def train_model(sym_net,
model_prefix,
dataset,
input_conf,
clip_length=16,
train_frame_interval=2,
resume_epoch=-1,
batch_size=4,
save_frequency=1,
lr_base=0.01,
lr_factor=0.1,
lr_steps=[400000, 800000],
end_epoch=1000,
distributed=False,
fine_tune=False,
**kwargs):
assert torch.cuda.is_available(), "Currently, we only support CUDA version"
# data iterator
iter_seed = torch.initial_seed() + 100 + max(0, resume_epoch) * 100
train_iter = iter_fac.creat(name=dataset,
batch_size=batch_size,
clip_length=clip_length,
train_interval=train_frame_interval,
mean=input_conf['mean'],
std=input_conf['std'],
seed=iter_seed)
# wapper (dynamic model)
net = model(
net=sym_net,
criterion=nn.CrossEntropyLoss().cuda(),
model_prefix=model_prefix,
step_callback_freq=50,
save_checkpoint_freq=save_frequency,
opt_batch_size=batch_size,
)
net.net.cuda()
# config optimization
param_base_layers = []
param_new_layers = []
name_base_layers = []
for name, param in net.net.named_parameters():
if fine_tune:
if ('classifier' in name) or ('fc' in name):
param_new_layers.append(param)
else:
param_base_layers.append(param)
name_base_layers.append(name)
else:
param_new_layers.append(param)
if name_base_layers:
out = "[\'" + '\', \''.join(name_base_layers) + "\']"
logging.info(
"Optimizer:: >> recuding the learning rate of {} params: {}".
format(
len(name_base_layers),
out if len(out) < 300 else out[0:150] + " ... " + out[-150:]))
net.net = torch.nn.DataParallel(net.net).cuda()
optimizer = torch.optim.SGD([{
'params': param_base_layers,
'lr_mult': 0.2
}, {
'params': param_new_layers,
'lr_mult': 1.0
}],
lr=lr_base,
momentum=0.9,
weight_decay=0.0001,
nesterov=True)
# load params from pretrained 3d network
if resume_epoch > 0:
logging.info("Initializer:: resuming model from previous training")
# resume training: model and optimizer
if resume_epoch < 0:
epoch_start = 0
step_counter = 0
else:
net.load_checkpoint(epoch=resume_epoch, optimizer=optimizer)
epoch_start = resume_epoch
step_counter = epoch_start * train_iter.__len__()
# set learning rate scheduler
num_worker = dist.get_world_size() if torch.distributed.is_initialized(
) else 1
lr_scheduler = MultiFactorScheduler(
base_lr=lr_base,
steps=[int(x / (batch_size * num_worker)) for x in lr_steps],
factor=lr_factor,
step_counter=step_counter)
# define evaluation metric
metrics = metric.MetricList(
metric.Loss(name="loss-ce"),
metric.Accuracy(name="top1", topk=1),
metric.Accuracy(name="top5", topk=5),
)
net.fit(
train_iter=train_iter,
optimizer=optimizer,
lr_scheduler=lr_scheduler,
metrics=metrics,
epoch_start=epoch_start,
epoch_end=end_epoch,
)
def train_model(sym_net, model_prefix, dataset, input_conf,
clip_length=16, train_frame_interval=2, val_frame_interval=2,
resume_epoch=-1, batch_size=4, save_frequency=1,
lr_base=0.01, lr_factor=0.1, lr_steps=[400000, 800000],
end_epoch=1000, distributed=False,
pretrained_3d=None, fine_tune=False,
load_from_frames=False, use_flow=False, triplet_loss=False,
**kwargs):
assert torch.cuda.is_available(), "Currently, we only support CUDA version"
# data iterator
iter_seed = torch.initial_seed() \
+ (torch.distributed.get_rank() * 10 if distributed else 100) \
+ max(0, resume_epoch) * 100
train_iter, eval_iter = iterator_factory.creat(name=dataset,
batch_size=batch_size,
clip_length=clip_length,
train_interval=train_frame_interval,
val_interval=val_frame_interval,
mean=input_conf['mean'],
std=input_conf['std'],
seed=iter_seed,
load_from_frames=load_from_frames,
use_flow=use_flow)
# wapper (dynamic model)
if use_flow:
class LogNLLLoss(torch.nn.Module):
def __init__(self):
super(LogNLLLoss, self).__init__()
self.loss = torch.nn.NLLLoss()
def forward(self, output, target):
output = torch.log(output)
loss = self.loss(output, target)
return loss
# criterion = LogNLLLoss().cuda()
criterion = torch.nn.CrossEntropyLoss().cuda()
elif triplet_loss:
logging.info("Using triplet loss")
criterion=torch.nn.MarginRankingLoss().cuda()
else:
criterion = torch.nn.CrossEntropyLoss().cuda()
net = model(net=sym_net,
criterion=criterion,
triplet_loss=triplet_loss,
model_prefix=model_prefix,
step_callback_freq=50,
save_checkpoint_freq=save_frequency,
