ensemble/load.py

import torch
import pathlib
import importlib.util
import safetensors.torch
import matplotlib.pyplot as plt
import math

from typing import Literal


def load_model_module(model_path: pathlib.Path):
    model_path = model_path.resolve()

    spec = importlib.util.spec_from_file_location("model", model_path)
    model = importlib.util.module_from_spec(spec)
    spec.loader.exec_module(model)

    return model

class EnsembleModel(torch.nn.Module):
    def __init__(self, model1, model2, model3, model4, model5, model6, mode="max"):
        super(EnsembleModel, self).__init__()
        self.model1 = model1
        self.model2 = model2
        self.model3 = model3
        self.model4 = model4
        self.model5 = model5
        self.model6 = model6
        self.models = [model1, model2, model3, model4, model5, model6]
        self.mode = mode
        if mode not in ["min", "mean", "max", "none"]:
            raise ValueError("Mode must be 'none', 'min', 'mean', or 'max'.")
        
    def forward(self, x):
        outputs = []
        for model in self.models:
            output = model(x)
            outputs.append(output)
        
        # Average the outputs
        if self.mode == "max":
            output_probs = torch.max(torch.cat(outputs, dim=1), dim=1)[0].squeeze()
        elif self.mode == "mean":
            output_probs = torch.mean(torch.cat(outputs, dim=1), dim=1)[0].squeeze()
        elif self.mode == "min":
            output_probs = torch.min(torch.cat(outputs, dim=1), dim=1)[0].squeeze()
        elif self.mode == "none":
            return torch.cat(outputs, dim=1)
        else:
            raise ValueError("Mode must be 'min', 'mean', or 'max'.")

        # Kind of uncertainty
        std_output = torch.std(torch.cat(outputs, dim=1), dim=1)[0].squeeze()

        # Normalize the standard deviation [0 - 1]
        N = len(outputs)
        std_max = math.sqrt(0.25 * N / (N - 1))
        std_output = std_output / std_max

        return output_probs, std_output


# MLSTAC API -----------------------------------------------------------------------
def example_data(path: pathlib.Path, device = "cpu", *args, **kwargs):
    data_f = path / "example_data.safetensor"
    sample = safetensors.torch.load_file(data_f)
    return  sample["image"].float().unsqueeze(0).to(device)

def trainable_model(*args, **kwargs):
    print("The model is not available in training mode.")
    return None

def compiled_model(path, device: str = "cpu", mode: Literal["min", "mean", "max"] ="max",*args, **kwargs):
    
    model1_f = path / "1dpwdeeplabv3.safetensor"
    model2_f = path / "1dpwseg.safetensor"
    model3_f = path / "1dpwunetpp.safetensor"
    model4_f = path / "unet.safetensor"
    model5_f = path / "unetpp.safetensor"
    model6_f = path / "c2r1km.safetensor"

    # Load model parameters
    model1_weights = safetensors.torch.load_file(model1_f)
    model2_weights = safetensors.torch.load_file(model2_f)
    model3_weights = safetensors.torch.load_file(model3_f)
    model4_weights = safetensors.torch.load_file(model4_f)
    model5_weights = safetensors.torch.load_file(model5_f)
    model6_weights = safetensors.torch.load_file(model6_f)

    # Load all models

    # Model 1 (DeepLabV3Branch + PixelWise)
    model1 = load_model_module(path / "model.py").CombinedNet4(
        classes=1, benchmark=True, in_channels=4
    )
    model1.load_state_dict(model1_weights)
    model1 = model1.to(device)
    for param in model1.parameters():
        param.requires_grad = False
    model1 = model1.eval()
    
    
    # Model 2 (SegformerBranch + PixelWise)
    model2 = load_model_module(path / "model.py").CombinedNet(
        classes=1, benchmark=True
    )
    model2.load_state_dict(model2_weights)
    model2 = model2.to(device)
    for param in model2.parameters():
        param.requires_grad = False
    model2 = model2.eval()
    
    # Model 3 (UNetPlusPlusBranch + PixelWise)
    model3 = load_model_module(path / "model.py").CombinedNet3(
        classes=1, benchmark=True
    )
    model3.load_state_dict(model3_weights)
    model3 = model3.to(device)
    for param in model3.parameters():
        param.requires_grad = False
    model3 = model3.eval()

    # Model 4 (UNetBranch)
    model4 = load_model_module(path / "model.py").UNetBranch(
        classes=1, benchmark=True
    )    
    model4.load_state_dict(model4_weights)
    model4 = model4.to(device)
    for param in model4.parameters():
        param.requires_grad = False
    model4 = model4.eval()

    # Model 5 (UNetPlusPlusBranch)
    model5 = load_model_module(path / "model.py").UNetPlusPlusBranch(
        classes=1, benchmark=True
    )
    model5.load_state_dict(model5_weights)
    model5 = model5.to(device)
    for param in model5.parameters():
        param.requires_grad = False
    model5 = model5.eval()

    # Model 6 (C2R1KM)
    model6 = load_model_module(path / "c2r1km.py").CloudMaskOne(
        hidden_layer_sizes=(128, 112),
        activation='relu',
        last_activation='sigmoid',
        dropout_rate=0.1,
        input_dim=40,
        batch_norm=False
    )
    model6.load_state_dict(model6_weights)
    model6 = model6.to(device)
    for param in model6.parameters():
        param.requires_grad = False
    model6 = model6.eval()
    
    # Create ensemble model
    cloud_model = EnsembleModel(model1, model2, model3, model4, model5, model6, mode=mode)

    return cloud_model



def display_results(path: pathlib.Path, device: str = "cpu", mode: Literal["min", "mean", "max"] ="max", *args, **kwargs):
    # Load model
    model = compiled_model(path, device, mode=mode)

    # Load data
    probav = example_data(path)

    # Run model
    cloudprob, uncertainty = model(probav.float().to(device))

    #Display results
    fig, ax = plt.subplots(1, 3, figsize=(12, 4))
    ax[0].imshow(probav[0, [2, 1, 0]].cpu().detach().numpy().transpose(1, 2, 0))
    ax[0].set_title("Input")
    ax[1].imshow(cloudprob.cpu().detach().numpy(), cmap="gray")
    ax[1].set_title("Cloud Probability")
    ax[2].imshow(uncertainty.cpu().detach().numpy(), cmap="gray")
    ax[2].set_title("Uncertainty")
    for a in ax:
        a.axis("off")    
    fig.tight_layout()    
    return fig