def inference(cfg):
# # Set up environment.
du.init_distributed_training(cfg)
# Set random seed from configs.
np.random.seed(cfg.RNG_SEED)
torch.manual_seed(cfg.RNG_SEED)
# Build the video model and print model statistics.
model = build_model(cfg)
if du.is_master_proc() and cfg.LOG_MODEL_INFO:
misc.log_model_info(model, cfg, use_train_input=False)
cu.load_test_checkpoint(cfg, model)
# Create video loaders.
video_loader = loader.construct_loader(cfg, "test")
# Create saver
saver = Saver(cfg.DATA.PATH_TO_DATA_DIR, video_loader.dataset)
model.eval()
for i, (inputs, index) in tqdm(enumerate(video_loader),
total=len(video_loader)):
for i in range(len(inputs)):
inputs[i] = inputs[i].cuda(non_blocking=True)
index = index.cuda()
feats = model(inputs)
# Gather all the predictions across all the devices to perform ensemble.
if cfg.NUM_GPUS > 1:
feats, index = du.all_gather([feats, index])
saver.save(feats, index)
saver.merge()
def saveOnnxModel(self):
model = build_model(self.cfg)
optimizer = optim.construct_optimizer(model, self.cfg)
start_epoch = cu.load_train_checkpoint(self.cfg, model, optimizer,
self.logger)
self.cfg.TRAIN['BATCH_SIZE'] = self.cfg.ONNX.BATCH_SIZE
dl = loader.construct_loader(self.cfg, "train")
inputs, labels, _, _ = next(iter(dl))
if isinstance(inputs, (list, )):
for i in range(len(inputs)):
inputs[i].to(self.onnxDevice)
model.to(torch.device(self.onnxDevice))
model.eval()
onnxPath, _ = self.getOnnxModelPath()
with torch.no_grad():
torch.onnx.export(
model,
inputs,
onnxPath,
opset_version=self.cfg.ONNX.OPSET_VER,
verbose=True,
input_names=self.cfg.ONNX.INPUT_NAMES,
output_names=self.cfg.ONNX.OUTPUT_NAMES,
)
self.logger.info("Exported {}".format(onnxPath))
def test(cfg):
"""
Perform multi-view testing on the pretrained video model.
Args:
cfg (CfgNode): configs. Details can be found in
slowfast/config/defaults.py
"""
# Set up environment.
du.init_distributed_training(cfg)
# Set random seed from configs.
np.random.seed(cfg.RNG_SEED)
torch.manual_seed(cfg.RNG_SEED)
# Setup logging format.
logging.setup_logging(cfg.OUTPUT_DIR)
# Print config.
logger.info("Test with config:")
logger.info(cfg)
# Build the video model and print model statistics.
model = build_model(cfg)
if du.is_master_proc() and cfg.LOG_MODEL_INFO:
misc.log_model_info(model, cfg, use_train_input=False)
cu.load_test_checkpoint(cfg, model)
# Create video testing loaders.
test_loader = loader.construct_loader(cfg, "test")
logger.info("Testing model for {} iterations".format(len(test_loader)))
if cfg.DETECTION.ENABLE:
assert cfg.NUM_GPUS == cfg.TEST.BATCH_SIZE or cfg.NUM_GPUS == 0
test_meter = AVAMeter(len(test_loader), cfg, mode="test")
else:
assert (
test_loader.dataset.num_videos %
(cfg.TEST.NUM_ENSEMBLE_VIEWS * cfg.TEST.NUM_SPATIAL_CROPS) == 0)
# Create meters for multi-view testing.
test_meter = TestMeter(
test_loader.dataset.num_videos //
(cfg.TEST.NUM_ENSEMBLE_VIEWS * cfg.TEST.NUM_SPATIAL_CROPS),
cfg.TEST.NUM_ENSEMBLE_VIEWS * cfg.TEST.NUM_SPATIAL_CROPS,
cfg.MODEL.NUM_CLASSES,
len(test_loader),
cfg.DATA.MULTI_LABEL,
cfg.DATA.ENSEMBLE_METHOD,
)
# Set up writer for logging to Tensorboard format.
if cfg.TENSORBOARD.ENABLE and du.is_master_proc(
cfg.NUM_GPUS * cfg.NUM_SHARDS):
writer = tb.TensorboardWriter(cfg)
else:
writer = None
# # Perform multi-view test on the entire dataset.
test_meter = perform_test(test_loader, model, test_meter, cfg, writer)
if writer is not None:
writer.close()
def test_from_train(model, cfg, cnt=-1):
test_loader = loader.construct_loader(cfg, "test")
logger.info("Testing model for {} iterations".format(len(test_loader)))
if cfg.DETECTION.ENABLE:
assert cfg.NUM_GPUS == cfg.TEST.BATCH_SIZE
test_meter = AVAMeter(len(test_loader), cfg, mode="test")
else:
assert (
len(test_loader.dataset)
% (cfg.TEST.NUM_ENSEMBLE_VIEWS * cfg.TEST.NUM_SPATIAL_CROPS)
== 0
)
# Create meters for multi-view testing.
if cfg.TEST.DATASET == 'epickitchens':
test_meter = EPICTestMeter(
len(test_loader.dataset)
// (cfg.TEST.NUM_ENSEMBLE_VIEWS * cfg.TEST.NUM_SPATIAL_CROPS),
cfg.TEST.NUM_ENSEMBLE_VIEWS * cfg.TEST.NUM_SPATIAL_CROPS,
cfg.MODEL.NUM_CLASSES,
len(test_loader),
)
else:
test_meter = TestMeter(
len(test_loader.dataset)
// (cfg.TEST.NUM_ENSEMBLE_VIEWS * cfg.TEST.NUM_SPATIAL_CROPS),
cfg.TEST.NUM_ENSEMBLE_VIEWS * cfg.TEST.NUM_SPATIAL_CROPS,
cfg.MODEL.NUM_CLASSES,
len(test_loader),
)
# # Perform multi-view test on the entire dataset.
scores_path = os.path.join(cfg.OUTPUT_DIR, 'scores')
if not os.path.exists(scores_path):
os.makedirs(scores_path)
filename_root = cfg.EPICKITCHENS.TEST_LIST.split('.')[0]
if cnt >= 0:
file_name = '{}_{}_{}.pkl'.format(filename_root, cnt, cfg.MODEL.MODEL_NAME)
else:
file_name = '{}_{}_{}.pkl'.format(filename_root, 'test_only', cfg.MODEL.MODEL_NAME)
file_path = os.path.join(scores_path, file_name)
logger.info(file_path)
pickle.dump([], open(file_path, 'wb+'))
preds, labels, metadata = perform_test(test_loader, model, test_meter, cfg)
if du.is_master_proc():
if cfg.TEST.DATASET == 'epickitchens':
results = {'verb_output': preds[0],
'noun_output': preds[1],
'verb_gt': labels[0],
'noun_gt': labels[1],
'narration_id': metadata}
scores_path = os.path.join(cfg.OUTPUT_DIR, 'scores')
if not os.path.exists(scores_path):
os.makedirs(scores_path)
pickle.dump(results, open(file_path, 'wb'))
def build_trainer(cfg):
"""
Build training model and its associated tools, including optimizer,
dataloaders and meters.
Args:
cfg (CfgNode): configs. Details can be found in
slowfast/config/defaults.py
Returns:
model (nn.Module): training model.
optimizer (Optimizer): optimizer.
train_loader (DataLoader): training data loader.
val_loader (DataLoader): validatoin data loader.
precise_bn_loader (DataLoader): training data loader for computing
precise BN.
train_meter (TrainMeter): tool for measuring training stats.
val_meter (ValMeter): tool for measuring validation stats.
"""
# Build the video model and print model statistics.
model = build_model(cfg)
if du.is_master_proc() and cfg.LOG_MODEL_INFO:
misc.log_model_info(model, cfg, use_train_input=True)
# Construct the optimizer.
optimizer = optim.construct_optimizer(model, cfg)
# Create the video train and val loaders.
train_loader = loader.construct_loader(cfg, "train")
val_loader = loader.construct_loader(cfg, "val")
precise_bn_loader = loader.construct_loader(cfg,
"train",
is_precise_bn=True)
# Create meters.
train_meter = TrainMeter(len(train_loader), cfg)
val_meter = ValMeter(len(val_loader), cfg)
return (
model,
optimizer,
train_loader,
val_loader,
precise_bn_loader,
train_meter,
val_meter,
)
def train():
# Create model.
model = build_model(configs.backbone)
# Construct the optimizer.
optimizer = torch.optim.SGD(model.parameters(),
0.1,
0.9,
weight_decay=0.0001,
nesterov=True)
start_epoch = cu.load(model, optimizer)
# Create the video train and val loaders.
train_loader = loader.construct_loader('train')
val_loader = loader.construct_loader('val')
# Train.
for epoch in range(start_epoch, configs.max_epoch):
train_epoch(train_loader, model, optimizer, epoch)
eval_epoch(val_loader, model, epoch)
cu.save(model, optimizer, epoch)
def visualize(cfg):
"""
Perform layer weights and activations visualization on the model.
Args:
cfg (CfgNode): configs. Details can be found in
slowfast/config/defaults.py
"""
if cfg.TENSORBOARD.ENABLE and cfg.TENSORBOARD.MODEL_VIS.ENABLE:
# Set up environment.
du.init_distributed_training(cfg)
# Set random seed from configs.
np.random.seed(cfg.RNG_SEED)
torch.manual_seed(cfg.RNG_SEED)
# Setup logging format.
logging.setup_logging(cfg.OUTPUT_DIR)
# Print config.
logger.info("Model Visualization with config:")
logger.info(cfg)
# Build the video model and print model statistics.
model = build_model(cfg)
if du.is_master_proc() and cfg.LOG_MODEL_INFO:
misc.log_model_info(model, cfg, is_train=False)
cu.load_test_checkpoint(cfg, model)
# Create video testing loaders.
vis_loader = loader.construct_loader(cfg, "test")
logger.info(
"Visualize model for {} data points".format(len(vis_loader))
)
if cfg.DETECTION.ENABLE:
assert cfg.NUM_GPUS == cfg.TEST.BATCH_SIZE
# Set up writer for logging to Tensorboard format.
if du.is_master_proc(cfg.NUM_GPUS * cfg.NUM_SHARDS):
writer = tb.TensorboardWriter(cfg)
else:
writer = None
# Run visualization on the model
run_visualization(vis_loader, model, cfg, writer)
if writer is not None:
writer.close()
def test(cfg):
"""
Perform multi-view testing on the pretrained video model.