opt_batch_size=batch_size, # optional
)
net.net.cuda()
# config optimization
param_base_layers = []
param_new_layers = []
name_base_layers = []
for name, param in net.net.named_parameters():
if fine_tune:
# if name.startswith('classifier'):
if 'classifier' in name or 'fc' in name:
param_new_layers.append(param)
else:
param_base_layers.append(param)
name_base_layers.append(name)
else:
param_new_layers.append(param)
if name_base_layers:
out = "[\'" + '\', \''.join(name_base_layers) + "\']"
logging.info("Optimizer:: >> recuding the learning rate of {} params: {}".format(len(name_base_layers),
out if len(out) < 300 else out[0:150] + " ... " + out[-150:]))
if distributed:
net.net = torch.nn.parallel.DistributedDataParallel(net.net).cuda()
else:
net.net = torch.nn.DataParallel(net.net).cuda()
optimizer = torch.optim.SGD([{'params': param_base_layers, 'lr_mult': 0.2},
{'params': param_new_layers, 'lr_mult': 1.0}],
lr=lr_base,
momentum=0.9,
weight_decay=0.0001,
nesterov=True)
# load params from pretrained 3d network
if pretrained_3d:
if resume_epoch < 0:
if os.path.exists(pretrained_3d):
# assert os.path.exists(pretrained_3d), "cannot locate: '{}'".format(pretrained_3d)
logging.info("Initializer:: loading model states from: `{}'".format(pretrained_3d))
checkpoint = torch.load(pretrained_3d)
net.load_state(checkpoint['state_dict'], mode='ada')
else:
logging.warning("cannot locate: '{}'".format(pretrained_3d))
else:
logging.info("Initializer:: skip loading model states from: `{}'"
+ ", since it's going to be overwrited by the resumed model".format(pretrained_3d))
# resume training: model and optimizer
if resume_epoch < 0:
epoch_start = 0
step_counter = 0
else:
net.load_checkpoint(epoch=resume_epoch, optimizer=optimizer)
epoch_start = resume_epoch
step_counter = epoch_start * train_iter.__len__()
# set learning rate scheduler
num_worker = 1
lr_scheduler = MultiFactorScheduler(base_lr=lr_base,
steps=[int(x/(batch_size*num_worker)) for x in lr_steps],
factor=lr_factor,
step_counter=step_counter)
# define evaluation metric
if triplet_loss:
metrics = metric.MetricList(metric.Loss(name="loss-triplet"),
metric.TripletAccuracy(name="acc"), )
else:
metrics = metric.MetricList(metric.Loss(name="loss-ce"),
metric.Accuracy(name="top1", topk=1),
metric.Accuracy(name="top5", topk=5), )
# enable cudnn tune
cudnn.benchmark = True
net.fit(train_iter=train_iter,
eval_iter=eval_iter,
optimizer=optimizer,
lr_scheduler=lr_scheduler,
metrics=metrics,
epoch_start=epoch_start,
epoch_end=end_epoch,)
def train_model(net_name,
sym_net,
model_prefix,
dataset,
input_conf,
modality='rgb',
split=1,
clip_length=16,
train_frame_interval=2,
val_frame_interval=2,
resume_epoch=-1,
batch_size=4,
save_frequency=1,
lr_base=0.01,
lr_base2=0.01,
lr_d=None,
lr_factor=0.1,
lr_steps=[400000, 800000],
end_epoch=1000,
distributed=False,
pretrained_3d=None,
fine_tune=False,
iter_size=1,
optim='sgd',
accumulate=True,
ds_factor=16,
epoch_thre=1,
score_dir=None,
mv_minmaxnorm=False,
mv_loadimg=False,
detach=False,
adv=0,
new_classifier=False,
**kwargs):
assert torch.cuda.is_available(), "Currently, we only support CUDA version"
torch.multiprocessing.set_sharing_strategy('file_system')
import resource
rlimit = resource.getrlimit(resource.RLIMIT_NOFILE)
resource.setrlimit(resource.RLIMIT_NOFILE, (2048, rlimit[1]))
# data iterator
iter_seed = torch.initial_seed() \
+ (torch.distributed.get_rank() * 10 if distributed else 100) \
+ max(0, resume_epoch) * 100
train_iter, eval_iter = iterator_factory.creat(
name=dataset,
batch_size=batch_size,
clip_length=clip_length,
train_interval=train_frame_interval,
val_interval=val_frame_interval,
mean=input_conf['mean'],
std=input_conf['std'],
seed=iter_seed,
modality=modality,
split=split,
net_name=net_name,
accumulate=accumulate,
ds_factor=ds_factor,
mv_minmaxnorm=mv_minmaxnorm,
mv_loadimg=mv_loadimg)
#define an instance of class model
net = model(
net=sym_net,
criterion=torch.nn.CrossEntropyLoss().cuda(),
model_prefix=model_prefix,
step_callback_freq=50,
save_checkpoint_freq=save_frequency,
opt_batch_size=batch_size, # optional
criterion2=torch.nn.MSELoss().cuda()
