Model Gaussian Variational Process

Adding the final version of the model with Semisupervised and supervised methods.

Co-Authored-By: Dong Han <dong.han@uconn.edu>

BML_project/active_learning/ss_active_learning.py

@@ -0,0 +1,120 @@
+# -*- coding: utf-8 -*-
+"""
+Created on Mon Dec 18 18:23:23 2023
+
+@author: lrm22005
+"""
+import numpy as np
+import random
+import torch
+from torch.utils.data import DataLoader
+from sklearn.cluster import MiniBatchKMeans
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+def label_samples(uncertain_samples, validation_data):
+    labels = [validation_data[sample_id]['label'] for sample_id in uncertain_samples]
+    return uncertain_samples, labels
+
+def stochastic_uncertainty_sampling(gp_model, gp_likelihood, val_loader, n_samples, n_batches, n_components=2):
+    gp_model.eval()
+    gp_likelihood.eval()
+    uncertain_sample_indices = []
+    sampled_batches = random.sample(list(val_loader), n_batches)  # Randomly sample n_batches from val_loader
+
+    with torch.no_grad():
+        for batch in sampled_batches:
+            # reduced_data = apply_tsne(batch['data'].reshape(batch['data'].size(0), -1), n_components=n_components)
+            # reduced_data_tensor = torch.Tensor(reduced_data).to(device)
+            reduced_data_tensor = batch['data'].view(batch['data'].size(0), -1).to(device)
+            predictions = gp_likelihood(gp_model(reduced_data_tensor))
+            var = predictions.variance
+            top_indices = torch.argsort(-var.flatten())[:n_samples]
+            uncertain_sample_indices.extend(top_indices.cpu().numpy())
+
+    return uncertain_sample_indices[:n_samples]
+
+# def uncertainty_sampling(gp_model, gp_likelihood, val_loader, n_samples, n_components=2):
+#     gp_model.eval()
+#     gp_likelihood.eval()
+#     uncertain_sample_indices = []
+#     with torch.no_grad():
+#         for batch_idx, batch in tqdm(enumerate(val_loader), desc='Uncertainty Sampling', unit='batch'):
+#             reduced_data_tensor = batch['data'].view(batch['data'].size(0), -1).to(device)
+#             predictions = gp_likelihood(gp_model(reduced_data_tensor))
+#             var = predictions.variance
+#             top_indices = torch.argsort(-var.flatten())[:n_samples]
+#             batch_uncertain_indices = [batch_idx * val_loader.batch_size + idx for idx in top_indices]
+#             uncertain_sample_indices.extend(batch_uncertain_indices)
+#     return uncertain_sample_indices[:n_samples]
+
+def run_minibatch_kmeans(data_loader, n_clusters, device, batch_size=100):
+    # Initialize MiniBatchKMeans
+    minibatch_kmeans = MiniBatchKMeans(n_clusters=n_clusters, random_state=0, batch_size=batch_size)
+
+    # Iterate through data_loader and fit MiniBatchKMeans
+    for batch in data_loader:
+        data = batch['data'].view(batch['data'].size(0), -1).to(device).cpu().numpy()
+        minibatch_kmeans.partial_fit(data)
+
+    return minibatch_kmeans
+
+# def compare_kmeans_gp_predictions(kmeans_model, gp_model, data_loader, device):
+#     # Compare K-Means with GP model predictions
+#     all_data, all_labels = [], []
+#     for batch in data_loader:
+#         data = batch['data'].view(batch['data'].size(0), -1).to(device)
+#         labels = batch['label'].to(device)
+#         gp_predictions = gp_model(data).mean.argmax(dim=0).cpu().numpy()
+#         kmeans_predictions = kmeans_model.predict(data.cpu().numpy())
+#         all_labels.append(labels.cpu().numpy())
+#         all_data.append((gp_predictions, kmeans_predictions))
+#     return all_data, np.concatenate(all_labels)
+
+def stochastic_compare_kmeans_gp_predictions(kmeans_model, gp_model, data_loader, n_batches, device):
+    all_data, all_labels = [], []
+    sampled_batches = random.sample(list(data_loader), n_batches)  # Randomly sample n_batches from data_loader
+
+    for batch in sampled_batches:
+        data = batch['data'].view(batch['data'].size(0), -1).to(device)
+        labels = batch['label'].to(device)
+        gp_predictions = gp_model(data).mean.argmax(dim=0).cpu().numpy()
+        kmeans_predictions = kmeans_model.predict(data.cpu().numpy())
+        all_labels.append(labels.cpu().numpy())
+        all_data.append((gp_predictions, kmeans_predictions))
+
+    return all_data, np.concatenate(all_labels)
+
+import random
+
+def refined_uncertainty_sampling(gp_model, gp_likelihood, kmeans_model, data_loader, n_samples, n_batches, uncertainty_threshold=0.2):
+    gp_model.eval()
+    gp_likelihood.eval()
+    uncertain_sample_indices = []
+
+    # Calculate the total number of batches in the DataLoader
+    total_batches = len(data_loader)
+
+    # Ensure that n_batches does not exceed total_batches
+    n_batches = min(n_batches, total_batches)
+
+    # Randomly sample n_batches from data_loader
+    sampled_batches = random.sample(list(data_loader), n_batches)
+
+    with torch.no_grad():
+        for batch in sampled_batches:
+            data_tensor = batch['data'].view(batch['data'].size(0), -1).to(device)
+            gp_predictions = gp_likelihood(gp_model(data_tensor))
+            kmeans_predictions = kmeans_model.predict(data_tensor.cpu().numpy())
+
+            # Calculate the difference between K-means and GP predictions
+            disagreement = (gp_predictions.mean.argmax(dim=-1).cpu().numpy() != kmeans_predictions).astype(int)
+
+            # Calculate uncertainty based on variance of GP predictions
+            uncertainty = gp_predictions.variance.cpu().numpy()
+
+            # Select samples where the disagreement is high and the model is uncertain
+            uncertain_indices = np.where((disagreement > 0) & (uncertainty > uncertainty_threshold))[0]
+            uncertain_sample_indices.extend(uncertain_indices)
+
+    return uncertain_sample_indices[:n_samples]

