def loadCheckpoint(self):
opts = self.opts
previous_checkpoint = opts['temp_var']['checkpoint_model']
pretrained_model = opts['temp_var']['pretrained_model']
num_xpus = opts['distributed']['num_xpus']
if (previous_checkpoint is not None):
if os.path.exists(previous_checkpoint):
log.info(
'Load previous checkpoint:{}'.format(previous_checkpoint))
start_epoch, prev_checkpointed_lr, _best_metric = \
checkpoint.initialize_params_from_file(
model=self.train_model,
weights_file=previous_checkpoint,
num_xpus=num_xpus,
opts=opts,
broadcast_computed_param=True,
reset_epoch=False,
)
elif pretrained_model is not None and os.path.exists(pretrained_model):
log.info("Load pretrained model: {}".format(pretrained_model))
start_epoch, prev_checkpointed_lr, best_metric = \
checkpoint.initialize_params_from_file(
model=self.train_model,
weights_file=pretrained_model,
num_xpus=num_xpus,
opts=opts,
broadcast_computed_param=True,
reset_epoch=opts['model_param']['reset_epoch'],
)
data_parallel_model.FinalizeAfterCheckpoint(self.train_model)
def Train(args):
# Either use specified device list or generate one
if args.gpus is not None:
gpus = [int(x) for x in args.gpus.split(',')]
num_gpus = len(gpus)
else:
gpus = list(range(args.num_gpus))
num_gpus = args.num_gpus
log.info("Running on GPUs: {}".format(gpus))
# Verify valid batch size
total_batch_size = args.batch_size
batch_per_device = total_batch_size // num_gpus
assert \
total_batch_size % num_gpus == 0, \
"Number of GPUs must divide batch size"
# Round down epoch size to closest multiple of batch size across machines
global_batch_size = total_batch_size * args.num_shards
epoch_iters = int(args.epoch_size / global_batch_size)
assert \
epoch_iters > 0, \
"Epoch size must be larger than batch size times shard count"
args.epoch_size = epoch_iters * global_batch_size
log.info("Using epoch size: {}".format(args.epoch_size))
# Create ModelHelper object
train_arg_scope = {
'order': 'NCHW',
'use_cudnn': True,
'cudnn_exhaustive_search': True,
'ws_nbytes_limit': (args.cudnn_workspace_limit_mb * 1024 * 1024),
}
train_model = model_helper.ModelHelper(
name="ban-pc-resnet50", arg_scope=train_arg_scope
)
num_shards = args.num_shards
shard_id = args.shard_id
# Expect interfaces to be comma separated.
# Use of multiple network interfaces is not yet complete,
# so simply use the first one in the list.
interfaces = args.distributed_interfaces.split(",")
# Rendezvous using MPI when run with mpirun
if os.getenv("OMPI_COMM_WORLD_SIZE") is not None:
num_shards = int(os.getenv("OMPI_COMM_WORLD_SIZE", 1))
shard_id = int(os.getenv("OMPI_COMM_WORLD_RANK", 0))
if num_shards > 1:
rendezvous = dict(
kv_handler=None,
num_shards=num_shards,
shard_id=shard_id,
engine="GLOO",
transport=args.distributed_transport,
interface=interfaces[0],
mpi_rendezvous=True,
exit_nets=None)
elif num_shards > 1:
# Create rendezvous for distributed computation
store_handler = "store_handler"
if args.redis_host is not None:
# Use Redis for rendezvous if Redis host is specified
workspace.RunOperatorOnce(
core.CreateOperator(
"RedisStoreHandlerCreate", [], [store_handler],
host=args.redis_host,
port=args.redis_port,
prefix=args.run_id,
)
)
else:
# Use filesystem for rendezvous otherwise
workspace.RunOperatorOnce(
core.CreateOperator(
"FileStoreHandlerCreate", [], [store_handler],
path=args.file_store_path,
prefix=args.run_id,
)
)
rendezvous = dict(
kv_handler=store_handler,
shard_id=shard_id,
num_shards=num_shards,
engine="GLOO",
transport=args.distributed_transport,
interface=interfaces[0],
exit_nets=None)
else:
rendezvous = None
# Model configs for constructing model
with open(args.model_config) as f:
model_config = yaml.load(f)
# Model building functions
def create_target_model_ops(model, loss_scale):
initializer = (PseudoFP16Initializer if args.dtype == 'float16'
else Initializer)
with brew.arg_scope([brew.conv, brew.fc],
WeightInitializer=initializer,
BiasInitializer=initializer,
enable_tensor_core=args.enable_tensor_core,
float16_compute=args.float16_compute):
pred = add_se_model(model, model_config, "data", is_test=False)
if args.dtype == 'float16':
pred = model.net.HalfToFloat(pred, pred + '_fp32')
loss = add_softmax_loss(model, pred, 'label')
brew.accuracy(model, ['softmax', 'label'], 'accuracy')
return [loss]
def add_optimizer(model):
'''
stepsz = int(30 * args.epoch_size / total_batch_size / num_shards)
optimizer.add_weight_decay(model, args.weight_decay)
opt = optimizer.build_multi_precision_sgd(
model,
args.base_learning_rate,
momentum=0.9,
nesterov=1,
policy="step",
stepsize=stepsz,
gamma=0.1
)
'''
optimizer.add_weight_decay(model, args.weight_decay)
opt = optimizer.build_multi_precision_sgd(
model,
base_learning_rate = args.base_learning_rate,
momentum = model_config['solver']['momentum'],
nesterov = model_config['solver']['nesterov'],
policy = model_config['solver']['lr_policy'],
power = model_config['solver']['power'],
max_iter = model_config['solver']['max_iter'],
)
return opt
# Define add_image_input function.
# Depends on the "train_data" argument.
# Note that the reader will be shared with between all GPUS.
reader = train_model.CreateDB(
"reader",
db=args.train_data,
db_type=args.db_type,
num_shards=num_shards,
shard_id=shard_id,
)
def add_image_input(model):
AddImageInput(
model,
reader,
batch_size=batch_per_device,
img_size=args.image_size,
dtype=args.dtype,
is_test=False,
)
def add_post_sync_ops(model):
"""Add ops applied after initial parameter sync."""
for param_info in model.GetOptimizationParamInfo(model.GetParams()):
if param_info.blob_copy is not None:
model.param_init_net.HalfToFloat(
param_info.blob,
param_info.blob_copy[core.DataType.FLOAT]
)
# Create parallelized model
data_parallel_model.Parallelize(
train_model,
input_builder_fun=add_image_input,
forward_pass_builder_fun=create_target_model_ops,
optimizer_builder_fun=add_optimizer,
post_sync_builder_fun=add_post_sync_ops,
devices=gpus,
rendezvous=rendezvous,
optimize_gradient_memory=False,
cpu_device=args.use_cpu,
shared_model=args.use_cpu,
combine_spatial_bn=args.use_cpu,
)
if args.model_parallel:
# Shift half of the activations to another GPU
assert workspace.NumCudaDevices() >= 2 * args.num_gpus
activations = data_parallel_model_utils.GetActivationBlobs(train_model)
data_parallel_model_utils.ShiftActivationDevices(
train_model,
activations=activations[len(activations) // 2:],
shifts={g: args.num_gpus + g for g in range(args.num_gpus)},
)
data_parallel_model.OptimizeGradientMemory(train_model, {}, set(), False)
workspace.RunNetOnce(train_model.param_init_net)
workspace.CreateNet(train_model.net)
# Add test model, if specified
test_model = None
if (args.test_data is not None):
log.info("----- Create test net ----")
test_arg_scope = {
'order': "NCHW",
'use_cudnn': True,
'cudnn_exhaustive_search': True,
}
test_model = model_helper.ModelHelper(
name="ban-pc-resnet50_test", arg_scope=test_arg_scope, init_params=False
)
test_reader = test_model.CreateDB(
"test_reader",
db=args.test_data,
db_type=args.db_type,
)
def test_input_fn(model):
AddImageInput(
model,
test_reader,
batch_size=batch_per_device,
img_size=args.image_size,
dtype=args.dtype,
is_test=True,
)
data_parallel_model.Parallelize(
test_model,
input_builder_fun=test_input_fn,
forward_pass_builder_fun=create_target_model_ops,
post_sync_builder_fun=add_post_sync_ops,
param_update_builder_fun=None,
devices=gpus,
cpu_device=args.use_cpu,
)
workspace.RunNetOnce(test_model.param_init_net)
workspace.CreateNet(test_model.net)
epoch = 0
# load the pre-trained model and reset epoch
if args.load_model_path is not None:
LoadModel(args.load_model_path, train_model)
# Sync the model params
data_parallel_model.FinalizeAfterCheckpoint(train_model)
# reset epoch. load_model_path should end with *_X.mdl,
# where X is the epoch number
last_str = args.load_model_path.split('_')[-1]
if last_str.endswith('.mdl'):
epoch = int(last_str[:-4])
log.info("Reset epoch to {}".format(epoch))
else:
log.warning("The format of load_model_path doesn't match!")
expname = "log/{}/resnet50_gpu{}_b{}_L{}_lr{:.2f}_v2".format(
args.dataset_name,
args.num_gpus,
total_batch_size,
args.num_labels,
args.base_learning_rate,
)
explog = experiment_util.ModelTrainerLog(expname, args)
# Load pretrained param_init_net
load_init_net_multigpu(args)
# Run the training one epoch a time
best_accuracy = 0
while epoch < args.num_epochs:
epoch, best_accuracy = RunEpoch(
args,
epoch,
train_model,
test_model,
total_batch_size,
num_shards,
expname,
explog,
best_accuracy,
)
# Save the model for each epoch
SaveModel(args, train_model, epoch)
model_path = "%s/%s_" % (
args.file_store_path,
args.save_model_name
)
# remove the saved model from the previous epoch if it exists
if os.path.isfile(model_path + str(epoch - 1) + ".mdl"):
os.remove(model_path + str(epoch - 1) + ".mdl")
def Train(args):
if args.model == "resnext":
model_name = "resnext" + str(args.num_layers)
elif args.model == "shufflenet":
model_name = "shufflenet"
# Either use specified device list or generate one
if args.gpus is not None:
gpus = [int(x) for x in args.gpus.split(',')]
num_gpus = len(gpus)
else:
gpus = list(range(args.num_gpus))
num_gpus = args.num_gpus
log.info("Running on GPUs: {}".format(gpus))
# Verify valid batch size
total_batch_size = args.batch_size
batch_per_device = total_batch_size // num_gpus
assert \
total_batch_size % num_gpus == 0, \
"Number of GPUs must divide batch size"
# Verify valid image mean/std per channel
if args.image_mean_per_channel:
assert \
len(args.image_mean_per_channel) == args.num_channels, \
"The number of channels of image mean doesn't match input"
if args.image_std_per_channel:
assert \
len(args.image_std_per_channel) == args.num_channels, \
"The number of channels of image std doesn't match input"
# Round down epoch size to closest multiple of batch size across machines
global_batch_size = total_batch_size * args.num_shards
epoch_iters = int(args.epoch_size / global_batch_size)
assert \
epoch_iters > 0, \
"Epoch size must be larger than batch size times shard count"
args.epoch_size = epoch_iters * global_batch_size
log.info("Using epoch size: {}".format(args.epoch_size))
# Create ModelHelper object
if args.use_ideep:
train_arg_scope = {
'use_cudnn': False,
'cudnn_exhaustive_search': False,
'training_mode': 1
}
else:
train_arg_scope = {
'order': 'NCHW',
'use_cudnn': True,
'cudnn_exhaustive_search': True,
'ws_nbytes_limit': (args.cudnn_workspace_limit_mb * 1024 * 1024),
}
train_model = model_helper.ModelHelper(
name=model_name, arg_scope=train_arg_scope
)
num_shards = args.num_shards
shard_id = args.shard_id
# Expect interfaces to be comma separated.
# Use of multiple network interfaces is not yet complete,
# so simply use the first one in the list.
interfaces = args.distributed_interfaces.split(",")
# Rendezvous using MPI when run with mpirun
if os.getenv("OMPI_COMM_WORLD_SIZE") is not None:
num_shards = int(os.getenv("OMPI_COMM_WORLD_SIZE", 1))
shard_id = int(os.getenv("OMPI_COMM_WORLD_RANK", 0))
if num_shards > 1:
rendezvous = dict(
kv_handler=None,
num_shards=num_shards,
shard_id=shard_id,
engine="GLOO",
transport=args.distributed_transport,
interface=interfaces[0],
mpi_rendezvous=True,
exit_nets=None)
elif num_shards > 1:
# Create rendezvous for distributed computation
store_handler = "store_handler"
if args.redis_host is not None:
# Use Redis for rendezvous if Redis host is specified
workspace.RunOperatorOnce(
core.CreateOperator(
"RedisStoreHandlerCreate", [], [store_handler],
host=args.redis_host,
port=args.redis_port,
prefix=args.run_id,
)
)
else:
# Use filesystem for rendezvous otherwise
workspace.RunOperatorOnce(
core.CreateOperator(
"FileStoreHandlerCreate", [], [store_handler],
path=args.file_store_path,
prefix=args.run_id,
)
)
rendezvous = dict(
kv_handler=store_handler,
shard_id=shard_id,
num_shards=num_shards,
engine="GLOO",
transport=args.distributed_transport,
interface=interfaces[0],
exit_nets=None)
else:
rendezvous = None
# Model building functions
def create_resnext_model_ops(model, loss_scale):
initializer = (PseudoFP16Initializer if args.dtype == 'float16'
else Initializer)
with brew.arg_scope([brew.conv, brew.fc],
WeightInitializer=initializer,
BiasInitializer=initializer,
enable_tensor_core=args.enable_tensor_core,
float16_compute=args.float16_compute):
pred = resnet.create_resnext(
model,
"data",
num_input_channels=args.num_channels,
num_labels=args.num_labels,
num_layers=args.num_layers,
num_groups=args.resnext_num_groups,
num_width_per_group=args.resnext_width_per_group,
no_bias=True,
no_loss=True,
)
if args.dtype == 'float16':
pred = model.net.HalfToFloat(pred, pred + '_fp32')
softmax, loss = model.SoftmaxWithLoss([pred, 'label'],
['softmax', 'loss'])
loss = model.Scale(loss, scale=loss_scale)
brew.accuracy(model, [softmax, "label"], "accuracy", top_k=1)
brew.accuracy(model, [softmax, "label"], "accuracy_top5", top_k=5)
return [loss]
def create_shufflenet_model_ops(model, loss_scale):
initializer = (PseudoFP16Initializer if args.dtype == 'float16'
else Initializer)
with brew.arg_scope([brew.conv, brew.fc],
WeightInitializer=initializer,
BiasInitializer=initializer,
enable_tensor_core=args.enable_tensor_core,
float16_compute=args.float16_compute):
pred = shufflenet.create_shufflenet(
model,
"data",
num_input_channels=args.num_channels,
num_labels=args.num_labels,
no_loss=True,
)
if args.dtype == 'float16':
pred = model.net.HalfToFloat(pred, pred + '_fp32')
softmax, loss = model.SoftmaxWithLoss([pred, 'label'],
['softmax', 'loss'])
loss = model.Scale(loss, scale=loss_scale)
brew.accuracy(model, [softmax, "label"], "accuracy", top_k=1)
brew.accuracy(model, [softmax, "label"], "accuracy_top5", top_k=5)
return [loss]
def add_optimizer(model):
stepsz = int(30 * args.epoch_size / total_batch_size / num_shards)
if args.float16_compute:
# TODO: merge with multi-precision optimizer
opt = optimizer.build_fp16_sgd(
model,
args.base_learning_rate,
momentum=0.9,
nesterov=1,
weight_decay=args.weight_decay, # weight decay included
policy="step",
stepsize=stepsz,
gamma=0.1
)
else:
optimizer.add_weight_decay(model, args.weight_decay)
opt = optimizer.build_multi_precision_sgd(
model,
args.base_learning_rate,
momentum=0.9,
nesterov=1,
policy="step",
stepsize=stepsz,
gamma=0.1
)