if modality == 'flow+mp4' else None,
criterion3=torch.nn.CrossEntropyLoss().cuda() if adv > 0. else None,
adv=adv,
)
net.net.cuda()
print(torch.cuda.current_device(), torch.cuda.device_count())
# config optimization
param_base_layers = []
param_new_layers = []
name_base_layers = []
params_gf = []
params_d = []
for name, param in net.net.named_parameters():
if modality == 'flow+mp4':
if name.startswith('gen_flow_model'):
params_gf.append(param)
elif name.startswith('discriminator'):
params_d.append(param)
else:
if (name.startswith('conv3d_0c_1x1')
or name.startswith('classifier')):
#if name.startswith('classifier'):
param_new_layers.append(param)
else:
param_base_layers.append(param)
name_base_layers.append(name)
#else:
# #print(name)
# param_new_layers.append(param)
else:
if fine_tune:
if name.startswith('classifier') or name.startswith(
'conv3d_0c_1x1'):
#if name.startswith('classifier'):
param_new_layers.append(param)
else:
param_base_layers.append(param)
name_base_layers.append(name)
else:
param_new_layers.append(param)
if modality == 'flow+mp4':
if fine_tune:
lr_mul = 0.2
else:
lr_mul = 0.5
else:
lr_mul = 0.2
#print(params_d)
if name_base_layers:
out = "[\'" + '\', \''.join(name_base_layers) + "\']"
logging.info(
"Optimizer:: >> recuding the learning rate of {} params: {} by factor {}"
.format(
len(name_base_layers),
out if len(out) < 300 else out[0:150] + " ... " + out[-150:],
lr_mul))
if net_name == 'I3D':
weight_decay = 0.0001
else:
raise ValueError('UNKOWN net_name', net_name)
logging.info("Train_Model:: weight_decay: `{}'".format(weight_decay))
if distributed:
net.net = torch.nn.parallel.DistributedDataParallel(net.net).cuda()
else:
net.net = torch.nn.DataParallel(net.net).cuda()
if optim == 'adam':
optimizer = torch.optim.Adam([{
'params': param_base_layers,
'lr_mult': lr_mul
}, {
'params': param_new_layers,
'lr_mult': 1.0
}],
lr=lr_base,
weight_decay=weight_decay)
optimizer_2 = torch.optim.Adam([{
'params': param_base_layers,
'lr_mult': lr_mul
}, {
'params': param_new_layers,
'lr_mult': 1.0
}],
lr=lr_base2,
weight_decay=weight_decay)
else:
optimizer = torch.optim.SGD([{
'params': param_base_layers,
'lr_mult': lr_mul
}, {
'params': param_new_layers,
'lr_mult': 1.0
}],
lr=lr_base,
momentum=0.9,
weight_decay=weight_decay,
nesterov=True)
optimizer_2 = torch.optim.SGD([{
'params': param_base_layers,
'lr_mult': lr_mul
}, {
'params': param_new_layers,
'lr_mult': 1.0
}],
lr=lr_base2,
momentum=0.9,
weight_decay=weight_decay,
nesterov=True)
if adv > 0.:
optimizer_3 = torch.optim.Adam(params_d,
lr=lr_base,
weight_decay=weight_decay,
eps=0.001)
else:
optimizer_3 = None
if modality == 'flow+mp4':
if optim == 'adam':
optimizer_mse = torch.optim.Adam(params_gf,
lr=lr_base,
weight_decay=weight_decay,
eps=1e-08)
optimizer_mse_2 = torch.optim.Adam(params_gf,
lr=lr_base2,
weight_decay=weight_decay,
eps=0.001)
else:
optimizer_mse = torch.optim.SGD(params_gf,
lr=lr_base,
momentum=0.9,
weight_decay=weight_decay,
nesterov=True)
optimizer_mse_2 = torch.optim.SGD(params_gf,
lr=lr_base2,
momentum=0.9,
weight_decay=weight_decay,
nesterov=True)
else:
optimizer_mse = None
optimizer_mse_2 = None
# load params from pretrained 3d network
if pretrained_3d and not pretrained_3d == 'False':
if resume_epoch < 0:
assert os.path.exists(pretrained_3d), "cannot locate: `{}'".format(
pretrained_3d)
logging.info(
"Initializer:: loading model states from: `{}'".format(
pretrained_3d))
if net_name == 'I3D':
checkpoint = torch.load(pretrained_3d)
keys = list(checkpoint.keys())
state_dict = {}
for name in keys:
state_dict['module.' + name] = checkpoint[name]
del checkpoint
net.load_state(state_dict, strict=False)
if new_classifier:
checkpoint = torch.load(
'./network/pretrained/model_flow.pth')
keys = list(checkpoint.keys())
state_dict = {}
for name in keys:
state_dict['module.' + name] = checkpoint[name]
del checkpoint
net.load_state(state_dict, strict=False)
else:
checkpoint = torch.load(pretrained_3d)
net.load_state(checkpoint['state_dict'], strict=False)
else:
logging.info(
"Initializer:: skip loading model states from: `{}'" +
", since it's going to be overwrited by the resumed model".
