我正在尝试为 R ( github ) 中的 keras 包实现自定义层。
我正在实现的层基于此处可用的 AttentionWithContext 层：gist
这是我的代码：
AttentionWithContext <- R6::R6Class("AttentionWithContext",
inherit = KerasLayer,
public = list(
W_regularizer = NULL,
b_regularizer = NULL,
u_regularizer = NULL,
W_constraint=NULL,
b_constraint=NULL,
u_constraint=NULL,
bias=NULL,
b=NULL,
W=NULL,
u=NULL,
supports_masking=NULL,
init=NULL,
name = NULL,
initialize = function(name = 'attention',
W_regularizer = NULL,
b_regularizer = NULL,
u_regularizer = NULL,
W_constraint=NULL,
b_constraint=NULL,
u_constraint=NULL,
bias=TRUE ) {
self$supports_masking = TRUE
self$init = keras::initializer_glorot_uniform()
self$W_regularizer = W_regularizer
self$b_regularizer = b_regularizer
self$u_regularizer = u_regularizer
self$W_constraint = W_constraint
self$b_constraint = b_constraint
self$u_constraint = u_constraint
self$bias = bias
self$name = name
},
build = function(input_shape) {
assertthat::assert_that(length(input_shape) == 3)
self$W = self$add_weight(shape = reticulate::tuple(input_shape[[3]],input_shape[[3]], NULL),
initializer = self$init,
name=stringr::str_interp('${self$name}_W'),
regularizer = self$W_regularizer,
constraint = self$W_constraint)
if (self$bias) {
self$b = self$add_weight(shape = reticulate::tuple(input_shape[[3]]),
initializer='zero',
name = stringr::str_interp('${self$name}_b'),
regularizer = self$b_regularizer,
constraint = self$b_constraint)
}
self$u = self$add_weight(shape = reticulate::tuple(input_shape[[3]]),
initializer=self$init,
name = stringr::str_interp('${self$name}_u'),
regularizer = self$u_regularizer,
constraint = self$u_constraint)
},
compute_mask = function(input, input_mask=NULL) {
return(NULL)
},
call = function(x, mask = NULL) {
uit = keras::k_squeeze(keras::k_dot(x, keras::k_expand_dims(self$W)), axis=-1)
if (self$bias) {
uit = uit + self$b
}
uit = keras::k_tanh(uit)
ait = keras::k_dot(uit, self$u)
a = keras::k_exp(ait)
if (!is.null(mask)) {
a = a * keras::k_cast(mask, keras::k_floatx())
}
a = a/keras::k_cast(keras::k_sum(a, axis = 1, keepdims = TRUE) + keras::k_epsilon(), keras::k_floatx())
weighted_input = x * keras::k_expand_dims(a)
keras::k_sum(weighted_input, axis=1)
},
compute_output_shape = function(input_shape) {
list(input_shape[[1]], input_shape[[3]])
}
)
)
# define layer wrapper function
layer_attention_with_context <- function(object, W_regularizer = NULL,
b_regularizer = NULL,
u_regularizer = NULL,
W_constraint=NULL,
b_constraint=NULL,
u_constraint=NULL,
bias=TRUE,
name = 'attention_with_context') {
create_layer(AttentionWithContext, object, list(W_regularizer = W_regularizer,
b_regularizer = b_regularizer,
u_regularizer = u_regularizer,
W_constraint= W_constraint,
b_constraint=b_constraint,
u_constraint=u_constraint,
bias=bias,
name = name
))
}
# Example
model <- keras_model_sequential()
model %>%
layer_embedding(input_dim = 20000,
output_dim = 128,
input_length = 30) %>%
layer_lstm(64, return_sequences = TRUE) %>%
layer_attention_with_context() %>%
time_distributed(layer_dense(units=10))
当我运行它时，我收到一条神秘的错误消息：
Error in py_call_impl(callable, dots$args, dots$keywords) :
RuntimeError: Evaluation error: TypeError: unsupported operand type(s) for *: 'NoneType' and 'int'.
我试图探索这个错误，我认为它可能来自这一行：
reticulate::tuple(input_shape[[3]],input_shape[[3]], NULL)
在原始代码中，在 python 中，我们可以看到：
(input_shape[-1], input_shape[-1],)
我找不到在 R 中创建这种结构的方法。
有任何想法吗 ？