Args:
cfg (CfgNode): configs. Details can be found in
slowfast/config/defaults.py
"""
# Set up environment.
du.init_distributed_training(cfg)
# Set random seed from configs.
np.random.seed(cfg.RNG_SEED)
torch.manual_seed(cfg.RNG_SEED)
# Setup logging format.
logging.setup_logging(cfg.OUTPUT_DIR)
# Print config.
logger.info("Test with config:")
logger.info(cfg)
# Build the video model and print model statistics.
model = build_model(cfg)
cu.load_test_checkpoint(cfg, model)
# Create video testing loaders.
test_loader = loader.construct_loader(cfg, "test")
logger.info("Testing model for {} iterations".format(len(test_loader)))
# Create meters for loss tracking
test_meter = TrainMeter(test_loader.dataset.num_videos, cfg)
# Set up writer for logging to Tensorboard format.
if cfg.TENSORBOARD.ENABLE and du.is_master_proc(
cfg.NUM_GPUS * cfg.NUM_SHARDS
):
writer = tb.TensorboardWriter(cfg)
else:
writer = None
# # Perform multi-view test on the entire dataset.
test_meter = perform_test(test_loader, model, test_meter, cfg, writer)
if writer is not None:
writer.close()
def infer(cfg):
# Setup logging format.
logging.setup_logging()
# Print config.
logger.info("Infer with config:")
logger.info(cfg)
# Build the SlowFast model and print its statistics
model = build_model(cfg)
if du.is_master_proc():
misc.log_model_info(model, cfg, is_train=False)
# load weights
if cfg.INFERENCE.WEIGHTS_FILE_PATH != "":
cu.load_checkpoint(cfg.INFERENCE.WEIGHTS_FILE_PATH, model, cfg.NUM_GPUS > 1, None,
inflation=False, convert_from_caffe2=cfg.INFERENCE.WEIGHTS_TYPE == "caffe2")
else:
raise FileNotFoundError("Model weights file could not be found")
inference_loader = loader.construct_loader(cfg, "inference")
perform_inference(inference_loader, model, cfg)
def test(cfg):
"""
Perform multi-view testing on the pretrained video model.
Args:
cfg (CfgNode): configs. Details can be found in
slowfast/config/defaults.py
"""
# Set up environment.
du.init_distributed_training(cfg)
# Set random seed from configs.
np.random.seed(cfg.RNG_SEED)
torch.manual_seed(cfg.RNG_SEED)
# Setup logging format.
logging.setup_logging(cfg.OUTPUT_DIR)
# Print config.
logger.info("Test with config:")
logger.info(cfg)
# Build the video model and print model statistics.
model = build_model(cfg)
out_str_prefix = "lin" if cfg.MODEL.DETACH_FINAL_FC else ""
if du.is_master_proc() and cfg.LOG_MODEL_INFO:
misc.log_model_info(model, cfg, use_train_input=False)
if (cfg.TASK == "ssl" and cfg.MODEL.MODEL_NAME == "ContrastiveModel"
and cfg.CONTRASTIVE.KNN_ON):
train_loader = loader.construct_loader(cfg, "train")
out_str_prefix = "knn"
if hasattr(model, "module"):
model.module.init_knn_labels(train_loader)
else:
model.init_knn_labels(train_loader)
cu.load_test_checkpoint(cfg, model)
# Create video testing loaders.
test_loader = loader.construct_loader(cfg, "test")
logger.info("Testing model for {} iterations".format(len(test_loader)))
if cfg.DETECTION.ENABLE:
assert cfg.NUM_GPUS == cfg.TEST.BATCH_SIZE or cfg.NUM_GPUS == 0
test_meter = AVAMeter(len(test_loader), cfg, mode="test")
else:
assert (
test_loader.dataset.num_videos %
(cfg.TEST.NUM_ENSEMBLE_VIEWS * cfg.TEST.NUM_SPATIAL_CROPS) == 0)
# Create meters for multi-view testing.
test_meter = TestMeter(
test_loader.dataset.num_videos //
(cfg.TEST.NUM_ENSEMBLE_VIEWS * cfg.TEST.NUM_SPATIAL_CROPS),
cfg.TEST.NUM_ENSEMBLE_VIEWS * cfg.TEST.NUM_SPATIAL_CROPS,
cfg.MODEL.NUM_CLASSES if not cfg.TASK == "ssl" else
cfg.CONTRASTIVE.NUM_CLASSES_DOWNSTREAM,
len(test_loader),
cfg.DATA.MULTI_LABEL,
cfg.DATA.ENSEMBLE_METHOD,
)
# Set up writer for logging to Tensorboard format.
if cfg.TENSORBOARD.ENABLE and du.is_master_proc(
cfg.NUM_GPUS * cfg.NUM_SHARDS):
writer = tb.TensorboardWriter(cfg)
else:
writer = None
# # Perform multi-view test on the entire dataset.
test_meter = perform_test(test_loader, model, test_meter, cfg, writer)
if writer is not None:
writer.close()
result_string = (
"_a{}{}{} Top1 Acc: {} Top5 Acc: {} MEM: {:.2f} dataset: {}{}"
"".format(
out_str_prefix,
cfg.TEST.DATASET[0],
test_meter.stats["top1_acc"],
test_meter.stats["top1_acc"],
test_meter.stats["top5_acc"],
misc.gpu_mem_usage(),
cfg.TEST.DATASET[0],
cfg.MODEL.NUM_CLASSES,
))
logger.info("testing done: {}".format(result_string))
return result_string
def train(cfg):
"""
Train a video model for many epochs on train set and evaluate it on val set.
Args:
cfg (CfgNode): configs. Details can be found in
slowfast/config/defaults.py
"""
# Set random seed from configs.
np.random.seed(cfg.RNG_SEED)
torch.manual_seed(cfg.RNG_SEED)
# Setup logging format.
logging.setup_logging(cfg)
# Print config.
logger.info("Train with config:")
logger.info(pprint.pformat(cfg))
if du.get_rank()==0 and du.is_master_proc(num_gpus=cfg.NUM_GPUS):
writer = SummaryWriter(log_dir=cfg.OUTPUT_DIR)
else:
writer = None
if du.get_rank()==0 and du.is_master_proc(num_gpus=cfg.NUM_GPUS) and not cfg.DEBUG:
tags = []
if 'TAGS' in cfg and cfg.TAGS !=[]:
tags=list(cfg.TAGS)
neptune.set_project('Serre-Lab/motion')
######################
overrides = sys.argv[1:]
overrides_dict = {}
for i in range(len(overrides)//2):
overrides_dict[overrides[2*i]] = overrides[2*i+1]
overrides_dict['dir'] = cfg.OUTPUT_DIR
######################
if 'NEP_ID' in cfg and cfg.NEP_ID != "":
session = Session()
project = session.get_project(project_qualified_name='Serre-Lab/motion')
nep_experiment = project.get_experiments(id=cfg.NEP_ID)[0]
else:
nep_experiment = neptune.create_experiment (name=cfg.NAME,
params=overrides_dict,
tags=tags)
else:
nep_experiment=None
# Build the video model and print model statistics.
model = build_model(cfg)
if du.is_master_proc(num_gpus=cfg.NUM_GPUS):
misc.log_model_info(model, cfg, is_train=True)
# Construct the optimizer.
optimizer = optim.construct_optimizer(model, cfg)
# Load a checkpoint to resume training if applicable.
if cfg.TRAIN.AUTO_RESUME and cu.has_checkpoint(cfg.OUTPUT_DIR):
logger.info("Load from last checkpoint.")
last_checkpoint = cu.get_last_checkpoint(cfg.OUTPUT_DIR)
checkpoint_epoch = cu.load_checkpoint(
last_checkpoint, model, cfg.NUM_GPUS > 1, optimizer
)
start_epoch = checkpoint_epoch + 1
elif cfg.TRAIN.CHECKPOINT_FILE_PATH != "":
logger.info("Load from given checkpoint file.")
checkpoint_epoch = cu.load_checkpoint(
cfg.TRAIN.CHECKPOINT_FILE_PATH,
model,
cfg.NUM_GPUS > 1,
optimizer,
inflation=cfg.TRAIN.CHECKPOINT_INFLATE,
convert_from_caffe2=cfg.TRAIN.CHECKPOINT_TYPE == "caffe2",
)
start_epoch = checkpoint_epoch + 1
else:
start_epoch = 0
# Create the video train and val loaders.
train_loader = loader.construct_loader(cfg, "train")
val_loader = loader.construct_loader(cfg, "val")
# Create meters.
if cfg.DETECTION.ENABLE:
train_meter = AVAMeter(len(train_loader), cfg, mode="train")
val_meter = AVAMeter(len(val_loader), cfg, mode="val")
else:
train_meter = TrainMeter(len(train_loader), cfg)
val_meter = ValMeter(len(val_loader), cfg)
# Perform the training loop.
logger.info("Start epoch: {}".format(start_epoch + 1))
for cur_epoch in range(start_epoch, cfg.SOLVER.MAX_EPOCH):
# Shuffle the dataset.
loader.shuffle_dataset(train_loader, cur_epoch)
# Train for one epoch.
train_epoch(train_loader, model, optimizer, train_meter, cur_epoch, writer, nep_experiment, cfg)
# Compute precise BN stats.
# if cfg.BN.USE_PRECISE_STATS and len(get_bn_modules(model)) > 0:
# calculate_and_update_precise_bn(
# train_loader, model, cfg.BN.NUM_BATCHES_PRECISE
# )
# Save a checkpoint.
if cu.is_checkpoint_epoch(cur_epoch, cfg.TRAIN.CHECKPOINT_PERIOD):
cu.save_checkpoint(cfg.OUTPUT_DIR, model, optimizer, cur_epoch, cfg)
# Evaluate the model on validation set.
if misc.is_eval_epoch(cfg, cur_epoch):
eval_epoch(val_loader, model, val_meter, cur_epoch, nep_experiment, cfg)
if du.get_rank()==0 and du.is_master_proc(num_gpus=cfg.NUM_GPUS) and not cfg.DEBUG:
nep_experiment.log_metric('epoch', cur_epoch)
writer = None
# # Perform multi-view test on the entire dataset.
perform_test(test_loader, model, test_meter, cfg, writer)
if writer is not None:
writer.close()
from slowfast.utils.misc import launch_job
from slowfast.utils.parser import load_config, parse_args
import numpy as np
import torch
import slowfast.utils.checkpoint as cu
import slowfast.utils.distributed as du
import slowfast.utils.logging as logging
import slowfast.utils.misc as misc
import slowfast.visualization.tensorboard_vis as tb
from slowfast.datasets import loader
from slowfast.models import build_model
from slowfast.utils.meters import AVAMeter, TestMeter
logger = logging.get_logger(__name__)
args = parse_args()
cfg = load_config(args)
test_loader = loader.construct_loader(cfg, "test")
def test(cfg):
"""
Perform multi-view testing on the pretrained video model.
