Updates to CorrDiffTaiwan model wrapper

earth2studio/models/dx/corrdiff.py

@@ -880,7 +880,7 @@ def load_default_package(cls) -> Package:
 
     @classmethod
     @check_optional_dependencies()
-    def load_model(cls, package: Package) -> DiagnosticModel:
+    def load_model(cls, package: Package, device: str | None = None) -> DiagnosticModel:
         """Load diagnostic from package"""
 
         if StackedRandomGenerator is None or deterministic_sampler is None:
@@ -897,15 +897,33 @@ def load_model(cls, package: Package) -> DiagnosticModel:
             str(
                 checkpoint_zip.parent
                 / Path("corrdiff_inference_package/checkpoints/diffusion.mdlus")
-            )
-        ).eval()
+            ),
+            override_args={"use_apex_gn": True}
+        )
+        residual.use_fp16, residual.profile_mode = True, False
+        residual = residual.eval()
+        if device is not None:
+            residual = residual.to(device)
+        residual = residual.to(memory_format=torch.channels_last)
 
         regression = PhysicsNemoModule.from_checkpoint(
             str(
                 checkpoint_zip.parent
                 / Path("corrdiff_inference_package/checkpoints/regression.mdlus")
-            )
-        ).eval()
+            ),
+            override_args={"use_apex_gn": True}
+        )
+        regression.use_fp16, regression.profile_mode = True, False
+        regression = regression.eval()
+        if device is not None:
+            regression = regression.to(device)
+        regression = regression.to(memory_format=torch.channels_last)
+
+        # Compile models
+        torch._dynamo.config.cache_size_limit = 264
+        torch._dynamo.reset()
+        residual = torch.compile(residual)
+        regression = torch.compile(regression)
 
         # Get dataset for lat/lon grid info and centers/stds'
         try:
@@ -1011,10 +1029,11 @@ def _forward(self, x: torch.Tensor) -> torch.Tensor:
         x = self._interpolate(x)
 
         # Add sample dimension
-        x = x.unsqueeze(0)
+        x = x.unsqueeze(0).to(memory_format=torch.channels_last)
         x = (x - self.in_center) / self.in_scale
 
         # Create grid channels
+        # TODO: why do we need this grid concatenated to the input?
         x1 = np.sin(np.linspace(0, 2 * np.pi, 448))
         x2 = np.cos(np.linspace(0, 2 * np.pi, 448))
         y1 = np.sin(np.linspace(0, 2 * np.pi, 448))
@@ -1032,43 +1051,44 @@ def _forward(self, x: torch.Tensor) -> torch.Tensor:
         # Concat Grids
         x = torch.cat((x, grid), dim=1)
 
-        # Repeat for sample size
-        sample_seeds = torch.arange(self.number_of_samples)
-        x = x.repeat(self.number_of_samples, 1, 1, 1)
+        # Create seeds for each sample
+        seed = self.seed if self.seed is not None else np.random.randint(2**32)
+        gen = torch.Generator(device=x.device)
+        gen.manual_seed(seed)
+        sample_seeds = torch.randint(0, 2**32, (self.number_of_samples,), device=x.device, generator=gen)
 
-        # Create latents
-        rnd = StackedRandomGenerator(x.device, sample_seeds)
+        sampler_fn = partial(
+            deterministic_sampler,
+            num_steps=self.number_of_steps,
+            solver=self.solver,
+        )
 
+        # Get high-res image shape
         coord = self.output_coords(self.input_coords())
-        img_resolution_x = coord["lat"].shape[0]
-        img_resolution_y = coord["lon"].shape[1]
-        latents = rnd.randn(
-            [
-                self.number_of_samples,
-                self.regression_model.img_out_channels,
-                img_resolution_x,
-                img_resolution_y,
-            ],
-            device=x.device,
-        )
+        H_hr = coord["lat"].shape[0]
+        W_hr = coord["lon"].shape[1]
 
