gengeneeval 0.2.0__tar.gz → 0.3.0__tar.gz

{gengeneeval-0.2.0 → gengeneeval-0.3.0}/PKG-INFO

@@ -1,10 +1,10 @@
 Metadata-Version: 2.4
 Name: gengeneeval
-Version: 0.2.0
-Summary: Comprehensive evaluation of generated gene expression data. Computes metrics between real and generated datasets with support for condition matching, train/test splits, and publication-quality visualizations.
+Version: 0.3.0
+Summary: Comprehensive evaluation of generated gene expression data. Computes metrics between real and generated datasets with support for condition matching, train/test splits, memory-efficient lazy loading, CPU parallelization, GPU acceleration, and publication-quality visualizations.
 License: MIT
 License-File: LICENSE
-Keywords: gene expression,evaluation,metrics,single-cell,generative models,benchmarking
+Keywords: gene expression,evaluation,metrics,single-cell,generative models,benchmarking,memory-efficient
 Author: GenEval Team
 Author-email: geneval@example.com
 Requires-Python: >=3.8,<4.0
@@ -24,6 +24,7 @@ Provides-Extra: full
 Provides-Extra: gpu
 Requires-Dist: anndata (>=0.8.0)
 Requires-Dist: geomloss (>=0.2.1) ; extra == "full" or extra == "gpu"
+Requires-Dist: joblib (>=1.0.0)
 Requires-Dist: matplotlib (>=3.5.0)
 Requires-Dist: numpy (>=1.21.0)
 Requires-Dist: pandas (>=1.3.0)
@@ -46,7 +47,7 @@ Description-Content-Type: text/markdown
 
 **Comprehensive evaluation of generated gene expression data against real datasets.**
 
-GenEval is a modular, object-oriented Python framework for computing metrics between real and generated gene expression datasets stored in AnnData (h5ad) format. It supports condition-based matching, train/test splits, and generates publication-quality visualizations.
+GenEval is a modular, object-oriented Python framework for computing metrics between real and generated gene expression datasets stored in AnnData (h5ad) format. It supports condition-based matching, train/test splits, memory-efficient lazy loading for large datasets, and generates publication-quality visualizations.
 