Args:
cfg (CfgNode): configs. Details can be found in
slowfast/config/defaults.py
"""
# Set up environment.
du.init_distributed_training(cfg)
# Set random seed from configs.
np.random.seed(cfg.RNG_SEED)
torch.manual_seed(cfg.RNG_SEED)
# Setup logging format.
logging.setup_logging(cfg.OUTPUT_DIR)
# Print config.
logger.info("Test with config:")
logger.info(cfg)
# Build the video model and print model statistics.
model = build_model(cfg)
if du.is_master_proc() and cfg.LOG_MODEL_INFO:
misc.log_model_info(model, cfg, is_train=False)
# Load a checkpoint to test if applicable.
if cfg.TEST.CHECKPOINT_FILE_PATH != "":
cu.load_checkpoint(
cfg.TEST.CHECKPOINT_FILE_PATH,
model,
cfg.NUM_GPUS > 1,
None,
inflation=False,
convert_from_caffe2=cfg.TEST.CHECKPOINT_TYPE == "caffe2",
)
elif cu.has_checkpoint(cfg.OUTPUT_DIR):
last_checkpoint = cu.get_last_checkpoint(cfg.OUTPUT_DIR)
cu.load_checkpoint(last_checkpoint, model, cfg.NUM_GPUS > 1)
elif cfg.TRAIN.CHECKPOINT_FILE_PATH != "":
# If no checkpoint found in TEST.CHECKPOINT_FILE_PATH or in the current
# checkpoint folder, try to load checkpoint from
# TRAIN.CHECKPOINT_FILE_PATH and test it.
cu.load_checkpoint(
cfg.TRAIN.CHECKPOINT_FILE_PATH,
model,
cfg.NUM_GPUS > 1,
None,
inflation=False,
convert_from_caffe2=cfg.TRAIN.CHECKPOINT_TYPE == "caffe2",
)
else:
# raise NotImplementedError("Unknown way to load checkpoint.")
logger.info("Testing with random initialization. Only for debugging.")
# Create video testing loaders.
test_loader = loader.construct_loader(cfg, "test")
logger.info("Testing model for {} iterations".format(len(test_loader)))
assert (len(test_loader.dataset) %
(cfg.TEST.NUM_ENSEMBLE_VIEWS * cfg.TEST.NUM_SPATIAL_CROPS) == 0)
# Create meters for multi-view testing.
test_meter = TestMeter(
len(test_loader.dataset) //
(cfg.TEST.NUM_ENSEMBLE_VIEWS * cfg.TEST.NUM_SPATIAL_CROPS),
cfg.TEST.NUM_ENSEMBLE_VIEWS * cfg.TEST.NUM_SPATIAL_CROPS,
cfg.MODEL.NUM_CLASSES,
len(test_loader),
cfg.DATA.MULTI_LABEL,
cfg.DATA.ENSEMBLE_METHOD,
)
# Set up writer for logging to Tensorboard format.
if cfg.TENSORBOARD.ENABLE and du.is_master_proc(
cfg.NUM_GPUS * cfg.NUM_SHARDS):
writer = tb.TensorboardWriter(cfg)
else:
writer = None
# # Perform multi-view test on the entire dataset.
perform_test(test_loader, model, test_meter, cfg, writer)
if writer is not None:
writer.close()
def benchmark_data_loading(cfg):
"""
Benchmark the speed of data loading in PySlowFast.
Args:
cfg (CfgNode): configs. Details can be found in
slowfast/config/defaults.py
"""
# Set up environment.
setup_environment()
# Set random seed from configs.
np.random.seed(cfg.RNG_SEED)
torch.manual_seed(cfg.RNG_SEED)
# Setup logging format.
logging.setup_logging()
# Print config.
logger.info("Benchmark data loading with config:")
logger.info(pprint.pformat(cfg))
timer = Timer()
dataloader = loader.construct_loader(cfg, "train")
logger.info("Initialize loader using {:.2f} seconds.".format(
timer.seconds()))
# Total batch size across different machines.
batch_size = cfg.TRAIN.BATCH_SIZE * cfg.NUM_SHARDS
log_period = cfg.BENCHMARK.LOG_PERIOD
epoch_times = []
# Test for a few epochs.
for cur_epoch in range(cfg.BENCHMARK.NUM_EPOCHS):
timer = Timer()
timer_epoch = Timer()
iter_times = []
for cur_iter, _ in enumerate(tqdm.tqdm(dataloader)):
if cur_iter > 0 and cur_iter % log_period == 0:
iter_times.append(timer.seconds())
ram_usage, ram_total = misc.cpu_mem_usage()
logger.info(
"Epoch {}: {} iters ({} videos) in {:.2f} seconds. "
"RAM Usage: {:.2f}/{:.2f} GB.".format(
cur_epoch,
log_period,
log_period * batch_size,
iter_times[-1],
ram_usage,
ram_total,
))
timer.reset()
epoch_times.append(timer_epoch.seconds())
ram_usage, ram_total = misc.cpu_mem_usage()
logger.info(
"Epoch {}: in total {} iters ({} videos) in {:.2f} seconds. "
"RAM Usage: {:.2f}/{:.2f} GB.".format(
cur_epoch,
len(dataloader),
len(dataloader) * batch_size,
epoch_times[-1],
ram_usage,
ram_total,
))
logger.info(
"Epoch {}: on average every {} iters ({} videos) take {:.2f}/{:.2f} "
"(avg/std) seconds.".format(
cur_epoch,
log_period,
log_period * batch_size,
np.mean(iter_times),
np.std(iter_times),
))
logger.info("On average every epoch ({} videos) takes {:.2f}/{:.2f} "
"(avg/std) seconds.".format(
len(dataloader) * batch_size,
np.mean(epoch_times),
np.std(epoch_times),
))
def test(cfg, cnt=-1):
"""
Perform multi-view testing on the pretrained video model.
Args:
cfg (CfgNode): configs. Details can be found in
slowfast/config/defaults.py
"""
# Set random seed from configs.
np.random.seed(cfg.RNG_SEED)
torch.manual_seed(cfg.RNG_SEED)
# Setup logging format.
logging.setup_logging()
# # Perform multi-view test on the entire dataset.
scores_path = os.path.join(cfg.OUTPUT_DIR, 'scores')
if not os.path.exists(scores_path):
os.makedirs(scores_path)
filename_root = cfg.EPICKITCHENS.TEST_LIST.split('.')[0]
if cnt >= 0:
file_name = '{}_{}_{}.pkl'.format(filename_root, cnt, cfg.MODEL.MODEL_NAME)
else:
file_name = '{}_{}_{}.pkl'.format(filename_root, 'test_only', cfg.MODEL.MODEL_NAME)
file_path = os.path.join(scores_path, file_name)
logger.info(file_path)
# Print config.
# if cnt < 0:
# logger.info("Test with config:")
# logger.info(cfg)
# Build the video model and print model statistics.
model = build_model(cfg)
if cfg.EPICKITCHENS.USE_BBOX:
model.module.load_weight_slowfast()
# if du.is_master_proc():
# misc.log_model_info(model, cfg, is_train=False)
# Load a checkpoint to test if applicable.
if cfg.TEST.CHECKPOINT_FILE_PATH != "":
logger.info("Load from given checkpoint file.")
cu.load_checkpoint(
cfg.TEST.CHECKPOINT_FILE_PATH,
model,
cfg.NUM_GPUS > 1,
None,
inflation=False,
convert_from_caffe2=cfg.TEST.CHECKPOINT_TYPE == "caffe2",
)
elif cu.has_checkpoint(cfg.OUTPUT_DIR):
logger.info("Load from last checkpoint.")
last_checkpoint = cu.get_last_checkpoint(cfg.OUTPUT_DIR)
cu.load_checkpoint(last_checkpoint, model, cfg.NUM_GPUS > 1)
elif cfg.TRAIN.CHECKPOINT_FILE_PATH != "":
# If no checkpoint found in TEST.CHECKPOINT_FILE_PATH or in the current
# checkpoint folder, try to load checkpint from
# TRAIN.CHECKPOINT_FILE_PATH and test it.
cu.load_checkpoint(
cfg.TRAIN.CHECKPOINT_FILE_PATH,
model,
cfg.NUM_GPUS > 1,
None,
inflation=False,
convert_from_caffe2=cfg.TRAIN.CHECKPOINT_TYPE == "caffe2",
)
else:
# raise NotImplementedError("Unknown way to load checkpoint.")
logger.info("Testing with random initialization. Only for debugging.")
# Create video testing loaders.
if cfg.TEST.EXTRACT_FEATURES_MODE != "" and cfg.TEST.EXTRACT_FEATURES_MODE in ["test","train","val"]:
test_loader = loader.construct_loader(cfg, cfg.TEST.EXTRACT_FEATURES_MODE)
else:
test_loader = loader.construct_loader(cfg, "test")
logger.info("Testing model for {} iterations".format(len(test_loader)))
if cfg.DETECTION.ENABLE:
assert cfg.NUM_GPUS == cfg.TEST.BATCH_SIZE
test_meter = AVAMeter(len(test_loader), cfg, mode="test")
else:
assert (
len(test_loader.dataset)
% (cfg.TEST.NUM_ENSEMBLE_VIEWS * cfg.TEST.NUM_SPATIAL_CROPS)
== 0
)
# Create meters for multi-view testing.
if cfg.TEST.DATASET == 'epickitchens':
test_meter = EPICTestMeter(
len(test_loader.dataset)
// (cfg.TEST.NUM_ENSEMBLE_VIEWS * cfg.TEST.NUM_SPATIAL_CROPS),
cfg.TEST.NUM_ENSEMBLE_VIEWS * cfg.TEST.NUM_SPATIAL_CROPS,
cfg.MODEL.NUM_CLASSES,
len(test_loader),
)
else:
test_meter = TestMeter(
len(test_loader.dataset)
// (cfg.TEST.NUM_ENSEMBLE_VIEWS * cfg.TEST.NUM_SPATIAL_CROPS),
cfg.TEST.NUM_ENSEMBLE_VIEWS * cfg.TEST.NUM_SPATIAL_CROPS,
cfg.MODEL.NUM_CLASSES,
len(test_loader),
)
pickle.dump([], open(file_path, 'wb+'))
if cfg.TEST.EXTRACT_FEATURES:
preds, labels, metadata, x_feat_list = perform_test(test_loader, model, test_meter, cfg)
else:
preds, labels, metadata = perform_test(test_loader, model, test_meter, cfg)
if du.is_master_proc():
if cfg.TEST.DATASET == 'epickitchens':
results = {'verb_output': preds[0],
'noun_output': preds[1],
'verb_gt': labels[0],
'noun_gt': labels[1],
'narration_id': metadata}
scores_path = os.path.join(cfg.OUTPUT_DIR, 'scores')
if not os.path.exists(scores_path):
os.makedirs(scores_path)
pickle.dump(results, open(file_path, 'wb'))
if cfg.TEST.EXTRACT_FEATURES:
pid = cfg.EPICKITCHENS.FEATURE_VID.split("_")[0]
if not os.path.exists(os.path.join(cfg.TEST.EXTRACT_FEATURES_PATH, pid)):
os.mkdir(os.path.join(cfg.TEST.EXTRACT_FEATURES_PATH, pid))
if not cfg.TEST.EXTRACT_MSTCN_FEATURES and cfg.TEST.EXTRACT_FEATURES:
arr_slow = torch.cat(x_feat_list[0], dim=0).numpy()
arr_fast = torch.cat(x_feat_list[1], dim=0).numpy()
print(arr_slow.shape, arr_fast.shape)
fpath_feat = os.path.join(cfg.TEST.EXTRACT_FEATURES_PATH, pid, '{}.pkl'.format(cfg.EPICKITCHENS.FEATURE_VID))
with open(fpath_feat,'wb+') as f:
pickle.dump([arr_slow, arr_fast], f)
elif cfg.TEST.EXTRACT_MSTCN_FEATURES and cfg.TEST.EXTRACT_FEATURES:
fpath_feat = os.path.join(cfg.TEST.EXTRACT_FEATURES_PATH, pid, '{}.npy'.format(cfg.EPICKITCHENS.FEATURE_VID))
with open(fpath_feat,'wb+') as f:
arr = torch.cat(x_feat_list, dim=0).numpy()
print(arr.shape)
np.save(f, arr)
def feature_extract(kwargs):
"""
Perform multi-view testing on the pretrained video model.
