AOT-biomaps 2.9.186__py3-none-any.whl → 2.9.294__py3-none-any.whl

AOT_biomaps/AOT_Recon/AOT_Optimizers/MAPEM.py

@@ -2,13 +2,94 @@ from AOT_biomaps.AOT_Recon.ReconEnums import PotentialType
 from AOT_biomaps.AOT_Recon.AOT_PotentialFunctions.Quadratic import _Omega_QUADRATIC_CPU, _Omega_QUADRATIC_GPU
 from AOT_biomaps.AOT_Recon.AOT_PotentialFunctions.RelativeDifferences import _Omega_RELATIVE_DIFFERENCE_CPU, _Omega_RELATIVE_DIFFERENCE_GPU
 from AOT_biomaps.AOT_Recon.AOT_PotentialFunctions.Huber import _Omega_HUBER_PIECEWISE_CPU, _Omega_HUBER_PIECEWISE_GPU
-from AOT_biomaps.AOT_Recon.ReconTools import _build_adjacency_sparse
+from AOT_biomaps.AOT_Recon.ReconTools import _build_adjacency_sparse, check_gpu_memory, calculate_memory_requirement
 from AOT_biomaps.Config import config
+
+import warnings
 import numpy as np
 import torch
 from tqdm import trange
 
-def _MAPEM_CPU_STOP(SMatrix, y, Omega, numIterations, beta, delta, gamma, sigma, isSavingEachIteration, withTumor, max_saves=5000):
+def MAPEM(
+    SMatrix, 
+    y, 
+    Omega,
+    beta,
+    delta=None,
+    gamma=None,
+    sigma=None,
+    numIterations=100,
+    isSavingEachIteration=True,
+    withTumor=True,
+    device=None,
+    max_saves=5000,
+    show_logs=True):
+    """
+    This method implements the MAPEM algorithm using either CPU or single-GPU PyTorch acceleration.
+    Multi-GPU and Multi-CPU modes are not implemented for this algorithm.
+    """
+    try:
+        tumor_str = "WITH" if withTumor else "WITHOUT"
+        # Auto-select device and method
+        if device is None:
+            if torch.cuda.is_available() and check_gpu_memory(config.select_best_gpu(), calculate_memory_requirement(SMatrix, y), show_logs=show_logs):
+                device = torch.device(f"cuda:{config.select_best_gpu()}")
+                use_gpu = True
+            else:
+                device = torch.device("cpu")
+                use_gpu = False
+        else:
+            use_gpu = device.type == "cuda"
+        # Dispatch to the appropriate implementation
+        if use_gpu:
+                return _MAPEM_GPU(SMatrix, y, Omega, beta, delta, gamma, sigma, numIterations, isSavingEachIteration, tumor_str, device, max_saves, show_logs=True)
+        else:
+                return _MAPEM_CPU(SMatrix, y, Omega, beta, delta, gamma, sigma, numIterations, isSavingEachIteration, tumor_str, max_saves, show_logs=True)
+    except Exception as e:
+        print(f"Error in MLEM: {type(e).__name__}: {e}")
+        return None, None
+
+def MAPEM_STOP(
+    SMatrix, 
+    y, 
+    Omega,
+    beta,
+    delta=None,
+    gamma=None,
+    sigma=None,
+    numIterations=100,
+    isSavingEachIteration=True,
+    withTumor=True,
+    device=None,
+    max_saves=5000,
+    show_logs=True):
+    """
+    This method implements the MAPEM_STOP algorithm using either CPU or single-GPU PyTorch acceleration.
+    Multi-GPU and Multi-CPU modes are not implemented for this algorithm.
+    """
+    try:
+        tumor_str = "WITH" if withTumor else "WITHOUT"
+        # Auto-select device and method
+        if device is None:
+            if torch.cuda.is_available() and check_gpu_memory(config.select_best_gpu(), calculate_memory_requirement(SMatrix, y), show_logs=show_logs):
+                device = torch.device(f"cuda:{config.select_best_gpu()}")
+                use_gpu = True
+            else:
+                device = torch.device("cpu")
+                use_gpu = False
+        else:
+            use_gpu = device.type == "cuda"
+        # Dispatch to the appropriate implementation
+        if use_gpu:
+                return _MAPEM_GPU_STOP(SMatrix, y, Omega, beta, delta, gamma, sigma, numIterations, isSavingEachIteration, tumor_str, device, max_saves, show_logs=True)
+        else:
+                return _MAPEM_CPU_STOP(SMatrix, y, Omega, beta, delta, gamma, sigma, numIterations, isSavingEachIteration, tumor_str, max_saves, show_logs=True)
+    except Exception as e:
+        print(f"Error in MLEM: {type(e).__name__}: {e}")
+        return None, None
+
+
+def _MAPEM_CPU_STOP(SMatrix, y, Omega, beta, delta, gamma, sigma, numIterations, isSavingEachIteration, tumor_str, max_saves, show_logs=True):
     """
     MAPEM version CPU simple - sans GPU - torch uniquement
     """
@@ -42,7 +123,6 @@ def _MAPEM_CPU_STOP(SMatrix, y, Omega, numIterations, beta, delta, gamma, sigma,
         theta_list = [I_0]
         results = [I_0.numpy()]
         saved_indices = [0]
-        previous = -np.inf
         normalization_factor = SMatrix.sum(dim=(0, 3)).reshape(-1)
         adj_index, adj_values = _build_adjacency_sparse(Z, X)
 
