FilesLast commitLast update
2 months ago
3 months ago
30 days ago
26 days ago
30 days ago
19 days ago
30 days ago
26 days ago
18 days ago
README

MaxPool3DGrad

English | 简体中文

Applicable Products

Product Support
Ascend 950PR&950DT products √
Atlas A3 products ×
Atlas A2 products ×
Atlas 200I/500 A2 inference products ×
Atlas inference products ×
Atlas training products ×

Function

Operator function: Performs backpropagation after the forward max pooling, backfills the gradient at the coordinates of the maximum value of each window, and accumulates the gradients at the same coordinates.

Formula for deriving the forward output shape

The output shape of the forward MaxPool3D (that is, the shape of orig_y and grads) is jointly determined by the input shape, ksize, strides, pads, and the padding mode. Taking the NCDHW format as an example, the shape of the input orig_x is (N,C,Din,Hin,Win)(N, C, D_{in}, H_{in}, W_{in}), and the forward output shape is (N,C,Dout,Hout,Wout)(N, C, D_{out}, H_{out}, W_{out}). The formulas for computing the dimensions are as follows (ceil_mode defaults to False, rounding down):

CALCULATED mode:

Dout=⌊Din+2×pads[0]−(kd−1)−1sd⌋+1D_{out} = \left\lfloor \frac{D_{in} + 2 \times pads[0] - (k_d - 1) - 1}{s_d} \right\rfloor + 1

Hout=⌊Hin+2×pads[2]−(kh−1)−1sh⌋+1H_{out} = \left\lfloor \frac{H_{in} + 2 \times pads[2] - (k_h - 1) - 1}{s_h} \right\rfloor + 1

Wout=⌊Win+2×pads[4]−(kw−1)−1sw⌋+1W_{out} = \left\lfloor \frac{W_{in} + 2 \times pads[4] - (k_w - 1) - 1}{s_w} \right\rfloor + 1

VALID mode:

It is equivalent to the CALCULATED mode with all pads being 0, that is:

Dout=⌊Din−(kd−1)−1sd⌋+1D_{out} = \lfloor{\frac{D_{in} - {(k_d - 1) - 1}}{s_d}}\rfloor + 1

Hout=⌊Hin−(kh−1)−1sh⌋+1H_{out} = \lfloor{\frac{H_{in} - {(k_h - 1) - 1}}{s_h}}\rfloor + 1

Wout=⌊Win−(kw−1)−1sw⌋+1W_{out} = \lfloor{\frac{W_{in} - {(k_w - 1) - 1}}{s_w}}\rfloor + 1

SAME mode:

Dout=⌊Din+sd−1sd⌋D_{out} = \lfloor{\frac{D_{in} + {s_d} - 1}{s_d}}\rfloor

Hout=⌊Hin+sh−1sh⌋H_{out} = \lfloor{\frac{H_{in} + {s_h} - 1}{s_h}}\rfloor

Wout=⌊Win+sw−1sw⌋W_{out} = \lfloor{\frac{W_{in} + {s_w} - 1}{s_w}}\rfloor

Output shape:

The output y of the backpropagation has exactly the same shape as the input orig_x:

y.shape=orig_x.shape=(N,C,Din,Hin,Win)y.shape = orig\_x.shape = (N, C, D_{in}, H_{in}, W_{in})

Parameter description:

  • kd,kh,kwk_d, k_h, k_w: window sizes of ksize in the D, H, and W dimensions
  • sd,sh,sws_d, s_h, s_w: strides in the D, H, and W dimensions

Parameters

Parameter name Input/Output/Attribute Description Data type Data format
orig_x Input Input parameter for the MaxPool3DGrad computation, representing the forward input tensor. FLOAT, FLOAT16, BFLOAT16 NCDHW, NDHWC
orig_y Input Index positions of the maximum elements in the forward input. The data format must be the same as that of `orig_x`, and the shape must be the same as that of `grads`. FLOAT, FLOAT16, BFLOAT16 NCDHW, NDHWC
grads Input Input parameter for the MaxPool3DGrad computation, representing the gradient of the current node. FLOAT, FLOAT16, BFLOAT16 NCDHW, NDHWC
ksize Attribute Window size of the max pooling. INT64 -
strides Attribute Stride of the pooling operation. INT64 -
padding Optional attribute Pad mode. SAME, VALID, and CALCULATED are supported. The padded data is -inf, and the default value is SAME. - -
pads Attribute Takes effect when the pad mode is CALCULATED. It specifies the padding amounts at the front and back in the D/H/W dimensions, in the format of [d_front, d_back, h_front, h_back, w_front, w_back]. INT64 -
data_format Optional attribute Supported data formats. - NCDHW and NDHWC are supported
y Output Output parameter of the MaxPool3DGrad computation. The shape and data format must be the same as those of `orig_x`. FLOAT, FLOAT16, BFLOAT16 NCDHW, NDHWC

Constraints

  • ksize: The array length must be 5, and the values corresponding to the N and C dimensions must be 1, that is, (1, 1, k_d, k_h, k_w).
  • strides: The array length must be 5, and the values corresponding to the N and C dimensions must be 1, that is, (1, 1, s_d, s_h, s_w).
  • pads: It takes effect only when padding="CALCULATED". The array length must be 6, in the format of [d_front, d_back, h_front, h_back, w_front, w_back], and each value must be >= 0. The front pads (pads[0], pads[2], and pads[4]) must satisfy <= the corresponding dimension kernel_size / 2.
  • The shapes of orig_y and grads must be the same and must match the output shape of the forward MaxPool3D.
  • The shape of y must be the same as the shape of orig_x.

Invocation Description

Invocation method Sample code Description
Graph-mode API test_geir_max_pool3d_grad Invoke the MaxPool3DGrad operator by building a graph through the IR MaxPool3DGrad.