ConvTranspose3d — PyTorch 2.7 documentation (original) (raw)

class torch.ao.nn.quantized.ConvTranspose3d(in_channels, out_channels, kernel_size, stride=1, padding=0, output_padding=0, groups=1, bias=True, dilation=1, padding_mode='zeros', device=None, dtype=None)[source][source]¶

Applies a 3D transposed convolution operator over an input image composed of several input planes. For details on input arguments, parameters, and implementation seeConvTranspose3d.

Note

Currently only the FBGEMM engine is implemented. Please, set the torch.backends.quantized.engine = ‘fbgemm’

For special notes, please, see Conv3d

Variables

• weight (Tensor) – packed tensor derived from the learnable weight parameter.
• scale (Tensor) – scalar for the output scale
• zero_point (Tensor) – scalar for the output zero point

See ConvTranspose3d for other attributes.

Examples:

torch.backends.quantized.engine = 'fbgemm' from torch.ao.nn import quantized as nnq

With cubic kernels and equal stride

m = nnq.ConvTranspose3d(16, 33, 3, stride=2)

non-cubic kernels and unequal stride and with padding

m = nnq.ConvTranspose3d(16, 33, (3, 3, 5), stride=(2, 1, 1), padding=(4, 2, 2)) input = torch.randn(20, 16, 50, 100, 100) q_input = torch.quantize_per_tensor(input, scale=1.0, zero_point=0, dtype=torch.quint8) output = m(q_input)

exact output size can be also specified as an argument

input = torch.randn(1, 16, 12, 12, 12) q_input = torch.quantize_per_tensor(input, scale=1.0, zero_point=0, dtype=torch.quint8) downsample = nnq.Conv3d(16, 16, 3, stride=2, padding=1) upsample = nnq.ConvTranspose3d(16, 16, 3, stride=2, padding=1) h = downsample(q_input) h.size() torch.Size([1, 16, 6, 6, 6]) output = upsample(h, output_size=input.size()) output.size() torch.Size([1, 16, 12, 12, 12])