@@ -56,13 +136,14 @@ def _MAPEM_CPU_STOP(SMatrix, y, Omega, numIterations, beta, delta, gamma, sigma,
                 save_indices.append(numIterations - 1)
 
         if Omega == PotentialType.HUBER_PIECEWISE:
-            description = f"AOT-BioMaps -- Bayesian Reconstruction Tomography: MAP-EM (Sparse HUBER β:{beta:.4f}, δ:{delta:.4f}) + STOP condition (penalized log-likelihood) ---- {'WITH' if withTumor else 'WITHOUT'} TUMOR ---- processing on single CPU ----"
+            description = f"AOT-BioMaps -- Bayesian Reconstruction Tomography: MAP-EM (Sparse HUBER β:{beta:.4f}, δ:{delta:.4f}) + STOP condition (penalized log-likelihood) ---- {tumor_str} TUMOR ---- processing on single CPU ----"
         elif Omega == PotentialType.RELATIVE_DIFFERENCE:
-            description = f"AOT-BioMaps -- Bayesian Reconstruction Tomography: MAP-EM (Sparse RD β:{beta:.4f}, γ:{gamma:.4f}) + STOP condition (penalized log-likelihood) ---- {'WITH' if withTumor else 'WITHOUT'} TUMOR ---- processing on single CPU ----"
+            description = f"AOT-BioMaps -- Bayesian Reconstruction Tomography: MAP-EM (Sparse RD β:{beta:.4f}, γ:{gamma:.4f}) + STOP condition (penalized log-likelihood) ---- {tumor_str} TUMOR ---- processing on single CPU ----"
         elif Omega == PotentialType.QUADRATIC:
-            description = f"AOT-BioMaps -- Bayesian Reconstruction Tomography: MAP-EM (Sparse QUADRATIC β:{beta:.4f}, σ:{sigma:.4f}) + STOP condition (penalized log-likelihood) ---- {'WITH' if withTumor else 'WITHOUT'} TUMOR ---- processing on single CPU ----"
+            description = f"AOT-BioMaps -- Bayesian Reconstruction Tomography: MAP-EM (Sparse QUADRATIC β:{beta:.4f}, σ:{sigma:.4f}) + STOP condition (penalized log-likelihood) ---- {tumor_str} TUMOR ---- processing on single CPU ----"
 
