check_shape_forward Class — pytorch Architecture

Source Code

aten/src/ATen/native/Convolution.cpp lines 633–723

template <typename T>
static void check_shape_forward(const at::Tensor& input,
                                const c10::ArrayRef<T>& weight_sizes, const at::Tensor& bias,
                                const ConvParams<T>& params) {
  int64_t k = input.ndimension();
  int64_t weight_dim = weight_sizes.size();
  auto groups = params.groups;
  const auto& padding = params.padding;
  const auto& dilation = params.dilation;
  bool transposed = params.transposed;

  TORCH_CHECK(!params.is_padding_neg(), "negative padding is not supported");
  TORCH_CHECK(!params.is_output_padding_neg(), "negative output_padding is not supported");
  TORCH_CHECK(!params.is_stride_nonpos(), "non-positive stride is not supported");
  TORCH_CHECK(!params.is_dilation_neg(), "dilation should be greater than zero");
  TORCH_CHECK(groups > 0, "expected groups to be greater than 0, but got groups=", groups);

  TORCH_CHECK(weight_dim == k,
           "Expected ", weight_dim, "-dimensional input for ", weight_dim,
           "-dimensional weight ", weight_sizes, ", but got ", k, "-dimensional input of size ",
           at::symint::sizes<T>(input), " instead");
  TORCH_CHECK(weight_sizes[0] >= groups,
           "Given groups=", groups, ", expected weight to be at least ", groups,
           " at dimension 0, but got weight of size ", weight_sizes, " instead");
  TORCH_CHECK(weight_sizes[0] % groups == 0,
           "Given groups=", groups, ", expected weight to be divisible by ",
           groups, " at dimension 0, but got weight of size [", weight_sizes,
           "] instead");

  if (!transposed) {
    std::vector<T> input_shape;
    std::vector<T> kernel_shape;
    bool kernel_size_correct = true;

    TORCH_CHECK(at::symint::size<T>(input, 1) == (weight_sizes[1] * groups),
                "Given groups=", groups, ", weight of size ", weight_sizes,
                ", expected input", at::symint::sizes<T>(input), " to have ",
                (weight_sizes[1] * groups), " channels, but got ", at::symint::size<T>(input, 1),
                " channels instead");

    TORCH_CHECK(!bias.defined() || (bias.ndimension() == 1 && at::symint::size<T>(bias, 0) == weight_sizes[0]),
             "Given weight of size ", weight_sizes,
             ", expected bias to be 1-dimensional with ", weight_sizes[0], " elements",
             ", but got bias of size ", at::symint::sizes<T>(bias), " instead");

    for (const auto i : c10::irange(2, k)) {
      // T could be int64_t or SymInt, Specialized numeric_limts<SymInt> in c10/core/SymInt.h
      TORCH_CHECK(padding[i-2] <= (std::numeric_limits<T>::max() - padding[i-2]),
                  "Given padding=", padding[i-2], " at dimension ", i-2, " , expected padding to be at most ",
                  (std::numeric_limits<T>::max() / 2));
      input_shape.push_back(at::symint::size<T>(input, i) + 2 * padding[i-2]);
      // log new kernel size considering dilation
      kernel_shape.push_back(dilation[i-2] * (weight_sizes[i]-1) + 1);
      if (input_shape.back() < kernel_shape.back()) {
        kernel_size_correct = false;
      }
    }

    TORCH_CHECK(input_shape.size() == kernel_shape.size(), "Inconsistent shape between Input and Kernel");

    if (!kernel_size_correct) {
      // If kernel size is incorrect
      std::ostringstream input_ss;
      std::ostringstream kernel_ss;
      std::string separator;

      for (int i = 0, len = input_shape.size(); i < len; ++i) {
        input_ss << separator << input_shape[i];
        kernel_ss << separator << kernel_shape[i];
        separator = " x ";
      }

      TORCH_CHECK(false, "Calculated padded input size per channel: (", input_ss.str(), "). "
               "Kernel size: (", kernel_ss.str(), "). Kernel size can't be greater than actual input size");
    }
  } else { // transposed
    for (const auto i : c10::irange(2, k)) {
      TORCH_CHECK(padding[i-2] <= (std::numeric_limits<T>::max() - padding[i-2]),
                  "Given padding=", padding[i-2], " at dimension ", i-2, " , expected padding to be at most ",
                  (std::numeric_limits<T>::max() / 2));
    }
    TORCH_CHECK(at::symint::size<T>(input, 1) == weight_sizes[0],
             "Given transposed=", transposed, ", weight of size ", weight_sizes,
             ", expected input", at::symint::sizes<T>(input), " to have ", weight_sizes[0],
             " channels, but got ", at::symint::size<T>(input, 1), " channels instead");
    TORCH_CHECK(!bias.defined() || (bias.ndimension() == 1 && at::symint::size<T>(bias, 0) == weight_sizes[1] * groups),
             "Given transposed=", transposed, ", weight of size ", weight_sizes,
             ", expected bias to be 1-dimensional with ", weight_sizes[1] * groups, " elements",
             ", but got bias of size ", at::symint::sizes<T>(bias), " instead");
  }
}