format(pretrained_3d))
# resume training: model and optimizer
if resume_epoch < 0:
epoch_start = 0
step_counter = 0
else:
net.load_checkpoint(epoch=resume_epoch,
optimizer=optimizer,
optimizer_mse=optimizer_mse)
epoch_start = resume_epoch
step_counter = epoch_start * train_iter.__len__()
# set learning rate scheduler
num_worker = dist.get_world_size() if torch.distributed._initialized else 1
lr_scheduler = MultiFactorScheduler(
base_lr=lr_base,
steps=[int(x / (batch_size * num_worker)) for x in lr_steps],
factor=lr_factor,
step_counter=step_counter)
if modality == 'flow+mp4':
lr_scheduler2 = MultiFactorScheduler(
base_lr=lr_base2,
steps=[int(x / (batch_size * num_worker)) for x in lr_steps],
factor=lr_factor,
step_counter=step_counter)
if lr_d == None:
lr_scheduler3 = MultiFactorScheduler(
base_lr=lr_d,
steps=[int(x / (batch_size * num_worker)) for x in lr_steps],
factor=lr_factor,
step_counter=step_counter)
else:
print("_____________", lr_d)
lr_scheduler3 = MultiFactorScheduler(
base_lr=lr_d,
steps=[int(x / (batch_size * num_worker)) for x in lr_steps],
factor=lr_factor,
step_counter=step_counter)
else:
lr_scheduler2 = None
lr_scheduler3 = None
# define evaluation metric
metrics_D = None
if modality == 'flow+mp4':
metrics = metric.MetricList(
metric.Loss(name="loss-ce"),
metric.Loss(name="loss-mse"),
metric.Accuracy(name="top1", topk=1),
metric.Accuracy(name="top5", topk=5),
)
if adv > 0:
metrics_D = metric.MetricList(metric.Loss(name="classi_D"),
metric.Loss(name="adv_D"))
else:
metrics = metric.MetricList(
metric.Loss(name="loss-ce"),
metric.Accuracy(name="top1", topk=1),
metric.Accuracy(name="top5", topk=5),
)
# enable cudnn tune
cudnn.benchmark = True
net.fit(train_iter=train_iter,
eval_iter=eval_iter,
optimizer=optimizer,
lr_scheduler=lr_scheduler,
metrics=metrics,
epoch_start=epoch_start,
epoch_end=end_epoch,
iter_size=iter_size,
optimizer_mse=optimizer_mse,
optimizer_2=optimizer_2,
optimizer_3=optimizer_3,
optimizer_mse_2=optimizer_mse_2,
lr_scheduler2=lr_scheduler2,
lr_scheduler3=lr_scheduler3,
metrics_D=metrics_D,
epoch_thre=epoch_thre,
score_dir=score_dir,
detach=detach)
import datetime
from network.srtg_resnet import srtg_r2plus1d_50
from data import iterator_factory
from train import metric
from train.model import model
#%% Define params
# define path
dataset_path = '/nfs/s2/userhome/zhouming/workingdir/Video/HACS/dataset'
working_path = '/nfs/s2/userhome/zhouming/workingdir/Video/HACS/train_model'
output_path = f'{working_path}/out'
#%% prepare model
net = srtg_r2plus1d_50(num_classes=200)
net = torch.nn.DataParallel(net, device_ids=[0])
model = model(net=net,
criterion=torch.nn.CrossEntropyLoss().cuda())
info = torch.load(f'{working_path}/models/srtg_r2plus1d_50_best.pth')
model.load_state(info['state_dict'],strict=True)
del info
#%% prepare dataset
# data iterator - randomisation based on date and time values
iter_seed = torch.initial_seed() + 100
now = datetime.datetime.now()
iter_seed += now.year + now.month + now.day + now.hour + now.minute + now.second
# Create custom loaders for validation
eval_loader = iterator_factory.create(
name='HACS',
batch_size=8,
val_clip_length=16,
def train_model(sym_net, name, model_prefix, input_conf,
clip_length=32, clip_size=224, train_frame_interval=2, val_frame_interval=2,
resume_epoch=-1, batch_size=16, save_frequency=1,
lr_base=0.01, lr_factor=0.1, lr_steps=[50,100,150],
enable_long_cycles=True, enable_short_cycles=True, end_epoch=300,
pretrained_3d=None, fine_tune=False, dataset_location='Kinetics', net_name='r3d_50', gpus=4,
**kwargs):
assert torch.cuda.is_available(), "Only support CUDA devices."
# Make results directory for .csv files if it does not exist
results_path = str('./results/'+str(name)+'/'+str(net_name))
if not os.path.exists(results_path):
os.makedirs(results_path)
# data iterator - randomisation based on date and time values
iter_seed = torch.initial_seed() + 100 + max(0, resume_epoch) * 100
now = datetime.datetime.now()