-        for p in trange(numIterations, desc=description):
+        iterator = trange(numIterations, desc=description) if show_logs else range(numIterations)
+        for it in iterator:
             theta_p = theta_list[-1]
             theta_p_flat = theta_p.reshape(-1)
             q_flat = A_flat @ theta_p_flat
@@ -80,14 +161,13 @@ def _MAPEM_CPU_STOP(SMatrix, y, Omega, numIterations, beta, delta, gamma, sigma,
             theta_next_flat = torch.clamp(theta_next_flat, min=0)
             theta_next = theta_next_flat.reshape(Z, X)
             theta_list[-1] = theta_next
-            if isSavingEachIteration and p in save_indices:
+            if isSavingEachIteration and it in save_indices:
                 results.append(theta_next.numpy())
-                saved_indices.append(p+1)
+                saved_indices.append(it+1)
             log_likelihood = (y_flat * torch.log(q_flat + 1e-8) - (q_flat + 1e-8)).sum()
             penalized_log_likelihood = log_likelihood - beta * U_value
-            current = penalized_log_likelihood.item()
-            if (p + 1) % 100 == 0:
-                print(f"Iter {p+1}: logL={log_likelihood:.3e}, U={U_value:.3e}, penalized logL={penalized_log_likelihood:.3e}")
+            if (it + 1) % 100 == 0:
+                print(f"Iter {it+1}: logL={log_likelihood:.3e}, U={U_value:.3e}, penalized logL={penalized_log_likelihood:.3e}")
 
         if isSavingEachIteration:
             return results, saved_indices
@@ -97,7 +177,7 @@ def _MAPEM_CPU_STOP(SMatrix, y, Omega, numIterations, beta, delta, gamma, sigma,
         print(f"An error occurred in _MAPEM_CPU_STOP: {e}")
         return None, None
 
-def _MAPEM_GPU_STOP(SMatrix, y, Omega, numIterations, beta, delta, gamma, sigma, isSavingEachIteration, withTumor, max_saves=5000):
+def _MAPEM_GPU_STOP(SMatrix, y, Omega, beta, delta, gamma, sigma, numIterations, isSavingEachIteration, tumor_str, device, max_saves, show_logs=True):
     """
     Maximum A Posteriori (MAP) estimation for Bayesian reconstruction.
     This method computes the MAP estimate of the parameters given the data.
@@ -123,7 +203,6 @@ def _MAPEM_GPU_STOP(SMatrix, y, Omega, numIterations, beta, delta, gamma, sigma,
         else:
             raise ValueError(f"Unknown potential type: {Omega}")
 
-        device = torch.device(f"cuda:{config.select_best_gpu()}")
         A_matrix_torch = torch.tensor(SMatrix, dtype=torch.float32).to(device)
         y_torch = torch.tensor(y, dtype=torch.float32).to(device)
         T, Z, X, N = SMatrix.shape
@@ -150,13 +229,14 @@ def _MAPEM_GPU_STOP(SMatrix, y, Omega, numIterations, beta, delta, gamma, sigma,
                 save_indices.append(numIterations - 1)
 
         if Omega == PotentialType.HUBER_PIECEWISE:
-            description = f"AOT-BioMaps -- Bayesian Reconstruction Tomography: MAP-EM (Sparse HUBER β:{beta:.4f}, δ:{delta:.4f}) + STOP condition (penalized log-likelihood) ---- {'WITH' if withTumor else 'WITHOUT'} TUMOR ---- processing on single GPU no.{torch.cuda.current_device()} ----"
+            description = f"AOT-BioMaps -- Bayesian Reconstruction Tomography: MAP-EM (Sparse HUBER β:{beta:.4f}, δ:{delta:.4f}) + STOP condition (penalized log-likelihood) ---- {tumor_str} TUMOR ---- processing on single GPU no.{torch.cuda.current_device()} ----"
         elif Omega == PotentialType.RELATIVE_DIFFERENCE:
-            description = f"AOT-BioMaps -- Bayesian Reconstruction Tomography: MAP-EM (Sparse RD β:{beta:.4f}, γ:{gamma:.4f}) + STOP condition (penalized log-likelihood) ---- {'WITH' if withTumor else 'WITHOUT'} TUMOR ---- processing on single GPU no.{torch.cuda.current_device()} ----"
+            description = f"AOT-BioMaps -- Bayesian Reconstruction Tomography: MAP-EM (Sparse RD β:{beta:.4f}, γ:{gamma:.4f}) + STOP condition (penalized log-likelihood) ---- {tumor_str} TUMOR ---- processing on single GPU no.{torch.cuda.current_device()} ----"
         elif Omega == PotentialType.QUADRATIC:
-            description = f"AOT-BioMaps -- Bayesian Reconstruction Tomography: MAP-EM (Sparse QUADRATIC β:{beta:.4f}, σ:{sigma:.4f}) + STOP condition (penalized log-likelihood) ---- {'WITH' if withTumor else 'WITHOUT'} TUMOR ---- processing on single GPU no.{torch.cuda.current_device()} ----"
+            description = f"AOT-BioMaps -- Bayesian Reconstruction Tomography: MAP-EM (Sparse QUADRATIC β:{beta:.4f}, σ:{sigma:.4f}) + STOP condition (penalized log-likelihood) ---- {tumor_str} TUMOR ---- processing on single GPU no.{torch.cuda.current_device()} ----"
 
