args["overwrite_opts"])
networks[net_name], optimizers["opt_{}".format(
net_name)] = utils.load_network(**net_kwargs[net_name])
# load the loss functions
loss_info = setup_kwargs_loss(args, pt2=False)
losses = {}
for k in loss_info:
losses[k] = model_utils.load_loss(loss_info[k]["name"],
loss_info[k]["kwargs"])
if not os.path.exists(args["save_dir"]):
os.makedirs(args["save_dir"])
print(args)
model = trainer_module.Model(data_provider=dp,
**networks,
**optimizers,
**losses,
**trainer_kwargs)
model.train()
#######
# DONE TRAINING REFERENCE MODEL
#######
embeddings_path = "{}/embeddings.pkl".format(args["save_dir"])
embeddings = model_utils.load_embeddings(embeddings_path, model.enc, dp)
model.load(opt.save_dir)
model.train()
#######
### DONE TRAINING REFERENCE MODEL
#######
#######
### TRAIN STRUCTURE MODEL
#######
embeddings_path = opt.save_dir + os.sep + 'embeddings.pkl'
if opt.skip_pt1:
embeddings = model_utils.load_embeddings(embeddings_path, None, dp, opt)
else:
embeddings = model_utils.load_embeddings(embeddings_path, model.enc, dp,
opt)
models = None
optimizers = None
dp.embeddings = embeddings
dp.channelInds = opt.channels_pt2
opt.save_dir = os.path.join(opt.save_parent, opt.struct_dir)
if not os.path.exists(opt.save_dir):
os.makedirs(opt.save_dir)
kwargs_model = setup_kwargs_model(opt)
logger.add((epoch, this_iter) + errors + (deltaT, ))
if model_utils.maybe_save(model, epoch, epoch_next, models, optimizers,
logger, zAll, dp, opt):
zAll = list()
#######
### DONE TRAINING REFERENCE MODEL
#######
#######
### TRAIN STRUCTURE MODEL
#######
embeddings_path = opt.save_dir + os.sep + 'embeddings.pkl'
embeddings = model_utils.load_embeddings(embeddings_path, models['enc'], dp,
opt)
models = None
optimizers = None
def get_ref(self, inds, train_or_test='train'):
inds = torch.LongTensor(inds)
return self.embeddings[train_or_test][inds]
dp.embeddings = embeddings
# do this thing to bind the get_ref method to the dataprovider object
import types
dp.get_ref = types.MethodType(get_ref, dp)