@@ -2964,6 +2964,94 @@ def space_to_depth(data, block_size, layout="NCHW"):
29642964 return _make .space_to_depth (data , block_size , layout )
29652965
29662966
2967+ def adaptive_max_pool1d (data , output_size = None , layout = "NCW" ):
2968+ r"""1D adaptive max pooling operator. This operator is experimental.
2969+
2970+ This operator takes data as input and does 1D max value calculation
2971+ across each window represented by W.
2972+
2973+
2974+ In the default case, where the data_layout is `NCW`
2975+ a data Tensor with shape `(batch_size, in_channels, width)`,
2976+ to produce an output Tensor with shape
2977+ (batch_size, in_channels, output_width).
2978+
2979+ The pooling kernel and stride sizes are automatically chosen for
2980+ desired output sizes.
2981+
2982+ For output_size:
2983+ If this argument is not provided, input height and width will be used
2984+ as output height and width.
2985+
2986+ If a single integer is provided for output_size, the output size is
2987+ (N x C x output_size) for any input (NCW).
2988+
2989+ Parameters
2990+ ----------
2991+ data : tvm.relay.Expr
2992+ The input data to the operator.
2993+
2994+ output_size : tuple of int. optional
2995+ Output height and width.
2996+
2997+ layout : str, optional
2998+ Layout of the input.
2999+
3000+ Returns
3001+ -------
3002+ result : tvm.relay.Expr
3003+ The computed result.
3004+ """
3005+ output_size = [] or output_size
3006+ if isinstance (output_size , int ):
3007+ output_size = [output_size ]
3008+ return _make .adaptive_max_pool1d (data , output_size , layout )
3009+
3010+
3011+ def adaptive_avg_pool1d (data , output_size = None , layout = "NCW" ):
3012+ r"""1D adaptive average pooling operator. This operator is experimental.
3013+
3014+ This operator takes data as input and does 1D average value calculation
3015+ across each window represented by W.
3016+
3017+
3018+ In the default case, where the data_layout is `NCW`
3019+ a data Tensor with shape `(batch_size, in_channels, width)`,
3020+ to produce an output Tensor with shape
3021+ (batch_size, in_channels, output_width).
3022+
3023+ The pooling kernel and stride sizes are automatically chosen for
3024+ desired output sizes.
3025+
3026+ For output_size:
3027+ If this argument is not provided, input height and width will be used
3028+ as output width.
3029+
3030+ If a single integer is provided for output_size, the output size is
3031+ (N x C x output_size) for any input (NCW).
3032+
3033+ Parameters
3034+ ----------
3035+ data : tvm.relay.Expr
3036+ The input data to the operator.
3037+
3038+ output_size : tuple of int. optional
3039+ Output height and width.
3040+
3041+ layout : str, optional
3042+ Layout of the input.
3043+
3044+ Returns
3045+ -------
3046+ result : tvm.relay.Expr
3047+ The computed result.
3048+ """
3049+ output_size = [] or output_size
3050+ if isinstance (output_size , int ):
3051+ output_size = [output_size ]
3052+ return _make .adaptive_avg_pool1d (data , output_size , layout )
3053+
3054+
29673055def adaptive_max_pool2d (data , output_size = None , layout = "NCHW" ):
29683056 r"""2D adaptive max pooling operator. This operator is experimental.
29693057
@@ -3142,6 +3230,71 @@ def adaptive_avg_pool3d(data, output_size=None, layout="NCDHW"):
31423230 return _make .adaptive_avg_pool3d (data , output_size , layout )
31433231
31443232
3233+ def global_max_pool1d (data , layout = "NCW" ):
3234+ r"""1D global maximum pooling operator.
3235+
3236+ This operator takes data as input and does 1D max value calculation
3237+ across each window represented by W.
3238+
3239+ In the default case, where the data_layout is `NCW`
3240+ a data Tensor with shape `(batch_size, in_channels, width)`,
3241+ to produce an output Tensor with the following rule:
3242+
3243+ with data of shape (b, c, w)
3244+ .. math::
3245+
3246+ \mbox{out}(b, c, 1) = \max_{n=0, \ldots, w} \mbox{data}(b, c, n)
3247+
3248+ Parameters
3249+ ----------
3250+ data : tvm.relay.Expr
3251+ The input data to the operator.
3252+
3253+ layout : str, optional
3254+ Layout of the input.
3255+
3256+ Returns
3257+ -------
3258+ result : tvm.relay.Expr
3259+ The computed result.
3260+ """
3261+ output_size = [1 ]
3262+ return _make .adaptive_max_pool1d (data , output_size , layout )
3263+
3264+
3265+ def global_avg_pool1d (data , layout = "NCW" ):
3266+ r"""1D global average pooling operator.
3267+
3268+ This operator takes data as input and does 1D average value calculation
3269+ across each window represented by W.
3270+
3271+ In the default case, where the data_layout is `NCW`
3272+ a data Tensor with shape `(batch_size, in_channels, width)`,
3273+ to produce an output Tensor with the following rule:
3274+
3275+ with data of shape (b, c, w)
3276+
3277+ .. math::
3278+
3279+ \mbox{out}(b, c, 1) = \frac{1}{w} \sum_{n=0}^{w-1} \mbox{data}(b, c, n)
3280+
3281+ Parameters
3282+ ----------
3283+ data : tvm.relay.Expr
3284+ The input data to the operator.
3285+
3286+ layout : str, optional
3287+ Layout of the input.
3288+
3289+ Returns
3290+ -------
3291+ result : tvm.relay.Expr
3292+ The computed result.
3293+ """
3294+ output_size = [1 ]
3295+ return _make .adaptive_avg_pool1d (data , output_size , layout )
3296+
3297+
31453298def global_max_pool3d (data , layout = "NCDHW" ):
31463299 r"""3D global maximum pooling operator.
31473300
0 commit comments