-        for p in trange(numIterations, desc=description):
+        iterator = trange(numIterations, desc=description) if show_logs else range(numIterations)
+        for it in iterator:
             theta_p = matrix_theta_torch[-1]
             theta_p_flat = theta_p.reshape(-1)
             q_flat = A_flat @ theta_p_flat
@@ -176,18 +256,18 @@ def _MAPEM_GPU_STOP(SMatrix, y, Omega, numIterations, beta, delta, gamma, sigma,
             theta_p_plus_1_flat = torch.clamp(theta_p_plus_1_flat, min=0)
             theta_next = theta_p_plus_1_flat.reshape(Z, X)
             matrix_theta_torch[-1] = theta_next
-            if isSavingEachIteration and p in save_indices:
+            if isSavingEachIteration and it in save_indices:
                 matrix_theta_from_gpu_MAPEM.append(theta_next.cpu().numpy())
-                saved_indices.append(p+1)
+                saved_indices.append(it+1)
             log_likelihood = (y_flat * (torch.log(q_flat + torch.finfo(torch.float32).tiny)) - (q_flat + torch.finfo(torch.float32).tiny)).sum()
             penalized_log_likelihood = log_likelihood - beta * U_value
-            if p == 0 or (p + 1) % 100 == 0:
+            if it == 0 or (it + 1) % 100 == 0:
                 current = penalized_log_likelihood.item()
                 if current <= previous:
                     nb_false_successive += 1
                 else:
                     nb_false_successive = 0
-                print(f"Iter {p + 1}: lnL without term ln(m_i !) inside={log_likelihood.item():.8e}, Gibbs energy function U={U_value.item():.4e}, penalized lnL without term ln(m_i !) inside={penalized_log_likelihood.item():.8e}, p lnL (current {current:.8e} - previous {previous:.8e} > 0)={(current - previous > 0)}, nb_false_successive={nb_false_successive}")
+                print(f"Iter {it + 1}: lnL without term ln(m_i !) inside={log_likelihood.item():.8e}, Gibbs energy function U={U_value.item():.4e}, penalized lnL without term ln(m_i !) inside={penalized_log_likelihood.item():.8e}, p lnL (current {current:.8e} - previous {previous:.8e} > 0)={(current - previous > 0)}, nb_false_successive={nb_false_successive}")
                 previous = current
 
         del A_matrix_torch, y_torch, A_flat, y_flat, I_0, normalization_factor, normalization_factor_flat
@@ -202,7 +282,7 @@ def _MAPEM_GPU_STOP(SMatrix, y, Omega, numIterations, beta, delta, gamma, sigma,
         torch.cuda.empty_cache()
         return None, None
 
-def _MAPEM_CPU(SMatrix, y, Omega, numIterations, beta, delta, gamma, sigma, isSavingEachIteration, withTumor, max_saves=5000):
+def _MAPEM_CPU(SMatrix, y, Omega, beta, delta, gamma, sigma, numIterations, isSavingEachIteration, tumor_str, max_saves, show_logs=True):
     try:
         if not isinstance(Omega, PotentialType):
             raise TypeError(f"Omega must be of type PotentialType, got {type(Omega)}")
@@ -225,7 +305,6 @@ def _MAPEM_CPU(SMatrix, y, Omega, numIterations, beta, delta, gamma, sigma, isSa
             raise ValueError(f"Unknown potential type: {Omega}")
 
