fix the bug that parameters are note transposed. #387
Conversation
lcy-seso
force-pushed
the
fix_transpose_in_hsigmoid_nce
branch
4 times, most recently
from
645f237
to
249aa15
Compare
lcy-seso
force-pushed
the
fix_transpose_in_hsigmoid_nce
branch
from
249aa15
to
f4cdd71
Compare
hsigmoid/README.md
Outdated
|
|
||
| ## 训练阶段 | ||
| 训练比较简单，直接运行``` python train.py ```。程序第一次运行会检测用户缓存文件夹中是否包含imikolov数据集，如果未包含，则自动下载。运行过程中，每100个iteration会打印模型训练信息，主要包含训练损失和测试损失，每个pass会保存一次模型。 | ||
|
|
||
| ## 预测阶段 | ||
| 预测时，直接运行``` python infer.py ```，程序会首先load模型，然后按照batch方式进行预测，并打印预测结果。预测阶段最重要的就是根据概率得到编码路径，然后遍历路径获取最终的预测类别，这部分逻辑如下： | ||
| 预测时，在命令行运行 `infer.py --model_path XX`，通过`model_path`指定训练好的模型所在的路径。程序会首先load模型，然后按照batch方式进行预测，并打印预测结果。预测阶段最重要的就是根据概率得到编码路径，然后遍历路径获取最终的预测类别，这部分逻辑如下： |
There was a problem hiding this comment.
python infer.py --model_path ...
nce_cost/README.md
Outdated
| input=hidden_layer, | ||
| param_attr=paddle.attr.Param(name="nce_w")) | ||
| ``` | ||
| 1. 在命令行运行 `infer.py --model_path XX`，通过`model_path`指定训练好的模型所在的路径。 |
There was a problem hiding this comment.
python infer.py --model_path ...
There was a problem hiding this comment.
The develop branch has been updated. If the parameter num_classes is not set, it is automatically determined by the size of the inputlabel https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/trainer_config_helpers/layers.py#L2344.
This may cause the problem if the size of label is not correctly set (If the size of label is not correctly set, PaddlePaddle will not throw an error.), maybe explicitly set the num_classes is better.
hsigmoid/network_conf.py
Outdated
| input=hidden_layer, | ||
| label=target_word, | ||
| num_classes=dict_size, |
There was a problem hiding this comment.
According to the code: https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/gserver/layers/HierarchicalSigmoidLayer.cpp#L27, maybe num_classes is necessary?
lcy-seso
force-pushed
the
fix_transpose_in_hsigmoid_nce
branch
2 times, most recently
from
b52c5c1
to
6d39971
Compare
lcy-seso
force-pushed
the
fix_transpose_in_hsigmoid_nce
branch
from
6d39971
to
406761e
Compare
fix the bug that when loading the parameter trained by
hsigmoidandNCElayer in inferring, the parameters are not transposed.