return opt
# Define add_image_input function.
# Depends on the "train_data" argument.
# Note that the reader will be shared with between all GPUS.
if args.train_data == "null":
def add_image_input(model):
AddNullInput(
model,
None,
batch_size=batch_per_device,
img_size=args.image_size,
dtype=args.dtype,
)
else:
reader = train_model.CreateDB(
"reader",
db=args.train_data,
db_type=args.db_type,
num_shards=num_shards,
shard_id=shard_id,
)
def add_image_input(model):
AddImageInput(
model,
reader,
batch_size=batch_per_device,
img_size=args.image_size,
dtype=args.dtype,
is_test=False,
mean_per_channel=args.image_mean_per_channel,
std_per_channel=args.image_std_per_channel,
)
def add_post_sync_ops(model):
"""Add ops applied after initial parameter sync."""
for param_info in model.GetOptimizationParamInfo(model.GetParams()):
if param_info.blob_copy is not None:
model.param_init_net.HalfToFloat(
param_info.blob,
param_info.blob_copy[core.DataType.FLOAT]
)
data_parallel_model.Parallelize(
train_model,
input_builder_fun=add_image_input,
forward_pass_builder_fun=create_resnext_model_ops
if args.model == "resnext" else create_shufflenet_model_ops,
optimizer_builder_fun=add_optimizer,
post_sync_builder_fun=add_post_sync_ops,
devices=gpus,
rendezvous=rendezvous,
optimize_gradient_memory=False,
use_nccl=args.use_nccl,
cpu_device=args.use_cpu,
ideep=args.use_ideep,
shared_model=args.use_cpu,
combine_spatial_bn=args.use_cpu,
)
data_parallel_model.OptimizeGradientMemory(train_model, {}, set(), False)
workspace.RunNetOnce(train_model.param_init_net)
workspace.CreateNet(train_model.net)
# Add test model, if specified
test_model = None
if (args.test_data is not None):
log.info("----- Create test net ----")
if args.use_ideep:
test_arg_scope = {
'use_cudnn': False,
'cudnn_exhaustive_search': False,
}
else:
test_arg_scope = {
'order': "NCHW",
'use_cudnn': True,
'cudnn_exhaustive_search': True,
}
test_model = model_helper.ModelHelper(
name=model_name + "_test",
arg_scope=test_arg_scope,
init_params=False,
)
test_reader = test_model.CreateDB(
"test_reader",
db=args.test_data,
db_type=args.db_type,
)
def test_input_fn(model):
AddImageInput(
model,
test_reader,
batch_size=batch_per_device,
img_size=args.image_size,
dtype=args.dtype,
is_test=True,
mean_per_channel=args.image_mean_per_channel,
std_per_channel=args.image_std_per_channel,
)
data_parallel_model.Parallelize(
test_model,
input_builder_fun=test_input_fn,
forward_pass_builder_fun=create_resnext_model_ops
if args.model == "resnext" else create_shufflenet_model_ops,
post_sync_builder_fun=add_post_sync_ops,
param_update_builder_fun=None,
devices=gpus,
use_nccl=args.use_nccl,
cpu_device=args.use_cpu,
)
workspace.RunNetOnce(test_model.param_init_net)
workspace.CreateNet(test_model.net)
epoch = 0
# load the pre-trained model and reset epoch
if args.load_model_path is not None:
LoadModel(args.load_model_path, train_model, args.use_ideep)
# Sync the model params
data_parallel_model.FinalizeAfterCheckpoint(train_model)
# reset epoch. load_model_path should end with *_X.mdl,
# where X is the epoch number
last_str = args.load_model_path.split('_')[-1]
if last_str.endswith('.mdl'):
epoch = int(last_str[:-4])
log.info("Reset epoch to {}".format(epoch))
else:
log.warning("The format of load_model_path doesn't match!")
expname = "%s_gpu%d_b%d_L%d_lr%.2f_v2" % (
model_name,
args.num_gpus,
total_batch_size,
args.num_labels,
args.base_learning_rate,
)
explog = experiment_util.ModelTrainerLog(expname, args)
# Run the training one epoch a time
while epoch < args.num_epochs:
epoch = RunEpoch(
args,
epoch,
train_model,
test_model,
total_batch_size,
num_shards,
expname,
explog
)
# Save the model for each epoch
SaveModel(args, train_model, epoch, args.use_ideep)
model_path = "%s/%s_" % (
args.file_store_path,
args.save_model_name
)
# remove the saved model from the previous epoch if it exists
if os.path.isfile(model_path + str(epoch - 1) + ".mdl"):
os.remove(model_path + str(epoch - 1) + ".mdl")
def Train(args):
if args.gpus is not None:
gpus = [int(x) for x in args.gpus.split(',')]
num_gpus = len(gpus)
else:
gpus = range(args.num_gpus)
num_gpus = args.num_gpus
log.info("Running on GPUs: {}".format(gpus))
# Modify to make it consistent with the distributed trainer
total_batch_size = args.batch_size * num_gpus
batch_per_device = args.batch_size
# Round down epoch size to closest multiple of batch size across machines
epoch_iters = int(args.epoch_size / total_batch_size)
args.epoch_size = epoch_iters * total_batch_size
log.info("Using epoch size: {}".format(args.epoch_size))
# Create CNNModeLhelper object
train_model = cnn.CNNModelHelper(
order="NCHW",
name='{}_train'.format(args.model_name),
use_cudnn=(True if args.use_cudnn == 1 else False),
cudnn_exhaustive_search=True,
ws_nbytes_limit=(args.cudnn_workspace_limit_mb * 1024 * 1024),
)
# Model building functions
def create_model_ops(model, loss_scale):
return model_builder.build_model(
model=model,
model_name=args.model_name,
model_depth=args.model_depth,
num_labels=args.num_labels,
batch_size=args.batch_size,
num_channels=args.num_channels,
crop_size=args.crop_size,
clip_length=(args.clip_length_of
if args.input_type else args.clip_length_rgb),
loss_scale=loss_scale,
pred_layer_name=args.pred_layer_name,
multi_label=args.multi_label,
channel_multiplier=args.channel_multiplier,
bottleneck_multiplier=args.bottleneck_multiplier,
use_dropout=args.use_dropout,
conv1_temporal_stride=args.conv1_temporal_stride,
conv1_temporal_kernel=args.conv1_temporal_kernel,
use_pool1=args.use_pool1,
audio_input_3d=args.audio_input_3d,
g_blend=args.g_blend,
audio_weight=args.audio_weight,
visual_weight=args.visual_weight,
av_weight=args.av_weight,
)
# SGD
def add_parameter_update_ops(model):
model.AddWeightDecay(args.weight_decay)
ITER = model.Iter("ITER")
stepsz = args.step_epoch * args.epoch_size / args.batch_size / num_gpus
LR = model.net.LearningRate(
[ITER],
"LR",
base_lr=args.base_learning_rate * num_gpus,
policy="step",
stepsize=int(stepsz),
gamma=args.gamma,
)
AddMomentumParameterUpdate(model, LR)
# Input. Note that the reader must be shared with all GPUS.
train_reader, train_examples = reader_utils.create_data_reader(
train_model,
name="train_reader",
input_data=args.train_data,
)
log.info("Training set has {} examples".format(train_examples))
def add_video_input(model):
model_helper.AddVideoInput(
model,
train_reader,
batch_size=batch_per_device,
length_rgb=args.clip_length_rgb,
clip_per_video=1,
random_mirror=True,
decode_type=0,
sampling_rate_rgb=args.sampling_rate_rgb,
scale_h=args.scale_h,
scale_w=args.scale_w,
crop_size=args.crop_size,
video_res_type=args.video_res_type,
short_edge=min(args.scale_h, args.scale_w),
num_decode_threads=args.num_decode_threads,
do_multi_label=args.multi_label,
num_of_class=args.num_labels,
random_crop=True,
input_type=args.input_type,
length_of=args.clip_length_of,
sampling_rate_of=args.sampling_rate_of,
frame_gap_of=args.frame_gap_of,
do_flow_aggregation=args.do_flow_aggregation,
flow_data_type=args.flow_data_type,
get_rgb=(args.input_type == 0 or args.input_type >= 3),
get_optical_flow=(args.input_type == 1 or args.input_type >= 4),
get_logmels=(args.input_type >= 2),
get_video_id=args.get_video_id,
jitter_scales=[int(n) for n in args.jitter_scales.split(',')],
use_local_file=args.use_local_file,
)
# Create parallelized model
data_parallel_model.Parallelize_GPU(
train_model,
input_builder_fun=add_video_input,
forward_pass_builder_fun=create_model_ops,
param_update_builder_fun=add_parameter_update_ops,
devices=gpus,
rendezvous=None,
net_type=('prof_dag' if args.profiling == 1 else 'dag'),
optimize_gradient_memory=True,
)
# Add test model, if specified
test_model = None
if args.test_data is not None:
log.info("----- Create test net ----")
test_model = cnn.CNNModelHelper(
order="NCHW",
name='{}_test'.format(args.model_name),
use_cudnn=(True if args.use_cudnn == 1 else False),
cudnn_exhaustive_search=True)
test_reader, test_examples = reader_utils.create_data_reader(
test_model,
name="test_reader",
input_data=args.test_data,
)
log.info("Testing set has {} examples".format(test_examples))
def test_input_fn(model):
model_helper.AddVideoInput(
model,
test_reader,
batch_size=batch_per_device,
length_rgb=args.clip_length_rgb,
clip_per_video=1,
decode_type=0,
random_mirror=False,
random_crop=False,
sampling_rate_rgb=args.sampling_rate_rgb,
scale_h=args.scale_h,
scale_w=args.scale_w,
crop_size=args.crop_size,
video_res_type=args.video_res_type,
short_edge=min(args.scale_h, args.scale_w),
num_decode_threads=args.num_decode_threads,
do_multi_label=args.multi_label,
num_of_class=args.num_labels,
input_type=args.input_type,
length_of=args.clip_length_of,
sampling_rate_of=args.sampling_rate_of,
frame_gap_of=args.frame_gap_of,
do_flow_aggregation=args.do_flow_aggregation,
flow_data_type=args.flow_data_type,
get_rgb=(args.input_type == 0),
get_optical_flow=(args.input_type == 1),
get_video_id=args.get_video_id,
use_local_file=args.use_local_file,
)
data_parallel_model.Parallelize_GPU(
test_model,
input_builder_fun=test_input_fn,
forward_pass_builder_fun=create_model_ops,
param_update_builder_fun=None,
devices=gpus,
optimize_gradient_memory=True,
)
workspace.RunNetOnce(test_model.param_init_net)
workspace.CreateNet(test_model.net)
workspace.RunNetOnce(train_model.param_init_net)
workspace.CreateNet(train_model.net)
epoch = 0
# load the pre-trained model and reset epoch
if args.load_model_path is not None:
if args.db_type == 'pickle':
model_loader.LoadModelFromPickleFile(train_model,
args.load_model_path,
use_gpu=True,
root_gpu_id=gpus[0])
else:
model_helper.LoadModel(args.load_model_path, args.db_type)
# Sync the model params
data_parallel_model.FinalizeAfterCheckpoint(
train_model,
GetCheckpointParams(train_model),
)
if args.is_checkpoint:
# reset epoch. load_model_path should end with *_X.mdl,
# where X is the epoch number
last_str = args.load_model_path.split('_')[-1]
if last_str.endswith('.mdl'):
epoch = int(last_str[:-4])
log.info("Reset epoch to {}".format(epoch))
else:
log.warning("The format of load_model_path doesn't match!")