         T, Z, X, N = SMatrix.shape
-        J = Z * X
         A_flat = np.transpose(SMatrix, (0, 3, 1, 2)).reshape(T * N, Z * X)
         y_flat = y.reshape(-1)
         theta_0 = np.ones((Z, X), dtype=np.float32)
@@ -246,13 +325,14 @@ def _MAPEM_CPU(SMatrix, y, Omega, numIterations, beta, delta, gamma, sigma, isSa
                 save_indices.append(numIterations - 1)
 
         if Omega == PotentialType.HUBER_PIECEWISE:
-            description = f"AOT-BioMaps -- Bayesian Reconstruction Tomography: MAP-EM (Sparse HUBER β:{beta:.4f}, δ:{delta:.4f}) ---- {'WITH' if withTumor else 'WITHOUT'} TUMOR ---- processing on single CPU ----"
+            description = f"AOT-BioMaps -- Bayesian Reconstruction Tomography: MAP-EM (Sparse HUBER β:{beta:.4f}, δ:{delta:.4f}) ---- {tumor_str} TUMOR ---- processing on single CPU ----"
         elif Omega == PotentialType.RELATIVE_DIFFERENCE:
-            description = f"AOT-BioMaps -- Bayesian Reconstruction Tomography: MAP-EM (Sparse RD β:{beta:.4f}, γ:{gamma:.4f}) ---- {'WITH' if withTumor else 'WITHOUT'} TUMOR ---- processing on single CPU ----"
+            description = f"AOT-BioMaps -- Bayesian Reconstruction Tomography: MAP-EM (Sparse RD β:{beta:.4f}, γ:{gamma:.4f}) ---- {tumor_str} TUMOR ---- processing on single CPU ----"
         elif Omega == PotentialType.QUADRATIC:
-            description = f"AOT-BioMaps -- Bayesian Reconstruction Tomography: MAP-EM (Sparse QUADRATIC β:{beta:.4f}, σ:{sigma:.4f}) ---- {'WITH' if withTumor else 'WITHOUT'} TUMOR ---- processing on single CPU ----"
+            description = f"AOT-BioMaps -- Bayesian Reconstruction Tomography: MAP-EM (Sparse QUADRATIC β:{beta:.4f}, σ:{sigma:.4f}) ---- {tumor_str} TUMOR ---- processing on single CPU ----"
 
-        for p in trange(numIterations, desc=description):
+        iterator = trange(numIterations, desc=description) if show_logs else range(numIterations)
+        for it in iterator:
             theta_p = matrix_theta_np[-1]
             theta_p_flat = theta_p.reshape(-1)
             q_flat = A_flat @ theta_p_flat
@@ -270,9 +350,9 @@ def _MAPEM_CPU(SMatrix, y, Omega, numIterations, beta, delta, gamma, sigma, isSa
             theta_p_plus_1_flat = np.clip(theta_p_plus_1_flat, 0, None)
             theta_next = theta_p_plus_1_flat.reshape(Z, X)
             matrix_theta_np.append(theta_next)
-            if isSavingEachIteration and p in save_indices:
+            if isSavingEachIteration and it in save_indices:
                 I_reconMatrix.append(theta_next.copy())
-                saved_indices.append(p+1)
+                saved_indices.append(it+1)
 
         if isSavingEachIteration:
             return I_reconMatrix, saved_indices
@@ -282,7 +362,7 @@ def _MAPEM_CPU(SMatrix, y, Omega, numIterations, beta, delta, gamma, sigma, isSa
         print(f"An error occurred in _MAPEM_CPU: {e}")
         return None, None
 
