Method input_gradient

include/tiny-cuda-nn/object.h:592–616

	}

	void input_gradient(
		cudaStream_t stream,
		uint32_t dim,
		const GPUMatrix<T>& input,
		GPUMatrix<T>& d_dinput,
		float backprop_scale = default_loss_scale<PARAMS_T>() // Prevents underflows during half-precision backprop. Same reason for loss_scale to exist.
	) {
		// Make sure our temporary buffers have the correct size for the given batch size
		uint32_t batch_size = input.n();

		GPUMatrix<COMPUTE_T> d_doutput = {padded_output_width(), batch_size, stream};
		GPUMatrix<COMPUTE_T> output = {padded_output_width(), batch_size, stream};

		if (dim >= padded_output_width()) {
			throw std::runtime_error{"Invalid dimension to compute the input gradient for."};
		}

		// Set "loss gradient" at network outputs to 1 at the chosen dimension and 0 elsewhere.
		one_hot_batched(stream, output.n_elements(), padded_output_width(), dim, d_doutput.data(), backprop_scale);

		auto ctx = forward(stream, input, &output, true /* inference matrices */, true /* prep forward buffers for input gradients */);
		backward(stream, *ctx, input, output, d_doutput, &d_dinput, true /* inference matrices */, GradientMode::Ignore);

		mult(stream, d_dinput.n_elements(), d_dinput.data(), 1.0f / backprop_scale);
	}

	virtual uint32_t input_width() const = 0;