Args:
cfg (CfgNode): configs. Details can be found in
slowfast/config/defaults.py
"""
cfg = kwargs.pop('cfg')
path_to_video_list = kwargs.pop('path_to_video_list')
path_to_video_dir = kwargs.pop('path_to_video_dir')
path_to_feat_dir = kwargs.pop('path_to_feat_dir')
num_features = kwargs.pop('num_features')
if not os.path.isdir(path_to_feat_dir):
os.makedirs(path_to_feat_dir)
# Setup logging format.
logging.setup_logging()
# Print config.
logger.info("Extract feature with config:")
logger.info(cfg)
# Build the video model and print model statistics.
model = model_builder.build_model(cfg, feature_extraction=True)
if du.is_master_proc():
misc.log_model_info(model)
# Load a checkpoint to test if applicable.
if cfg.TEST.CHECKPOINT_FILE_PATH != "":
cu.load_checkpoint(
cfg.TEST.CHECKPOINT_FILE_PATH,
model,
cfg.NUM_GPUS > 1,
None,
inflation=False,
convert_from_caffe2=cfg.TEST.CHECKPOINT_TYPE == "caffe2",
)
elif cu.has_checkpoint(cfg.OUTPUT_DIR):
last_checkpoint = cu.get_last_checkpoint(cfg.OUTPUT_DIR)
cu.load_checkpoint(last_checkpoint, model, cfg.NUM_GPUS > 1)
elif cfg.TRAIN.CHECKPOINT_FILE_PATH != "":
# If no checkpoint found in TEST.CHECKPOINT_FILE_PATH or in the current
# checkpoint folder, try to load checkpint from
# TRAIN.CHECKPOINT_FILE_PATH and test it.
cu.load_checkpoint(
cfg.TRAIN.CHECKPOINT_FILE_PATH,
model,
cfg.NUM_GPUS > 1,
None,
inflation=False,
convert_from_caffe2=cfg.TRAIN.CHECKPOINT_TYPE == "caffe2",
)
else:
# raise NotImplementedError("Unknown way to load checkpoint.")
logger.info("Extracting features with random initialization. Only for debugging.")
feature_extract_fn = perform_bbox_feature_extract if cfg.DETECTION.ENABLE else perform_feature_extract
# Create video feature extraction loaders.
if osp.isfile(path_to_video_list):
path_to_videos = []
with open(path_to_video_list, 'r') as f:
video_list = json.load(f)
for video_name in video_list:
path_to_video = osp.join(path_to_video_dir, video_name)
if osp.isfile(path_to_video):
path_to_videos.append(path_to_video)
else:
path_to_videos = glob.glob(osp.join(path_to_video_dir, '*'))
for video_idx, path_to_video in enumerate(tqdm(path_to_videos)):
path_to_feature = osp.join(path_to_feat_dir, osp.splitext(osp.basename(path_to_video))[0] + '.json')
if osp.isfile(path_to_feature) and json.load(open(path_to_feature))['num_features'] == num_features:
continue
video_extraction_loader = loader.construct_loader(cfg, path_to_video)
video_data = {
'num_features': 0,
'video_features': {}
}
if len(video_extraction_loader) > 0:
video_features = feature_extract_fn(video_extraction_loader, model, cfg)
assert all([feature is not None for feature in video_features]), 'Missing some features !'
video_data['num_features'] = len(video_features)
video_data['video_features'] = video_features
if video_data['num_features'] != num_features:
print('Warning! Video %s has %d features.' % (path_to_video, video_data['num_features']))
with open(path_to_feature, 'w') as f:
json.dump(video_data, f)
def train(cfg):
"""
Train a video model for many epochs on train set and evaluate it on val set.
Args:
cfg (CfgNode): configs. Details can be found in
slowfast/config/defaults.py
"""
# Set up environment.
du.init_distributed_training(cfg)
# Set random seed from configs.
np.random.seed(cfg.RNG_SEED)
torch.manual_seed(cfg.RNG_SEED)
# Setup logging format.
logging.setup_logging(cfg.OUTPUT_DIR)
# Init multigrid.
multigrid = None
if cfg.MULTIGRID.LONG_CYCLE or cfg.MULTIGRID.SHORT_CYCLE:
multigrid = MultigridSchedule()
cfg = multigrid.init_multigrid(cfg)
if cfg.MULTIGRID.LONG_CYCLE:
cfg, _ = multigrid.update_long_cycle(cfg, cur_epoch=0)
# Print config.
logger.info("Train with config:")
logger.info(pprint.pformat(cfg))
# Build the video model and print model statistics.
model = build_model(cfg)
if du.is_master_proc() and cfg.LOG_MODEL_INFO:
misc.log_model_info(model, cfg, use_train_input=True)
# Construct the optimizer.
optimizer = optim.construct_optimizer(model, cfg)
# Create a GradScaler for mixed precision training
scaler = torch.cuda.amp.GradScaler(enabled=cfg.TRAIN.MIXED_PRECISION)
# Load a checkpoint to resume training if applicable.
if cfg.TRAIN.AUTO_RESUME and cu.has_checkpoint(cfg.OUTPUT_DIR):
logger.info("Load from last checkpoint.")
last_checkpoint = cu.get_last_checkpoint(cfg.OUTPUT_DIR, task=cfg.TASK)
if last_checkpoint is not None:
checkpoint_epoch = cu.load_checkpoint(
last_checkpoint,
model,
cfg.NUM_GPUS > 1,
optimizer,
scaler if cfg.TRAIN.MIXED_PRECISION else None,
)
start_epoch = checkpoint_epoch + 1
elif "ssl_eval" in cfg.TASK:
last_checkpoint = cu.get_last_checkpoint(cfg.OUTPUT_DIR,
task="ssl")
checkpoint_epoch = cu.load_checkpoint(
last_checkpoint,
model,
cfg.NUM_GPUS > 1,
optimizer,
scaler if cfg.TRAIN.MIXED_PRECISION else None,
epoch_reset=True,
clear_name_pattern=cfg.TRAIN.CHECKPOINT_CLEAR_NAME_PATTERN,
)
start_epoch = checkpoint_epoch + 1
else:
start_epoch = 0
elif cfg.TRAIN.CHECKPOINT_FILE_PATH != "":
logger.info("Load from given checkpoint file.")
checkpoint_epoch = cu.load_checkpoint(
cfg.TRAIN.CHECKPOINT_FILE_PATH,
model,
cfg.NUM_GPUS > 1,
optimizer,
scaler if cfg.TRAIN.MIXED_PRECISION else None,
inflation=cfg.TRAIN.CHECKPOINT_INFLATE,
convert_from_caffe2=cfg.TRAIN.CHECKPOINT_TYPE == "caffe2",
epoch_reset=cfg.TRAIN.CHECKPOINT_EPOCH_RESET,
clear_name_pattern=cfg.TRAIN.CHECKPOINT_CLEAR_NAME_PATTERN,
)
start_epoch = checkpoint_epoch + 1
else:
start_epoch = 0
# Create the video train and val loaders.
train_loader = loader.construct_loader(cfg, "train")
val_loader = loader.construct_loader(cfg, "val")
precise_bn_loader = (loader.construct_loader(
cfg, "train", is_precise_bn=True)
if cfg.BN.USE_PRECISE_STATS else None)
# if (
# cfg.TASK == "ssl"
# and cfg.MODEL.MODEL_NAME == "ContrastiveModel"
# and cfg.CONTRASTIVE.KNN_ON
# ):
# if hasattr(model, "module"):
# model.module.init_knn_labels(train_loader)
# else:
# model.init_knn_labels(train_loader)
# Create meters.
if cfg.DETECTION.ENABLE:
train_meter = AVAMeter(len(train_loader), cfg, mode="train")
val_meter = AVAMeter(len(val_loader), cfg, mode="val")
else:
train_meter = TrainMeter(1e6, cfg)
val_meter = ValMeter(1e6, cfg)
# set up writer for logging to Tensorboard format.
if cfg.TENSORBOARD.ENABLE and du.is_master_proc(
cfg.NUM_GPUS * cfg.NUM_SHARDS):
writer = tb.TensorboardWriter(cfg)
else:
writer = None
# Perform the training loop.
logger.info("Start epoch: {}".format(start_epoch + 1))
epoch_timer = EpochTimer()
for cur_epoch in range(start_epoch, cfg.SOLVER.MAX_EPOCH):
if cur_epoch > 0 and cfg.DATA.LOADER_CHUNK_SIZE > 0:
num_chunks = math.ceil(cfg.DATA.LOADER_CHUNK_OVERALL_SIZE /
cfg.DATA.LOADER_CHUNK_SIZE)
skip_rows = (cur_epoch) % num_chunks * cfg.DATA.LOADER_CHUNK_SIZE
logger.info(
f"=================+++ num_chunks {num_chunks} skip_rows {skip_rows}"
)
cfg.DATA.SKIP_ROWS = skip_rows
logger.info(f"|===========| skip_rows {skip_rows}")
train_loader = loader.construct_loader(cfg, "train")
loader.shuffle_dataset(train_loader, cur_epoch)
if cfg.MULTIGRID.LONG_CYCLE:
cfg, changed = multigrid.update_long_cycle(cfg, cur_epoch)
if changed:
(
model,
optimizer,
train_loader,
val_loader,
precise_bn_loader,
train_meter,
val_meter,
) = build_trainer(cfg)
# Load checkpoint.
if cu.has_checkpoint(cfg.OUTPUT_DIR):
last_checkpoint = cu.get_last_checkpoint(cfg.OUTPUT_DIR,
task=cfg.TASK)
assert "{:05d}.pyth".format(cur_epoch) in last_checkpoint
else:
last_checkpoint = cfg.TRAIN.CHECKPOINT_FILE_PATH
logger.info("Load from {}".format(last_checkpoint))
cu.load_checkpoint(last_checkpoint, model, cfg.NUM_GPUS > 1,
optimizer)
# Shuffle the dataset.
loader.shuffle_dataset(train_loader, cur_epoch)
if hasattr(train_loader.dataset, "_set_epoch_num"):
train_loader.dataset._set_epoch_num(cur_epoch)
# Train for one epoch.
epoch_timer.epoch_tic()
train_epoch(
train_loader,
model,
optimizer,
scaler,
train_meter,
cur_epoch,
cfg,
writer,
)
epoch_timer.epoch_toc()
logger.info(
f"Epoch {cur_epoch} takes {epoch_timer.last_epoch_time():.2f}s. Epochs "
f"from {start_epoch} to {cur_epoch} take "
f"{epoch_timer.avg_epoch_time():.2f}s in average and "
f"{epoch_timer.median_epoch_time():.2f}s in median.")
logger.info(
f"For epoch {cur_epoch}, each iteraction takes "
f"{epoch_timer.last_epoch_time()/len(train_loader):.2f}s in average. "
f"From epoch {start_epoch} to {cur_epoch}, each iteraction takes "
f"{epoch_timer.avg_epoch_time()/len(train_loader):.2f}s in average."
