The pre-trained model 'model-y1-test2'(provided in Baidu netdisk) and the normal trained model have different 'fc1_weight' in model-symbol.json

The file model-symbol.json in two model(training the y1 network with the [recognition/arcface](https://github.com/deepinsight/insightface/tree/master/recognition/arcface) and pre-trained model 'model-y1-test2') are different in name 'fc1_weight'.

The pre-trained model 'model-y1-test2'(provided in Baidu netdisk) is called 'fc1_weight':

```json
{
"op": "null",
"name": "fc1_weight",
"attrs": {
"__lr_mult__": "1.0",
"__shape__": "(128, 512)",
"__wd_mult__": "10.0"
},
"inputs": []
},
```
while the normal trained model is called 'pre_fc1_weight':
```json
{
"op": "null",
"name": "pre_fc1_weight",
"attrs": {"num_hidden": "128"},
"inputs": []
},
```