BML_project/models/ss_gp_model.py

@@ -0,0 +1,198 @@
+# -*- coding: utf-8 -*-
+"""
+Created on Mon Dec 18 18:01:41 2023
+
+@author: lrm22005
+"""
+import numpy as np
+from tqdm import tqdm
+import torch
+import gpytorch
+from sklearn.metrics import precision_recall_fscore_support, roc_auc_score
+from sklearn.preprocessing import label_binarize
+
+num_latents = 6  # This should match the complexity of your data or the number of tasks
+num_tasks = 4    # This should match the number of output classes or tasks
+num_inducing_points = 50  # This is independent and should be sufficient for the input space
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+class MultitaskGPModel(gpytorch.models.ApproximateGP):
+    def __init__(self):
+        # Let's use a different set of inducing points for each latent function
+        inducing_points = torch.rand(num_latents, num_inducing_points, 127 * 128)  # Assuming flattened 128x128 images
+
+        # We have to mark the CholeskyVariationalDistribution as batch
+        # so that we learn a variational distribution for each task
+        variational_distribution = gpytorch.variational.CholeskyVariationalDistribution(
+            inducing_points.size(-2), batch_shape=torch.Size([num_latents])
+        )
+
+        # We have to wrap the VariationalStrategy in a LMCVariationalStrategy
+        # so that the output will be a MultitaskMultivariateNormal rather than a batch output
+        variational_strategy = gpytorch.variational.LMCVariationalStrategy(
+            gpytorch.variational.VariationalStrategy(
+                self, inducing_points, variational_distribution, learn_inducing_locations=True
+            ),
+            num_tasks=num_tasks,
+            num_latents=num_latents,
+            latent_dim=-1
+        )
+
+        super().__init__(variational_strategy)
+
+        # The mean and covariance modules should be marked as batch
+        # so we learn a different set of hyperparameters
+        self.mean_module = gpytorch.means.ConstantMean(batch_shape=torch.Size([num_latents]))
+        self.covar_module = gpytorch.kernels.ScaleKernel(
+            gpytorch.kernels.RBFKernel(batch_shape=torch.Size([num_latents])),
+            batch_shape=torch.Size([num_latents])
+        )
+
+    def forward(self, x):
+        # The forward function should be written as if we were dealing with each output
+        # dimension in batch
+        # Ensure x is correctly shaped. It should have the same last dimension size as inducing_points
+        # x should be reshaped or sliced to have the shape [?, 1] where ? can be any size
+        # For example, if x originally has shape [N, D], and D != 1, you need to modify x accordingly
+        # print(f"Input shape: {x.shape}")
+        # x = x.view(x.size(0), -1)  # Flattening the images
+        # print(f"Input shape after flattening: {x.shape}")  # Debugging input shape
+        mean_x = self.mean_module(x)
+        covar_x = self.covar_module(x)
+
+        # Debugging: Print shapes of intermediate outputs
+        # print(f"Mean shape: {mean_x.shape}, Covariance shape: {covar_x.shape}")
+        latent_pred = gpytorch.distributions.MultivariateNormal(mean_x, covar_x)
+        # print(f"Latent prediction shape: {latent_pred.mean.shape}, {latent_pred.covariance_matrix.shape}")
+
+        return latent_pred
+
+
+def train_gp_model(train_loader, val_loader, num_iterations=50, n_classes=4, patience=10, checkpoint_path='model_checkpoint_full.pt'):
+    model = MultitaskGPModel().to(device)
+    likelihood = gpytorch.likelihoods.SoftmaxLikelihood(num_features=4, num_classes=4).to(device)
+    optimizer = torch.optim.Adam(model.parameters(), lr=0.1)
+    mll = gpytorch.mlls.VariationalELBO(likelihood, model, num_data=len(train_loader.dataset))
+    best_val_loss = float('inf')
+    epochs_no_improve = 0
+
+    metrics = {
+        'precision': [],
+        'recall': [],
+        'f1_score': [],
+        'auc_roc': [],
+        'train_loss': []  # Add a list to store training losses
+    }
+
+    for epoch in tqdm(range(num_iterations), desc='Training', unit='epoch', leave=False):
+        for train_batch in train_loader:
+            model.train()
+            likelihood.train()
+            optimizer.zero_grad()
+            train_x = train_batch['data'].reshape(train_batch['data'].size(0), -1).to(device)  # Use reshape here
+            train_y = train_batch['label'].to(device)
+            output = model(train_x)