 ## Features
 
@@ -77,6 +78,10 @@ All metrics are computed **per-gene** (returning a vector) and **aggregated**:
 - ✅ Condition-based matching (perturbation, cell type, etc.)
 - ✅ Train/test split support
 - ✅ Per-gene and aggregate metrics
+- ✅ **Memory-efficient lazy loading** for large datasets
+- ✅ **Batched evaluation** to avoid OOM errors
+- ✅ **CPU parallelization** via joblib (multi-core speedup)
+- ✅ **GPU acceleration** via PyTorch (10-100x speedup)
 - ✅ Modular, extensible architecture
 - ✅ Command-line interface
 - ✅ Publication-quality visualizations
@@ -140,6 +145,108 @@ geneval --real real.h5ad --generated generated.h5ad \
         --output results/
 ```
 
+### Memory-Efficient Mode (for Large Datasets)
+
+For datasets too large to fit in memory, use the lazy evaluation API:
+
+```python
+from geneval import evaluate_lazy, load_data_lazy
+
+# Memory-efficient evaluation (streams data one condition at a time)
+results = evaluate_lazy(
+    real_path="large_real.h5ad",
+    generated_path="large_generated.h5ad",
+    condition_columns=["perturbation"],
+    batch_size=256,                    # Process in batches
+    use_backed=True,                   # Memory-mapped file access
+    output_dir="eval_output/",
+    save_per_condition=True,           # Save each condition to disk
+)
+
+# Get summary statistics
+print(results.get_summary())
+
+# Or use the lazy loader directly for custom workflows
+with load_data_lazy("real.h5ad", "gen.h5ad", ["perturbation"]) as loader:
+    print(f"Memory estimate: {loader.estimate_memory_usage()}")
+    
+    # Process one condition at a time
+    for key, real, gen, info in loader.iterate_conditions():
+        # Your custom evaluation logic
+        pass
+```
+
+### Accelerated Evaluation (CPU Parallelization & GPU)
+
+GenEval supports CPU parallelization and GPU acceleration for significant speedups:
+
+```python
+from geneval import evaluate, get_available_backends
+
+# Check available backends
+print(get_available_backends())
+# {'joblib': True, 'torch': True, 'geomloss': True, 'cuda': True, 'mps': False}
+
+# Parallel CPU evaluation (use all cores)
+results = evaluate(
+    real_path="real.h5ad",
+    generated_path="generated.h5ad",
+    condition_columns=["perturbation"],
+    n_jobs=-1,  # Use all available CPU cores
+)
+
+# GPU-accelerated evaluation
+results = evaluate(
+    real_path="real.h5ad",
+    generated_path="generated.h5ad",
+    condition_columns=["perturbation"],
+    device="cuda",  # Use NVIDIA GPU
+)
+
+# Combined: parallel CPU + auto device selection
+results = evaluate(..., n_jobs=8, device="auto")
+```
+
+#### Low-level Accelerated API
+
+For custom workflows, use the accelerated metrics directly:
+
+```python
+from geneval.metrics.accelerated import (
+    compute_metrics_accelerated,
+    GPUWasserstein1,
+    GPUMMD,
+    vectorized_wasserstein1,
+)
+import numpy as np
+
+# Load your data
+real = np.random.randn(1000, 5000)      # 1000 cells, 5000 genes
+generated = np.random.randn(1000, 5000)
+
+# Compute multiple metrics with acceleration
+results = compute_metrics_accelerated(
+    real, generated,
+    metrics=["wasserstein_1", "wasserstein_2", "mmd", "energy"],
+    n_jobs=8,          # CPU parallelization
+    device="cuda",     # GPU acceleration
+    verbose=True,
+)
+
+# Access results
+print(f"W1: {results['wasserstein_1'].aggregate_value:.4f}")
+print(f"MMD: {results['mmd'].aggregate_value:.4f}")
+```
+
+#### Performance Tips
+
+| Optimization | Speedup | When to Use |
+|--------------|---------|-------------|
+| `n_jobs=-1` (all cores) | 4-16x | Always (if joblib available) |
+| `device="cuda"` | 10-100x | Large datasets, NVIDIA GPU available |
+| `device="mps"` | 5-20x | Apple Silicon Macs |
+| Vectorized NumPy | 2-5x | Automatic fallback |
+
 ## Expected Data Format
 
 GenEval expects AnnData (h5ad) files with:

{gengeneeval-0.2.0 → gengeneeval-0.3.0}/README.md

@@ -7,7 +7,7 @@
 
 **Comprehensive evaluation of generated gene expression data against real datasets.**
 
-GenEval is a modular, object-oriented Python framework for computing metrics between real and generated gene expression datasets stored in AnnData (h5ad) format. It supports condition-based matching, train/test splits, and generates publication-quality visualizations.
+GenEval is a modular, object-oriented Python framework for computing metrics between real and generated gene expression datasets stored in AnnData (h5ad) format. It supports condition-based matching, train/test splits, memory-efficient lazy loading for large datasets, and generates publication-quality visualizations.
 
 ## Features
 
@@ -38,6 +38,10 @@ All metrics are computed **per-gene** (returning a vector) and **aggregated**:
 - ✅ Condition-based matching (perturbation, cell type, etc.)
 - ✅ Train/test split support
 - ✅ Per-gene and aggregate metrics
+- ✅ **Memory-efficient lazy loading** for large datasets
+- ✅ **Batched evaluation** to avoid OOM errors
+- ✅ **CPU parallelization** via joblib (multi-core speedup)
+- ✅ **GPU acceleration** via PyTorch (10-100x speedup)
 - ✅ Modular, extensible architecture
 - ✅ Command-line interface
 - ✅ Publication-quality visualizations
@@ -101,6 +105,108 @@ geneval --real real.h5ad --generated generated.h5ad \
         --output results/
 ```
 