expname = "%s_gpu%d_b%d_L%d_lr%.2f" % (
args.model_name,
args.num_gpus,
total_batch_size,
args.num_labels,
args.base_learning_rate,
)
explog = experiment_util.ModelTrainerLog(expname, args)
# Run the training one epoch a time
while epoch < args.num_epochs:
epoch = RunEpoch(args, epoch, train_model, test_model,
total_batch_size, 1, expname, explog)
# Save the model for each epoch
SaveModel(args, train_model, epoch)
def Train(args):
# Either use specified device list or generate one
if args.gpus is not None:
gpus = [int(x) for x in args.gpus.split(',')]
num_gpus = len(gpus)
else:
gpus = range(args.num_gpus)
num_gpus = args.num_gpus
log.info("Running on GPUs: {}".format(gpus))
# Verify valid batch size
total_batch_size = args.batch_size
batch_per_device = total_batch_size // num_gpus
assert \
total_batch_size % num_gpus == 0, \
"Number of GPUs must divide batch size"
# Round down epoch size to closest multiple of batch size across machines
global_batch_size = total_batch_size * args.num_shards
epoch_iters = int(args.epoch_size / global_batch_size)
args.epoch_size = epoch_iters * global_batch_size
log.info("Using epoch size: {}".format(args.epoch_size))
# Create CNNModeLhelper object
train_model = cnn.CNNModelHelper(
order="NCHW",
name="resnet50",
use_cudnn=True,
cudnn_exhaustive_search=True,
ws_nbytes_limit=(args.cudnn_workspace_limit_mb * 1024 * 1024),
)
num_shards = args.num_shards
shard_id = args.shard_id
if num_shards > 1:
# Create rendezvous for distributed computation
store_handler = "store_handler"
if args.redis_host is not None:
# Use Redis for rendezvous if Redis host is specified
workspace.RunOperatorOnce(
core.CreateOperator(
"RedisStoreHandlerCreate",
[],
[store_handler],
host=args.redis_host,
port=args.redis_port,
prefix=args.run_id,
))
else:
# Use filesystem for rendezvous otherwise
workspace.RunOperatorOnce(
core.CreateOperator(
"FileStoreHandlerCreate",
[],
[store_handler],
path=args.file_store_path,
))
rendezvous = dict(kv_handler=store_handler,
shard_id=shard_id,
num_shards=num_shards,
engine="GLOO",
exit_nets=None)
else:
rendezvous = None
# Model building functions
def create_resnet50_model_ops(model, loss_scale):
[softmax, loss] = resnet.create_resnet50(
model,
"data",
num_input_channels=args.num_channels,
num_labels=args.num_labels,
label="label",
no_bias=True,
)
loss = model.Scale(loss, scale=loss_scale)
model.Accuracy([softmax, "label"], "accuracy")
return [loss]
# SGD
def add_parameter_update_ops(model):
model.AddWeightDecay(args.weight_decay)
ITER = model.Iter("ITER")
stepsz = int(30 * args.epoch_size / total_batch_size / num_shards)
LR = model.net.LearningRate(
[ITER],
"LR",
base_lr=args.base_learning_rate,
policy="step",
stepsize=stepsz,
gamma=0.1,
)
AddMomentumParameterUpdate(model, LR)
# Input. Note that the reader must be shared with all GPUS.
reader = train_model.CreateDB(
"reader",
db=args.train_data,
db_type=args.db_type,
num_shards=num_shards,
shard_id=shard_id,
)
def add_image_input(model):
AddImageInput(
model,
reader,
batch_size=batch_per_device,
img_size=args.image_size,
)
# Create parallelized model
data_parallel_model.Parallelize_GPU(
train_model,
input_builder_fun=add_image_input,
forward_pass_builder_fun=create_resnet50_model_ops,
param_update_builder_fun=add_parameter_update_ops,
devices=gpus,
rendezvous=rendezvous,
optimize_gradient_memory=True,
)
# Add test model, if specified
test_model = None
if (args.test_data is not None):
log.info("----- Create test net ----")
test_model = cnn.CNNModelHelper(order="NCHW",
name="resnet50_test",
use_cudnn=True,
cudnn_exhaustive_search=True)
test_reader = test_model.CreateDB(
"test_reader",
db=args.test_data,
db_type=args.db_type,
)
def test_input_fn(model):
AddImageInput(
model,
test_reader,
batch_size=batch_per_device,
img_size=args.image_size,
)
data_parallel_model.Parallelize_GPU(
test_model,
input_builder_fun=test_input_fn,
forward_pass_builder_fun=create_resnet50_model_ops,
param_update_builder_fun=None,
devices=gpus,
)
workspace.RunNetOnce(test_model.param_init_net)
workspace.CreateNet(test_model.net)
workspace.RunNetOnce(train_model.param_init_net)
workspace.CreateNet(train_model.net)
epoch = 0
# load the pre-trained model and reset epoch
if args.load_model_path is not None:
LoadModel(args.load_model_path, train_model)
# Sync the model params
data_parallel_model.FinalizeAfterCheckpoint(
train_model,
GetCheckpointParams(train_model),
)
# reset epoch. load_model_path should end with *_X.mdl,
# where X is the epoch number
last_str = args.load_model_path.split('_')[-1]
if last_str.endswith('.mdl'):
epoch = int(last_str[:-4])
log.info("Reset epoch to {}".format(epoch))
else:
log.warning("The format of load_model_path doesn't match!")
expname = "resnet50_gpu%d_b%d_L%d_lr%.2f_v2" % (
args.num_gpus,
total_batch_size,
args.num_labels,
args.base_learning_rate,
)
explog = experiment_util.ModelTrainerLog(expname, args)
# Run the training one epoch a time
while epoch < args.num_epochs:
epoch = RunEpoch(args, epoch, train_model, test_model,
total_batch_size, num_shards, expname, explog)
# Save the model for each epoch
SaveModel(args, train_model, epoch)
model_path = "%s/%s_" % (args.file_store_path, args.save_model_name)
# remove the saved model from the previous epoch if it exists
if os.path.isfile(model_path + str(epoch - 1) + ".mdl"):
os.remove(model_path + str(epoch - 1) + ".mdl")
def Train(args):
# Either use specified device list or generate one
if args.gpus is not None:
gpus = [int(x) for x in args.gpus.split(',')]
num_gpus = len(gpus)
else:
gpus = list(range(args.num_gpus))
num_gpus = args.num_gpus
log.info("Running on GPUs: {}".format(gpus))
# Verify valid batch size
total_batch_size = args.batch_size
batch_per_device = total_batch_size // num_gpus
global_batch_size = total_batch_size * args.num_shards
epoch_iters = int(args.epoch_size / global_batch_size)
args.epoch_size = epoch_iters * global_batch_size
log.info("Using epoch size: {}".format(args.epoch_size))
train_arg_scope = {
'order': 'NCHW',
'use_cudnn': True,
'cudnn_exhaustive_search': True,
'ws_nbytes_limit': (args.cudnn_workspace_limit_mb * 1024 * 1024),
}
train_model = model_helper.ModelHelper(name="resnet101",
arg_scope=train_arg_scope)
num_shards = args.num_shards
shard_id = args.shard_id
interfaces = args.distributed_interfaces.split(",")
if os.getenv("OMPI_COMM_WORLD_SIZE") is not None:
num_shards = int(os.getenv("OMPI_COMM_WORLD_SIZE", 1))
shard_id = int(os.getenv("OMPI_COMM_WORLD_RANK", 0))
if num_shards > 1:
rendezvous = dict(kv_handler=None,
num_shards=num_shards,
shard_id=shard_id,
engine="GLOO",
transport=args.distributed_transport,
interface=interfaces[0],
mpi_rendezvous=True,
exit_nets=None)
elif num_shards > 1:
store_handler = "store_handler"
if args.redis_host is not None:
workspace.RunOperatorOnce(
core.CreateOperator(
"RedisStoreHandlerCreate",
[],
[store_handler],
host=args.redis_host,
port=args.redis_port,
prefix=args.run_id,
))
else:
workspace.RunOperatorOnce(
core.CreateOperator(
"FileStoreHandlerCreate",
[],
[store_handler],
path=args.file_store_path,
prefix=args.run_id,
))
rendezvous = dict(kv_handler=store_handler,
shard_id=shard_id,
num_shards=num_shards,
engine="GLOO",
transport=args.distributed_transport,
interface=interfaces[0],
exit_nets=None)
else:
rendezvous = None
def create_resnet101_model_ops(model, loss_scale):
initializer = (pFP16Initializer
if args.dtype == 'float16' else Initializer)
with brew.arg_scope([brew.conv, brew.fc],
WeightInitializer=initializer,
BiasInitializer=initializer,
enable_tensor_core=args.enable_tensor_core,
float16_compute=args.float16_compute):
pred = resnet.create_resnet101(
model,
"data",
num_input_channels=args.num_channels,
num_labels=args.num_labels,
no_bias=True,
no_loss=True,
)
if args.dtype == 'float16':
pred = model.net.HalfToFloat(pred, pred + '_fp32')
softmax, loss = model.SoftmaxWithLoss([pred, 'label'],
['softmax', 'loss'])
loss = model.Scale(loss, scale=loss_scale)
brew.accuracy(model, [softmax, "label"], "accuracy")
return [loss]
def add_optimizer(model):
stepsz = int(30 * args.epoch_size / total_batch_size / num_shards)
if args.float16_compute:
opt = optimizer.build_fp16_sgd(model,
args.base_learning_rate,
momentum=0.9,
nesterov=1,
weight_decay=args.weight_decay,
policy="step",
stepsize=stepsz,
gamma=0.1)
else:
optimizer.add_weight_decay(model, args.weight_decay)
opt = optimizer.build_multi_precision_sgd(model,
args.base_learning_rate,
momentum=0.9,
nesterov=1,
policy="step",
stepsize=stepsz,
gamma=0.1)
return opt
if args.train_data == "null":
def add_image_input(model):
AddNullInput(
model,
None,
batch_size=batch_per_device,
img_size=args.image_size,
dtype=args.dtype,
)
else:
reader = train_model.CreateDB(
"reader",
db=args.train_data,
db_type=args.db_type,
num_shards=num_shards,
shard_id=shard_id,
)
def add_image_input(model):
AddImageInput(
model,
reader,
batch_size=batch_per_device,
img_size=args.image_size,
dtype=args.dtype,
is_test=False,
)
def add_post_sync_ops(model):
for param_info in model.GetOptimizationParamInfo(model.GetParams()):
if param_info.blob_copy is not None:
model.param_init_net.HalfToFloat(
param_info.blob, param_info.blob_copy[core.DataType.FLOAT])
data_parallel_model.Parallelize(
train_model,
input_builder_fun=add_image_input,
forward_pass_builder_fun=create_resnet101_model_ops,
optimizer_builder_fun=add_optimizer,
post_sync_builder_fun=add_post_sync_ops,
devices=gpus,
rendezvous=rendezvous,
optimize_gradient_memory=False,
cpu_device=args.use_cpu,
shared_model=args.use_cpu,
)
if args.model_parallel:
activations = data_parallel_model_utils.GetActivationBlobs(train_model)
data_parallel_model_utils.ShiftActivationDevices(
train_model,
activations=activations[len(activations) // 2:],
shifts={g: args.num_gpus + g
for g in range(args.num_gpus)},
)
data_parallel_model.OptimizeGradientMemory(train_model, {}, set(), False)
workspace.RunNetOnce(train_model.param_init_net)
workspace.CreateNet(train_model.net)
test_model = None
if (args.test_data is not None):
log.info("----- Create test net ----")
test_arg_scope = {
'order': "NCHW",
'use_cudnn': True,
'cudnn_exhaustive_search': True,
}
test_model = model_helper.ModelHelper(name="resnet101_test",
arg_scope=test_arg_scope,
init_params=False)
test_reader = test_model.CreateDB(
"test_reader",
db=args.test_data,
db_type=args.db_type,
)
def test_input_fn(model):
AddImageInput(
model,
test_reader,
batch_size=batch_per_device,
img_size=args.image_size,
dtype=args.dtype,
is_test=True,
)
data_parallel_model.Parallelize(
test_model,
input_builder_fun=test_input_fn,
forward_pass_builder_fun=create_resnet101_model_ops,
post_sync_builder_fun=add_post_sync_ops,
param_update_builder_fun=None,
devices=gpus,
cpu_device=args.use_cpu,
)
workspace.RunNetOnce(test_model.param_init_net)
workspace.CreateNet(test_model.net)
epoch = 0
if args.load_model_path is not None:
LoadModel(args.load_model_path, train_model)
data_parallel_model.FinalizeAfterCheckpoint(train_model)
last_str = args.load_model_path.split('_')[-1]
if last_str.endswith('.mdl'):
epoch = int(last_str[:-4])
log.info("Reset epoch to {}".format(epoch))
else:
log.warning("The format of load_model_path doesn't match!")