iter_seed += now.year + now.month + now.day + now.hour + now.minute + now.second
# Get parent location
# - `data` folder should include all the dataset examples.
# - `labels` folder should inclde all labels in .csv format.
# We use a global label formating - you can have a look at the link in the `README.md` to download the files.
data_location = dataset_location.split('/data/')[0]
clip_length = int(clip_length)
clip_size = int(clip_size)
train_loaders = {}
# Create custom loaders for train and validation
train_data, eval_loader, train_length = iterator_factory.create(
name=name,
batch_size=batch_size,
return_len=True,
clip_length=clip_length,
clip_size=clip_size,
val_clip_length=clip_length,
val_clip_size=clip_size,
train_interval=train_frame_interval,
val_interval=val_frame_interval,
mean=input_conf['mean'],
std=input_conf['std'],
seed=iter_seed,
data_root=data_location)
# Create model
net = model(net=sym_net,
criterion=torch.nn.CrossEntropyLoss().cuda(),
model_prefix=model_prefix,
step_callback_freq=1,
save_checkpoint_freq=save_frequency,
opt_batch_size=batch_size, # optional
)
net.net.cuda()
# Parameter LR configuration for optimiser
# Base layers are based on the layers as loaded to the model
param_base_layers = []
base_layers_mult = 1.0
# New layers are based on fine-tuning
param_new_layers = []
new_layers_mult = 1.0
name_base_layers = []
param_transpose_layers = []
transpose_layers_mult = 1.0
param_rec_layers = []
rec_layers_mult = 1.0
# Iterate over all parameters
for name, param in net.net.named_parameters():
if fine_tune:
if 'transpose' in name.lower():
param_transpose_layers.append(param)
transpose_layers_mult = .2
elif name.lower().startswith('classifier'):
new_layers_mult = .1
param_new_layers.append(param)
elif ('lstm' or 'gru') in name.lower():
param_transpose_layers.append(param)
rec_layers_mult = .5
else:
param_base_layers.append(param)
base_layers_mult = .6
name_base_layers.append(name)
else:
if 'transpose' in name.lower():
param_transpose_layers.append(param)
transpose_layers_mult = .8
elif ('lstm' or 'gru') in name.lower():
param_transpose_layers.append(param)
rec_layers_mult = 1.
else:
param_new_layers.append(param)
# User feedback
if name_base_layers:
out = "[\'" + '\', \''.join(name_base_layers) + "\']"
logging.info("Optimiser:: >> recuding the learning rate of {} params: {}".format(len(name_base_layers),
out if len(out) < 300 else out[0:150] + " ... " + out[-150:]))
optimiser = torch.optim.SGD([
{'params': param_base_layers, 'lr_mult': base_layers_mult},
{'params': param_new_layers, 'lr_mult': new_layers_mult},
{'params': param_rec_layers, 'lr_mult': rec_layers_mult},
{'params': param_transpose_layers, 'lr_mult': transpose_layers_mult},],
lr=lr_base,
momentum=0.9,
weight_decay=0.0001,
nesterov=True
)
# Use Apex for Mixed precidion - Note: Please comment out any apex code on `train_model.py` and `model.py`
# in case you wish to switch back to standard float32. "O0" opt_level still has some bugs when also using DataParallel
net.net, optimiser = amp.initialize(net.net, optimiser, opt_level="O1")
# Create DataParallel wrapper
net.net = torch.nn.DataParallel(net.net, device_ids=[i for i in range(int(gpus))])
# load params from pretrained 3d network
if pretrained_3d:
assert os.path.exists(pretrained_3d), "cannot locate: `{}'".format(pretrained_3d)
logging.info("Initialiser:: loading model states from: `{}'".format(pretrained_3d))
checkpoint = torch.load(pretrained_3d)
net.load_state(checkpoint['state_dict'], strict=False)
num_steps = train_length // batch_size
# Long Cycle steps
if (enable_long_cycles):
count = 0
index = 0
iter_sizes = [8, 4, 2, 1]
initial_num = num_steps
# Expected to find the number of batches that fit exactly to the number of iterations.
# So the sum of the floowing batch sizes should be less or equal to the number of batches left.
while sum(iter_sizes[index:]) <= num_steps:
# Case 1: 8 x B
if iter_sizes[index] == 8:
count += 1
index = 1
num_steps -= 8
# Case 2: 4 x B
elif iter_sizes[index] == 4:
count += 1
index = 2
num_steps -= 4
# Case 3: 2 x B
elif iter_sizes[index] == 2:
count += 1
index = 3
num_steps -= 2
# Base case
elif iter_sizes[index] == 1:
count += 1
index = 0
num_steps -= 1
print ("New number of batches per epoch is {:d} being equivalent to {:1.3f} of original number of batches with Long cycles".format(count,float(count)/float(initial_num)))
num_steps = count
# Short Cycle steps
if (enable_short_cycles):
# Iterate for *every* batch
i = 0
while i <= num_steps:
m = i%3
# Case 1: Base case
if (m==0):
num_steps -= 1
# Case 2: b = 2 x B
if (m==1):
num_steps -= 2
# Case 3: b = 4 x B
else:
num_steps -= 4
i += 1
# Update new number of batches
print ("New number of batches per epoch is {:d} being equivalent to {:1.3f} of original number of batches with Short cycles".format(i,float(i)/float(initial_num)))
num_steps = i
# Split the batch number to four for every change in the long cycles
long_steps = None
if (enable_long_cycles):
step = num_steps//4
long_steps = list(range(num_steps))[0::step]
num_steps = long_steps[-1]
# Create full list of long steps (for all batches)
for epoch in range(1,end_epoch):
end = long_steps[-1]
long_steps = long_steps + [x.__add__(end) for x in long_steps[-4:]]
# Fool-proofing
if (long_steps[0]==0):
long_steps[0]=1
# resume training: model and optimiser - (account of various batch sizes)
if resume_epoch < 0:
epoch_start = 0
step_counter = 0
else:
# Try to load previous state dict in case `pretrained_3d` is None
if not pretrained_3d:
try:
net.load_checkpoint(epoch=resume_epoch, optimizer=optimiser)
except Exception:
logging.warning('Initialiser:: No previous checkpoint found in the directory! You can specify the path explicitly with `pretrained_3d` argument.')