+### Memory-Efficient Mode (for Large Datasets)
+
+For datasets too large to fit in memory, use the lazy evaluation API:
+
+```python
+from geneval import evaluate_lazy, load_data_lazy
+
+# Memory-efficient evaluation (streams data one condition at a time)
+results = evaluate_lazy(
+    real_path="large_real.h5ad",
+    generated_path="large_generated.h5ad",
+    condition_columns=["perturbation"],
+    batch_size=256,                    # Process in batches
+    use_backed=True,                   # Memory-mapped file access
+    output_dir="eval_output/",
+    save_per_condition=True,           # Save each condition to disk
+)
+
+# Get summary statistics
+print(results.get_summary())
+
+# Or use the lazy loader directly for custom workflows
+with load_data_lazy("real.h5ad", "gen.h5ad", ["perturbation"]) as loader:
+    print(f"Memory estimate: {loader.estimate_memory_usage()}")
+    
+    # Process one condition at a time
+    for key, real, gen, info in loader.iterate_conditions():
+        # Your custom evaluation logic
+        pass
+```
+
+### Accelerated Evaluation (CPU Parallelization & GPU)
+
+GenEval supports CPU parallelization and GPU acceleration for significant speedups:
+
+```python
+from geneval import evaluate, get_available_backends
+
+# Check available backends
+print(get_available_backends())
+# {'joblib': True, 'torch': True, 'geomloss': True, 'cuda': True, 'mps': False}
+
+# Parallel CPU evaluation (use all cores)
+results = evaluate(
+    real_path="real.h5ad",
+    generated_path="generated.h5ad",
+    condition_columns=["perturbation"],
+    n_jobs=-1,  # Use all available CPU cores
+)
+
+# GPU-accelerated evaluation
+results = evaluate(
+    real_path="real.h5ad",
+    generated_path="generated.h5ad",
+    condition_columns=["perturbation"],
+    device="cuda",  # Use NVIDIA GPU
+)
+
+# Combined: parallel CPU + auto device selection
+results = evaluate(..., n_jobs=8, device="auto")
+```
+
+#### Low-level Accelerated API
+
+For custom workflows, use the accelerated metrics directly:
+
+```python
+from geneval.metrics.accelerated import (
+    compute_metrics_accelerated,
+    GPUWasserstein1,
+    GPUMMD,
+    vectorized_wasserstein1,
+)
+import numpy as np
+
+# Load your data
+real = np.random.randn(1000, 5000)      # 1000 cells, 5000 genes
+generated = np.random.randn(1000, 5000)
+
+# Compute multiple metrics with acceleration
+results = compute_metrics_accelerated(
+    real, generated,
+    metrics=["wasserstein_1", "wasserstein_2", "mmd", "energy"],
+    n_jobs=8,          # CPU parallelization
+    device="cuda",     # GPU acceleration
+    verbose=True,
+)
+
+# Access results
+print(f"W1: {results['wasserstein_1'].aggregate_value:.4f}")
+print(f"MMD: {results['mmd'].aggregate_value:.4f}")
+```
+
+#### Performance Tips
+
+| Optimization | Speedup | When to Use |
+|--------------|---------|-------------|
+| `n_jobs=-1` (all cores) | 4-16x | Always (if joblib available) |
+| `device="cuda"` | 10-100x | Large datasets, NVIDIA GPU available |
+| `device="mps"` | 5-20x | Apple Silicon Macs |
+| Vectorized NumPy | 2-5x | Automatic fallback |
+
 ## Expected Data Format
 
 GenEval expects AnnData (h5ad) files with:

{gengeneeval-0.2.0 → gengeneeval-0.3.0}/pyproject.toml

@@ -1,13 +1,13 @@
 [tool.poetry]
 name = "gengeneeval"
-version = "0.2.0"
-description = "Comprehensive evaluation of generated gene expression data. Computes metrics between real and generated datasets with support for condition matching, train/test splits, and publication-quality visualizations."
+version = "0.3.0"
+description = "Comprehensive evaluation of generated gene expression data. Computes metrics between real and generated datasets with support for condition matching, train/test splits, memory-efficient lazy loading, CPU parallelization, GPU acceleration, and publication-quality visualizations."
 authors = ["GenEval Team <geneval@example.com>"]
 license = "MIT"
 readme = "README.md"
 homepage = "https://github.com/AndreaRubbi/GenGeneEval"
 repository = "https://github.com/AndreaRubbi/GenGeneEval"
-keywords = ["gene expression", "evaluation", "metrics", "single-cell", "generative models", "benchmarking"]
+keywords = ["gene expression", "evaluation", "metrics", "single-cell", "generative models", "benchmarking", "memory-efficient"]
 classifiers = [
     "Development Status :: 4 - Beta",
     "Intended Audience :: Science/Research",
@@ -29,6 +29,7 @@ scipy = ">=1.7.0"
 torch = ">=1.9.0"
 matplotlib = ">=3.5.0"
 seaborn = ">=0.11.0"
+joblib = ">=1.0.0"
 geomloss = {version = ">=0.2.1", optional = true}
 pykeops = {version = ">=1.4.0", optional = true}
 umap-learn = {version = ">=0.5.0", optional = true}

