def __init__(self):
is_pair = True
class_labels = ['not_entailment', 'entailment']
metric = CompositeEvalMetric()
metric.add(Accuracy())
super(AXgTask, self).__init__(class_labels,
metric, is_pair, output_format="jsonl")
def __init__(self):
is_pair = True
class_labels = ['0', '1']
metric = CompositeEvalMetric()
metric.add(F1())
metric.add(Accuracy())
super(QQPTask, self).__init__(class_labels, metric, is_pair)
def __init__(self):
is_pair = True
class_labels = ['0', '1']
metric = CompositeEvalMetric()
metric.add(F1(average='micro'))
super(MultiRCTask, self).__init__(class_labels,
metric, is_pair, output_format="jsonl")
def train():
if pretrain_weight_path is not None:
net_args, net_auxs = load_params_from_file(pretrain_weight_path)
else:
net_args, net_auxs = None, None
to_model = osp.join(model_save_dir, '{}_ep'.format('FCN'))
mod = get_module()
opt = get_optimizer()
dataiter = CustomIter(data_lst=data_list,
dataset='pascal',
data_root=data_root,
batch_size=batch_size,
crop_h=input_h,
crop_w=input_w,
label_stride=8,
sampler='random')
custom_eval = CompositeEvalMetric()
custom_eval.add(Custom_Accuracy())
dataiter.reset()
mod.fit(
dataiter,
eval_metric=custom_eval,
batch_end_callback=mx.callback.Speedometer(batch_size, 1),
epoch_end_callback=mx.callback.do_checkpoint(to_model),
kvstore='local',
begin_epoch=from_epoch,
num_epoch=num_epochs,
optimizer=opt,
initializer=mx.init.Xavier(),
arg_params=net_args,
aux_params=net_auxs,
allow_missing=True,
)
def __init__(self):
is_pair = True
class_labels = ['0', '1']
metric = CompositeEvalMetric()
metric.add(F1())
metric.add(Accuracy())
super(ReCoRDTask, self).__init__(class_labels,
metric, is_pair, output_format="jsonl")
def __init__(self, *args, **kwargs): # passthrough arguments to TSVDataset
# (filename, field_separator=nlp.data.Splitter(','), num_discard_samples=1, field_indices=[2,1])
self.args = args
self.kwargs = kwargs
is_pair = False
class_labels = ['0', '1']
metric = CompositeEvalMetric()
metric.add(F1())
metric.add(Accuracy())
super(TSVClassificationTask, self).__init__(class_labels, metric,
is_pair)
dataset = nlp.data.TSVDataset(*self.args, **self.kwargs)
# do the split
train_sampler, val_sampler = get_split_samplers(dataset,
split_ratio=0.8)
self.trainset = SampledDataset(dataset, train_sampler)
self.valset = SampledDataset(dataset, val_sampler)
def get_metric():
"""Get metrics Accuracy and F1"""
metric = CompositeEvalMetric()
for child_metric in [Accuracy(), F1()]:
metric.add(child_metric)
return metric
t_modules.append(t_module)
s_module = KTModule(symbol=s_symbol,
context=devices,
logger=logger,
data_names=data_names,
data_shapes=data_shapes,
label_names=label_names,
label_shapes=label_shapes,
is_transfer=False,
teacher_module=t_modules)
# eval_metric = MultiMetric(loss_types=args.eval_metric)
eval_metric = CompositeEvalMetric()
eval_metric.add(
Accuracy(output_names=["softmax_output"],
label_names=["softmax_label"]))
eval_metric.add(Loss(output_names=["lmnn_output"], label_names=[]))
s_module.fit(train_data=train,
eval_metric=eval_metric,
kvstore=kv,
initializer=init,
optimizer=sgd,
num_epoch=args.num_epochs,
arg_params=s_arg_params,
aux_params=s_aux_params,
epoch_end_callback=epoch_end_callback,
batch_end_callback=batch_end_callback,
allow_missing=True)
def get_metric(cls):
"""Get metrics Accuracy and F1"""
metric = CompositeEvalMetric()
for child_metric in [Accuracy(), F1(average='micro')]:
metric.add(child_metric)
return metric