expname = "resnet101_gpu%d_b%d_L%d_lr%.2f_v2" % (
args.num_gpus,
total_batch_size,
args.num_labels,
args.base_learning_rate,
)
explog = experiment_util.ModelTrainerLog(expname, args)
while epoch < args.num_epochs:
epoch = RunEpoch(args, epoch, train_model, test_model,
total_batch_size, num_shards, expname, explog)
# final save
SaveModel(workspace, train_model)
def extract_features(args):
if args.gpus is not None:
gpus = [int(x) for x in args.gpus.split(',')]
num_gpus = len(gpus)
else:
gpus = range(args.num_gpus)
num_gpus = args.num_gpus
if num_gpus > 0:
log.info("Running on GPUs: {}".format(gpus))
else:
log.info("Running on CPU")
my_arg_scope = {
'order': 'NCHW',
'use_cudnn': True,
'cudnn_exhaustive_search': True
}
model = cnn.CNNModelHelper(
name="Extract Features",
**my_arg_scope
)
video_input_args = dict(
batch_size=args.batch_size,
clip_per_video=args.clip_per_video,
decode_type=args.decode_type,
length_rgb=args.clip_length_rgb,
sampling_rate_rgb=args.sampling_rate_rgb,
scale_h=args.scale_h,
scale_w=args.scale_w,
crop_size=args.crop_size,
video_res_type=args.video_res_type,
short_edge=min(args.scale_h, args.scale_w),
num_decode_threads=args.num_decode_threads,
do_multi_label=args.multi_label,
num_of_class=args.num_labels,
random_mirror=False,
random_crop=False,
input_type=args.input_type,
length_of=args.clip_length_of,
sampling_rate_of=args.sampling_rate_of,
frame_gap_of=args.frame_gap_of,
do_flow_aggregation=args.do_flow_aggregation,
flow_data_type=args.flow_data_type,
get_rgb=args.input_type == 0,
get_optical_flow=args.input_type == 1,
get_video_id=args.get_video_id,
get_start_frame=args.get_start_frame,
use_local_file=args.use_local_file,
crop_per_clip=args.crop_per_clip,
)
reader_args = dict(
name="extract_features" + '_reader',
input_data=os.path.join(vmz_data.VMZ_DIR, args.test_data),
)
reader, num_examples = reader_utils.create_data_reader(
model,
**reader_args
)
def input_fn(model):
model_helper.AddVideoInput(
model,
reader,
**video_input_args)
def create_model_ops(model, loss_scale):
return model_builder.build_model(
model=model,
model_name=args.model_name,
model_depth=args.model_depth,
num_labels=args.num_labels,
batch_size=args.batch_size,
num_channels=args.num_channels,
crop_size=args.crop_size,
clip_length=(
args.clip_length_of if args.input_type == 1
else args.clip_length_rgb
),
loss_scale=loss_scale,
is_test=1,
multi_label=args.multi_label,
channel_multiplier=args.channel_multiplier,
bottleneck_multiplier=args.bottleneck_multiplier,
use_dropout=args.use_dropout,
use_convolutional_pred=args.use_convolutional_pred,
use_pool1=args.use_pool1,
)
if num_gpus > 0:
data_parallel_model.Parallelize_GPU(
model,
input_builder_fun=input_fn,
forward_pass_builder_fun=create_model_ops,
param_update_builder_fun=None, # 'None' since we aren't training
devices=gpus,
optimize_gradient_memory=True,
)
else:
model._device_type = caffe2_pb2.CPU
model._devices = [0]
device_opt = core.DeviceOption(model._device_type, 0)
with core.DeviceScope(device_opt):
# Because our loaded models are named with "gpu_x", keep the naming for now.
# TODO: Save model using `data_parallel_model.ExtractPredictorNet`
# to extract the model for "gpu_0". It also renames
# the input and output blobs by stripping the "gpu_x/" prefix
with core.NameScope("{}_{}".format("gpu", 0)):
input_fn(model)
create_model_ops(model, 1.0)
workspace.RunNetOnce(model.param_init_net)
workspace.CreateNet(model.net)
if args.db_type == 'pickle':
load_model_path = os.path.join(vmz_data.PRETRAINED_MODEL_DIR, args.load_model_path)
model_loader.LoadModelFromPickleFile(model, load_model_path)
elif args.db_type == 'minidb':
if num_gpus > 0:
model_helper.LoadModel(args.load_model_path, args.db_type)
else:
with core.DeviceScope(core.DeviceOption(caffe2_pb2.CPU, 0)):
model_helper.LoadModel(args.load_model_path, args.db_type)
else:
log.warning("Unsupported db_type: {}".format(args.db_type))
data_parallel_model.FinalizeAfterCheckpoint(model)
def fetchActivations(model, outputs, num_iterations):
all_activations = {}
for counter in range(num_iterations):
workspace.RunNet(model.net.Proto().name)
num_devices = 1 # default for cpu
if num_gpus > 0:
num_devices = num_gpus
for g in range(num_devices):
for output_name in outputs:
blob_name = 'gpu_{}/'.format(g) + output_name
activations = workspace.FetchBlob(blob_name)
if output_name not in all_activations:
all_activations[output_name] = []
all_activations[output_name].append(activations)
if counter % 20 == 0:
log.info('{}/{} iterations'.format(counter, num_iterations))
# each key holds a list of activations obtained from each minibatch.
# we now concatenate these lists to get the final arrays.
# concatenating during the loop requires a realloc and can get slow.
for key in all_activations:
all_activations[key] = np.concatenate(all_activations[key])
return all_activations
outputs = [name.strip() for name in args.features.split(',')]
assert len(outputs) > 0
if args.num_iterations > 0:
num_iterations = args.num_iterations
else:
if num_gpus > 0:
examples_per_iteration = args.batch_size * num_gpus
else:
examples_per_iteration = args.batch_size
num_iterations = int(num_examples / examples_per_iteration)
activations = fetchActivations(model, outputs, num_iterations)
# saving extracted features
for index in range(len(outputs)):
log.info(
"Read '{}' with shape {}".format(
outputs[index],
activations[outputs[index]].shape
)
)
output_path = os.path.join(vmz_data.VMZ_FEATURE_DIR, args.output_path)
if not os.path.exists(os.path.dirname(output_path)):
os.makedirs(os.path.dirname(output_path))
log.info('Writing to {}'.format(output_path))
if args.save_h5:
with h5py.File(output_path, 'w') as handle:
for name, activation in activations.items():
handle.create_dataset(name, data=activation)
else:
with open(output_path, 'wb') as handle:
pickle.dump(activations, handle)
# perform sanity check
if args.sanity_check == 1: # check clip accuracy
assert args.multi_label == 0
clip_acc = 0
softmax = activations['softmax']
label = activations['label']
for i in range(len(softmax)):
sorted_preds = \
np.argsort(softmax[i])
sorted_preds[:] = sorted_preds[::-1]
if sorted_preds[0] == label[i]:
clip_acc += 1
log.info('Sanity check --- clip accuracy: {}'.format(
clip_acc / len(softmax))
)
elif args.sanity_check == 2: # check auc
assert args.multi_label == 1
prob = activations['prob']
label = activations['label']
mean_auc, mean_ap, mean_wap, _ = metric.mean_ap_metric(prob, label)
log.info('Sanity check --- AUC: {}, mAP: {}, mWAP: {}'.format(
mean_auc, mean_ap, mean_wap)
)
def Train(args):
# Either use specified device list or generate one
if args.gpus is not None:
gpus = [int(x) for x in args.gpus.split(',')]
num_gpus = len(gpus)
else:
gpus = list(range(args.num_gpus))
num_gpus = args.num_gpus
log.info("Running on GPUs: {}".format(gpus))
# Verify valid batch size
total_batch_size = args.batch_size
batch_per_device = total_batch_size // num_gpus
assert \
total_batch_size % num_gpus == 0, \
"Number of GPUs must divide batch size"
# Round down epoch size to closest multiple of batch size across machines
global_batch_size = total_batch_size * args.num_shards
epoch_iters = int(args.epoch_size / global_batch_size)
assert \
epoch_iters > 0, \
"Epoch size must be larger than batch size times shard count"
args.epoch_size = epoch_iters * global_batch_size
log.info("Using epoch size: {}".format(args.epoch_size))
# Create ModelHelper object
train_arg_scope = {
'order': 'NCHW',
'use_cudnn': True,
'cudnn_exhaustive_search': True,
'ws_nbytes_limit': (args.cudnn_workspace_limit_mb * 1024 * 1024),
}
train_model = model_helper.ModelHelper(name="resnet50",
arg_scope=train_arg_scope)
num_shards = args.num_shards
shard_id = args.shard_id
# Expect interfaces to be comma separated.
# Use of multiple network interfaces is not yet complete,
# so simply use the first one in the list.
interfaces = args.distributed_interfaces.split(",")
# Rendezvous using MPI when run with mpirun
if os.getenv("OMPI_COMM_WORLD_SIZE") is not None:
num_shards = int(os.getenv("OMPI_COMM_WORLD_SIZE", 1))
shard_id = int(os.getenv("OMPI_COMM_WORLD_RANK", 0))
if num_shards > 1:
rendezvous = dict(kv_handler=None,
num_shards=num_shards,
shard_id=shard_id,
engine="GLOO",
transport=args.distributed_transport,
interface=interfaces[0],
mpi_rendezvous=True,
exit_nets=None)
elif num_shards > 1:
# Create rendezvous for distributed computation
store_handler = "store_handler"
if args.redis_host is not None:
# Use Redis for rendezvous if Redis host is specified
workspace.RunOperatorOnce(
core.CreateOperator(
"RedisStoreHandlerCreate",
[],
[store_handler],
host=args.redis_host,
port=args.redis_port,
prefix=args.run_id,
))
else:
# Use filesystem for rendezvous otherwise
workspace.RunOperatorOnce(
core.CreateOperator(
"FileStoreHandlerCreate",
[],
[store_handler],
path=args.file_store_path,
prefix=args.run_id,
))
rendezvous = dict(kv_handler=store_handler,
shard_id=shard_id,
num_shards=num_shards,
engine="GLOO",
transport=args.distributed_transport,
interface=interfaces[0],
exit_nets=None)
else:
rendezvous = None
# Model building functions
# def create_resnet50_model_ops(model, loss_scale):
# initializer = (PseudoFP16Initializer if args.dtype == 'float16'
# else Initializer)
# with brew.arg_scope([brew.conv, brew.fc],
# WeightInitializer=initializer,
# BiasInitializer=initializer,
# enable_tensor_core=args.enable_tensor_core,
# float16_compute=args.float16_compute):
# pred = resnet.create_resnet50(
# #args.layers,
# model,
# "data",
# num_input_channels=args.num_channels,
# num_labels=args.num_labels,
# no_bias=True,
# no_loss=True,
# )
# if args.dtype == 'float16':
# pred = model.net.HalfToFloat(pred, pred + '_fp32')
# softmax, loss = model.SoftmaxWithLoss([pred, 'label'],
# ['softmax', 'loss'])
# loss = model.Scale(loss, scale=loss_scale)
# brew.accuracy(model, [softmax, "label"], "accuracy")
# return [loss]
def create_model_ops(model, loss_scale):
return create_model_ops_testable(model, loss_scale, is_test=False)
def create_model_ops_test(model, loss_scale):
return create_model_ops_testable(model, loss_scale, is_test=True)
# Model building functions
def create_model_ops_testable(model, loss_scale, is_test=False):
initializer = (PseudoFP16Initializer
if args.dtype == 'float16' else Initializer)
with brew.arg_scope([brew.conv, brew.fc],
WeightInitializer=initializer,
BiasInitializer=initializer,
enable_tensor_core=args.enable_tensor_core,
float16_compute=args.float16_compute):
if args.model == "cifar10":
if args.image_size != 32:
log.warn("Cifar10 expects a 32x32 image.")
pred = models.cifar10.create_cifar10(
model,
"data",
image_channels=args.num_channels,
num_classes=args.num_labels,
image_height=args.image_size,
image_width=args.image_size,
)
elif args.model == "resnet32x32":
if args.image_size != 32:
log.warn("ResNet32x32 expects a 32x32 image.")
pred = models.resnet.create_resnet32x32(
model,
"data",
num_layers=args.num_layers,
num_input_channels=args.num_channels,
num_labels=args.num_labels,
is_test=is_test)
elif args.model == "resnet":
if args.image_size != 224:
log.warn(
"ResNet expects a 224x224 image. input image = %d" %
args.image_size)
pred = resnet.create_resnet50(
#args.layers,
model,
"data",
num_input_channels=args.num_channels,
num_labels=args.num_labels,
no_bias=True,
no_loss=True,
)
elif args.model == "vgg":
if args.image_size != 224:
log.warn("VGG expects a 224x224 image.")
pred = vgg.create_vgg(model,
"data",
num_input_channels=args.num_channels,
num_labels=args.num_labels,
num_layers=args.num_layers,
is_test=is_test)
elif args.model == "googlenet":
if args.image_size != 224:
log.warn("GoogLeNet expects a 224x224 image.")
pred = googlenet.create_googlenet(
model,
"data",
num_input_channels=args.num_channels,
num_labels=args.num_labels,
is_test=is_test)
elif args.model == "alexnet":
if args.image_size != 224:
log.warn("Alexnet expects a 224x224 image.")
pred = alexnet.create_alexnet(
model,
"data",
num_input_channels=args.num_channels,
num_labels=args.num_labels,
is_test=is_test)
elif args.model == "alexnetv0":
if args.image_size != 224:
log.warn("Alexnet v0 expects a 224x224 image.")
pred = alexnet.create_alexnetv0(
model,
"data",
num_input_channels=args.num_channels,
num_labels=args.num_labels,
is_test=is_test)
else:
raise NotImplementedError("Network {} not found.".format(
args.model))
if args.dtype == 'float16':
pred = model.net.HalfToFloat(pred, pred + '_fp32')
softmax, loss = model.SoftmaxWithLoss([pred, 'label'],
['softmax', 'loss'])
loss = model.Scale(loss, scale=loss_scale)
brew.accuracy(model, [softmax, "label"], "accuracy")
return [loss]
def add_optimizer(model):
stepsz = int(30 * args.epoch_size / total_batch_size / num_shards)
if args.float16_compute:
# TODO: merge with multi-prceision optimizer
opt = optimizer.build_fp16_sgd(
model,
args.base_learning_rate,
momentum=0.9,
nesterov=1,
weight_decay=args.weight_decay, # weight decay included
policy="step",
stepsize=stepsz,
gamma=0.1)
else:
optimizer.add_weight_decay(model, args.weight_decay)
opt = optimizer.build_multi_precision_sgd(model,
args.base_learning_rate,
momentum=0.9,
nesterov=1,
policy="step",
stepsize=stepsz,
gamma=0.1)
print("info:===============================" + str(opt))
return opt
# Define add_image_input function.
# Depends on the "train_data" argument.
# Note that the reader will be shared with between all GPUS.
if args.train_data == "null":
def add_image_input(model):
AddNullInput(
model,
None,
batch_size=batch_per_device,
img_size=args.image_size,
dtype=args.dtype,
)
else:
reader = train_model.CreateDB(
"reader",
db=args.train_data,
db_type=args.db_type,
num_shards=num_shards,
shard_id=shard_id,
)
def add_image_input(model):
AddImageInput(
model,
reader,
batch_size=batch_per_device,
img_size=args.image_size,
dtype=args.dtype,
is_test=False,
)
def add_post_sync_ops(model):
"""Add ops applied after initial parameter sync."""
for param_info in model.GetOptimizationParamInfo(model.GetParams()):
if param_info.blob_copy is not None:
model.param_init_net.HalfToFloat(
param_info.blob, param_info.blob_copy[core.DataType.FLOAT])
# Create parallelized model
data_parallel_model.Parallelize(train_model,
input_builder_fun=add_image_input,
forward_pass_builder_fun=create_model_ops,
optimizer_builder_fun=add_optimizer,
post_sync_builder_fun=add_post_sync_ops,
devices=gpus,
rendezvous=rendezvous,
optimize_gradient_memory=False,
cpu_device=args.use_cpu,
shared_model=args.use_cpu,
combine_spatial_bn=args.use_cpu,
use_nccl=args.use_nccl)
if args.model_parallel:
# Shift half of the activations to another GPU
assert workspace.NumCudaDevices() >= 2 * args.num_gpus
activations = data_parallel_model_utils.GetActivationBlobs(train_model)
data_parallel_model_utils.ShiftActivationDevices(
train_model,
activations=activations[len(activations) // 2:],
shifts={g: args.num_gpus + g
for g in range(args.num_gpus)},
)
data_parallel_model.OptimizeGradientMemory(train_model, {}, set(), False)
workspace.RunNetOnce(train_model.param_init_net)
workspace.CreateNet(train_model.net)
if "GLOO_ALGORITHM" in os.environ and os.environ[
"GLOO_ALGORITHM"] == "PHUB":
#i need to communicate to PHub about the elements that need aggregation,
#as well as their sizes.