)
is_checkp_epoch = (cu.is_checkpoint_epoch(
cfg,
cur_epoch,
None if multigrid is None else multigrid.schedule,
) or cur_epoch == cfg.SOLVER.MAX_EPOCH - 1)
is_eval_epoch = misc.is_eval_epoch(
cfg, cur_epoch, None if multigrid is None else multigrid.schedule)
# Compute precise BN stats.
if ((is_checkp_epoch or is_eval_epoch) and cfg.BN.USE_PRECISE_STATS
and len(get_bn_modules(model)) > 0):
calculate_and_update_precise_bn(
precise_bn_loader,
model,
min(cfg.BN.NUM_BATCHES_PRECISE, len(precise_bn_loader)),
cfg.NUM_GPUS > 0,
)
_ = misc.aggregate_sub_bn_stats(model)
# Save a checkpoint.
if is_checkp_epoch:
cu.save_checkpoint(
cfg.OUTPUT_DIR,
model,
optimizer,
cur_epoch,
cfg,
scaler if cfg.TRAIN.MIXED_PRECISION else None,
)
# Evaluate the model on validation set.
if is_eval_epoch:
eval_epoch(
val_loader,
model,
val_meter,
cur_epoch,
cfg,
train_loader,
writer,
)
if writer is not None:
writer.close()
result_string = "Top1 Acc: {:.2f} Top5 Acc: {:.2f} MEM: {:.2f}" "".format(
100 - val_meter.min_top1_err,
100 - val_meter.min_top5_err,
misc.gpu_mem_usage(),
)
logger.info("training done: {}".format(result_string))
return result_string
def evalOnnxOne(self, dataSet):
self.cfg.TRAIN['BATCH_SIZE'] = 1
dl = loader.construct_loader(self.cfg, dataSet)
meter = ValMeter(len(dl), self.cfg, self.logger)
totTime = 0
results = []
predsAll = []
labelsAll = []
GB = 1024**3
meta = {'id': ['s'] * 100}
sess, inNames, outNames = self.loadOnnx()
for cur_iter, (frames, labels, idx, meta) in enumerate(dl):
# if sess is None or cur_iter % 50000 == 0:
# if sess is not None:
# del sess
# sess, inNames, outNames = self.loadOnnx()
startTime = time.time()
meter.iter_tic()
idx = torch.tensor(cur_iter)
# assert isinstance(frames, (list,)), 'Expect a list as input'
# assert len(inNames) == len(frames), "Num Onnx inputs {} [{}] Data has {}".format(len(inNames), inNames, len(frames))
inputs = {
n: d.numpy().astype(np.float32)
for n, d in zip(inNames, frames)
}
pred = sess.run([outNames], inputs).pop().squeeze()
# extract the track start and end from meta data
# m = re.search('.*tr_([\d]+)_st_([\d]+)_end_([\d]+).pt', meta['id'][0])
# tr, st, end = [int(x) for x in m.groups()]
meter.iter_toc()
results.append((meta['id'], pred, labels))
if np.isnan(pred).any():
self.logger.info("NAN Iter {} Idx {} {} {}".format(
cur_iter, idx.item(), meta, pred))
continue
if labels != 0.0 and labels != 1.0:
self.logger.info(
"Bad Label Iter {} Idx {} {} Label {}".format(
cur_iter, idx.item(), meta, labels))
continue
# num_topks_correct = metrics.topks_correct(torch.tensor(pred), labels, (1, min(5, self.cfg.MODEL.NUM_CLASSES)))
# top1_err, _ = [(1.0 - x / pred.size(0)) * 100.0 for x in num_topks_correct ]
meter.update_stats(0, 0, 1)
meter.update_predictions(pred, labels)
labels = labels.numpy().astype(np.float32)[0]
predsAll.append(pred[-1])
labelsAll.append(labels)
meter.log_iter_stats(0, cur_iter, predsAll, labelsAll)
totTime += time.time() - startTime
if len(results) % self.cfg.LOG_PERIOD == 0:
c = len(results)
mm = psutil.virtual_memory()
ss = psutil.swap_memory()
self.logger.info(
"{} {} eTime {:.4f} sec avg {:.4f} sec CPU {} used GB {:.3f} AVAIL {:.3f} FREE {:.3f} SWAP Tot {:.3f} SWAP USED {:.3f} SWAP Free {:.3f} "
.format(
dataSet,
c,
totTime,
totTime / max(c, 1),
psutil.cpu_percent(),
mm.used / GB,
mm.available / GB,
mm.free / GB,
ss.total / GB,
ss.used / GB,
ss.free / GB,
))
c = len(results)
eTime = totTime / max(c, 1)
self.logger.info("{} {} eTime {} sec avg {} sec".format(
dataSet, c, totTime, eTime))
meter.log_epoch_stats(0, predsAll, labelsAll)
if len(self.cfg.ONNX.SAVE_PREDS_PATH) > 0:
df = pd.DataFrame(results, columns=['clip', 'pred', 'label'])
dfPath = createFullPathTree(
self.cfg.OUTPUT_DIR,
addFilenameSuffix(self.cfg.ONNX.SAVE_PREDS_PATH, dataSet))
ensureDir(os.path.dirname(dfPath))
self.logger.info("Saving predictions results {}".format(dfPath))
df.to_pickle(dfPath)
if self.cfg.ONNX.PREC_RECALL_PLOT:
precision, recall, thresholds = precision_recall_curve(
labelsAll, predsAll)
self.plotPrCurve(
dataSet,
'PR AUC',
recall,
precision,
xLabel='Recall',
yLabel='Precision',
)
if self.cfg.ONNX.ROC_PLOT:
fpr, tpr, rocThresholds = roc_curve(labelsAll, predsAll)
self.plotPrCurve(dataSet,
'ROC AUC',
fpr,
tpr,
xLabel='FPR',
yLabel='TPR')
def train(cfg):
"""
Train a video model for many epochs on train set and evaluate it on val set.
Args:
cfg (CfgNode): configs. Details can be found in
slowfast/config/defaults.py
"""
# Set random seed from configs.
np.random.seed(cfg.RNG_SEED)
torch.manual_seed(cfg.RNG_SEED)
# Setup logging format.
logging.setup_logging()
# Print config.
logger.info("Train with config:")
logger.info(pprint.pformat(cfg))
# Build the video model and print model statistics.
model = model_builder.build_model(cfg)
if du.is_master_proc():
misc.log_model_info(model)
# Construct the optimizer.
optimizer = optim.construct_optimizer(model, cfg)
# Load a checkpoint to resume training if applicable.
if cfg.TRAIN.AUTO_RESUME and cu.has_checkpoint(cfg.OUTPUT_DIR):
logger.info("Load from last checkpoint.")
last_checkpoint = cu.get_last_checkpoint(cfg.OUTPUT_DIR)
checkpoint_epoch = cu.load_checkpoint(last_checkpoint, model,
cfg.NUM_GPUS > 1, optimizer)
start_epoch = checkpoint_epoch + 1
elif cfg.TRAIN.CHECKPOINT_FILE_PATH != "":
logger.info("Load from given checkpoint file.")
checkpoint_epoch = cu.load_checkpoint(
cfg.TRAIN.CHECKPOINT_FILE_PATH,
model,
cfg.NUM_GPUS > 1,
optimizer,
inflation=cfg.TRAIN.CHECKPOINT_INFLATE,
convert_from_caffe2=cfg.TRAIN.CHECKPOINT_TYPE == "caffe2",
)
start_epoch = checkpoint_epoch + 1
else:
start_epoch = 0
# Create the video train and val loaders.
train_loader = loader.construct_loader(cfg, "train")
val_loader = loader.construct_loader(cfg, "val")
# Create meters.
if cfg.DETECTION.ENABLE:
train_meter = AVAMeter(len(train_loader), cfg, mode="train")
val_meter = AVAMeter(len(val_loader), cfg, mode="val")
else:
train_meter = TrainMeter(len(train_loader), cfg)
val_meter = ValMeter(len(val_loader), cfg)
# Perform the training loop.
logger.info("Start epoch: {}".format(start_epoch + 1))
for cur_epoch in range(start_epoch, cfg.SOLVER.MAX_EPOCH):
# Shuffle the dataset.
loader.shuffle_dataset(train_loader, cur_epoch)
# Train for one epoch.
train_epoch(train_loader, model, optimizer, train_meter, cur_epoch,
cfg)
# Compute precise BN stats.
if cfg.BN.USE_PRECISE_STATS and len(get_bn_modules(model)) > 0:
calculate_and_update_precise_bn(train_loader, model,
cfg.BN.NUM_BATCHES_PRECISE)
# Save a checkpoint.
if cu.is_checkpoint_epoch(cur_epoch, cfg.TRAIN.CHECKPOINT_PERIOD):
cu.save_checkpoint(cfg.OUTPUT_DIR, model, optimizer, cur_epoch,
cfg)
# Evaluate the model on validation set.
if misc.is_eval_epoch(cfg, cur_epoch):
eval_epoch(val_loader, model, val_meter, cur_epoch, cfg)
def train(cfg):
"""
Train an audio model for many epochs on train set and evaluate it on val set.
Args:
cfg (CfgNode): configs. Details can be found in
slowfast/config/defaults.py
"""
# Set up environment.
du.init_distributed_training(cfg)
# Set random seed from configs.
np.random.seed(cfg.RNG_SEED)
torch.manual_seed(cfg.RNG_SEED)
# Setup logging format.
logging.setup_logging(cfg.OUTPUT_DIR)
# Print config.
logger.info("Train with config:")
logger.info(pprint.pformat(cfg))
# Build the audio model and print model statistics.
model = build_model(cfg)
if du.is_master_proc() and cfg.LOG_MODEL_INFO:
misc.log_model_info(model, cfg)
if cfg.BN.FREEZE:
model.module.freeze_fn(
'bn_parameters') if cfg.NUM_GPUS > 1 else model.freeze_fn(
'bn_parameters')
# Construct the optimizer.
optimizer = optim.construct_optimizer(model, cfg)
# Load a checkpoint to resume training if applicable.
start_epoch = cu.load_train_checkpoint(cfg, model, optimizer)
# Create the audio train and val loaders.
if cfg.TRAIN.DATASET != 'epickitchens' or not cfg.EPICKITCHENS.TRAIN_PLUS_VAL:
train_loader = loader.construct_loader(cfg, "train")
val_loader = loader.construct_loader(cfg, "val")
precise_bn_loader = (loader.construct_loader(cfg, "train")
if cfg.BN.USE_PRECISE_STATS else None)
else:
train_loader = loader.construct_loader(cfg, "train+val")
val_loader = loader.construct_loader(cfg, "val")
precise_bn_loader = (loader.construct_loader(cfg, "train+val")
if cfg.BN.USE_PRECISE_STATS else None)
# Create meters.
if cfg.TRAIN.DATASET == 'epickitchens':
train_meter = EPICTrainMeter(len(train_loader), cfg)
val_meter = EPICValMeter(len(val_loader), cfg)
else:
train_meter = TrainMeter(len(train_loader), cfg)
val_meter = ValMeter(len(val_loader), cfg)