-        mean = self.unet_regression(
+        mean_hr = self.unet_regression(
             self.regression_model,
-            torch.zeros_like(latents),
-            x,
-            num_steps=self.number_of_steps,
+            img_lr=x,
+            output_channels=len(self.output_variables),
+            number_of_samples=self.number_of_samples,
         )
-        res = deterministic_sampler(
-            self.residual_model,
-            latents,
-            x,
-            randn_like=rnd.randn_like,
-            num_steps=self.number_of_steps,
-            solver=self.solver,
+
+        res_hr = diffusion_step(
+            net=self.residual_model,
+            sampler_fn=sampler_fn,
+            img_shape=(H_hr, W_hr),
+            img_out_channels=len(self.output_variables),
+            rank_batches=sample_seeds,
+            img_lr=x,
+            device=x.device,
+            mean_hr=mean_hr,
         )
-        x = mean + res
-        x = self.out_scale * x + self.out_center
-        return x
+
+        x_hr = mean_hr + res_hr
+        x_hr = self.out_scale * x_hr + self.out_center
+        return x_hr
 
     @batch_func()
     def __call__(
@@ -1092,79 +1112,43 @@ def __call__(
     @staticmethod
     def unet_regression(
         net: torch.nn.Module,
-        latents: torch.Tensor,
         img_lr: torch.Tensor,
-        class_labels: torch.Tensor = None,
-        randn_like: Callable = torch.randn_like,
-        num_steps: int = 8,
-        sigma_min: float = 0.0,
-        sigma_max: float = 0.0,
-        rho: int = 7,
-        S_churn: float = 0,
-        S_min: float = 0,
-        S_max: float = float("inf"),
-        S_noise: float = 0.0,
+        lead_time_label: torch.Tensor,
+        output_channels: int,
+        number_of_samples: int,
     ) -> torch.Tensor:
         """
-        Perform U-Net regression with temporal sampling.
+        Perform U-Net regression.
 
         Parameters
         ----------
         net : torch.nn.Module
             U-Net model for regression.
-        latents : torch.Tensor
-            Latent representation.
-        img_lr : torch.Tensor)
-            Low-resolution input image.
-        class_labels : torch.Tensor, optional
-            Class labels for conditional generation.
-        randn_like : function, optional
-            Function for generating random noise.
-        num_steps : int, optional
-            Number of time steps for temporal sampling.
-        sigma_min : float, optional
-            Minimum noise level.
-        sigma_max : float, optional
-            Maximum noise level.
-        rho : int, optional
-            Exponent for noise level interpolation.
-        S_churn : float, optional
-            Churning parameter.
-        S_min : float, optional
-            Minimum churning value.
-        S_max : float, optional
-            Maximum churning value.
-        S_noise : float, optional
-            Noise level for churning.
+        img_lr : torch.Tensor
+            Low-resolution input image of shape (1, C_in, H_hr, W_hr).
+        lead_time_label : torch.Tensor
+            Lead time label of shape (1, 1, 1, 1), or (1,).
+        output_channels : int
+            Number of output channels C_out.
+        number_of_samples : int
+            Number of samples to generate for the single input batch element.
+            Only used to expand the shape of the output tensor.
 
         Returns
         -------
-        torch.Tensor: Predicted output at the next time step.
+        torch.Tensor: Predicted output with shape (number_of_samples, C_out,
+        H_hr, W_hr).
         """
-        # Adjust noise levels based on what's supported by the network.
-        sigma_min = max(sigma_min, 0)
-        sigma_max = min(sigma_max, np.inf)
-
-        # Time step discretization.
-        step_indices = torch.arange(
-            num_steps, dtype=torch.float64, device=latents.device
+        mean_hr = regression_step(
+            net=net,
+            img_lr=img_lr,
+            latents_shape=(
+                number_of_samples,
+                output_channels,
+                img_lr.shape[-2],
+                img_lr.shape[-1],
+            ),
+            lead_time_label=lead_time_label
         )
-        t_steps = (
-            sigma_max ** (1 / rho)
-            + step_indices
-            / (num_steps - 1)
-            * (sigma_min ** (1 / rho) - sigma_max ** (1 / rho))
-        ) ** rho
-        t_steps = torch.cat(
-            [net.round_sigma(t_steps), torch.zeros_like(t_steps[:1])]
-        )  # t_N = 0
-
-        # conditioning
-        x_lr = img_lr
-
-        # Main sampling loop.
-        x_hat = latents.to(torch.float64) * t_steps[0]
-
-        x_next = net(x_hat, x_lr).to(torch.float64)
 
-        return x_next
+        return mean_hr