ops.py 文件源码

def softmax_and_sample(logits):
    old_shape = logits.shape
    flattened_logits = logits.reshape((-1, logits.shape[logits.ndim-1]))
    samples = T.cast(
        srng.multinomial(pvals=T.nnet.softmax(flattened_logits)),
        theano.config.floatX
    ).reshape(old_shape)
    return T.argmax(samples, axis=samples.ndim-1)

# TODO: Have a look at this benchmark:
#       https://github.com/MaximumEntropy/cudnn_rnn_theano_benchmarks