#at this stage, all i need is the name of keys and my key ID.
grad_names = list(reversed(train_model._grad_names))
phubKeyNames = ["allreduce_{}_status".format(x) for x in grad_names]
caffe2GradSizes = dict(
zip([
data_parallel_model.stripBlobName(name) + "_grad"
for name in train_model._parameters_info.keys()
], [x.size for x in train_model._parameters_info.values()]))
phubKeySizes = [str(caffe2GradSizes[x]) for x in grad_names]
if rendezvous["shard_id"] == 0:
#only id 0 needs to send to rendezvous.
r = redis.StrictRedis()
#foreach key, I need to assign an ID
joinedStr = ",".join(phubKeyNames)
r.set("[PLink]IntegrationKeys", joinedStr)
joinedStr = ",".join(phubKeySizes)
r.set("[PLink]IntegrationKeySizes", joinedStr)
# Add test model, if specified
test_model = None
if (args.test_data is not None):
log.info("----- Create test net ----")
test_arg_scope = {
'order': "NCHW",
'use_cudnn': True,
'cudnn_exhaustive_search': True,
}
test_model = model_helper.ModelHelper(name="resnet50_test",
arg_scope=test_arg_scope,
init_params=False)
test_reader = test_model.CreateDB(
"test_reader",
db=args.test_data,
db_type=args.db_type,
)
def test_input_fn(model):
AddImageInput(
model,
test_reader,
batch_size=batch_per_device,
img_size=args.image_size,
dtype=args.dtype,
is_test=True,
)
data_parallel_model.Parallelize(
test_model,
input_builder_fun=test_input_fn,
forward_pass_builder_fun=create_model_ops_test,
post_sync_builder_fun=add_post_sync_ops,
param_update_builder_fun=None,
devices=gpus,
cpu_device=args.use_cpu,
)
workspace.RunNetOnce(test_model.param_init_net)
workspace.CreateNet(test_model.net)
epoch = 0
# load the pre-trained model and reset epoch
if args.load_model_path is not None:
LoadModel(args.load_model_path, train_model)
# Sync the model params
data_parallel_model.FinalizeAfterCheckpoint(train_model)
# reset epoch. load_model_path should end with *_X.mdl,
# where X is the epoch number
last_str = args.load_model_path.split('_')[-1]
if last_str.endswith('.mdl'):
epoch = int(last_str[:-4])
log.info("Reset epoch to {}".format(epoch))
else:
log.warning("The format of load_model_path doesn't match!")
expname = "resnet50_gpu%d_b%d_L%d_lr%.2f_v2" % (
args.num_gpus,
total_batch_size,
args.num_labels,
args.base_learning_rate,
)
explog = experiment_util.ModelTrainerLog(expname, args)
# Run the training one epoch a time
while epoch < args.num_epochs:
epoch = RunEpoch(args, epoch, train_model, test_model,
total_batch_size, num_shards, expname, explog)
# Save the model for each epoch
SaveModel(args, train_model, epoch)
model_path = "%s/%s_" % (args.file_store_path, args.save_model_name)
# remove the saved model from the previous epoch if it exists
if os.path.isfile(model_path + str(epoch - 1) + ".mdl"):
os.remove(model_path + str(epoch - 1) + ".mdl")
def Train(args):
subdir = datetime.strftime(datetime.now(), '%Y%m%d-%H%M%S')
save_dir = os.path.join(args.file_store_path, subdir)
if not os.path.exists(save_dir): # Create the model directory if it doesn't exist
os.mkdir(save_dir)
# Either use specified device list or generate one
if args.gpus is not None:
gpus = [int(x) for x in args.gpus.split(',')]
num_gpus = len(gpus)
else:
gpus = list(range(args.num_gpus))
num_gpus = args.num_gpus
log.info("Running on GPUs: {}".format(gpus))
# Verify valid batch size
total_batch_size = args.batch_size
batch_per_device = total_batch_size // num_gpus
assert \
total_batch_size % num_gpus == 0, \
"Number of GPUs must divide batch size"
# Round down epoch size to closest multiple of batch size across machines
global_batch_size = total_batch_size * args.num_shards
epoch_iters = int(args.epoch_size / global_batch_size)
args.epoch_size = epoch_iters * global_batch_size
log.info("Using epoch size: {}".format(args.epoch_size))
# Create ModelHelper object
train_arg_scope = {
'order': 'NCHW',
'use_cudnn': True,
'cudnn_exhaustive_search': True,
'ws_nbytes_limit': (args.cudnn_workspace_limit_mb * 1024 * 1024),
}
train_model = model_helper.ModelHelper(
name="sphereface", arg_scope=train_arg_scope
)
num_shards = args.num_shards
shard_id = args.shard_id
# Expect interfaces to be comma separated.
# Use of multiple network interfaces is not yet complete,
# so simply use the first one in the list.
interfaces = args.distributed_interfaces.split(",")
# Rendezvous using MPI when run with mpirun
if os.getenv("OMPI_COMM_WORLD_SIZE") is not None:
num_shards = int(os.getenv("OMPI_COMM_WORLD_SIZE", 1))
shard_id = int(os.getenv("OMPI_COMM_WORLD_RANK", 0))
if num_shards > 1:
rendezvous = dict(
kv_handler=None,
num_shards=num_shards,
shard_id=shard_id,
engine="GLOO",
transport=args.distributed_transport,
interface=interfaces[0],
mpi_rendezvous=True,
exit_nets=None)
elif num_shards > 1:
# Create rendezvous for distributed computation
store_handler = "store_handler"
if args.redis_host is not None:
# Use Redis for rendezvous if Redis host is specified
workspace.RunOperatorOnce(
core.CreateOperator(
"RedisStoreHandlerCreate", [], [store_handler],
host=args.redis_host,
port=args.redis_port,
prefix=args.run_id,
)
)
else:
# Use filesystem for rendezvous otherwise
workspace.RunOperatorOnce(
core.CreateOperator(
"FileStoreHandlerCreate", [], [store_handler],
path=args.file_store_path,
prefix=args.run_id,
)
)
rendezvous = dict(
kv_handler=store_handler,
shard_id=shard_id,
num_shards=num_shards,
engine="GLOO",
transport=args.distributed_transport,
interface=interfaces[0],
exit_nets=None)
else:
rendezvous = None
# Model building functions
def create_sphereface_model_ops(model, loss_scale):
initializer = (pFP16Initializer if args.dtype == 'float16'
else Initializer)
with brew.arg_scope([brew.conv, brew.fc],
WeightInitializer=initializer,
BiasInitializer=initializer,
enable_tensor_core=args.enable_tensor_core):
pred = sphereface.create_net(
model,
"data",
"label",
in_dim=args.num_channels,
class_num=args.num_labels,
feature_dim=args.feature_dim,
is_test=False,
no_loss=True,
fp16_data=True if args.dtype == 'float16' else False,
)
if args.dtype == 'float16':
pred = model.net.HalfToFloat(pred, pred + '_fp32')
softmax, loss = model.SoftmaxWithLoss([pred, 'label'],
['softmax', 'loss'])
loss = model.Scale(loss, scale=loss_scale)
brew.accuracy(model, [softmax, "label"], "accuracy")
return [loss]
def add_optimizer(model):
# stepsz = int(30 * args.epoch_size / total_batch_size / num_shards)
stepsz = 1
if args.dtype == 'float16':
opt = optimizer.build_fp16_sgd(
model,
args.base_learning_rate,
momentum=0.9,
nesterov=1,
weight_decay=args.weight_decay, # weight decay included
policy="step",
stepsize=stepsz,
gamma=0.9999
)
else:
optimizer.add_weight_decay(model, args.weight_decay)
opt = optimizer.build_multi_precision_sgd(
model,
args.base_learning_rate,
momentum=0.9,
nesterov=1,
policy="step",
stepsize=stepsz,
gamma=0.9999
)
return opt
# Define add_image_input function.
# Depends on the "train_data" argument.
# Note that the reader will be shared with between all GPUS.
if args.train_data == "null":
def add_image_input(model):
AddNullInput(
model,
None,
batch_size=batch_per_device,
img_size=args.image_size,
dtype=args.dtype,
)
else:
reader = train_model.CreateDB(
"reader",
db=args.train_data,
db_type=args.db_type,
num_shards=num_shards,
shard_id=shard_id,
)
def add_image_input(model):
AddImageInput(
model,
reader,
batch_size=batch_per_device,
img_size=args.image_size,
dtype=args.dtype,
is_test=False,
)
def add_post_sync_ops(model):
"""Add ops applied after initial parameter sync."""
for param_info in model.GetOptimizationParamInfo(model.GetParams()):
if param_info.blob_copy is not None:
model.param_init_net.HalfToFloat(
param_info.blob,
param_info.blob_copy[core.DataType.FLOAT]
)
# Create parallelized model
data_parallel_model.Parallelize(
train_model,
input_builder_fun=add_image_input,
forward_pass_builder_fun=create_sphereface_model_ops,
optimizer_builder_fun=add_optimizer,
post_sync_builder_fun=add_post_sync_ops,
devices=gpus,
rendezvous=rendezvous,
optimize_gradient_memory=True,
cpu_device=args.use_cpu,
shared_model=args.use_cpu,
)
workspace.RunNetOnce(train_model.param_init_net)
workspace.CreateNet(train_model.net)
# Add test model, if specified
test_model = None
if (args.test_data is not None):
log.info("----- Create test net ----")
test_arg_scope = {
'order': "NCHW",
'use_cudnn': True,
'cudnn_exhaustive_search': True,
}
test_model = model_helper.ModelHelper(
name="sphereface_test", arg_scope=test_arg_scope, init_params=False
)
test_reader = test_model.CreateDB(
"test_reader",
db=args.test_data,
db_type=args.db_type,
)
def test_input_fn(model):
AddImageInput(
model,
test_reader,
batch_size=batch_per_device,
img_size=args.image_size,
dtype=args.dtype,
is_test=True,
)
data_parallel_model.Parallelize(
test_model,
input_builder_fun=test_input_fn,
forward_pass_builder_fun=create_sphereface_model_ops,
post_sync_builder_fun=add_post_sync_ops,
param_update_builder_fun=None,
devices=gpus,
cpu_device=args.use_cpu,
)
workspace.RunNetOnce(test_model.param_init_net)
workspace.CreateNet(test_model.net)
graph = net_drawer2.GetPydotGraphMinimal(test_model.net.Proto(),
"sphereface",
rankdir="TB")
graph.write(os.path.join(save_dir, "sphereface.pdf"), format='pdf')
epoch = 0
# load the pre-trained model and reset epoch
if args.load_model_path is not None:
LoadModel(args.load_model_path, train_model)
# Sync the model params
data_parallel_model.FinalizeAfterCheckpoint(train_model)
# reset epoch. load_model_path should end with *_X.mdl,
# where X is the epoch number
last_str = args.load_model_path.split('_')[-1]
if last_str.endswith('.mdl'):
epoch = int(last_str[:-4])
log.info("Reset epoch to {}".format(epoch))
else:
log.warning("The format of load_model_path doesn't match!")
