Paddlepaddle 19 dynamically modify the last layer of the model

  During transfer learning , You usually need to modify the last layer of the model , Redefine as needed . however , Transfer learning needs comparative experiments in many models , And define the last one layer when , Need to know layer Of name, and layer The input data is shape, That is, we need to be right model Of layer Traversal .

1、 Traverse model

Through the following code, you can traverse model

import paddle
#print(' Built in model of propeller frame ：', paddle.vision.models.__all__)
model=paddle.vision.resnet18()
layer_list=[(name,atom_layer) for name,atom_layer in model.named_sublayers()]
#print(model)

feature_layer_name,feature_layer=layer_list[-1]
feature_layer_shape=feature_layer.weight.shape
feature_layer_type=feature_layer.__class__
print("feature layer:",feature_layer_name,feature_layer_shape,feature_layer_type)

  The code execution output is as follows ：

feature layer: fc [512, 1000] <class 'paddle.nn.layer.common.Linear'>

2、 Dynamically modify the last layer of the model

Dynamically modify the last layer of the model , Support semantic segmentation model and image classification model . But here's the thing , The last layer of the model must be layer object （ Such as resnet series ,densenet series ,mobilenet system