def __init__(self,
bert: BERTLM,
optimizer: Optimizer,
with_cuda: bool = True,
log_freq: int = 10):
assert paddle.is_compiled_with_cuda, 'the version of padddle is not compiled with the cuda'
assert with_cuda is True, 'warning: do you want to train bert using cpu? are you kidding me?'
self.model = bert
self.opt = optimizer
self.criterion_sentence = nn.NLLLoss()
self.criterion_tokens = nn.NLLLoss(ignore_index=0)
self.log_freq = log_freq
list_n_elemnets = []
for p in self.model.parameters():
n = 1
for i in p.shape:
n *= i
list_n_elemnets.append(n)
print('Total Parameters:', sum(list_n_elemnets))
def __init__(self,
feature_dim: int,
n_classes: int,
eps: float = 1e-5,
margin: float = 0.3,
scale: float = 30.0):
super(AMSoftmaxLoss, self).__init__()
self.w = paddle.create_parameter((feature_dim, n_classes), 'float32')
self.eps = eps
self.scale = scale
self.margin = margin
self.nll_loss = nn.NLLLoss()
self.n_classes = n_classes
def __init__(self,
margin=0.0,
scale=1.0,
feature_dim=256,
n_classes=1000,
easy_margin=False):
super(AdditiveAngularMargin, self).__init__(margin, scale)
self.easy_margin = easy_margin
self.w = paddle.create_parameter((feature_dim, n_classes), 'float32')
self.cos_m = math.cos(self.margin)
self.sin_m = math.sin(self.margin)
self.th = math.cos(math.pi - self.margin)
self.mm = math.sin(math.pi - self.margin) * self.margin
self.nll_loss = nn.NLLLoss()
self.n_classes = n_classes
def loss(self, logit_dict, label_dict, loss_func_dict=None):
if loss_func_dict is None:
loss_func_dict = defaultdict(list)
loss_func_dict['glance'].append(nn.NLLLoss())
loss_func_dict['focus'].append(MRSD())
loss_func_dict['cm'].append(MRSD())
loss_func_dict['err'].append(paddleseg.models.MSELoss())
loss_func_dict['refine'].append(paddleseg.models.L1Loss())
loss = {}
# glance loss computation
# get glance label
glance_label = F.interpolate(label_dict['trimap'],
logit_dict['glance'].shape[2:],
mode='nearest',
align_corners=False)
glance_label_trans = (glance_label == 128).astype('int64')
glance_label_bg = (glance_label == 0).astype('int64')
glance_label = glance_label_trans + glance_label_bg * 2
loss_glance = loss_func_dict['glance'][0](
paddle.log(logit_dict['glance'] + 1e-6), glance_label.squeeze(1))
loss['glance'] = loss_glance
# focus loss computation
focus_label = F.interpolate(label_dict['alpha'],
logit_dict['focus'].shape[2:],
mode='bilinear',
align_corners=False)
loss_focus = loss_func_dict['focus'][0](logit_dict['focus'],
focus_label,
glance_label_trans)
loss['focus'] = loss_focus
# collaborative matting loss
loss_cm_func = loss_func_dict['cm']
# fusion_sigmoid loss
loss_cm = loss_cm_func[0](logit_dict['fusion'], focus_label)
loss['cm'] = loss_cm
# error loss
err = F.interpolate(logit_dict['error'],
label_dict['alpha'].shape[2:],
mode='bilinear',
align_corners=False)
err_label = (F.interpolate(logit_dict['fusion'],
label_dict['alpha'].shape[2:],
mode='bilinear',
align_corners=False) -
label_dict['alpha']).abs()
loss_err = loss_func_dict['err'][0](err, err_label)
loss['err'] = loss_err
loss_all = 0.25 * loss_glance + 0.25 * loss_focus + 0.25 * loss_cm + loss_err
# refine loss
if self.if_refine:
loss_refine = loss_func_dict['refine'][0](logit_dict['refine'],
label_dict['alpha'])
loss['refine'] = loss_refine
loss_all = loss_all + loss_refine
loss['all'] = loss_all
return loss
def __init__(self, s=20.0, eps=1e-8):
super(ProtoTypical, self).__init__()
self.nll_loss = nn.NLLLoss()
self.eps = eps
self.s = s
def forward(self, logit, label, semantic_weights=None):
temp = F.log_softmax(logit, axis=1)
loss_func = nn.NLLLoss(ignore_index=self.ignore_index,
weight=paddle.to_tensor(self.weights))
loss = loss_func(temp, label)
return loss
def main(args):
if args.seed is not None:
paddle.fluid.Program.random_seed = args.seed
np.random.seed(args.seed)
if args.gpu < 0:
device = "cpu"
else:
device = "cuda:{args.gpu}"
# Load dataset
# data = load_data(device, args)
# g, labels, num_classes, train_nid, val_nid, test_nid = data
labels = np.load("./data/lables.npy")
num_classes = np.load("./data/num_classes.npy")
train_nid = np.load("./data/train_nid.npy")
val_nid = np.load("./data/val_nid.npy")
test_nid = np.load("./data/test_nid.npy")
evaluator = get_evaluator(args.dataset)
# Preprocess neighbor-averaged features over sampled relation subgraphs
rel_subsets = []
with paddle.no_grad():
feats = []
for i in range(args.R + 1):
#数据集请自行在OGB官网下载，并按照官网教程生产训练集，或者在AiStudio上查询data88697
feature = np.load(f'../data/data88697/feat{i}.npy')
feats.append(paddle.to_tensor(feature))
# feats = preprocess_features(g, rel_subsets, args, device)
print("Done preprocessing")
# labels = labels.to(device)
# Release the graph since we are not going to use it later
g = None
# Set up logging
logging.basicConfig(format='[%(levelname)s] %(message)s',
level=logging.INFO)
logging.info(str(args))
_, num_feats, in_feats = feats[0].shape
logging.info("new input size: {} {}".format(num_feats, in_feats))
# Create model
num_hops = args.R + 1 # include self feature hop 0
model = nn.Sequential(
WeightedAggregator(num_feats, in_feats, num_hops),
SIGN(in_feats, args.num_hidden, num_classes, num_hops, args.ff_layer,
args.dropout, args.input_dropout))
if len(labels.shape) == 1:
# single label multi-class
loss_fcn = nn.NLLLoss()
else:
# multi-label multi-class
loss_fcn = nn.KLDivLoss(reduction='batchmean')
print('!' * 100)
optimizer = paddle.optimizer.Adam(parameters=model.parameters(),
weight_decay=args.weight_decay)
# optimizer = paddle.optimizer.Adam(parameters=model.parameters())
# Start training
best_epoch = 0
best_val = 0
f = open('log.txt', 'w+')
for epoch in range(1, args.num_epochs + 1):
start = time.time()
print(epoch)
train(model, feats, labels, train_nid, loss_fcn, optimizer,
args.batch_size)
if epoch % args.eval_every == 0:
with paddle.no_grad():
train_res, val_res, test_res = test(model, feats, labels,
train_nid, val_nid,
test_nid, evaluator,
args.eval_batch_size)
end = time.time()
val_acc = val_res[0]
log = "Epoch {}, Times(s): {:.4f}".format(epoch, end - start)
log += ", Accuracy: Train {:.4f}, Val {:.4f}".format(
train_res[0], val_res[0])
log += f", best_acc:{best_val}"
logging.info(log)
print(log, file=f, flush=True)
if val_acc > best_val:
best_val = val_acc
best_epoch = epoch
f.close()
logging.info("Best Epoch {}, Val {:.4f}".format(best_epoch, best_val))