expname = "sphereface_gpu%d_b%d_L%d_lr%.2f_v2" % (
args.num_gpus,
total_batch_size,
args.num_labels,
args.base_learning_rate,
)
explog = experiment_util.ModelTrainerLog(os.path.join(save_dir, expname), args)
kernel_fig, plt_kernel = plt.subplots(nrows=4, ncols=5, figsize=(14, 14))
loss_fig, plt_loss = plt.subplots(1)
plt.tight_layout(h_pad=0, w_pad=0)
plt.ion()
iterations = 0
old_x = 0
old_loss = 0
old_acc = 0
while epoch < args.num_epochs or args.test_data_type != 'VAL':
epoch, epoch_loss, epoch_accuracy = RunEpoch(
args,
epoch,
train_model,
test_model,
total_batch_size,
num_shards,
explog,
plt_kernel
)
x = list(range(iterations, iterations + len(epoch_loss)))
x.insert(0, old_x)
epoch_loss.insert(0, old_loss)
epoch_accuracy.insert(0, old_acc)
plt_loss.plot(x, epoch_loss, 'b')
plt_loss.plot(x, epoch_accuracy, 'r')
iterations += len(epoch_loss)
old_x = iterations - 2
old_loss = epoch_loss[-1]
old_acc = epoch_accuracy[-1]
log.info('Save checkpoint {}'.format(epoch))
model_path = '{:s}/{:s}_{:d}.mdl'.format(save_dir, args.save_model_name, epoch)
SaveModel(args, train_model, model_path)
if DEBUG_TRAINING:
kernel_fig_path = '%s/%s_%d.jpg' % (save_dir, 'activation', epoch)
loss_fig_path = '%s/%s_%d.jpg' % (save_dir, 'loss', epoch)
kernel_fig.savefig(kernel_fig_path)
loss_fig.savefig(loss_fig_path)
def Train(args):
# Either use specified device list or generate one
if args.gpus is not None:
gpus = [int(x) for x in args.gpus.split(',')]
num_gpus = len(gpus)
else:
gpus = list(range(args.num_gpus))
num_gpus = args.num_gpus
log.info("Running on GPUs: {}".format(gpus))
# Verify valid batch size
total_batch_size = args.batch_size
batch_per_device = total_batch_size // num_gpus
assert \
total_batch_size % num_gpus == 0, \
"Number of GPUs must divide batch size"
# Round down epoch size to closest multiple of batch size across machines
global_batch_size = total_batch_size * args.num_shards
epoch_iters = int(args.epoch_size / global_batch_size)
args.epoch_size = epoch_iters * global_batch_size
log.info("Using epoch size: {}".format(args.epoch_size))
# Create ModelHelper object
train_arg_scope = {
'order': 'NCHW',
'use_cudnn': True,
'cudnn_exhaustive_search': True,
'ws_nbytes_limit': (args.cudnn_workspace_limit_mb * 1024 * 1024),
}
train_model = model_helper.ModelHelper(name="resnet50",
arg_scope=train_arg_scope)
num_shards = args.num_shards
shard_id = args.shard_id
if num_shards > 1:
# Create rendezvous for distributed computation
store_handler = "store_handler"
if args.redis_host is not None:
# Use Redis for rendezvous if Redis host is specified
workspace.RunOperatorOnce(
core.CreateOperator(
"RedisStoreHandlerCreate",
[],
[store_handler],
host=args.redis_host,
port=args.redis_port,
prefix=args.run_id,
))
else:
# Use filesystem for rendezvous otherwise
workspace.RunOperatorOnce(
core.CreateOperator(
"FileStoreHandlerCreate",
[],
[store_handler],
path=args.file_store_path,
prefix=args.run_id,
))
rendezvous = dict(kv_handler=store_handler,
shard_id=shard_id,
num_shards=num_shards,
engine="GLOO",
exit_nets=None)
else:
rendezvous = None
# Model building functions
def create_resnet50_model_ops(model, loss_scale):
initializer = (pFP16Initializer
if args.dtype == 'float16' else Initializer)
with brew.arg_scope([brew.conv, brew.fc],
WeightInitializer=initializer,
BiasInitializer=initializer,
enable_tensor_core=args.enable_tensor_core):
pred = resnet.create_resnet50(
model,
"data",
num_input_channels=args.num_channels,
num_labels=args.num_labels,
no_bias=True,
no_loss=True,
)
if args.dtype == 'float16':
pred = model.net.HalfToFloat(pred, pred + '_fp32')
softmax, loss = model.SoftmaxWithLoss([pred, 'label'],
['softmax', 'loss'])
loss = model.Scale(loss, scale=loss_scale)
brew.accuracy(model, [softmax, "label"], "accuracy")
return [loss]
def add_optimizer(model):
stepsz = int(30 * args.epoch_size / total_batch_size / num_shards)
optimizer.add_weight_decay(model, args.weight_decay)
opt = optimizer.build_multi_precision_sgd(model,
args.base_learning_rate,
momentum=0.9,
nesterov=1,
policy="step",
stepsize=stepsz,
gamma=0.1)
return opt
# Input. Note that the reader must be shared with all GPUS.
reader = train_model.CreateDB(
"reader",
db=args.train_data,
db_type=args.db_type,
num_shards=num_shards,
shard_id=shard_id,
)
def add_image_input(model):
AddImageInput(
model,
reader,
batch_size=batch_per_device,
img_size=args.image_size,
dtype=args.dtype,
)
def add_post_sync_ops(model):
"""Add ops applied after initial parameter sync."""
for param_info in model.GetOptimizationParamInfo(model.GetParams()):
if param_info.blob_copy is not None:
model.param_init_net.HalfToFloat(
param_info.blob, param_info.blob_copy[core.DataType.FLOAT])
# Create parallelized model
data_parallel_model.Parallelize(
train_model,
input_builder_fun=add_image_input,
forward_pass_builder_fun=create_resnet50_model_ops,
optimizer_builder_fun=add_optimizer,
post_sync_builder_fun=add_post_sync_ops,
devices=gpus,
rendezvous=rendezvous,
optimize_gradient_memory=True,
cpu_device=args.use_cpu,
)
# Add test model, if specified
test_model = None
if (args.test_data is not None):
log.info("----- Create test net ----")
test_arg_scope = {
'order': "NCHW",
'use_cudnn': True,
'cudnn_exhaustive_search': True,
}
test_model = model_helper.ModelHelper(name="resnet50_test",
arg_scope=test_arg_scope,
init_params=False)
test_reader = test_model.CreateDB(
"test_reader",
db=args.test_data,
db_type=args.db_type,
)
def test_input_fn(model):
AddImageInput(
model,
test_reader,
batch_size=batch_per_device,
img_size=args.image_size,
dtype=args.dtype,
)
data_parallel_model.Parallelize(
test_model,
input_builder_fun=test_input_fn,
forward_pass_builder_fun=create_resnet50_model_ops,
post_sync_builder_fun=add_post_sync_ops,
param_update_builder_fun=None,
devices=gpus,
cpu_device=args.use_cpu,
)
workspace.RunNetOnce(test_model.param_init_net)
workspace.CreateNet(test_model.net)
workspace.RunNetOnce(train_model.param_init_net)
workspace.CreateNet(train_model.net)
epoch = 0
# load the pre-trained model and reset epoch
if args.load_model_path is not None:
LoadModel(args.load_model_path, train_model)
# Sync the model params
print("=-=111=")
data_parallel_model.FinalizeAfterCheckpoint(train_model)
print("=-=====")
# reset epoch. load_model_path should end with *_X.mdl,
# where X is the epoch number
last_str = args.load_model_path.split('_')[-1]
if last_str.endswith('.mdl'):
epoch = int(last_str[:-4])
log.info("Reset epoch to {}".format(epoch))
else:
log.warning("The format of load_model_path doesn't match!")
expname = "resnet50_gpu%d_b%d_L%d_lr%.2f_v2" % (
args.num_gpus,
total_batch_size,
args.num_labels,
args.base_learning_rate,
)
explog = experiment_util.ModelTrainerLog(expname, args)
# Run the training one epoch a time
while epoch < args.num_epochs:
epoch = RunEpoch(args, epoch, train_model, test_model,
total_batch_size, num_shards, expname, explog)
# Save the model for each epoch
SaveModel(args, train_model, epoch)
model_path = "%s/%s_" % (args.file_store_path, args.save_model_name)
def Test(args):
if args.gpus is not None:
gpus = [int(x) for x in args.gpus.split(',')]
num_gpus = len(gpus)
else:
gpus = range(args.num_gpus)
num_gpus = args.num_gpus
if num_gpus > 0:
total_batch_size = args.batch_size * num_gpus
log.info("Running on GPUs: {}".format(gpus))
log.info("total_batch_size: {}".format(total_batch_size))
else:
total_batch_size = args.batch_size
log.info("Running on CPU")
log.info("total_batch_size: {}".format(total_batch_size))
# Model building functions
def create_model_ops(model, loss_scale):
return model_builder.build_model(
model=model,
model_name=args.model_name,
model_depth=args.model_depth,
num_labels=args.num_labels,
num_channels=args.num_channels,
crop_size=args.crop_size,
clip_length=(
args.clip_length_of if args.input_type == 1
else args.clip_length_rgb
),
loss_scale=loss_scale,
is_test=1,
pred_layer_name=args.pred_layer_name,
)
test_model = cnn.CNNModelHelper(
order="NCHW",
name="video_model_test",
use_cudnn=(True if args.use_cudnn == 1 else False),
cudnn_exhaustive_search=True,
)
test_reader, number_of_examples = model_builder.create_data_reader(
test_model,
name="test_reader",
input_data=args.test_data,
)
if args.num_iter <= 0:
num_iter = int(number_of_examples / total_batch_size)
else:
num_iter = args.num_iter
def test_input_fn(model):
model_helper.AddVideoInput(
test_model,
test_reader,
batch_size=args.batch_size,
clip_per_video=args.clip_per_video,
decode_type=1,
length_rgb=args.clip_length_rgb,
sampling_rate_rgb=args.sampling_rate_rgb,
scale_h=args.scale_h,
scale_w=args.scale_w,
crop_size=args.crop_size,
num_decode_threads=4,
num_of_class=args.num_labels,
random_mirror=False,
random_crop=False,
input_type=args.input_type,
length_of=args.clip_length_of,
sampling_rate_of=args.sampling_rate_of,
frame_gap_of=args.frame_gap_of,
do_flow_aggregation=args.do_flow_aggregation,
flow_data_type=args.flow_data_type,
get_rgb=(args.input_type == 0),
get_optical_flow=(args.input_type == 1),
get_video_id=args.get_video_id,
use_local_file=args.use_local_file,
)
if num_gpus > 0:
data_parallel_model.Parallelize_GPU(
test_model,
input_builder_fun=test_input_fn,
forward_pass_builder_fun=create_model_ops,
param_update_builder_fun=None,
devices=gpus
)
else:
test_model._device_type = caffe2_pb2.CPU
test_model._devices = [0]
device_opt = core.DeviceOption(test_model._device_type, 0)
with core.DeviceScope(device_opt):
# Because our loaded models are named with "gpu_x", keep the naming for now.
# TODO: Save model using `data_parallel_model.ExtractPredictorNet`
# to extract the model for "gpu_0". It also renames
# the input and output blobs by stripping the "gpu_x/" prefix
with core.NameScope("{}_{}".format("gpu", 0)):
test_input_fn(test_model)
create_model_ops(test_model, 1.0)
workspace.RunNetOnce(test_model.param_init_net)
workspace.CreateNet(test_model.net)
if args.db_type == 'minidb':
if num_gpus > 0:
model_helper.LoadModel(args.load_model_path, args.db_type)
data_parallel_model.FinalizeAfterCheckpoint(test_model)
else:
with core.DeviceScope(core.DeviceOption(caffe2_pb2.CPU, 0)):
model_helper.LoadModel(args.load_model_path, args.db_type)
elif args.db_type == 'pickle':
if num_gpus > 0:
model_loader.LoadModelFromPickleFile(
test_model,
args.load_model_path,
use_gpu=True,
root_gpu_id=gpus[0]
)
data_parallel_model.FinalizeAfterCheckpoint(test_model)
else:
model_loader.LoadModelFromPickleFile(
test_model,
args.load_model_path,
use_gpu=False
)
else:
log.warning("Unsupported db_type: {}".format(args.db_type))
# metric counters for classification
clip_acc = 0
video_top1 = 0
video_topk = 0
video_count = 0
clip_count = 0
for i in range(num_iter):
workspace.RunNet(test_model.net.Proto().name)
num_devices = 1 # default for cpu
if args.num_gpus > 0:
num_devices = args.num_gpus
for g in range(num_devices):
# get labels
label = workspace.FetchBlob(
"gpu_{}".format(g) + '/label'
)
# get predictions
predicts = workspace.FetchBlob("gpu_{}".format(g) + '/softmax')
assert predicts.shape[0] == args.batch_size * args.clip_per_video
for j in range(args.batch_size):
# get label for one video
sample_label = label[j * args.clip_per_video]
# get clip accuracy
for k in range(args.clip_per_video):
c1, _ = metric.accuracy_metric(
predicts[j * args.clip_per_video + k, :],
label[j * args.clip_per_video + k])
clip_acc = clip_acc + c1
# get all clip predictions for one video
all_clips = predicts[
j * args.clip_per_video:(j + 1) * args.clip_per_video, :]
# aggregate predictions into one
video_pred = PredictionAggregation(all_clips, args.aggregation)
c1, ck = metric.accuracy_metric(
video_pred, sample_label, args.top_k)
video_top1 = video_top1 + c1
video_topk = video_topk + ck
video_count = video_count + args.batch_size
clip_count = clip_count + label.shape[0]
if i > 0 and i % args.display_iter == 0:
log.info('Iter {}/{}: clip: {}, top1: {}, top 5: {}'.format(
i,
num_iter,
clip_acc / clip_count,
video_top1 / video_count,
video_topk / video_count))
log.info("Test accuracy: clip: {}, top 1: {}, top{}: {}".format(
clip_acc / clip_count,
video_top1 / video_count,
args.top_k,
video_topk / video_count
))
if num_gpus > 0:
flops, params = model_helper.GetFlopsAndParams(test_model, gpus[0])
else:
flops, params = model_helper.GetFlopsAndParams(test_model)
log.info('FLOPs: {}, params: {}'.format(flops, params))
def ExtractFeatures(args):
if args.gpus is not None:
gpus = [int(x) for x in args.gpus.split(',')]
num_gpus = len(gpus)
else:
gpus = range(args.num_gpus)
num_gpus = args.num_gpus
if num_gpus > 0:
log.info("Running on GPUs: {}".format(gpus))
else:
log.info("Running on CPU")
log.info("Running on GPUs: {}".format(gpus))
my_arg_scope = {
'order': 'NCHW',
'use_cudnn': True,
'cudnn_exhaustive_search': True
}
model = cnn.CNNModelHelper(name="Extract Features", **my_arg_scope)
reader, num_examples = model_builder.create_data_reader(
model,
name="reader",
input_data=args.test_data,
)
def input_fn(model):
model_helper.AddVideoInput(
model,
reader,
batch_size=args.batch_size,
clip_per_video=args.clip_per_video,
decode_type=args.decode_type,
length_rgb=args.clip_length_rgb,
sampling_rate_rgb=args.sampling_rate_rgb,
scale_h=args.scale_h,
scale_w=args.scale_w,
crop_size=args.crop_size,
num_decode_threads=args.num_decode_threads,
num_of_class=args.num_labels,
random_mirror=False,
random_crop=False,
input_type=args.input_type,
length_of=args.clip_length_of,
sampling_rate_of=args.sampling_rate_of,
frame_gap_of=args.frame_gap_of,
do_flow_aggregation=args.do_flow_aggregation,
flow_data_type=args.flow_data_type,
get_rgb=(args.input_type == 0),
get_optical_flow=(args.input_type == 1),
get_video_id=args.get_video_id,
use_local_file=args.use_local_file,
)
def create_model_ops(model, loss_scale):
return model_builder.build_model(
model=model,
model_name=args.model_name,
model_depth=args.model_depth,
num_labels=args.num_labels,
num_channels=args.num_channels,
crop_size=args.crop_size,
clip_length=(args.clip_length_of
if args.input_type == 1 else args.clip_length_rgb),
loss_scale=loss_scale,
is_test=1,
)
if num_gpus > 0:
data_parallel_model.Parallelize_GPU(
model,
input_builder_fun=input_fn,
forward_pass_builder_fun=create_model_ops,
param_update_builder_fun=None, # 'None' since we aren't training
devices=gpus,
)
else:
model._device_type = caffe2_pb2.CPU
model._devices = [0]
device_opt = core.DeviceOption(model._device_type, 0)
with core.DeviceScope(device_opt):
with core.NameScope("{}_{}".format("gpu", 0)):
input_fn(model)
create_model_ops(model, 1.0)
workspace.RunNetOnce(model.param_init_net)
workspace.CreateNet(model.net)
if args.db_type == 'minidb':
if num_gpus > 0:
model_helper.LoadModel(args.load_model_path, args.db_type)
data_parallel_model.FinalizeAfterCheckpoint(model)
else:
with core.DeviceScope(core.DeviceOption(caffe2_pb2.CPU, 0)):
model_helper.LoadModel(args.load_model_path, args.db_type)
elif args.db_type == 'pickle':
if num_gpus > 0:
model_loader.LoadModelFromPickleFile(model,
args.load_model_path,
use_gpu=True,
root_gpu_id=gpus[0])
else:
model_loader.LoadModelFromPickleFile(
model,
args.load_model_path,
use_gpu=False,
)
else:
log.warning("Unsupported db_type: {}".format(args.db_type))
def fetchActivations(model, outputs, num_iterations):
all_activations = {}
for counter in range(num_iterations):
workspace.RunNet(model.net.Proto().name)
num_devices = 1 # default for cpu
for g in gpus if num_gpus > 0 else range(num_devices):
for output_name in outputs:
blob_name = 'gpu_{}/'.format(g) + output_name
activations = workspace.FetchBlob(blob_name)
if output_name not in all_activations:
all_activations[output_name] = []
all_activations[output_name].append(activations)
if counter % 20 == 0:
log.info('{}/{} iterations'.format(counter, num_iterations))
