Add deep learning resolution estimation model from Stanford

resonet/NeuralNetResPredictor.cpp

@@ -0,0 +1,93 @@
+// Copyright (2019-2024) Paul Scherrer Institute
+
+#include "NeuralNetResPredictor.h"
+
+#include <cmath>
+#include "../common/JFJochException.h"
+
+NeuralNetResPredictor::NeuralNetResPredictor(const DiffractionExperiment &in_experiment)
+        : experiment(in_experiment), model_input(512*512)
+#ifdef JFJOCH_USE_TORCH
+          , device(torch::kCUDA)
+#endif
+        {
+    float max_direction = std::max(in_experiment.GetXPixelsNum(), in_experiment.GetYPixelsNum()) / 2.0;
+    pool_factor = std::lround(max_direction / 512.0f);
+    if (pool_factor <= 0)
+        throw JFJochException(JFJochExceptionCategory::InputParameterInvalid, "Detector size is too small");
+
+    if (pool_factor > 8)
+        throw JFJochException(JFJochExceptionCategory::InputParameterInvalid, "Detector size is too large");
+#ifdef JFJOCH_USE_TORCH
+    module = torch::jit::load(experiment.GetNeuralNetModelPath());
+    module.to(device);
+#endif
+}
+
+template<class T>
+void NeuralNetResPredictor::PrepareInternal(const T* image) {
+    size_t xpixel = experiment.GetXPixelsNum();
+    size_t ypixel = experiment.GetYPixelsNum();
+    size_t start_x = std::lround(experiment.GetBeamX_pxl());
+    size_t start_y = std::lround(experiment.GetBeamY_pxl());
+
+    for (size_t y = 0; y < 512; y++) {
+        size_t y0 = y * pool_factor + start_y;
+        size_t min_yp = std::min(y0 + pool_factor, ypixel);
+        for (size_t x = 0; x < 512; x++) {
+            float val = 0.0;
+            size_t x0 = x * pool_factor + start_x;
+            size_t min_xp = std::min(x0 + pool_factor, ypixel);
+
+            for (size_t yp = y0; yp < min_yp; yp++) {
+                for (size_t xp = x0; xp < min_xp; xp++) {
+                    int16_t pxl = image[yp * xpixel + xp];
+                    if (val < pxl)
+                        val = pxl;
+                }
+            }
+            float max_pool = floorf(sqrtf(val));
+            model_input[512 * y + x] = max_pool;
+        }
+    }
+}
+
+void NeuralNetResPredictor::Prepare(const int16_t *image) {
+    PrepareInternal(image);
+}
+
+void NeuralNetResPredictor::Prepare(const int32_t *image) {
+    PrepareInternal(image);
+}
+
+size_t NeuralNetResPredictor::GetMaxPoolFactor() const {
+    return pool_factor;
+}
+
+float NeuralNetResPredictor::Inference(const void *image) {
+    if (experiment.GetPixelDepth() == 2)
+        Prepare((int16_t *) image);
+    else
+        Prepare((int32_t *) image);
+
+#ifdef JFJOCH_USE_TORCH
+    auto options = torch::TensorOptions().dtype(at::kFloat);
+    auto model_input_tensor = torch::from_blob(model_input.data(), {1,1,512,512}, options).to(device);
+
+    std::vector<torch::jit::IValue> pixels;
+    pixels.emplace_back(std::move(model_input_tensor));
+
+    auto output = module.forward(pixels).toTensor();
+    auto tensor_output = output[0].item<float>();
+    float two_theta = atanf(((2.0f * experiment.GetPixelSize_mm() / experiment.GetDetectorDistance_mm()) * tensor_output));
+    float stheta = sinf(two_theta * 0.5f);
+    float resolution = experiment.GetWavelength_A() / (2.0f * stheta);
+#else
+    float resolution = 50.0;
+#endif
+    return resolution;
+}
+
+const std::vector<float> &NeuralNetResPredictor::GetModelInput() const {
+    return model_input;
+}