-def _MAPEM_GPU(SMatrix, y, Omega, numIterations, beta, delta, gamma, sigma, isSavingEachIteration, withTumor, max_saves=5000):
+def _MAPEM_GPU(SMatrix, y, Omega, beta, delta, gamma, sigma, numIterations, isSavingEachIteration, tumor_str, device, max_saves, show_logs=True):
     """
     Maximum A Posteriori (MAP) estimation for Bayesian reconstruction using GPU.
     This method computes the MAP estimate of the parameters given the data.
@@ -308,7 +388,6 @@ def _MAPEM_GPU(SMatrix, y, Omega, numIterations, beta, delta, gamma, sigma, isSa
         else:
             raise ValueError(f"Unknown potential type: {Omega}")
 
-        device = torch.device(f"cuda:{config.select_best_gpu()}")
         A_matrix_torch = torch.tensor(SMatrix, dtype=torch.float32).to(device)
         y_torch = torch.tensor(y, dtype=torch.float32).to(device)
         T, Z, X, N = SMatrix.shape
@@ -333,13 +412,14 @@ def _MAPEM_GPU(SMatrix, y, Omega, numIterations, beta, delta, gamma, sigma, isSa
                 save_indices.append(numIterations - 1)
 
         if Omega == PotentialType.HUBER_PIECEWISE:
-            description = f"AOT-BioMaps -- Bayesian Reconstruction Tomography: MAP-EM (Sparse HUBER β:{beta:.4f}, δ:{delta:.4f}) ---- {'WITH' if withTumor else 'WITHOUT'} TUMOR ---- processing on single GPU no.{torch.cuda.current_device()} ----"
+            description = f"AOT-BioMaps -- Bayesian Reconstruction Tomography: MAP-EM (Sparse HUBER β:{beta:.4f}, δ:{delta:.4f}) ---- {tumor_str} TUMOR ---- processing on single GPU no.{torch.cuda.current_device()} ----"
         elif Omega == PotentialType.RELATIVE_DIFFERENCE:
-            description = f"AOT-BioMaps -- Bayesian Reconstruction Tomography: MAP-EM (Sparse RD β:{beta:.4f}, γ:{gamma:.4f}) ---- {'WITH' if withTumor else 'WITHOUT'} TUMOR ---- processing on single GPU no.{torch.cuda.current_device()} ----"
+            description = f"AOT-BioMaps -- Bayesian Reconstruction Tomography: MAP-EM (Sparse RD β:{beta:.4f}, γ:{gamma:.4f}) ---- {tumor_str} TUMOR ---- processing on single GPU no.{torch.cuda.current_device()} ----"
         elif Omega == PotentialType.QUADRATIC:
-            description = f"AOT-BioMaps -- Bayesian Reconstruction Tomography: MAP-EM (Sparse QUADRATIC σ:{sigma:.4f}) ---- {'WITH' if withTumor else 'WITHOUT'} TUMOR ---- processing on single GPU no.{torch.cuda.current_device()} ----"
+            description = f"AOT-BioMaps -- Bayesian Reconstruction Tomography: MAP-EM (Sparse QUADRATIC σ:{sigma:.4f}) ---- {tumor_str} TUMOR ---- processing on single GPU no.{torch.cuda.current_device()} ----"
 
-        for p in trange(numIterations, desc=description):
+        iterator = trange(numIterations, desc=description) if show_logs else range(numIterations)
+        for it in iterator:
             theta_p = matrix_theta_torch[-1]
             theta_p_flat = theta_p.reshape(-1)
             q_flat = A_flat @ theta_p_flat
@@ -359,9 +439,9 @@ def _MAPEM_GPU(SMatrix, y, Omega, numIterations, beta, delta, gamma, sigma, isSa
             theta_p_plus_1_flat = torch.clamp(theta_p_plus_1_flat, min=0)
             theta_next = theta_p_plus_1_flat.reshape(Z, X)
             matrix_theta_torch.append(theta_next)
-            if isSavingEachIteration and p in save_indices:
+            if isSavingEachIteration and it in save_indices:
                 I_reconMatrix.append(theta_next.cpu().numpy())
-                saved_indices.append(p+1)
+                saved_indices.append(it+1)
 
         del A_matrix_torch, y_torch, A_flat, y_flat, theta_0, normalization_factor, normalization_factor_flat
         torch.cuda.empty_cache()