# each key holds a list of activations obtained from each minibatch.
# we now concatenate these lists to get the final arrays.
# concatenating during the loop requires a realloc and can get slow.
for key in all_activations:
all_activations[key] = np.concatenate(all_activations[key])
return all_activations
outputs = [name.strip() for name in args.features.split(',')]
assert len(outputs) > 0
if args.num_iterations > 0:
num_iterations = args.num_iterations
else:
if num_gpus > 0:
examples_per_iteration = args.batch_size * num_gpus
else:
examples_per_iteration = args.batch_size
num_iterations = int(num_examples / examples_per_iteration)
activations = fetchActivations(model, outputs, num_iterations)
# saving extracted features
for index in range(len(outputs)):
log.info("Read '{}' with shape {}".format(
outputs[index], activations[outputs[index]].shape))
if args.output_path:
output_path = args.output_path
else:
output_path = os.path.dirname(args.test_data) + '/features.pickle'
log.info('Writing to {}'.format(output_path))
with open(output_path, 'wb') as handle:
pickle.dump(activations, handle)
# perform sanity check
if args.sanity_check == 1: # check clip accuracy
clip_acc = 0
softmax = activations['softmax']
label = activations['label']
for i in range(len(softmax)):
sorted_preds = \
np.argsort(softmax[i])
sorted_preds[:] = sorted_preds[::-1]
if sorted_preds[0] == label[i]:
clip_acc += 1
log.info('Sanity check --- clip accuracy: {}'.format(clip_acc /
len(softmax)))
def Test(args):
if args.gpus is not None:
gpus = [int(x) for x in args.gpus.split(',')]
num_gpus = len(gpus)
else:
gpus = range(args.num_gpus)
num_gpus = args.num_gpus
if num_gpus > 0:
total_batch_size = args.batch_size * num_gpus
log.info("Running on GPUs: {}".format(gpus))
log.info("total_batch_size: {}".format(total_batch_size))
else:
total_batch_size = args.batch_size
log.info("Running on CPU")
log.info("total_batch_size: {}".format(total_batch_size))
video_input_args = dict(
batch_size=args.batch_size,
clip_per_video=args.clip_per_video,
decode_type=1,
length_rgb=args.clip_length_rgb,
sampling_rate_rgb=args.sampling_rate_rgb,
scale_h=args.scale_h,
scale_w=args.scale_w,
crop_size=args.crop_size,
video_res_type=args.video_res_type,
short_edge=min(args.scale_h, args.scale_w),
num_decode_threads=args.num_decode_threads,
do_multi_label=args.multi_label,
num_of_class=args.num_labels,
random_mirror=False,
random_crop=False,
input_type=args.input_type,
length_of=args.clip_length_of,
sampling_rate_of=args.sampling_rate_of,
frame_gap_of=args.frame_gap_of,
do_flow_aggregation=args.do_flow_aggregation,
flow_data_type=args.flow_data_type,
get_rgb=(args.input_type == 0 or args.input_type >= 3),
get_optical_flow=(args.input_type == 1 or args.input_type >= 4),
use_local_file=args.use_local_file,
crop_per_clip=args.crop_per_clip,
)
reader_args = dict(
name="test_reader",
input_data=args.test_data,
)
# Model building functions
def create_model_ops(model, loss_scale):
return model_builder.build_model(
model=model,
model_name=args.model_name,
model_depth=args.model_depth,
num_labels=args.num_labels,
batch_size=args.batch_size * args.clip_per_video *
args.crop_per_clip,
num_channels=args.num_channels,
crop_size=args.crop_size,
clip_length=(args.clip_length_of
if args.input_type == 1 else args.clip_length_rgb),
loss_scale=loss_scale,
is_test=1,
pred_layer_name=args.pred_layer_name,
multi_label=args.multi_label,
channel_multiplier=args.channel_multiplier,
bottleneck_multiplier=args.bottleneck_multiplier,
use_dropout=args.use_dropout,
conv1_temporal_stride=args.conv1_temporal_stride,
conv1_temporal_kernel=args.conv1_temporal_kernel,
use_convolutional_pred=args.use_convolutional_pred,
use_pool1=args.use_pool1,
)
test_model = cnn.CNNModelHelper(
order="NCHW",
name="video_model_test",
use_cudnn=(True if args.use_cudnn == 1 else False),
cudnn_exhaustive_search=True,
)
test_reader, number_of_examples = reader_utils.create_data_reader(
test_model, **reader_args)
if args.num_iter <= 0:
num_iter = int(math.ceil(number_of_examples / total_batch_size))
else:
num_iter = args.num_iter
def test_input_fn(model):
model_helper.AddVideoInput(test_model, test_reader, **video_input_args)
if num_gpus > 0:
data_parallel_model.Parallelize_GPU(
test_model,
input_builder_fun=test_input_fn,
forward_pass_builder_fun=create_model_ops,
param_update_builder_fun=None,
devices=gpus,
optimize_gradient_memory=True,
)
else:
test_model._device_type = caffe2_pb2.CPU
test_model._devices = [0]
device_opt = core.DeviceOption(test_model._device_type, 0)
with core.DeviceScope(device_opt):
# Because our loaded models are named with "gpu_x", keep the naming for now.
# TODO: Save model using `data_parallel_model.ExtractPredictorNet`
# to extract the model for "gpu_0". It also renames
# the input and output blobs by stripping the "gpu_x/" prefix
with core.NameScope("{}_{}".format("gpu", 0)):
test_input_fn(test_model)
create_model_ops(test_model, 1.0)
workspace.RunNetOnce(test_model.param_init_net)
workspace.CreateNet(test_model.net)
if args.db_type == 'minidb':
if num_gpus > 0:
model_helper.LoadModel(args.load_model_path, args.db_type)
else:
with core.DeviceScope(core.DeviceOption(caffe2_pb2.CPU, 0)):
model_helper.LoadModel(args.load_model_path, args.db_type)
elif args.db_type == 'pickle':
if num_gpus > 0:
model_loader.LoadModelFromPickleFile(test_model,
args.load_model_path,
use_gpu=True,
root_gpu_id=gpus[0])
else:
model_loader.LoadModelFromPickleFile(test_model,
args.load_model_path,
use_gpu=False)
else:
log.warning("Unsupported db_type: {}".format(args.db_type))
data_parallel_model.FinalizeAfterCheckpoint(test_model)
# metric couters for multilabel
all_prob_for_map = np.empty(shape=[0, args.num_labels], dtype=np.float)
all_label_for_map = np.empty(shape=[0, args.num_labels], dtype=np.int32)
# metric counters for closed-world classification
clip_acc = 0
video_top1 = 0
video_topk = 0
video_count = 0
clip_count = 0
crop_per_video = args.clip_per_video * args.crop_per_clip
for i in range(num_iter):
workspace.RunNet(test_model.net.Proto().name)
num_devices = 1 # default for cpu
if num_gpus > 0:
num_devices = num_gpus
for g in range(num_devices):
# get labels
label = workspace.FetchBlob("gpu_{}".format(g) + '/label')
# get predictions
if args.multi_label:
predicts = workspace.FetchBlob("gpu_{}".format(g) + '/prob')
else:
predicts = workspace.FetchBlob("gpu_{}".format(g) + '/softmax')
assert predicts.shape[0] == args.batch_size * crop_per_video
for j in range(args.batch_size):
# get label for one video
if args.multi_label:
sample_label = label[j * crop_per_video, :]
else:
sample_label = label[j * crop_per_video]
# get clip accuracy
for k in range(crop_per_video):
sorted_preds = \
np.argsort(predicts[j * crop_per_video + k, :])
sorted_preds[:] = sorted_preds[::-1]
if sorted_preds[0] == label[j * crop_per_video + k]:
clip_acc = clip_acc + 1
# get all clip predictions for one video
all_clips = \
predicts[
j * crop_per_video:(j + 1) * crop_per_video, :
]
# aggregate predictions into one
video_pred = PredictionAggregation(all_clips, args.aggregation)
if args.multi_label:
video_pred = np.expand_dims(video_pred, axis=0)
sample_label = np.expand_dims(sample_label, axis=0)
all_prob_for_map = np.concatenate(
(all_prob_for_map, video_pred), axis=0)
all_label_for_map = np.concatenate(
(all_label_for_map, sample_label), axis=0)
else:
sorted_video_pred = np.argsort(video_pred)
sorted_video_pred[:] = sorted_video_pred[::-1]
if sorted_video_pred[0] == sample_label:
video_top1 = video_top1 + 1
if sample_label in sorted_video_pred[0:args.top_k]:
video_topk = video_topk + 1
video_count = video_count + args.batch_size
clip_count = clip_count + label.shape[0]
if i > 0 and i % args.display_iter == 0:
if args.multi_label:
# mAP
auc, ap, wap, aps = metric.mean_ap_metric(
all_prob_for_map, all_label_for_map)
log.info(
'Iter {}/{}: mAUC: {}, mAP: {}, mWAP: {}, mAP_all: {}'.
format(i, num_iter, auc, ap, wap, np.mean(aps)))
else:
# accuracy
log.info('Iter {}/{}: clip: {}, top1: {}, top 5: {}'.format(
i, num_iter, clip_acc / clip_count,
video_top1 / video_count, video_topk / video_count))
if args.multi_label:
# mAP
auc, ap, wap, aps = metric.mean_ap_metric(all_prob_for_map,
all_label_for_map)
log.info("Test mAUC: {}, mAP: {}, mWAP: {}, mAP_all: {}".format(
auc, ap, wap, np.mean(aps)))
if args.print_per_class_metrics:
log.info("Test mAP per class: {}".format(aps))
else:
# accuracy
log.info("Test accuracy: clip: {}, top 1: {}, top{}: {}".format(
clip_acc / clip_count, video_top1 / video_count, args.top_k,
video_topk / video_count))
if num_gpus > 0:
flops, params, inters = model_helper.GetFlopsAndParams(
test_model, gpus[0])
else:
flops, params, inters = model_helper.GetFlopsAndParams(test_model)
log.info('FLOPs: {}, params: {}, inters: {}'.format(flops, params, inters))
def Train(args):
# Either use specified device list or generate one
if args.gpus is not None:
gpus = [int(x) for x in args.gpus.split(',')]
num_gpus = len(gpus)
else:
gpus = list(range(args.num_gpus))
num_gpus = args.num_gpus
log.info("Running on GPUs: {}".format(gpus))
# Verify valid batch size
total_batch_size = args.batch_size
batch_per_device = total_batch_size // num_gpus
assert \
total_batch_size % num_gpus == 0, \
"Number of GPUs must divide batch size"
# Round down epoch size to closest multiple of batch size across machines
global_batch_size = total_batch_size * args.num_shards
epoch_iters = int(args.epoch_size / global_batch_size)
args.epoch_size = epoch_iters * global_batch_size
log.info("Using epoch size: {}".format(args.epoch_size))
# Create ModelHelper object
train_arg_scope = {
'order': 'NCHW',
'use_cudnn': True,
'cudnn_exhaustice_search': True,
'ws_nbytes_limit': (args.cudnn_workspace_limit_mb * 1024 * 1024),
}
train_model = model_helper.ModelHelper(name="mobilenet",
arg_scope=train_arg_scope)
num_shards = args.num_shards
shard_id = args.shard_id
if num_shards > 1:
# Create rendezvous for distributed computation
store_handler = "store_handler"
if args.redis_host is not None:
# Use Redis for rendezvous if Redis host is specified
workspace.RunOperatorOnce(
core.CreateOperator(
"RedisStoreHandlerCreate",
[],
[store_handler],
host=args.redis_host,
port=args.redis_port,
prefix=args.run_id,
))
else:
# Use filesystem for rendezvous otherwise
workspace.RunOperatorOnce(
core.CreateOperator(
"FileStoreHandlerCreate",
[],
[store_handler],
path=args.file_store_path,
))
rendezvous = dict(kv_handler=store_handler,
shard_id=shard_id,
num_shards=num_shards,
engine="GLOO",
exit_nets=None)
else:
rendezvous = None
# Model building functions
def create_mobilenet_model_ops(model, loss_scale):
[softmax, loss
] = mobilenet.create_mobilenet(model,
"data",
num_input_channels=args.num_channels,
num_labels=args.num_labels,
label="label")
loss = model.Scale(loss, scale=loss_scale)
brew.accuracy(model, [softmax, "label"], "accuracy")
return [loss]
def add_optimizer(model):
stepsz = int(30 * args.epoch_size / total_batch_size / num_shards)
optimizer.add_weight_decay(model, args.weight_decay)
optimizer.build_sgd(model,
args.base_learning_rate,
momentum=0.9,
nesterov=1,
policy="step",
stepsize=stepsz,
gamma=0.1)
