Class: DNN::Layers::GRUDense

Inherits:
Layer
  • Object
show all
Includes:
LayerNode
Defined in:
lib/dnn/core/layers/rnn_layers.rb

Instance Attribute Summary collapse

Attributes inherited from Layer

#input_shape

Instance Method Summary collapse

Methods included from LayerNode

#backward, #forward

Methods inherited from Layer

#build, #built?, #call, call, #clean, #forward, from_hash, #load_hash, #output_shape, #to_hash

Constructor Details

#initialize(weight, recurrent_weight, bias) ⇒ GRUDense

Returns a new instance of GRUDense.



377
378
379
380
381
382
383
384
385
 
# File 'lib/dnn/core/layers/rnn_layers.rb', line 377

def initialize(weight, recurrent_weight, bias)
  @weight = weight
  @recurrent_weight = recurrent_weight
  @bias = bias
  @update_sigmoid = Layers::Sigmoid.new
  @reset_sigmoid = Layers::Sigmoid.new
  @tanh = Layers::Tanh.new
  @trainable = true
end

Instance Attribute Details

#trainable ⇒ Object

Returns the value of attribute trainable.



375
376
377
 
# File 'lib/dnn/core/layers/rnn_layers.rb', line 375

def trainable
  @trainable
end

Instance Method Details

#backward_node(dh2) ⇒ Object 



410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
 
# File 'lib/dnn/core/layers/rnn_layers.rb', line 410

def backward_node(dh2)
  dtanh_h = @tanh.backward_node(dh2 * (1 - @update))
  dh = dh2 * @update

  if @trainable
    dweight_h = @x.transpose.dot(dtanh_h)
    dweight2_h = (@h * @reset).transpose.dot(dtanh_h)
    dbias_h = dtanh_h.sum(0) if @bias
  end
  dx = dtanh_h.dot(@weight_h.transpose)
  dh += dtanh_h.dot(@weight2_h.transpose) * @reset

  dreset = @reset_sigmoid.backward_node(dtanh_h.dot(@weight2_h.transpose) * @h)
  dupdate = @update_sigmoid.backward_node(dh2 * @h - dh2 * @tanh_h)
  da = Xumo::SFloat.hstack([dupdate, dreset])
  if @trainable
    dweight_a = @x.transpose.dot(da)
    dweight2_a = @h.transpose.dot(da)
    dbias_a = da.sum(0) if @bias
  end
  dx += da.dot(@weight_a.transpose)
  dh += da.dot(@weight2_a.transpose)

  if @trainable
    @weight.grad += Xumo::SFloat.hstack([dweight_a, dweight_h])
    @recurrent_weight.grad += Xumo::SFloat.hstack([dweight2_a, dweight2_h])
    @bias.grad += Xumo::SFloat.hstack([dbias_a, dbias_h]) if @bias
  end
  [dx, dh]
end

#forward_node(x, h) ⇒ Object 



387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
 
# File 'lib/dnn/core/layers/rnn_layers.rb', line 387

def forward_node(x, h)
  @x = x
  @h = h
  num_nodes = h.shape[1]
  @weight_a = @weight.data[true, 0...(num_nodes * 2)]
  @weight2_a = @recurrent_weight.data[true, 0...(num_nodes * 2)]
  a = x.dot(@weight_a) + h.dot(@weight2_a)
  a += @bias.data[0...(num_nodes * 2)] if @bias
  @update = @update_sigmoid.forward_node(a[true, 0...num_nodes])
  @reset = @reset_sigmoid.forward_node(a[true, num_nodes..-1])

  @weight_h = @weight.data[true, (num_nodes * 2)..-1]
  @weight2_h = @recurrent_weight.data[true, (num_nodes * 2)..-1]
  @tanh_h = if @bias
              bias_h = @bias.data[(num_nodes * 2)..-1]
              @tanh.forward_node(x.dot(@weight_h) + (h * @reset).dot(@weight2_h) + bias_h)
            else
              @tanh.forward_node(x.dot(@weight_h) + (h * @reset).dot(@weight2_h))
            end
  h2 = (1 - @update) * @tanh_h + @update * h
  h2
end