator from file {}'.format(flow_weights))
flow_generator = utils.load_weights(flow_generator, flow_weights)
_, data_info = get_datasets_from_cfg(
cfg,
model_type='feature_extractor',
input_images=cfg.feature_extractor.n_flows + 1)
model_parts = build_model_from_cfg(cfg,
pos=data_info['pos'],
neg=data_info['neg'])
_, spatial_classifier, flow_classifier, fusion, model = model_parts
# log.info('model: {}'.format(model))
num_classes = len(cfg.project.class_names)
utils.save_dict_to_yaml(data_info['split'],
os.path.join(cfg.run.dir, 'split.yaml'))
metrics = get_metrics(
cfg.run.dir,
num_classes=num_classes,
num_parameters=utils.get_num_parameters(spatial_classifier),
key_metric='f1_class_mean_nobg',
num_workers=cfg.compute.metrics_workers)
# cfg.compute.batch_size will be changed by the automatic batch size finder, possibly. store here so that
# with each step of the curriculum, we can auto-tune it
original_batch_size = cfg.compute.batch_size
original_lr = cfg.train.lr
# training in a curriculum goes as follows:
# first, we train the spatial classifier, which takes still images as input
# second, we train the flow classifier, which generates optic flow with the flow_generator model and then classifies
# it. Thirdly, we will train the whole thing end to end
# Without the curriculum we just train end to end from the start
if cfg.feature_extractor.curriculum:
# train spatial model, then flow model, then both end-to-end
# dataloaders = get_dataloaders_from_cfg(cfg, model_type='feature_extractor',
# input_images=cfg.feature_extractor.n_rgb)
datasets, data_info = get_datasets_from_cfg(
cfg,
model_type='feature_extractor',
input_images=cfg.feature_extractor.n_rgb)
stopper = get_stopper(cfg)
criterion = get_criterion(cfg, spatial_classifier, data_info)
lightning_module = HiddenTwoStreamLightning(spatial_classifier, cfg,
datasets, metrics,
criterion)
trainer = get_trainer_from_cfg(cfg, lightning_module, stopper)
# this horrible syntax is because we just changed our configuration's batch size and learning rate, if they are
# set to auto. so we need to re-instantiate our lightning module
# https://pytorch-lightning.readthedocs.io/en/latest/lr_finder.html?highlight=auto%20scale%20learning%20rate
# I tried to do this without re-creating module, but finding the learning rate increments the epoch??
# del lightning_module
# log.info('epoch num: {}'.format(trainer.current_epoch))
# lightning_module = HiddenTwoStreamLightning(spatial_classifier, cfg, datasets, metrics, criterion)
trainer.fit(lightning_module)
# free RAM. note: this doesn't do much
log.info('free ram')
del datasets, lightning_module, trainer, stopper, data_info
torch.cuda.empty_cache()
gc.collect()
# return
datasets, data_info = get_datasets_from_cfg(
cfg,
model_type='feature_extractor',
input_images=cfg.feature_extractor.n_flows + 1)
# re-initialize stopper so that it doesn't think we need to stop due to the previous model
stopper = get_stopper(cfg)
cfg.compute.batch_size = original_batch_size
cfg.train.lr = original_lr
# this class will freeze the flow generator
flow_generator_and_classifier = FlowOnlyClassifier(
flow_generator, flow_classifier)
criterion = get_criterion(cfg, flow_generator_and_classifier,
data_info)
lightning_module = HiddenTwoStreamLightning(
flow_generator_and_classifier, cfg, datasets, metrics, criterion)
trainer = get_trainer_from_cfg(cfg, lightning_module, stopper)
# lightning_module = HiddenTwoStreamLightning(flow_generator_and_classifier, cfg, datasets, metrics, criterion)
trainer.fit(lightning_module)
del datasets, lightning_module, trainer, stopper, data_info
torch.cuda.empty_cache()
gc.collect()
torch.cuda.empty_cache()
gc.collect()
model = HiddenTwoStream(flow_generator, spatial_classifier,
flow_classifier, fusion,
cfg.feature_extractor.arch)
model.set_mode('classifier')
datasets, data_info = get_datasets_from_cfg(
cfg,
model_type='feature_extractor',
input_images=cfg.feature_extractor.n_flows + 1)
criterion = get_criterion(cfg, model, data_info)
stopper = get_stopper(cfg)
cfg.compute.batch_size = original_batch_size
cfg.train.lr = original_lr
# log.warning('SETTING ANAOMALY DETECTION TO TRUE! WILL SLOW DOWN.')
# torch.autograd.set_detect_anomaly(True)
lightning_module = HiddenTwoStreamLightning(model, cfg, datasets, metrics,
criterion)
trainer = get_trainer_from_cfg(cfg, lightning_module, stopper)
# see above for horrible syntax explanation
# lightning_module = HiddenTwoStreamLightning(model, cfg, datasets, metrics, criterion)
trainer.fit(lightning_module)
# trainer.test(model=lightning_module)
return model