epoch_start = resume_epoch
step_counter = epoch_start * num_steps
# Step dictionary creation
iteration_steps = {'long_0':[],'long_1':[],'long_2':[],'long_3':[],'short_0':[],'short_1':[],'short_2':[]}
#Populate dictionary
for batch_i in range(0,num_steps):
# Long cycle cases
if batch_i>=0 and batch_i<num_steps//4:
iteration_steps['long_0'].append(batch_i)
elif batch_i>=num_steps//4 and batch_i<num_steps//2:
iteration_steps['long_1'].append(batch_i)
elif batch_i>=num_steps//2 and batch_i<(3*num_steps)//4:
iteration_steps['long_2'].append(batch_i)
else:
iteration_steps['long_3'].append(batch_i)
# Short cases
if (batch_i%3==0):
iteration_steps['short_0'].append(batch_i)
elif (batch_i%3==1):
iteration_steps['short_1'].append(batch_i)
else:
iteration_steps['short_2'].append(batch_i)
# set learning rate scheduler
lr_scheduler = MultiFactorScheduler(base_lr=lr_base,
steps=[x*num_steps for x in lr_steps],
iterations_per_epoch=num_steps,
iteration_steps=iteration_steps,
factor=lr_factor,
step_counter=step_counter)
# define evaluation metric
metrics = metric.MetricList(metric.Loss(name="loss-ce"),
metric.Accuracy(name="top1", topk=1),
metric.Accuracy(name="top5", topk=5),
metric.BatchSize(name="batch_size"),
metric.LearningRate(name="lr"))
# enable cudnn tune
cudnn.benchmark = True
# Main training happens here
net.fit(train_iter=train_data,
eval_iter=eval_loader,
batch_shape=(int(batch_size),int(clip_length),int(clip_size),int(clip_size)),
workers=8,
no_cycles=(not(enable_long_cycles) and not(enable_short_cycles)),
optimiser=optimiser,
long_short_steps_dir=iteration_steps,
lr_scheduler=lr_scheduler,
metrics=metrics,
iter_per_epoch=num_steps,
epoch_start=epoch_start,
epoch_end=end_epoch,
directory=results_path)
def train_model(Hash_center, sym_net, model_prefix, dataset, input_conf, hash_bit,
clip_length=16, train_frame_interval=2, val_frame_interval=2,
resume_epoch=-1, batch_size=4, save_frequency=1,
lr_base=0.01, lr_factor=0.1, lr_steps=[400000, 800000],
end_epoch=1000, distributed=False,
pretrained_3d=None, fine_tune=False,
**kwargs):
assert torch.cuda.is_available(), "Currently, we only support CUDA version"
# data iterator
iter_seed = torch.initial_seed() \
+ (torch.distributed.get_rank() * 10 if distributed else 100) \
+ max(0, resume_epoch) * 100
train_iter, eval_iter = iterator_factory.creat(name=dataset,
batch_size=batch_size,
clip_length=clip_length,
train_interval=train_frame_interval,
val_interval=val_frame_interval,
mean=input_conf['mean'],
std=input_conf['std'],
seed=iter_seed)
print(len(train_iter))
print(len(eval_iter))
# wapper (dynamic model)
net = model(net=sym_net,
criterion=torch.nn.BCELoss().cuda(),
model_prefix=model_prefix,
step_callback_freq=50,
save_checkpoint_freq=save_frequency,
opt_batch_size=batch_size, # optional
dataset=dataset, # dataset name
hash_bit=hash_bit,
)
net.net.cuda()
# config optimization
param_base_layers = []
param_new_layers = []
name_base_layers = []
for name, param in net.net.named_parameters():
if fine_tune:
#print(f'fine tune {fine_tune}')
if name.startswith('hash'):
param_new_layers.append(param)
else:
param_base_layers.append(param)
name_base_layers.append(name)
else:
param_new_layers.append(param)
if name_base_layers:
out = "[\'" + '\', \''.join(name_base_layers) + "\']"
logging.info("Optimizer:: >> recuding the learning rate of {} params: {}".format(len(name_base_layers),
out if len(out) < 300 else out[0:150] + " ... " + out[-150:]))
if distributed:
net.net = torch.nn.parallel.DistributedDataParallel(net.net).cuda()
else:
net.net = torch.nn.DataParallel(net.net).cuda()
optimizer = torch.optim.SGD([{'params': param_base_layers, 'lr_mult': 0.2},
{'params': param_new_layers, 'lr_mult': 1.0}],
lr=lr_base,
momentum=0.9,
weight_decay=0.0001,
nesterov=True)
# load params from pretrained 3d network
if pretrained_3d:
if resume_epoch < 0:
assert os.path.exists(pretrained_3d), "cannot locate: `{}'".format(pretrained_3d)
logging.info("Initializer:: loading model states from: `{}'".format(pretrained_3d))
checkpoint = torch.load(pretrained_3d)
