Keras: Implementation of 2D RNN，GRU, LSTM

[ypxie]

with dropout support. Currently only runs on theano.

Inputs to the model are of size (Nsample, rowsize, colsize, input_dim)
https://github.com/shampool/Multi-Dimensional-RNN

Any suggestions are welcome.

model can be created like this:

def buildmodel(input_dim, output_dim):
    hiddenDim = 128
    model = Graph()
    model.add_input(name='input', input_shape=(None,None,input_dim), dtype= 'float')
    #model.add_node(TwoDimRecurrentBatchNormalization(), name='bn1', input='input')
    model.add_node(TwoDimMasking(),
                   name='masked', input='input')

    model.add_node(TwoDimGRU(hiddenDim,droprate =  [0.25],init = 'glorot_normal',inner_init='orthogonal',
                                    activation='tanh', return_sequences = True,
                                    go_backwards=[False,False]), name='upperleft', input='masked')

    model.add_node(TwoDimGRU(hiddenDim,droprate = [0.25],
                                    init = 'glorot_normal',inner_init='orthogonal',
                                    activation='tanh', return_sequences = True,
                                    go_backwards=[True,True]), 
                                    name='lowerright',input= 'upperleft')

    model.add_node(TwoDimTimeDistributedDense(output_dim = output_dim,,droprate =  [0.25],init = 'glorot_normal',
                                    activation='hard_sigmoid'), name = 'last_mask',
                                    input= 'lowerright')

    model.add_output(name='output_mask', input='last_mask')
    opt = SGD(lr =1E-1, decay = 1E-7, momentum = 0.9, nesterov = True)
    #opt = RMSprop(lr=0.0001, rho=0.9, epsilon=1e-6) # start with 0.0001 then with 0.001
    model.compile(loss={'output_mask': 'mse' }, optimizer=opt)

    return model

[yudf2010]

@ypxie I search for 2D GRU implementation on theano and find this post. Do you have implemented this? I find the link in the post is lost, can you give a new one? Thanks