def __init__(self,
model,
opt,
loss_func,
data,
metrics=None,
callbacks=None,
device=None):
"""
Args:
model: pytorch 模型
opt: 优化器
loss_func: 损失函数
data: DataBunch
metrics: 性能评价指标或评价指标列表
callbacks: callbacks对象或类，或者其列表
device: cpu or gpu device
"""
self.model, self.opt, self.data, self.loss_func = model, opt, data, loss_func
self.state = 'train' # 'train', 'val', 'test'
self.messages = {} # 存放需要在不同callbacks之间共享的信息
self.epoch = 0
self.best_loss = float('inf')
cbs = tbox.listify(callbacks)
# 添加一些必要的回调
cbs.append(cbks.Recorder(as_attr=True))
cbs += [
cbks.StatesCallback(),
cbks.CudaCallback(device=device),
cbks.ProgressBarCallback(),
cbks.AvgStatsCallback(metrics=metrics)
]
cb_list = []
for cb in cbs:
if isinstance(cb, cbks.Callback):
cb_obj = cb
else:
cb_obj = cb()
if cb_obj.as_attr:
setattr(self, cb_obj.name, cb_obj)
cb_obj.set_learner(self)
cb_list.append(cb_obj)
self.cbs = cb_list
def __init__(self, metrics, state):
self.metrics, self.state = tbox.listify(metrics), state
self.metric_names = ['loss'] + [m.__name__
for m in self.metrics] # 构造metrics 名称
def __init__(self, items):
self.items = tbox.listify(items)