# set up writer for logging to Tensorboard format.
if cfg.TENSORBOARD.ENABLE and du.is_master_proc(
cfg.NUM_GPUS * cfg.NUM_SHARDS):
writer = tb.TensorboardWriter(cfg)
else:
writer = None
if cfg.WANDB.ENABLE and du.is_master_proc(cfg.NUM_GPUS * cfg.NUM_SHARDS):
wandb_log = True
if cfg.TRAIN.AUTO_RESUME and cfg.WANDB.RUN_ID != "":
wandb.init(project='slowfast',
config=cfg,
sync_tensorboard=True,
resume=cfg.WANDB.RUN_ID)
else:
wandb.init(project='slowfast', config=cfg, sync_tensorboard=True)
wandb.watch(model)
else:
wandb_log = False
# Perform the training loop.
logger.info("Start epoch: {}".format(start_epoch + 1))
for cur_epoch in range(start_epoch, cfg.SOLVER.MAX_EPOCH):
# Shuffle the dataset.
loader.shuffle_dataset(train_loader, cur_epoch)
# Train for one epoch.
train_epoch(train_loader, model, optimizer, train_meter, cur_epoch,
cfg, writer, wandb_log)
is_checkp_epoch = cu.is_checkpoint_epoch(
cfg,
cur_epoch,
)
is_eval_epoch = misc.is_eval_epoch(
cfg,
cur_epoch,
)
# Compute precise BN stats.
if ((is_checkp_epoch or is_eval_epoch) and cfg.BN.USE_PRECISE_STATS
and len(get_bn_modules(model)) > 0):
calculate_and_update_precise_bn(
precise_bn_loader,
model,
min(cfg.BN.NUM_BATCHES_PRECISE, len(precise_bn_loader)),
cfg.NUM_GPUS > 0,
)
_ = misc.aggregate_sub_bn_stats(model)
# Save a checkpoint.
if is_checkp_epoch:
cu.save_checkpoint(cfg.OUTPUT_DIR, model, optimizer, cur_epoch,
cfg)
# Evaluate the model on validation set.
if is_eval_epoch:
is_best_epoch, _ = eval_epoch(val_loader, model, val_meter,
cur_epoch, cfg, writer, wandb_log)
if is_best_epoch:
cu.save_checkpoint(cfg.OUTPUT_DIR,
model,
optimizer,
cur_epoch,
cfg,
is_best_epoch=is_best_epoch)
if writer is not None:
writer.close()
def benchmark_data(cfg):
# Set up environment.
setup_environment()
# Set random seed from configs.
np.random.seed(cfg.RNG_SEED)
torch.manual_seed(cfg.RNG_SEED)
# Setup logging format.
logging.setup_logging()
# Print config.
logger.info("Benchmark data loading with config:")
logger.info(pprint.pformat(cfg))
timer = Timer()
dataloader = loader.construct_loader(cfg, "train")
logger.info("Initialize loader using {:.2f} seconds.".format(
timer.seconds()))
batch_size = cfg.TRAIN.BATCH_SIZE
log_period = cfg.BENCHMARK.LOG_PERIOD
epoch_times = []
# Test for a few epochs.
for cur_epoch in range(cfg.BENCHMARK.NUM_EPOCHS):
timer = Timer()
timer_epoch = Timer()
iter_times = []
for cur_iter, _ in enumerate(tqdm.tqdm(dataloader)):
if cur_iter > 0 and cur_iter % log_period == 0:
iter_times.append(timer.seconds())
vram = psutil.virtual_memory()
logger.info(
"Epoch {}: {} iters ({} videos) in {:.2f} seconds. "
"RAM Usage: {:.2f}/{:.2f} GB.".format(
cur_epoch,
log_period,
log_period * batch_size,
iter_times[-1],
(vram.total - vram.available) / 1024**3,
vram.total / 1024**3,
))
timer.reset()
epoch_times.append(timer_epoch.seconds())
vram = psutil.virtual_memory()
logger.info(
"Epoch {}: in total {} iters ({} videos) in {:.2f} seconds. "
"RAM Usage: {:.2f}/{:.2f} GB.".format(
cur_epoch,
len(dataloader),
len(dataloader) * batch_size,
epoch_times[-1],
(vram.total - vram.available) / 1024**3,
vram.total / 1024**3,
))
logger.info(
"Epoch {}: on average every {} iters ({} videos) take {:.2f}/{:.2f} "
"(avg/std) seconds.".format(
cur_epoch,
log_period,
log_period * batch_size,
np.mean(iter_times),
np.std(iter_times),
))
logger.info("On average every epoch ({} videos) takes {:.2f}/{:.2f} "
"(avg/std) seconds.".format(
len(dataloader) * batch_size,
np.mean(epoch_times),
np.std(epoch_times),
))
def test(cfg):
"""
Perform multi-view testing on the pretrained audio model.
Args:
cfg (CfgNode): configs. Details can be found in
slowfast/config/defaults.py
"""
# Set up environment.
du.init_distributed_training(cfg)
# Set random seed from configs.
np.random.seed(cfg.RNG_SEED)
torch.manual_seed(cfg.RNG_SEED)
# Setup logging format.
logging.setup_logging(cfg.OUTPUT_DIR)
# Print config.
logger.info("Test with config:")
logger.info(cfg)
# Build the audio model and print model statistics.
model = build_model(cfg)
if du.is_master_proc() and cfg.LOG_MODEL_INFO:
misc.log_model_info(model, cfg)
cu.load_test_checkpoint(cfg, model)
# Create audio testing loaders.
test_loader = loader.construct_loader(cfg, "test")
logger.info("Testing model for {} iterations".format(len(test_loader)))
assert (
len(test_loader.dataset)
% cfg.TEST.NUM_ENSEMBLE_VIEWS
== 0
)
# Create meters for multi-view testing.
if cfg.TEST.DATASET == 'epickitchens':
test_meter = EPICTestMeter(
len(test_loader.dataset)
// cfg.TEST.NUM_ENSEMBLE_VIEWS,
cfg.TEST.NUM_ENSEMBLE_VIEWS,
cfg.MODEL.NUM_CLASSES,
len(test_loader),
cfg.DATA.ENSEMBLE_METHOD,
)
else:
test_meter = TestMeter(
len(test_loader.dataset)
// cfg.TEST.NUM_ENSEMBLE_VIEWS,
cfg.TEST.NUM_ENSEMBLE_VIEWS,
cfg.MODEL.NUM_CLASSES[0],
len(test_loader),
cfg.DATA.MULTI_LABEL,
cfg.DATA.ENSEMBLE_METHOD,
)
# Set up writer for logging to Tensorboard format.
if cfg.TENSORBOARD.ENABLE and du.is_master_proc(
cfg.NUM_GPUS * cfg.NUM_SHARDS
):
writer = tb.TensorboardWriter(cfg)
else:
writer = None
# # Perform multi-view test on the entire dataset.
test_meter, preds, preds_clips, labels, metadata = perform_test(test_loader, model, test_meter, cfg, writer)
if du.is_master_proc():
if cfg.TEST.DATASET == 'epickitchens':
results = {'verb_output': preds[0],
'noun_output': preds[1],
'narration_id': metadata}
scores_path = os.path.join(cfg.OUTPUT_DIR, 'scores')
if not os.path.exists(scores_path):
os.makedirs(scores_path)
file_path = os.path.join(scores_path, cfg.EPICKITCHENS.TEST_SPLIT+'.pkl')
pickle.dump(results, open(file_path, 'wb'))
else:
if cfg.TEST.DATASET == 'vggsound':
get_stats(preds, labels)
results = {'scores': preds, 'labels': labels}
scores_path = os.path.join(cfg.OUTPUT_DIR, 'scores')
if not os.path.exists(scores_path):
os.makedirs(scores_path)
file_path = os.path.join(scores_path, 'test.pkl')
pickle.dump(results, open(file_path, 'wb'))
if writer is not None:
writer.close()
def visualize(cfg):
"""
Perform layer weights and activations visualization on the model.
Args:
cfg (CfgNode): configs. Details can be found in
slowfast/config/defaults.py
"""
if cfg.TENSORBOARD.ENABLE and (cfg.TENSORBOARD.MODEL_VIS.ENABLE
or cfg.TENSORBOARD.WRONG_PRED_VIS.ENABLE):
# Set up environment.
du.init_distributed_training(cfg)
# Set random seed from configs.
np.random.seed(cfg.RNG_SEED)
torch.manual_seed(cfg.RNG_SEED)
# Setup logging format.
logging.setup_logging(cfg.OUTPUT_DIR)
# Print config.
logger.info("Model Visualization with config:")
logger.info(cfg)
# Build the video model and print model statistics.
model = build_model(cfg)
model.eval()
if du.is_master_proc() and cfg.LOG_MODEL_INFO:
misc.log_model_info(model, cfg, use_train_input=False)
cu.load_test_checkpoint(cfg, model)
# Create video testing loaders.
vis_loader = loader.construct_loader(cfg, "test")
if cfg.DETECTION.ENABLE:
assert cfg.NUM_GPUS == cfg.TEST.BATCH_SIZE or cfg.NUM_GPUS == 0
# Set up writer for logging to Tensorboard format.
if du.is_master_proc(cfg.NUM_GPUS * cfg.NUM_SHARDS):
writer = tb.TensorboardWriter(cfg)
else:
writer = None
if cfg.TENSORBOARD.PREDICTIONS_PATH != "":
assert not cfg.DETECTION.ENABLE, "Detection is not supported."
logger.info(
"Visualizing class-level performance from saved results...")
if writer is not None:
with g_pathmgr.open(cfg.TENSORBOARD.PREDICTIONS_PATH,
"rb") as f:
preds, labels = pickle.load(f, encoding="latin1")
writer.plot_eval(preds, labels)
if cfg.TENSORBOARD.MODEL_VIS.ENABLE:
if cfg.TENSORBOARD.MODEL_VIS.GRAD_CAM.ENABLE:
assert (
not cfg.DETECTION.ENABLE
), "Detection task is currently not supported for Grad-CAM visualization."
if cfg.MODEL.ARCH in cfg.MODEL.SINGLE_PATHWAY_ARCH:
assert (
len(cfg.TENSORBOARD.MODEL_VIS.GRAD_CAM.LAYER_LIST) == 1
), "The number of chosen CNN layers must be equal to the number of pathway(s), given {} layer(s).".format(
len(cfg.TENSORBOARD.MODEL_VIS.GRAD_CAM.LAYER_LIST))
elif cfg.MODEL.ARCH in cfg.MODEL.MULTI_PATHWAY_ARCH:
assert (
len(cfg.TENSORBOARD.MODEL_VIS.GRAD_CAM.LAYER_LIST) == 2
), "The number of chosen CNN layers must be equal to the number of pathway(s), given {} layer(s).".format(
len(cfg.TENSORBOARD.MODEL_VIS.GRAD_CAM.LAYER_LIST))
else:
raise NotImplementedError(
"Model arch {} is not in {}".format(
cfg.MODEL.ARCH,
cfg.MODEL.SINGLE_PATHWAY_ARCH +
cfg.MODEL.MULTI_PATHWAY_ARCH,
))
logger.info("Visualize model analysis for {} iterations".format(
len(vis_loader)))
# Run visualization on the model
run_visualization(vis_loader, model, cfg, writer)
if cfg.TENSORBOARD.WRONG_PRED_VIS.ENABLE:
logger.info("Visualize Wrong Predictions for {} iterations".format(
len(vis_loader)))
perform_wrong_prediction_vis(vis_loader, model, cfg)
if writer is not None:
writer.close()
def test(cfg):
"""
Perform multi-view testing on the pretrained video model.