net.load_state(checkpoint['state_dict'], strict=False)
else:
logging.info("Initializer:: skip loading model states from: `{}'"
+ ", since it's going to be overwrited by the resumed model".format(pretrained_3d))
# resume training: model and optimizer
if resume_epoch < 0:
epoch_start = 0
step_counter = 0
else:
net.load_checkpoint(epoch=resume_epoch, optimizer=optimizer)
epoch_start = resume_epoch
step_counter = epoch_start * train_iter.__len__()
# set learning rate scheduler
num_worker = dist.get_world_size() if torch.distributed._initialized else 1
lr_scheduler = MultiFactorScheduler(base_lr=lr_base,
steps=[int(x/(batch_size*num_worker)) for x in lr_steps],
factor=lr_factor,
step_counter=step_counter)
# define evaluation metric
metrics = metric.MetricList(metric.Loss(name="loss-ce"))
# enable cudnn tune
cudnn.benchmark = True
net.fit(train_iter=train_iter,
eval_iter=eval_iter,
optimizer=optimizer,
lr_scheduler=lr_scheduler,
metrics=metrics,
epoch_start=epoch_start,
epoch_end=end_epoch,
Hash_center=Hash_center,)
def train_model(sym_net,
model_prefix,
dataset,
fold,
clip_length=8,
train_frame_interval=2,
val_frame_interval=2,
resume_epoch=-1,
batch_size=4,
save_frequency=1,
lr_base=0.01,
lr_factor=0.1,
lr_steps=[400000, 800000],
end_epoch=1000,
distributed=False,
fine_tune=False,
epoch_div_factor=4,
precise_bn=False,
**kwargs):
assert torch.cuda.is_available(), "Currently, we only support CUDA version"
# data iterator
iter_seed = torch.initial_seed() \
+ (torch.distributed.get_rank() * 10 if distributed else 100) \
+ max(0, resume_epoch) * 100
train_iter, eval_iter = iterator_factory_brats.create(
name=dataset,
batch_size=batch_size,
fold=fold,
# clip_length=clip_length,
# train_interval=train_frame_interval,
# val_interval=val_frame_interval,
# mean=input_conf['mean'],
# std=input_conf['std'],
seed=iter_seed)
# model (dynamic)
net = model(
net=sym_net,
criterion=torch.nn.CrossEntropyLoss().cuda(),
model_prefix=model_prefix,
step_callback_freq=50,
save_checkpoint_freq=save_frequency,
opt_batch_size=batch_size, # optional
single_checkpoint=
precise_bn, # TODO: use shared filesystem to rsync running mean/var
)
# if True:
# for name, module in net.net.named_modules():
# if name.endswith("bn"): module.momentum = 0.005
net.net.cuda()
# config optimization, [[w/ wd], [w/o wd]]
param_base_layers = [[[], []], [[], []]]
param_new_layers = [[[], []], [[], []]]
name_freeze_layers, name_base_layers = [], []
for name, param in net.net.named_parameters():
idx_wd = 0 if name.endswith('.bias') else 1
idx_bn = 0 if name.endswith(('.bias', 'bn.weight')) else 1
if fine_tune:
if not name.startswith('classifier'):
param_base_layers[idx_bn][idx_wd].append(param)
name_base_layers.append(name)
else:
param_new_layers[idx_bn][idx_wd].append(param)
else:
if "conv_m2" in name:
param_base_layers[idx_bn][idx_wd].append(param)
name_base_layers.append(name)
else:
param_new_layers[idx_bn][idx_wd].append(param)
if name_freeze_layers:
out = "[\'" + '\', \''.join(name_freeze_layers) + "\']"
logging.info("Optimizer:: >> freezing {} params: {}".format(
len(name_freeze_layers),
out if len(out) < 300 else out[0:150] + " ... " + out[-150:]))
if name_base_layers:
out = "[\'" + '\', \''.join(name_base_layers) + "\']"
logging.info(
"Optimizer:: >> recuding the learning rate of {} params: {}".
format(
len(name_base_layers),
out if len(out) < 300 else out[0:150] + " ... " + out[-150:]))
if distributed:
net.net = torch.nn.parallel.DistributedDataParallel(net.net).cuda()
else:
net.net = torch.nn.DataParallel(net.net).cuda()
# optimizer = torch.optim.SGD(sym_net.parameters(),
wd = 0.0001
optimizer = custom_optim.SGD(
[
{
'params': param_base_layers[0][0],
'lr_mult': 0.5,
'weight_decay': 0.
},
{
'params': param_base_layers[0][1],
'lr_mult': 0.5,
'weight_decay': wd
},
{
'params': param_base_layers[1][0],
'lr_mult': 0.5,
'weight_decay': 0.,
'name': 'precise.bn'
}, # *.bias
{
'params': param_base_layers[1][1],
'lr_mult': 0.5,
'weight_decay': wd,
'name': 'precise.bn'
}, # bn.weight
{
'params': param_new_layers[0][0],
'lr_mult': 1.0,
'weight_decay': 0.