# Input. Note that the reader must be shared with all GPUS.
reader = train_model.CreateDB(
"reader",
db=args.train_data,
db_type=args.db_type,
num_shards=num_shards,
shard_id=shard_id,
)
def add_image_input(model):
AddImageInput(model,
reader,
batch_size=batch_per_device,
img_size=args.image_size,
is_test=False)
# Create parallelized model
data_parallel_model.Parallelize_GPU(
train_model,
input_builder_fun=add_image_input,
forward_pass_builder_fun=create_mobilenet_model_ops,
optimizer_builder_fun=add_optimizer,
devices=gpus,
rendezvous=rendezvous,
optimize_gradient_memory=True,
)
workspace.RunNetOnce(train_model.param_init_net)
workspace.CreateNet(train_model.net)
#
# # save network graph
# graph = net_drawer.GetPydotGraphMinimal(
# train_model.net.Proto().op, "mobilenet", rankdir="LR", minimal_dependency=True)
# with open("mobilenet.png", 'wb') as fid:
# fid.write(graph.create_png())
# Add test model, if specified
test_model = None
if (args.test_data is not None):
log.info("----- Create test net ----")
test_arg_scope = {
'order': "NCHW",
'use_cudnn': True,
'cudnn_exhaustive_search': True,
}
test_model = model_helper.ModelHelper(name="mobilenet_test",
arg_scope=test_arg_scope)
test_reader = test_model.CreateDB(
"test_reader",
db=args.test_data,
db_type=args.db_type,
)
def test_input_fn(model):
AddImageInput(model,
test_reader,
batch_size=batch_per_device,
img_size=args.image_size,
is_test=True)
data_parallel_model.Parallelize_GPU(
test_model,
input_builder_fun=test_input_fn,
forward_pass_builder_fun=create_mobilenet_model_ops,
param_update_builder_fun=None,
devices=gpus,
)
workspace.RunNetOnce(test_model.param_init_net)
workspace.CreateNet(test_model.net)
epoch = 0
# load the pre-trained model and mobilenet epoch
if args.load_model_path is not None:
LoadModel(args.load_model_path, train_model)
# Sync the model params
data_parallel_model.FinalizeAfterCheckpoint(train_model)
# mobilenet epoch. load_model_path should end with *_X.mdl,
# where X is the epoch number
last_str = args.load_model_path.split('_')[-1]
if last_str.endswith('.mdl'):
epoch = int(last_str[:-4])
log.info("mobilenet epoch to {}".format(epoch))
else:
log.warning("The format of load_model_path doesn't match!")
expname = "mobilenet_gpu%d_b%d_L%d_lr%.2f_v2" % (
args.num_gpus,
total_batch_size,
args.num_labels,
args.base_learning_rate,
)
explog = experiment_util.ModelTrainerLog(expname, args)
# Run the training one epoch a time
while epoch < args.num_epochs:
epoch = RunEpoch(args, epoch, train_model, test_model,
total_batch_size, num_shards, expname, explog)
# Save the model for each epoch
SaveModel(args, train_model, epoch)
model_path = "%s/%s_" % (args.file_store_path, args.save_model_name)
# remove the saved model from the previous epoch if it exists
if os.path.isfile(model_path + str(epoch - 1) + ".mdl"):
os.remove(model_path + str(epoch - 1) + ".mdl")
def feature_extractor(load_model_path=None,
test_data=None,
gpu_list=None,
num_gpus=0,
batch_size=4,
clip_per_video=1,
decode_type=2,
clip_length_rgb=4,
sampling_rate_rgb=1,
scale_h=128,
scale_w=171,
crop_size=112,
video_res_type=0,
num_decode_threads=4,
multi_label=0,
num_labels=101,
input_type=0,
clip_length_of=8,
sampling_rate_of=2,
frame_gap_of=2,
do_flow_aggregation=0,
flow_data_type=0,
get_video_id=1,
get_start_frame=0,
use_local_file=1,
crop_per_clip=1,
db_type='pickle',
model_name='r2plus1d',
model_depth=18,
num_channels=3,
output_path=None,
use_cudnn=1,
layers='final_avg',
num_iterations=1,
channel_multiplier=1.0,
bottleneck_multiplier=1.0,
use_pool1=0,
use_convolutional_pred=0,
use_dropout=0,
**kwargs):
"""
:param gpu_list: list of gpu ids to use
:param batch_size: batch size
:param clip_per_video: When clips_per_video > 1, sample this many clips uniformly in time
:param decode_type: 0: random, 1: uniform sampling, 2: use starting frame
:param clip_length_rgb: Length of input clips
:param sampling_rate_rgb: Frame sampling rate
:param scale_h: Scale image height to
:param scale_w: Scale image width to
:param crop_size: Input image size (to crop to)
:param video_res_type: Video frame scaling option, 0: scaled by height x width; 1: scaled by short edge
:param num_decode_threads: number of decoding threads
:param multi_label: Multiple label csv file input
:param num_labels: Number of labels
:param input_type: 0=rgb, 1=optical flow
:param clip_length_of: Frames of optical flow data
:param sampling_rate_of: Sampling rate for optial flows
:param frame_gap_of: Frame gap of optical flows
:param do_flow_aggregation: whether to aggregate optical flow across multiple frames
:param flow_data_type: 0=Flow2C, 1=Flow3C, 2=FlowWithGray, 3=FlowWithRGB
:param get_video_id: Output video id
:param get_start_frame: Output clip start frame
:param use_local_file: use local file
:param crop_per_clip: number of spatial crops per clip
:param db_type: Db type of the testing model
:param model_name: Model name
:param model_depth: Model depth
:param num_channels: Number of channels
:param load_model_path: Load saved model for testing
:param test_data: Path to output pickle; defaults to layers.pickle next to <test_data>
:param output_path: Path to output pickle; defaults to layers.pickle next to <test_data>
:param use_cudnn: Use CuDNN
:param layers: Comma-separated list of blob names to fetch
:param num_iterations: Run only this many iterations
:param channel_multiplier: Channel multiplier
:param bottleneck_multiplier: Bottleneck multiplier
:param use_pool1: use pool1 layer
:param use_convolutional_pred: using convolutional predictions
:param use_dropout: Use dropout at the prediction layer
"""
if load_model_path is None or test_data is None:
raise Exception('Model path AND test data need to be specified')
# Initialize Caffe2
workspace.GlobalInit(['caffe2', '--caffe2_log_level=2'])
if gpu_list is None:
if num_gpus == 0:
raise Exception('Must specify GPUs')
else:
gpus = [i for i in range(num_gpus)]
else:
gpus = gpu_list
num_gpus = len(gpus)
my_arg_scope = {
'order': 'NCHW',
'use_cudnn': True,
'cudnn_exhaustive_search': True
}
model = cnn.CNNModelHelper(name="Extract features", **my_arg_scope)
video_input_args = dict(
batch_size=batch_size,
clip_per_video=clip_per_video,
decode_type=decode_type,
length_rgb=clip_length_rgb,
sampling_rate_rgb=sampling_rate_rgb,
scale_h=scale_h,
scale_w=scale_w,
crop_size=crop_size,
video_res_type=video_res_type,
short_edge=min(scale_h, scale_w),
num_decode_threads=num_decode_threads,
do_multi_label=multi_label,
num_of_class=num_labels,
random_mirror=False,
random_crop=False,
input_type=input_type,
length_of=clip_length_of,
sampling_rate_of=sampling_rate_of,
frame_gap_of=frame_gap_of,
do_flow_aggregation=do_flow_aggregation,
flow_data_type=flow_data_type,
get_rgb=input_type == 0,
get_optical_flow=input_type == 1,
get_video_id=get_video_id,
get_start_frame=get_start_frame,
use_local_file=use_local_file,
crop_per_clip=crop_per_clip,
)
reader_args = dict(
name="extract_features" + '_reader',
input_data=test_data,
)
reader, num_examples = reader_utils.create_data_reader(
model, **reader_args)
def input_fn(model):
model_helper.AddVideoInput(model, reader, **video_input_args)
def create_model_ops(model, loss_scale):
return model_builder.build_model(
model=model,
model_name=model_name,
model_depth=model_depth,
num_labels=num_labels,
batch_size=batch_size,
num_channels=num_channels,
crop_size=crop_size,
clip_length=(clip_length_of
if input_type == 1 else clip_length_rgb),
loss_scale=loss_scale,
is_test=1,
multi_label=multi_label,
channel_multiplier=channel_multiplier,
bottleneck_multiplier=bottleneck_multiplier,
use_dropout=use_dropout,
use_convolutional_pred=use_convolutional_pred,
use_pool1=use_pool1,
)
##
if num_gpus > 0:
data_parallel_model.Parallelize_GPU(
model,
input_builder_fun=input_fn,
forward_pass_builder_fun=create_model_ops,
param_update_builder_fun=None, # 'None' since we aren't training
devices=gpus,
optimize_gradient_memory=True,
)
else:
model._device_type = caffe2_pb2.CPU
model._devices = [0]
device_opt = core.DeviceOption(model._device_type, 0)
with core.DeviceScope(device_opt):
# Because our loaded models are named with "gpu_x", keep the naming for now.
# TODO: Save model using `data_parallel_model.ExtractPredictorNet`
# to extract the model for "gpu_0". It also renames
# the input and output blobs by stripping the "gpu_x/" prefix
with core.NameScope("{}_{}".format("gpu", 0)):
input_fn(model)
create_model_ops(model, 1.0)
workspace.RunNetOnce(model.param_init_net)
workspace.CreateNet(model.net)
if db_type == 'pickle':
model_loader.LoadModelFromPickleFile(model, load_model_path)
elif db_type == 'minidb':
if num_gpus > 0:
model_helper.LoadModel(load_model_path, db_type)
else:
with core.DeviceScope(core.DeviceOption(caffe2_pb2.CPU, 0)):
model_helper.LoadModel(load_model_path, db_type)
else:
log.warning("Unsupported db_type: {}".format(db_type))
data_parallel_model.FinalizeAfterCheckpoint(model)
##
def fetchActivations(model, outputs, num_iterations):
all_activations = {}
for counter in range(num_iterations):
workspace.RunNet(model.net.Proto().name)
num_devices = 1 # default for cpu
if num_gpus > 0:
num_devices = num_gpus
for g in range(num_devices):
for output_name in outputs:
blob_name = 'gpu_{}/'.format(g) + output_name
activations = workspace.FetchBlob(blob_name)
if output_name not in all_activations:
all_activations[output_name] = []
all_activations[output_name].append(activations)
# each key holds a list of activations obtained from each minibatch.
# we now concatenate these lists to get the final arrays.
# concatenating during the loop requires a realloc and can get slow.
for key in all_activations:
all_activations[key] = np.concatenate(all_activations[key])
return all_activations
if not isinstance(layers, list):
layers = [layers]
if 'video_id' not in layers:
layers.append('video_id')
assert len(layers) > 0
examples_per_iteration = batch_size * num_gpus
num_iterations = int(num_examples / examples_per_iteration)
activations = fetchActivations(model, layers, num_iterations)
# saving extracted layers
for index in range(len(layers)):
log.info("Read '{}' with shape {}".format(
layers[index], activations[layers[index]].shape))
if output_path:
log.info('Writing to {}'.format(output_path))
if save_h5:
with h5py.File(output_path, 'w') as handle:
for name, activation in activations.items():
handle.create_dataset(name, data=activation)
else:
with open(output_path, 'wb') as handle:
pickle.dump(activations, handle)
else:
return activations
def Test(args):
# Either use specified device list or generate one
if args.gpus is not None:
gpus = [int(x) for x in args.gpus.split(',')]
num_gpus = len(gpus)
else:
gpus = list(range(args.num_gpus))
num_gpus = args.num_gpus
log.info("Running on GPUs: {}".format(gpus))
#========================================================
for db_name, db_size in data_base.items():
workspace.ResetWorkspace()
print(workspace.Blobs())
def create_mobilenet_model_ops(model, loss_scale):
[softmax, loss] = mobilenet.create_mobilenet(
model,
"data",
num_input_channels=args.num_channels,
num_labels=args.num_labels,
label="label")
# loss = model.Scale(loss, scale=loss_scale)
brew.accuracy(model, [softmax, "label"], "accuracy")
# return [loss]
log.info("----- Create test net ----")
test_arg_scope = {
'order': "NCHW",
'use_cudnn': True,
'cudnn_exhaustive_search': True,
}
test_model = model_helper.ModelHelper(
name="mobilenet_test",
arg_scope=test_arg_scope #, init_params=False
)
test_data_db = os.path.join(data_folder, db_name)
print(test_data_db, db_size)
test_reader = test_model.CreateDB(
"test_reader",
db=test_data_db,
db_type=args.db_type,
)
def test_input_fn(model):
AddImageInput(
model,
test_reader,
batch_size=args.batch_size,
img_size=args.image_size,
)
data_parallel_model.Parallelize_GPU(
test_model,
input_builder_fun=test_input_fn,
forward_pass_builder_fun=create_mobilenet_model_ops,
param_update_builder_fun=None,
devices=gpus,
)
workspace.RunNetOnce(test_model.param_init_net)
workspace.CreateNet(test_model.net)
# load the pre-trained model and mobilenet epoch
LoadModel(args.load_model_path, test_model)
data_parallel_model.FinalizeAfterCheckpoint(test_model)
(test_max_softmax, test_max_index,
test_label) = RunEpoch(args, db_size, test_model, args.batch_size)
# flag_accuracy = (test_label == test_max_index).astype(np.int)
# print(flag_accuracy.mean())
save_path = os.path.join(root_folder, db_name[:-5] + ".npz")
print(save_path)
np.savez(save_path,
max_softmax=test_max_softmax,
max_index=test_max_index,
label=test_label)
if os.path.exists(save_path):
print("OK")