Args:
cfg (CfgNode): configs. Details can be found in
slowfast/config/defaults.py
"""
# Set random seed from configs.
np.random.seed(cfg.RNG_SEED)
torch.manual_seed(cfg.RNG_SEED)
# Setup logging format.
logging.setup_logging()
# Print config.
logger.info("Test with config:")
logger.info(cfg)
# Build the video model and print model statistics.
model = build_model(cfg)
if du.is_master_proc():
misc.log_model_info(model, cfg, is_train=False)
# Load a checkpoint to test if applicable.
if cfg.TEST.CHECKPOINT_FILE_PATH != "":
cu.load_checkpoint(
cfg.TEST.CHECKPOINT_FILE_PATH,
model,
cfg.NUM_GPUS > 1,
None,
inflation=False,
convert_from_caffe2=cfg.TEST.CHECKPOINT_TYPE == "caffe2",
)
elif cu.has_checkpoint(cfg.OUTPUT_DIR):
last_checkpoint = cu.get_last_checkpoint(cfg.OUTPUT_DIR)
cu.load_checkpoint(last_checkpoint, model, cfg.NUM_GPUS > 1)
elif cfg.TRAIN.CHECKPOINT_FILE_PATH != "":
# If no checkpoint found in TEST.CHECKPOINT_FILE_PATH or in the current
# checkpoint folder, try to load checkpint from
# TRAIN.CHECKPOINT_FILE_PATH and test it.
cu.load_checkpoint(
cfg.TRAIN.CHECKPOINT_FILE_PATH,
model,
cfg.NUM_GPUS > 1,
None,
inflation=False,
convert_from_caffe2=cfg.TRAIN.CHECKPOINT_TYPE == "caffe2",
)
else:
# raise NotImplementedError("Unknown way to load checkpoint.")
logger.info("Testing with random initialization. Only for debugging.")
# Create video testing loaders.
test_loader = loader.construct_loader(cfg, "test")
logger.info("Testing model for {} iterations".format(len(test_loader)))
if cfg.DETECTION.ENABLE:
assert cfg.NUM_GPUS == cfg.TEST.BATCH_SIZE
test_meter = AVAMeter(len(test_loader), cfg, mode="test")
else:
assert (
len(test_loader.dataset) %
(cfg.TEST.NUM_ENSEMBLE_VIEWS * cfg.TEST.NUM_SPATIAL_CROPS) == 0)
# Create meters for multi-view testing.
if cfg.TEST.DATASET == 'epickitchens':
test_meter = EPICTestMeter(
len(test_loader.dataset) //
(cfg.TEST.NUM_ENSEMBLE_VIEWS * cfg.TEST.NUM_SPATIAL_CROPS),
cfg.TEST.NUM_ENSEMBLE_VIEWS * cfg.TEST.NUM_SPATIAL_CROPS,
cfg.MODEL.NUM_CLASSES,
len(test_loader),
)
else:
test_meter = TestMeter(
len(test_loader.dataset) //
(cfg.TEST.NUM_ENSEMBLE_VIEWS * cfg.TEST.NUM_SPATIAL_CROPS),
cfg.TEST.NUM_ENSEMBLE_VIEWS * cfg.TEST.NUM_SPATIAL_CROPS,
cfg.MODEL.NUM_CLASSES,
len(test_loader),
)
# # Perform multi-view test on the entire dataset.
preds, labels, metadata = perform_test(test_loader, model, test_meter, cfg)
if du.is_master_proc():
if cfg.TEST.DATASET == 'epickitchens':
results = {
'scores': {
'verb': preds[0],
'noun': preds[1]
},
'labels': {
'verb': labels[0],
'noun': labels[1]
},
'narration_id': metadata
}
scores_path = os.path.join(cfg.OUTPUT_DIR, 'scores')
if not os.path.exists(scores_path):
os.makedirs(scores_path)
file_path = os.path.join(scores_path,
cfg.EPICKITCHENS.TEST_SPLIT + '.pkl')
pickle.dump(results, open(file_path, 'wb'))
def test(cfg):
"""
Perform multi-view testing on the pretrained video model.
Args:
cfg (CfgNode): configs. Details can be found in
slowfast/config/defaults.py
"""
# Set random seed from configs.
np.random.seed(cfg.RNG_SEED)
torch.manual_seed(cfg.RNG_SEED)
# Setup logging format.
logging.setup_logging()
# Print config.
logger.info("Test with config:")
logger.info(cfg)
# Build the video model and print model statistics.
model = model_builder.build_model(cfg)
if du.is_master_proc():
misc.log_model_info(model)
# Load a checkpoint to test if applicable.
if cfg.TEST.CHECKPOINT_FILE_PATH != "":
cu.load_checkpoint(
cfg.TEST.CHECKPOINT_FILE_PATH,
model,
cfg.NUM_GPUS > 1,
None,
inflation=False,
convert_from_caffe2=cfg.TEST.CHECKPOINT_TYPE == "caffe2",
)
elif cu.has_checkpoint(cfg.OUTPUT_DIR):
last_checkpoint = cu.get_last_checkpoint(cfg.OUTPUT_DIR)
cu.load_checkpoint(last_checkpoint, model, cfg.NUM_GPUS > 1)
elif cfg.TRAIN.CHECKPOINT_FILE_PATH != "":
# If no checkpoint found in TEST.CHECKPOINT_FILE_PATH or in the current
# checkpoint folder, try to load checkpint from
# TRAIN.CHECKPOINT_FILE_PATH and test it.
cu.load_checkpoint(
cfg.TRAIN.CHECKPOINT_FILE_PATH,
model,
cfg.NUM_GPUS > 1,
None,
inflation=False,
convert_from_caffe2=cfg.TRAIN.CHECKPOINT_TYPE == "caffe2",
)
else:
# raise NotImplementedError("Unknown way to load checkpoint.")
logger.info("Testing with random initialization. Only for debugging.")
# Create video testing loaders.
test_loader = loader.construct_loader(cfg, "test")
logger.info("Testing model for {} iterations".format(len(test_loader)))
if cfg.DETECTION.ENABLE:
assert cfg.NUM_GPUS == cfg.TEST.BATCH_SIZE
test_meter = AVAMeter(len(test_loader), cfg, mode="test")
else:
assert (
len(test_loader.dataset)
% (cfg.TEST.NUM_ENSEMBLE_VIEWS * cfg.TEST.NUM_SPATIAL_CROPS)
== 0
)
# Create meters for multi-view testing.
test_meter = TestMeter(
len(test_loader.dataset)
// (cfg.TEST.NUM_ENSEMBLE_VIEWS * cfg.TEST.NUM_SPATIAL_CROPS),
cfg.TEST.NUM_ENSEMBLE_VIEWS * cfg.TEST.NUM_SPATIAL_CROPS,
cfg.MODEL.NUM_CLASSES,
len(test_loader),
)
# # Perform multi-view test on the entire dataset.
perform_test(test_loader, model, test_meter, cfg)
def train(cfg):
"""
Train a video model for many epochs on train set and evaluate it on val set.
Args:
cfg (CfgNode): configs. Details can be found in
slowfast/config/defaults.py
"""
# Set random seed from configs.
np.random.seed(cfg.RNG_SEED)
torch.manual_seed(cfg.RNG_SEED)
# Setup logging format.
logging.setup_logging()
if du.is_master_proc():
if cfg.ENABLE_WANDB:
wandb.login()
wandb.init(project='bbox', entity='slowfast')
# Print config.
# logger.info("Train with config:")
# logger.info(pprint.pformat(cfg))
# Build the video model and print model statistics.
model = build_model(cfg)
if cfg.EPICKITCHENS.USE_BBOX and not cu.has_checkpoint(cfg.OUTPUT_DIR):
slow_fast_model = SlowFast(cfg)
if cfg.EPICKITCHENS.USE_BBOX and cfg.EPICKITCHENS.LOAD_SLOWFAST_PRETRAIN:
if cfg.EPICKITCHENS.SLOWFAST_PRETRAIN_CHECKPOINT_FILE_PATH != "":
_ = cu.load_checkpoint(
cfg.EPICKITCHENS.SLOWFAST_PRETRAIN_CHECKPOINT_FILE_PATH,
slow_fast_model,
False,
optimizer=None,
inflation=cfg.TRAIN.CHECKPOINT_INFLATE,
convert_from_caffe2=False,
)
# cfg.TRAIN.CHECKPOINT_TYPE == "caffe2"
logger.info("Load from slowfast.")
if cfg.NUM_GPUS > 1:
model.module.load_weight_slowfast(slow_fast_model)
else:
model.load_weight_slowfast(slow_fast_model)
# if du.is_master_proc():
# misc.log_model_info(model, cfg, is_train=True)
if cfg.BN.FREEZE:
model.freeze_fn('bn_parameters')
if cfg.EPICKITCHENS.USE_BBOX and cfg.EPICKITCHENS.FREEZE_BACKBONE:
model.freeze_fn('slowfast_bbox')
# Construct the optimizer.
optimizer = optim.construct_optimizer(model, cfg)
# Load a checkpoint to resume training if applicable.
if cfg.TRAIN.AUTO_RESUME and cu.has_checkpoint(cfg.OUTPUT_DIR):
logger.info("Load from last checkpoint.")
last_checkpoint = cu.get_last_checkpoint(cfg.OUTPUT_DIR)
checkpoint_epoch = cu.load_checkpoint(last_checkpoint, model,
cfg.NUM_GPUS > 1, optimizer)
start_epoch = checkpoint_epoch + 1
elif cfg.TRAIN.CHECKPOINT_FILE_PATH != "" and not cfg.TRAIN.FINETUNE:
logger.info("Load from given checkpoint file.")
checkpoint_epoch = cu.load_checkpoint(
cfg.TRAIN.CHECKPOINT_FILE_PATH,
model,
cfg.NUM_GPUS > 1,
optimizer,
inflation=cfg.TRAIN.CHECKPOINT_INFLATE,
convert_from_caffe2=cfg.TRAIN.CHECKPOINT_TYPE == "caffe2",
)
start_epoch = checkpoint_epoch + 1
elif cfg.TRAIN.CHECKPOINT_FILE_PATH != "" and cfg.TRAIN.FINETUNE:
logger.info("Load from given checkpoint file. Finetuning.")
