Created by: guoshengCS

To support fine-tuning, add reset_optimizer, layers, weights arguements for model.load.

Currently, a feasible way for fine-tuning is given in YOLOv3 demo, which is implemented by integrating a backbone model backbone_model into a full model full_model and use full_model.backbone_model.load. However, this needs a way to ignore the saved optimizer states. Thus we add the reset_optimizer argument.

Additionally, add layers and weights arguments for partial loading. The usage is as follows:

guard = fluid.dygraph.guard() if True else null_guard()

inputs = [Input([None, in_size], "float32", name="x")]
labels = [Input([None, 1], "int64", name="label")]

class MyModel(Model):
    def __init__(self):
        super(MyModel, self).__init__()
        self._hid = Linear(in_size, hid_size, act="sigmoid")
        self._fc = Linear(hid_size, class_num, act="softmax")

    def forward(self, x):
        x = self._hid(x)
        return self._fc(x)

with guard:
    model = MyModel()
    optim = fluid.optimizer.SGD(0.001, parameter_list=model.parameters())
    model.prepare(optim, CrossEntropy(), inputs, labels)
    model.load("linear_dygraph_tmp/model", layers=model._hid))