},
{
'params': param_new_layers[0][1],
'lr_mult': 1.0,
'weight_decay': wd
},
{
'params': param_new_layers[1][0],
'lr_mult': 1.0,
'weight_decay': 0.,
'name': 'precise.bn'
}, # *.bias
{
'params': param_new_layers[1][1],
'lr_mult': 1.0,
'weight_decay': wd,
'name': 'precise.bn'
}
], # bn.weight
lr=lr_base,
momentum=0.9,
nesterov=True)
# resume: model and optimizer
if resume_epoch < 0:
epoch_start = 0
step_counter = 0
else:
net.load_checkpoint(epoch=resume_epoch, optimizer=optimizer)
epoch_start = resume_epoch
step_counter = epoch_start * int(
train_iter.__len__() / epoch_div_factor)
num_worker = torch.distributed.get_world_size(
) if torch.distributed.is_initialized() else 1
lr_scheduler = MultiFactorScheduler(
base_lr=lr_base,
steps=[int(x / (batch_size * num_worker)) for x in lr_steps],
factor=lr_factor,
step_counter=step_counter)
metrics = metric.MetricList(
metric.Loss(name="loss-ce"),
metric.Accuracy(name="top1", topk=1),
metric.Accuracy(name="top2", topk=2),
)
cudnn.benchmark = True
# cudnn.fastest = False
# cudnn.enabled = False
net.fit(
train_iter=train_iter,
eval_iter=eval_iter,
optimizer=optimizer,
lr_scheduler=lr_scheduler,
metrics=metrics,
epoch_start=epoch_start,
epoch_end=end_epoch,
epoch_div_factor=epoch_div_factor,
precise_bn=precise_bn,
)
def train_model(sym_net,
dataset,
input_conf,
train_list,
val_list,
clip_length=16,
resume_epoch=-1,
batch_size=32,
save_frequency=1,
lr_base=0.1,
lr_factor=0.1,
lr_steps=[400000, 800000],
end_epoch=120,
distributed=False,
pretrained_3d=False,
fine_tune=False,
**kwargs):
import argparse
parse = argparse.ArgumentParser(description="PyTorch resume checkpoint")
parse.add_argument(
'--resume',
default=
'/workspace/mnt/group/algorithm/kanghaidong/video_project/PyTorch-MFNet/checkpoint/best_model_mfnet_3d_120.pth.tar',
type=str,
metavar='PATH',
help='path to latest checkpoint (default: none)')
#parse.add_argument('--resume', default='', type=str, metavar='PATH',help='path to latest checkpoint (default: none)')
args = parse.parse_args()
assert torch.cuda.is_available(), "Currently, we only support CUDA version"
# data iterator
iter_seed = torch.initial_seed() \
+ (torch.distributed.get_rank() * 10 if distributed else 100) \
+ max(0, resume_epoch) * 100
train_iter, eval_iter = iterator_fac.creat(
name=dataset, # # enclosed DataLoader()
train_list=train_list,
val_list=val_list,
clip_length=clip_length,
batch_size=batch_size,
mean=input_conf['mean'],
std=input_conf['std'],
seed=iter_seed)
# wapper (dynamic model)
net = model(
net=sym_net,
criterion=torch.nn.CrossEntropyLoss().cuda(
), # using CrossEntropyLoss;
step_callback_freq=50,
save_checkpoint_freq=save_frequency,
opt_batch_size=batch_size, # optional
)
net.net.cuda()
# config optimization
param_base_layers = []
param_new_layers = []
name_base_layers = []
for name, param in net.net.named_parameters():
if fine_tune:
if name.startswith('classifier'):
param_new_layers.append(param)
else:
param_base_layers.append(param)
name_base_layers.append(name)
else:
param_new_layers.append(param)
if name_base_layers:
out = "[\'" + '\', \''.join(name_base_layers) + "\']"
logging.info(
"Optimizer:: >> recuding the learning rate of {} params: {}".
format(
len(name_base_layers),
out if len(out) < 300 else out[0:150] + " ... " + out[-150:]))
if distributed:
net.net = torch.nn.parallel.DistributedDataParallel(net.net).cuda()
else:
net.net = torch.nn.DataParallel(net.net).cuda()
optimizer = torch.optim.SGD([{
'params': param_base_layers,
'lr_mult': 0.2
}, {
'params': param_new_layers,
'lr_mult': 1.0
}],
lr=lr_base,
momentum=0.9,
weight_decay=0.0001,
nesterov=True)
# load params from pretrained 3d network
#忽略了一个问题就是，你要加载的模型，你想着全用人家的参数，但是fc层你不行，因为param的shape不一样；就是分类的label数目不一样；
#怎么办？1，去掉fc的param，然后加载之前的参数，这个你需要构造新的fc；并且在加载的时候要特定的写出来，一种就是和当前模型对比一下dict，去掉不一样的参数；
#即，删除与当前model不一样的key；这个比较高效！
if args.resume:
if os.path.isfile(args.resume):
print("loading checkpoint:'{}".format(args.resume))
checkpoint = torch.load(args.resume)
#for key, v in checkpoint.items():
# print(key, v)
if pretrained_3d:
#pretrained_state = checkpoint['state_dict']
model_state = net.net.state_dict()
pretrained_state = {
k: v
for k, v in checkpoint.items() if k in model_state
} # 删除与当前model不一致的key
#for key, v in pretrained_state.items():
# print(key,v)
model_state.update(pretrained_state)
net.net.load_state_dict(model_state)
else:
print('loading sus')
epoch_start = checkpoint['epoch']
net.net.load_state_dict(checkpoint['state_state'])
else:
epoch_start = 0
print("training start new scratch !")
print("no checkpoint found at '{}".format(args.resume))
# define evaluation metric
"""
metrics = metric.MetricList(metric.Loss(name="loss-ce"),
metric.Accuracy(name="top1", topk=1),
metric.Accuracy(name="top5", topk=5),)
"""
# enable cudnn tune
cudnn.benchmark = True
net.fit(
train_iter=train_iter,
eval_iter=eval_iter,
optimizer=optimizer,
#lr_scheduler=lr_scheduler,
#metrics=metrics,
epoch_start=epoch_start,
epoch_end=end_epoch,
)
