opt.py 文件源码

def local_abstractconv3d_gradweight_gemm(node):
    if theano.config.cxx == "" or not theano.config.blas.ldflags:
        return
    if not isinstance(node.op, AbstractConv3d_gradWeights):
        return None
    img, topgrad, shape = node.inputs
    if not isinstance(img.type, TensorType) or \
       not isinstance(topgrad.type, TensorType):
        return None

    rval = Corr3dMM_gradWeights(border_mode=node.op.border_mode,
                                subsample=node.op.subsample,
                                filter_dilation=node.op.filter_dilation)(img, topgrad, shape)
    copy_stack_trace(node.outputs[0], rval)

    # need to flip the kernel if necessary
    if node.op.filter_flip:
        rval = rval[:, :, ::-1, ::-1, ::-1]
    rval = theano.tensor.patternbroadcast(rval, node.outputs[0].broadcastable)
    copy_stack_trace(node.outputs[0], rval)

    return [rval]