{gengeneeval-0.2.0 → gengeneeval-0.3.0}/src/geneval/__init__.py

@@ -8,6 +8,7 @@ Features:
 - Multiple distance and correlation metrics (per-gene and aggregate)
 - Condition-based matching (perturbation, cell type, etc.)
 - Train/test split support
+- Memory-efficient lazy loading for large datasets
 - Publication-quality visualizations
 - Command-line interface
 
@@ -20,12 +21,22 @@ Quick Start:
     ...     output_dir="output/"
     ... )
 
+Memory-Efficient Mode (for large datasets):
+    >>> from geneval import evaluate_lazy
+    >>> results = evaluate_lazy(
+    ...     real_path="real.h5ad",
+    ...     generated_path="generated.h5ad",
+    ...     condition_columns=["perturbation"],
+    ...     batch_size=256,
+    ...     use_backed=True,  # Memory-mapped access
+    ... )
+
 CLI Usage:
     $ geneval --real real.h5ad --generated generated.h5ad \\
               --conditions perturbation cell_type --output results/
 """
 
-__version__ = "0.2.0"
+__version__ = "0.3.0"
 __author__ = "GenEval Team"
 
 # Main evaluation interface
@@ -35,12 +46,26 @@ from .evaluator import (
     MetricRegistry,
 )
 
+# Memory-efficient evaluation
+from .lazy_evaluator import (
+    evaluate_lazy,
+    MemoryEfficientEvaluator,
+    StreamingEvaluationResult,
+)
+
 # Data loading
 from .data.loader import (
     GeneExpressionDataLoader,
     load_data,
 )
 
+# Memory-efficient data loading
+from .data.lazy_loader import (
+    LazyGeneExpressionDataLoader,
+    load_data_lazy,
+    ConditionBatch,
+)
+
 # Results
 from .results import (
     EvaluationResult,
@@ -76,6 +101,14 @@ from .metrics.reconstruction import (
     R2Score,
 )
 
+# Accelerated computation
+from .metrics.accelerated import (
+    AccelerationConfig,
+    ParallelMetricComputer,
+    get_available_backends,
+    compute_metrics_accelerated,
+)
+
 # Visualization
 from .visualization.visualizer import (
     EvaluationVisualizer,
@@ -99,9 +132,17 @@ __all__ = [
     "evaluate",
     "GeneEvalEvaluator",
     "MetricRegistry",
+    # Memory-efficient evaluation
+    "evaluate_lazy",
+    "MemoryEfficientEvaluator",
+    "StreamingEvaluationResult",
     # Data loading
     "GeneExpressionDataLoader",
     "load_data",
+    # Memory-efficient data loading
+    "LazyGeneExpressionDataLoader",
+    "load_data_lazy",
+    "ConditionBatch",
     # Results
     "EvaluationResult",
     "SplitResult", 
@@ -123,6 +164,16 @@ __all__ = [
     "EnergyDistance",
     "MultivariateWasserstein",
     "MultivariateMMD",
+    # Reconstruction metrics
+    "MSEDistance",
+    "RMSEDistance",
+    "MAEDistance",
+    "R2Score",
+    # Acceleration
+    "AccelerationConfig",
+    "ParallelMetricComputer",
+    "get_available_backends",
+    "compute_metrics_accelerated",
     # Visualization
     "EvaluationVisualizer",
     "visualize",

{gengeneeval-0.2.0 → gengeneeval-0.3.0}/src/geneval/data/__init__.py

@@ -2,6 +2,7 @@
 Data loading module for gene expression evaluation.
 
 Provides data loaders for paired real and generated datasets.
+Includes both standard and memory-efficient lazy loading options.
 """
 
 from .loader import (
@@ -9,15 +10,28 @@ from .loader import (
     load_data,
     DataLoaderError,
 )
+from .lazy_loader import (
+    LazyGeneExpressionDataLoader,
+    load_data_lazy,
+    LazyDataLoaderError,
+    ConditionBatch,
+)
 from .gene_expression_datamodule import (
     GeneExpressionDataModule,
     DataModuleError,
 )
 
 __all__ = [
+    # Standard loader
     "GeneExpressionDataLoader",
     "load_data",
     "DataLoaderError",
+    # Lazy loader (memory-efficient)
+    "LazyGeneExpressionDataLoader",
+    "load_data_lazy",
+    "LazyDataLoaderError",
+    "ConditionBatch",
+    # DataModule
     "GeneExpressionDataModule",
     "DataModuleError",
 ]