conv2d_backprop_filter

Description

Computes 2D deconvolution of the float32 type with the given 5HD data and 5HD out_backprop.

The shape of the data tensor is 5HD, that is, (N, C1, H, W, C0). The shape of the out_backprop tensor is 5HD, that is, (N, C1, H, W, C0).

Prototype

conv2d_backprop_filter(input_x, out_backprop, filter_sizes, para_dict)

Parameters

• input_x: a 5HD tensor of type float16, for the feature map.
• out_backprop: backprop of the 2D convolution. Currently, the float16 type is supported.
• filter_size: weight matrix size.
• para_dict: a dictionary for the key-value pairs, including the following keys:
  • strides: a list of the strides along the H and W directions of the feature map.
  • padding: a list for the padding lines along the H and W directions of the feature map.
  • dilations: a list of the dilations along the H and W directions of the filter.
  • groups: group for 2D convolution filter. Defaults to 1.
  • res_dtype: output data type.
  • kernel_name: operator name.

Returns

res_tensor: result tensor.

Restrictions

This API cannot be used in conjunction with other TBE DSL APIs.

Applicability

Atlas 200/300/500 Inference Product

Atlas Training Series Product

Example

from tbe import tvm
from tbe import dsl

out_backprop_shape = (1, 1, 7, 7, 16)
out_backprop_dtype = "float16"
fmap_shape = (1, 1, 7, 7, 16)
fmap_dtype = "float16"
filter_sizes = (16, 16, 1, 1)

out_backprop = tvm.placeholder(out_backprop_shape, name="out_backprop", dtype=out_backprop_dtype)
fmap = tvm.placeholder(fmap_shape, name="fmap", dtype=fmap_dtype)

strides = [1, 1]
padding = [0, 0, 0, 0]
dilations = [1, 1, 1, 1]
groups = 1
res_dtype = "float32"
kernel_name = "conv2d_backprop_filter_dx_1_1_7_7_16_dy_1_1_7_7_16_dw_16_16_1_1_s_1_1_p_SAME"

para_dict = {
    "strides": strides,
    "padding": padding,
    "dilations": dilations,
    "groups": groups,
    "res_dtype": res_dtype,
    "kernel_name": kernel_name
}

filter_backprop = dsl.conv2d_backprop_filter(
    input_x=fmap,
    out_backprop=out_backprop,
    filter_sizes=filter_sizes,
    para_dict=para_dict
)