_ = cu.load_checkpoint(
cfg.TRAIN.CHECKPOINT_FILE_PATH,
model,
cfg.NUM_GPUS > 1,
None,
inflation=cfg.TRAIN.CHECKPOINT_INFLATE,
convert_from_caffe2=cfg.TRAIN.CHECKPOINT_TYPE == "caffe2",
)
start_epoch = 0
else:
start_epoch = 0
# Create the video train and val loaders.
if cfg.TRAIN.DATASET != 'epickitchens' or not cfg.EPICKITCHENS.TRAIN_PLUS_VAL:
train_loader = loader.construct_loader(cfg, "train")
val_loader = loader.construct_loader(cfg, "val")
logger.info("Train loader size: {}".format(len(train_loader)))
logger.info("Val loader size: {}".format(len(val_loader)))
else:
train_loader = loader.construct_loader(cfg, "train+val")
val_loader = loader.construct_loader(cfg, "val")
# Create meters.
if cfg.DETECTION.ENABLE:
train_meter = AVAMeter(len(train_loader), cfg, mode="train")
val_meter = AVAMeter(len(val_loader), cfg, mode="val")
else:
if cfg.TRAIN.DATASET == 'epickitchens':
train_meter = EPICTrainMeterSimple(len(train_loader), cfg)
val_meter = EPICValMeterSimple(len(val_loader), cfg)
else:
train_meter = TrainMeter(len(train_loader), cfg)
val_meter = ValMeter(len(val_loader), cfg)
# Perform the training loop.
logger.info("Start epoch: {}".format(start_epoch + 1))
cnt = 0
# eval_epoch(val_loader, model, val_meter, 0, cfg, cnt)
# test_from_train(model, cfg, cnt=cnt)
for cur_epoch in range(start_epoch, cfg.SOLVER.MAX_EPOCH):
# Shuffle the dataset.
loader.shuffle_dataset(train_loader, cur_epoch)
# Train for one epoch.
cnt = train_epoch(train_loader, model, optimizer, train_meter,
cur_epoch, cfg, cnt)
# Compute precise BN stats.
if cfg.BN.USE_PRECISE_STATS and len(get_bn_modules(model)) > 0:
calculate_and_update_precise_bn(train_loader, model,
cfg.BN.NUM_BATCHES_PRECISE)
# Save a checkpoint.
if cu.is_checkpoint_epoch(cur_epoch, cfg.TRAIN.CHECKPOINT_PERIOD):
cu.save_checkpoint(cfg.OUTPUT_DIR, model, optimizer, cur_epoch,
cfg)
# Evaluate the model on validation set.
if misc.is_eval_epoch(cfg, cur_epoch):
is_best_epoch = eval_epoch(val_loader, model, val_meter, cur_epoch,
cfg, cnt)
if is_best_epoch:
cu.save_checkpoint(cfg.OUTPUT_DIR,
model,
optimizer,
cur_epoch,
cfg,
is_best_epoch=is_best_epoch)
def train(cfg):
"""
Train a video model for many epochs on train set and evaluate it on val set.
Args:
cfg (CfgNode): configs. Details can be found in
slowfast/config/defaults.py
"""
# Set up environment.
du.init_distributed_training(cfg)
# Set random seed from configs.
np.random.seed(cfg.RNG_SEED)
torch.manual_seed(cfg.RNG_SEED)
# Setup logging format.
logging.setup_logging(cfg.OUTPUT_DIR)
# Init multigrid.
multigrid = None
if cfg.MULTIGRID.LONG_CYCLE or cfg.MULTIGRID.SHORT_CYCLE:
multigrid = MultigridSchedule()
cfg = multigrid.init_multigrid(cfg)
if cfg.MULTIGRID.LONG_CYCLE:
cfg, _ = multigrid.update_long_cycle(cfg, cur_epoch=0)
# Print config.
logger.info("Train with config:")
logger.info(pprint.pformat(cfg))
# Build the video model and print model statistics.
model = build_model(cfg)
if du.is_master_proc() and cfg.LOG_MODEL_INFO:
misc.log_model_info(model, cfg, use_train_input=True)
# Construct the optimizer.
optimizer = optim.construct_optimizer(model, cfg)
# Load a checkpoint to resume training if applicable.
start_epoch = cu.load_train_checkpoint(cfg, model, optimizer)
# Create the video train and val loaders.
train_loader = loader.construct_loader(cfg, "train")
val_loader = loader.construct_loader(cfg, "val")
precise_bn_loader = (loader.construct_loader(
cfg, "train", is_precise_bn=True)
if cfg.BN.USE_PRECISE_STATS else None)
# Create meters.
if cfg.DETECTION.ENABLE:
train_meter = AVAMeter(len(train_loader), cfg, mode="train")
val_meter = AVAMeter(len(val_loader), cfg, mode="val")
else:
train_meter = TrainMeter(len(train_loader), cfg)
val_meter = ValMeter(len(val_loader), cfg)
# set up writer for logging to Tensorboard format.
if cfg.TENSORBOARD.ENABLE and du.is_master_proc(
cfg.NUM_GPUS * cfg.NUM_SHARDS):
writer = tb.TensorboardWriter(cfg)
else:
writer = None
# Perform the training loop.
logger.info("Start epoch: {}".format(start_epoch + 1))
for cur_epoch in range(start_epoch, cfg.SOLVER.MAX_EPOCH):
if cfg.MULTIGRID.LONG_CYCLE:
cfg, changed = multigrid.update_long_cycle(cfg, cur_epoch)
if changed:
(
model,
optimizer,
train_loader,
val_loader,
precise_bn_loader,
train_meter,
val_meter,
) = build_trainer(cfg)
# Load checkpoint.
if cu.has_checkpoint(cfg.OUTPUT_DIR):
last_checkpoint = cu.get_last_checkpoint(cfg.OUTPUT_DIR)
assert "{:05d}.pyth".format(cur_epoch) in last_checkpoint
else:
last_checkpoint = cfg.TRAIN.CHECKPOINT_FILE_PATH
logger.info("Load from {}".format(last_checkpoint))
cu.load_checkpoint(last_checkpoint, model, cfg.NUM_GPUS > 1,
optimizer)
# Shuffle the dataset.
loader.shuffle_dataset(train_loader, cur_epoch)
# Train for one epoch.
train_epoch(train_loader, model, optimizer, train_meter, cur_epoch,
cfg, writer)
is_checkp_epoch = (cu.is_checkpoint_epoch(
cfg,
cur_epoch,
None if multigrid is None else multigrid.schedule,
))
is_eval_epoch = misc.is_eval_epoch(
cfg, cur_epoch, None if multigrid is None else multigrid.schedule)
# Compute precise BN stats.
if ((is_checkp_epoch or is_eval_epoch) and cfg.BN.USE_PRECISE_STATS
and len(get_bn_modules(model)) > 0):
calculate_and_update_precise_bn(
precise_bn_loader,
model,
min(cfg.BN.NUM_BATCHES_PRECISE, len(precise_bn_loader)),
cfg.NUM_GPUS > 0,
)
_ = misc.aggregate_sub_bn_stats(model)
# Save a checkpoint.
if is_checkp_epoch:
cu.save_checkpoint(cfg.OUTPUT_DIR, model, optimizer, cur_epoch,
cfg)
# Evaluate the model on validation set.
if is_eval_epoch:
eval_epoch(val_loader, model, val_meter, cur_epoch, cfg, writer)
if writer is not None:
writer.close()
def test(cfg):
"""
Perform multi-view testing on the pretrained video model.
Args:
cfg (CfgNode): configs. Details can be found in
slowfast/config/defaults.py
"""
# Set random seed from configs.
np.random.seed(cfg.RNG_SEED)
torch.manual_seed(cfg.RNG_SEED)
# Setup logging format.
logging.setup_logging()
# ADD Demo
if cfg.TEST.DEMO_PATH != "":
generate_subclip(cfg)
# Print config.
logger.info("Test with config:")
logger.info(cfg)
# Build the video model and print model statistics.
model = build_model(cfg)
if du.is_master_proc():
misc.log_model_info(model, cfg, is_train=False)
# Load a checkpoint to test if applicable.
if cfg.TEST.CHECKPOINT_FILE_PATH != "":
cu.load_checkpoint(
cfg.TEST.CHECKPOINT_FILE_PATH,
model,
cfg.NUM_GPUS > 1,
None,
inflation=False,
convert_from_caffe2=cfg.TEST.CHECKPOINT_TYPE == "caffe2",
)
elif cu.has_checkpoint(cfg.OUTPUT_DIR):
last_checkpoint = cu.get_last_checkpoint(cfg.OUTPUT_DIR)
cu.load_checkpoint(last_checkpoint, model, cfg.NUM_GPUS > 1)
elif cfg.TRAIN.CHECKPOINT_FILE_PATH != "":
# If no checkpoint found in TEST.CHECKPOINT_FILE_PATH or in the current
# checkpoint folder, try to load checkpint from
# TRAIN.CHECKPOINT_FILE_PATH and test it.
cu.load_checkpoint(
cfg.TRAIN.CHECKPOINT_FILE_PATH,
model,
cfg.NUM_GPUS > 1,
None,
inflation=False,
convert_from_caffe2=cfg.TRAIN.CHECKPOINT_TYPE == "caffe2",
)
else:
# raise NotImplementedError("Unknown way to load checkpoint.")
logger.info("Testing with random initialization. Only for debugging.")
# Create video testing loaders.
test_loader = loader.construct_loader(cfg, "test")
logger.info("Testing model for {} iterations".format(len(test_loader)))
if cfg.DETECTION.ENABLE:
assert cfg.NUM_GPUS == cfg.TEST.BATCH_SIZE
test_meter = AVAMeter(len(test_loader), cfg, mode="test")
else:
assert (
len(test_loader.dataset) %
(cfg.TEST.NUM_ENSEMBLE_VIEWS * cfg.TEST.NUM_SPATIAL_CROPS) == 0)
# Create meters for multi-view testing.
test_meter = TestMeter(
len(test_loader.dataset) //
(cfg.TEST.NUM_ENSEMBLE_VIEWS * cfg.TEST.NUM_SPATIAL_CROPS),
cfg.TEST.NUM_ENSEMBLE_VIEWS * cfg.TEST.NUM_SPATIAL_CROPS,
cfg.MODEL.NUM_CLASSES,
len(test_loader),
(cfg.TEST.DEMO_PATH != ""),
)
# # Perform multi-view test on the entire dataset.
perform_test(test_loader, model, test_meter, cfg)
if cfg.TEST.DEMO_PATH != "":
end_time_dir = os.path.join(cfg.DATA.PATH_TO_DATA_DIR, "end_time.npy")
end_time_array = np.load(end_time_dir)
cwd = os.getcwd()
probability_dir = os.path.join(cwd, "tmp/probability.npy")
p_array = np.load(probability_dir)
print(end_time_array)
print(p_array)
end_time_dir = os.path.join(cfg.DATA.PATH_TO_DATA_DIR)
if end_time_dir.split("/")[-1] == "demo":
try:
shutil.rmtree(end_time_dir)
except OSError as e:
print("Error: %s - %s." % (e.filename, e.strerror))
json_list = []
for i in range(len(end_time_array)):
json_list.append([end_time_array[i], p_array[i]])
print(json_list)
with open('timeLabel.json', 'w') as f:
json.dump({"jogging": str(json_list)}, f)
plt.plot(end_time_array, p_array, 'ro-')
plt.axis(
[0, end_time_array[len(end_time_array) - 1] + 0.5, -0.01, 1.01])
plt.xlabel('time(s)')
plt.ylabel('probability')
plt.suptitle("Jogging/Running")
plt.savefig("fig.png")
