def load_model_simclr(cfg):
# encoder = get_resnet("resnet50", cfg.dataset.name, False, False)
# projector = Projector(2048, 64)
# block = ClBlock(encoder,projector,cfg.device,cfg.N,cfg.batch_size)
# block = block.to(cfg.device)
# return block
proc_group = None
rank = 0
cfg.simcl.load = True
cfg.simcl.rank = rank
cfg.simcl.device = cfg.device
cfg.simcl.denoising_prep = True
simcl = simcl_load_model(cfg.simcl, rank, proc_group)
return simcl
def load_model_simclr(cfg):
# encoder = get_resnet("resnet50", cfg.dataset.name, False, False)
# projector = Projector(2048, 64)
# block = ClBlock(encoder,projector,cfg.device,cfg.N,cfg.batch_size)
# block = block.to(cfg.device)
# return block
proc_group = None
rank = cfg.gpuid
cfg.simcl.load = True
cfg.simcl.rank = rank
cfg.simcl.device = cfg.device
cfg.simcl.denoising_prep = True
simcl = simcl_load_model(cfg.simcl, rank, proc_group)
for name, param in simcl.named_parameters():
param = param.requires_grad_(False)
simcl_enc = ClBlockEnc(simcl, cfg.dynamic.frame_size, cfg.d_model_attn)
return simcl_enc
def load_models(cfg, rank, proc_group):
models = edict()
# HACK: just get the models loaded but parameters are
# saved together in block
load = cfg.load
cfg.load = False
if cfg.simcl.load:
cfg.simcl.rank = rank
cfg.simcl.device = cfg.device
cfg.simcl.denoising_prep = True
simcl = simcl_load_model(cfg.simcl, rank, proc_group)
encoder = simcl.encoder
projector = simcl.projector
# encoder.eval()
# projector.eval()
for param in encoder.parameters():
param.requires_grad = False
for param in projector.parameters():
param.requires_grad = False
else:
encoder = load_encoder(cfg, rank)
projector = load_projector(cfg, rank)
decoder = load_decoder(cfg, rank)
cfg.load = load
model = DenoisingBlock(encoder, decoder, projector, rank, cfg.N,
cfg.batch_size, cfg.agg_enc_fxn, cfg.agg_enc_type)
# if cfg.load:
# load_model_field(cfg,rank,model,"")
if cfg.use_ddp:
if cfg.sync_batchnorm:
fxn = SyncBatchNorm.convert_sync_batchnorm
model = fxn(model, proc_group)
if cfg.use_apex:
model = apex_DDP(model)
else:
model = th_DDP(model,
device_ids=[rank],
find_unused_parameters=True)
return model