[docs] tutorial/layers: clean up sample markdown

This commit is contained in:
Jonathan L Long 2014-09-06 21:42:35 -07:00
Родитель 15456286da
Коммит bd13f32123
1 изменённых файлов: 49 добавлений и 49 удалений

Просмотреть файл

@ -201,12 +201,12 @@ In general, activation / Neuron layers are element-wise operators, taking one bo
- `negative_slope` [default 0]: specifies whether to leak the negative part by multiplying it with the slope value rather than setting it to 0.
* Sample (as seen in `./examples/imagenet/imagenet_train_val.prototxt`)
layers {
name: "relu1"
type: RELU
bottom: "conv1"
top: "conv1"
}
layers {
name: "relu1"
type: RELU
bottom: "conv1"
top: "conv1"
}
Given an input value x, The `RELU` layer computes the output as x if x > 0 and negative_slope * x if x <= 0. When the negative slope parameter is not set, it is equivalent to the standard ReLU function of taking max(x, 0). It also supports in-place computation, meaning that the bottom and the top blob could be the same to preserve memory consumption.
@ -217,12 +217,12 @@ Given an input value x, The `RELU` layer computes the output as x if x > 0 and n
* CUDA GPU implementation: `./src/caffe/layers/sigmoid_layer.cu`
* Sample (as seen in `./examples/imagenet/mnist_autoencoder.prototxt`)
layers {
name: "encode1neuron"
bottom: "encode1"
top: "encode1neuron"
type: SIGMOID
}
layers {
name: "encode1neuron"
bottom: "encode1"
top: "encode1neuron"
type: SIGMOID
}
The `SIGMOID` layer computes the output as sigmoid(x) for each input element x.
@ -233,12 +233,12 @@ The `SIGMOID` layer computes the output as sigmoid(x) for each input element x.
* CUDA GPU implementation: `./src/caffe/layers/tanh_layer.cu`
* Sample
layers {
name: "layer"
bottom: "in"
top: "out"
type: TANH
}
layers {
name: "layer"
bottom: "in"
top: "out"
type: TANH
}
The `TANH` layer computes the output as tanh(x) for each input element x.
@ -249,12 +249,12 @@ The `TANH` layer computes the output as tanh(x) for each input element x.
* CUDA GPU implementation: `./src/caffe/layers/absval_layer.cu`
* Sample
layers {
name: "layer"
bottom: "in"
top: "out"
type: ABSVAL
}
layers {
name: "layer"
bottom: "in"
top: "out"
type: ABSVAL
}
The `ABSVAL` layer computes the output as abs(x) for each input element x.
@ -270,17 +270,17 @@ The `ABSVAL` layer computes the output as abs(x) for each input element x.
- `shift` [default 0]
* Sample
layers {
name: "layer"
bottom: "in"
top: "out"
type: POWER
power_param {
power: 1
scale: 1
shift: 0
}
layers {
name: "layer"
bottom: "in"
top: "out"
type: POWER
power_param {
power: 1
scale: 1
shift: 0
}
}
The `POWER` layer computes the output as (shift + scale * x) ^ power for each input element x.
@ -291,12 +291,12 @@ The `POWER` layer computes the output as (shift + scale * x) ^ power for each in
* CUDA GPU implementation: `./src/caffe/layers/bnll_layer.cu`
* Sample
layers {
name: "layer"
bottom: "in"
top: "out"
type: BNLL
}
layers {
name: "layer"
bottom: "in"
top: "out"
type: BNLL
}
The `BNLL` (binomial normal log likelihood) layer computes the output as log(1 + exp(x)) for each input element x.
@ -399,16 +399,16 @@ The `FLATTEN` layer is a utility layer that flattens an input of shape `n * c *
- if `concat_dim = 1`: `n_1 * (c_1 + c_2 + ... + c_K) * h * w`, and all input `n_i` should be the same.
* Sample
layers {
name: "concat"
bottom: "in1"
bottom: "in2"
top: "out"
type: CONCAT
concat_param {
concat_dim: 1
}
layers {
name: "concat"
bottom: "in1"
bottom: "in2"
top: "out"
type: CONCAT
concat_param {
concat_dim: 1
}
}
The `CONCAT` layer is a utility layer that concatenates its multiple input blobs to one single output blob. Currently, the layer supports concatenation along num or channels only.