Keras 模型导入: 支持的特性
鲜为人知的事实:DL4J的创始人,Skymind,在我们的团队中拥有前五名的Keras贡献者中的两个,使其成为继Keras的创始人Francois Chollet之后对Keras的最大贡献者。
虽然并非DL4J中的每个概念在Keras中都有等效的概念,反之亦然,但是许多关键概念可以匹配。将Keras模型导入DL4J是在我们的deeplearning4j-modelimport 模块中完成的。下面是当前支持的特性的综合列表。
- 层
- 损失
- 激活函数
- 初始化器
- 正则化器
- 约束
- 度量
- 优化器
层
将模型映射到DL4J层是在模型导入的层子模块中完成的。该项目的结构随意地反映了Keras的结构。
核心层
- Dense
- Activation
- Dropout
- Flatten
- Reshape
- Merge
- Permute
- RepeatVector
- Lambda
- ActivityRegularization
- Masking
- SpatialDropout1D
- SpatialDropout2D
- SpatialDropout3D
圈积层
- Conv1D
- Conv2D
- Conv3D
- AtrousConvolution1D
- AtrousConvolution2D
- SeparableConv1D
- SeparableConv2D
- Conv2DTranspose
- Conv3DTranspose
- Cropping1D
- Cropping2D
- Cropping3D
- UpSampling1D
- UpSampling2D
- UpSampling3D
- ZeroPadding1D
- ZeroPadding2D
- ZeroPadding3D
池化层
- MaxPooling1D
- MaxPooling2D
- MaxPooling3D
- AveragePooling1D
- AveragePooling2D
- AveragePooling3D
- GlobalMaxPooling1D
- GlobalMaxPooling2D
- GlobalMaxPooling3D
- GlobalAveragePooling1D
- GlobalAveragePooling2D
- GlobalAveragePooling3D
本地连接层
循环层
嵌入层
合并层
- Add / add
- Multiply / multiply
- Subtract / subtract
- Average / average
- Maximum / maximum
- Concatenate / concatenate
- Dot / dot
高级激活层
归一化层
噪声层
层包装器
- TimeDistributed
- Bidirectional
损失
- mean_squared_error
- mean_absolute_error
- mean_absolute_percentage_error
- mean_squared_logarithmic_error
- squared_hinge
- hinge
- categorical_hinge
- logcosh
- categorical_crossentropy
- sparse_categorical_crossentropy
- binary_crossentropy
- kullback_leibler_divergence
- poisson
- cosine_proximity
激活
- softmax
- elu
- selu
- softplus
- softsign
- relu
- tanh
- sigmoid
- hard_sigmoid
- linear
初始化器
- Zeros
- Ones
- Constant
- RandomNormal
- RandomUniform
- TruncatedNormal
- VarianceScaling
- Orthogonal
- Identity
- lecun_uniform
- lecun_normal
- glorot_normal
- glorot_uniform
- he_normal
- he_uniform
正则化器
- l1
- l2
- l1_l2
约束
- max_norm
- non_neg
- unit_norm
- min_max_norm
优化器
- SGD
- RMSprop
- Adagrad
- Adadelta
- Adam
- Adamax
- Nadam
- TFOptimizer