深度學習(框架)預設值列表
阿新 • • 發佈：2018-12-20
deeplearning-default-value
Functions
| Object names in Chainer | Parameter names in Chainer | Chainer | PyTorch | TensorFlow |
|---|---|---|---|---|
| dropout | ratio | 0.5 (drop ratio) | 0.5 (drop ratio) | Required (keep ratio) |
Links
| Object names in Chainer | Parameter names in Chainer | Chainer | PyTorch | TensorFlow |
|---|---|---|---|---|
| Link | initilizer | LecunNormal | Each link has a different initializer. | GlorotUniform |
Initializers
- PyTorch initializers are defined in each layer implementation, so skip it.
| Object names in Chainer | Parameter names in Chainer | Chainer | TensorFlow |
|---|---|---|---|
| GlorotUniform | scale | 1.0 | No args (1.0) |
| LecunNormal | scale | 1.0 | No args (1.0) |
| Normal | scale | 0.05 | 1.0 or 0.05 |
| Uniform | scale | 0.05 | min, max = (0.0, 1.0) or (-0.05, 0.05) |
Optimizers
| Object names in Chainer | Parameter names in Chainer | Chainer | PyTorch | TensorFlow |
|---|---|---|---|---|
| SGD | lr | 0.01 | Required | 0.01 |
| Adam | alpha | 0.001 | 0.001 | 0.001 |
| beta1 | 0.9 | 0.9 | 0.9 | |
| beta2 | 0.999 | 0.999 | 0.999 | |
| eps | 1e-08 | 1e-08 | 1e-08 | |
| eta | 1.0 | No args | No args | |
| weight_decay | 0.0 | 0.0 | No args | |
| amsgrad | False | False | No args |
Normalization methods
| Object names in Chainer | Parameter names in Chainer | Chainer | PyTorch | TensorFlow |
|---|---|---|---|---|
| BatchNormalization | eps | 2e-05 | 1e-05 | 1e-03 |
| LayerNormalization | eps | 1e-06 | 1e-05 | No args (1e-12) |
| GroupNormalization | eps | 1e-05 | 1e-05 | 1e-06 |
| groups | Required | Required | 32 | |
| local_response_normalization | n | 5 | Required | 5 |
| k | 2 | 1 | 1 | |
| alpha | 1e-04 | 1e-04 | 1 | |
| beta | 0.75 | 0.75 | 0.5 |
Rules
Kinds of symbols
- None: default value is None.
- x: No implementations.
- No args: Magic number. We can not specify the value.
- Required: No default value. We should specify the value.
轉載自：https://github.com/keisuke-umezawa/deeplearning-default-value