+            loss = -mll(output, train_y)
+            metrics['train_loss'].append(loss.item())  # Store the training loss
+            loss.backward()
+            optimizer.step()
+
+        # Stochastic validation
+        model.eval()
+        likelihood.eval()
+        with torch.no_grad():
+            val_indices = torch.randperm(len(val_loader.dataset))[:int(1 * len(val_loader.dataset))]
+            val_loss = 0.0
+            val_labels = []
+            val_predictions = []
+            for idx in val_indices:
+                val_batch = val_loader.dataset[idx]
+                val_x = val_batch['data'].reshape(-1).unsqueeze(0).to(device)  # Use reshape here
+                val_y = torch.tensor([val_batch['label']], device=device)
+                val_output = model(val_x)
+                val_loss_batch = -mll(val_output, val_y).sum()
+                val_loss += val_loss_batch.item()
+                val_labels.append(val_y.item())
+                val_predictions.append(val_output.mean.argmax(dim=-1).item())
+
+            precision, recall, f1, _ = precision_recall_fscore_support(val_labels, val_predictions, average='macro')
+            # auc_roc = roc_auc_score(label_binarize(val_labels, classes=np.arange(n_classes)),
+            #                         label_binarize(val_predictions, classes=np.arange(n_classes)),
+            #                         multi_class='ovr')
+
+            metrics['precision'].append(precision)
+            metrics['recall'].append(recall)
+            metrics['f1_score'].append(f1)
+            # metrics['auc_roc'].append(auc_roc)
+            val_loss /= len(val_indices)
+
+        if val_loss < best_val_loss:
+            best_val_loss = val_loss
+            epochs_no_improve = 0
+            torch.save({'model_state_dict': model.state_dict(),
+                        'likelihood_state_dict': likelihood.state_dict(),
+                        'optimizer_state_dict': optimizer.state_dict()}, checkpoint_path)
+        else:
+            epochs_no_improve += 1
+            if epochs_no_improve >= patience:
+                print(f"Early stopping triggered at epoch {epoch+1}")
+                break
+
+    checkpoint = torch.load(checkpoint_path)
+    model.load_state_dict(checkpoint['model_state_dict'])
+    likelihood.load_state_dict(checkpoint['likelihood_state_dict'])
+    optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
+
+    return model, likelihood, metrics
+
+def semi_supervised_labeling(kmeans_model, gp_model, gp_likelihood, data_loader, confidence_threshold=0.8):
+    gp_model.eval()
+    gp_likelihood.eval()
+    labeled_samples = []
+
+    with torch.no_grad():
+        for batch in data_loader:
+            data_tensor = batch['data'].view(batch['data'].size(0), -1).to(device)
+            kmeans_predictions = kmeans_model.predict(data_tensor.cpu().numpy())
+            gp_predictions = gp_likelihood(gp_model(data_tensor))
+
+            # Use GP predictions where the model is confident
+            confident_indices = gp_predictions.confidence().cpu().numpy() > confidence_threshold
+            for i, confident in enumerate(confident_indices):
+                if confident:
+                    labeled_samples.append((data_tensor[i], gp_predictions.mean.argmax(dim=-1)[i].item()))
+                else:
+                    labeled_samples.append((data_tensor[i], kmeans_predictions[i]))
+
+    return labeled_samples
+
+def calculate_elbo(model, likelihood, data_loader):
+    """
+    Calculates the ELBO (Evidence Lower Bound) score for the model on the given data.
+
+    Args:
+    - model: The trained Gaussian Process model.
+    - likelihood: The likelihood associated with the GP model.
+    - data_loader: DataLoader providing the data over which to calculate ELBO.
+
+    Returns:
+    - elbo_score: The calculated ELBO score.
+    """
+    model.eval()
+    likelihood.eval()
+    mll = gpytorch.mlls.VariationalELBO(likelihood, model, num_data=len(data_loader.dataset))
+
+    with torch.no_grad():
+        elbo_score = 0.0
+        for batch in data_loader:
+            train_x = batch['data'].reshape(batch['data'].size(0), -1).to(device)
+            train_y = batch['label'].to(device)
+            output = model(train_x)
+            # Calculate the ELBO as the negative loss
+            elbo_score += -mll(output, train_y).sum().item()
+
+        # Average the ELBO over all data samples
+        elbo_score /= len(data_loader.dataset)
+
+    return elbo_score