fasterbench
Overview
fasterbench is a comprehensive benchmarking library for PyTorch models that helps AI researchers and engineers evaluate model performance across five critical dimensions:
- Size: Model disk size and parameter count
- Speed: Latency and throughput on both GPU and CPU
- Compute: MACs (multiply-accumulate operations)
- Memory: Peak and average memory consumption
- Energy: Power consumption and carbon emissions
Whether you’re optimizing for edge deployment, comparing model architectures, or researching model efficiency, FasterBench provides the metrics you need with minimal setup.
Installation
pip install fasterbenchQuick Start
import torch
from torchvision.models import resnet18
from fasterbench import benchmark
# Load your model
model = resnet18()
dummy_input = torch.randn(1, 3, 224, 224)
# Run comprehensive benchmarks
result = benchmark(model, dummy_input, metrics=["size", "speed", "compute"])
# Typed access (IDE autocomplete!)
print(f"Size: {result.size.size_mib:.2f} MiB")
print(f"Params: {result.size.num_params:,}")
print(f"CPU Latency: {result.speed['cpu'].mean_ms:.2f} ms")
print(f"MACs: {result.compute.macs_m}M")
# Backward-compatible dict access still works
print(result["size_disk_bytes"])Features
BenchmarkResult: Typed Access + Backward Compatibility
The benchmark() function returns a BenchmarkResult object with both typed attribute access and dict-like access:
result = benchmark(model, dummy_input, metrics=["size", "speed", "compute"])
# Typed access - IDE autocomplete and type checking
result.size.size_mib # 44.59
result.size.num_params # 11689512
result.speed["cpu"].mean_ms # 45.23
result.speed["cpu"].throughput_s # 22.1
result.compute.macs_m # 1819.066
result.compute.macs_available # True (False if MACs couldn't be computed)
# Dict access - backward compatible with existing code
result["size_size_mib"] # 44.59
result["speed_cpu_mean_ms"] # 45.23
for key, value in result.items():
print(f"{key}: {value}")Human-Readable Summary
Get a quick overview with formatted output:
result.summary() # prints directly═══ Size ════════════════════════════════════
Disk: 44.59 MiB
Params: 11.69M
═══ Speed ═══════════════════════════════════
cpu: 45.23 ms │ 22.1 inf/s │ p99: 48.12 ms
═══ Compute ═════════════════════════════════
MACs: 1819.1 M
Params: 11.69 M
Export to DataFrame or JSON
# Convert to pandas DataFrame for analysis
df = result.to_dataframe()
df.to_csv("benchmark_results.csv")
# Serialize to JSON
json_str = result.to_json()
# Get formatted string (for logging, etc.)
summary_str = str(result)Selective Metrics
Only compute what you need:
# Fast: just size and compute (no inference runs)
result = benchmark(model, dummy_input, metrics=["size", "compute"])
# Full benchmark on specific devices
result = benchmark(model, dummy_input,
speed_devices=["cpu", "cuda"],
memory_devices=["cuda"])Individual Metric Functions
For fine-grained control, use the individual compute functions:
from fasterbench import compute_size, compute_speed_multi, compute_compute
# Size metrics
size = compute_size(model)
print(f"Disk Size: {size.size_mib:.2f} MiB")
print(f"Parameters: {size.num_params:,}")
# Speed metrics across devices
speed = compute_speed_multi(model, dummy_input)
for device, metrics in speed.items():
print(f"{device}: {metrics.mean_ms:.2f}ms, {metrics.throughput_s:.1f} inf/s")
# Compute metrics (MACs)
compute = compute_compute(model, dummy_input)
if compute.macs_available:
print(f"MACs: {compute.macs_m}M")Memory Metrics
Profile memory usage on CPU and GPU:
from fasterbench import compute_memory_multi
memory = compute_memory_multi(model, dummy_input)
for device, metrics in memory.items():
print(f"{device} peak: {metrics.peak_mib:.2f} MiB")Energy Metrics
Measure power consumption and carbon footprint (requires codecarbon):
from fasterbench import compute_energy_multi
energy = compute_energy_multi(model, dummy_input)
for device, metrics in energy.items():
print(f"{device}: {metrics.mean_watts:.1f}W, {metrics.co2_eq_g:.4f}g CO₂/inf")Thread Sweep for CPU Optimization
Find the optimal thread count for CPU inference:
from fasterbench import sweep_threads
results = sweep_threads(model, dummy_input, thread_counts=[1, 2, 4, 8])
for r in results:
print(f"{r['threads']} threads: {r['mean_ms']:.2f}ms")Visualize Results
Create radar plots to compare multiple models:
from fasterbench.benchmark import benchmark
from fasterbench.plot import *
from torchvision.models import resnet18, mobilenet_v3_large
import torch
dummy = torch.randn(8,3,224,224)
resnet = benchmark(resnet18(), dummy,
metrics=("size","speed","compute","energy"))
mobilenet= benchmark(mobilenet_v3_large(), dummy,
metrics=("size","speed","compute","energy"))fig = create_radar_plot([resnet, mobilenet],
model_names=["ResNet-18", "MobileNet-V3"])
fig.show()
Documentation
For more detailed usage examples and API documentation, visit our documentation.
Contributing
Contributions are welcome! Please feel free to submit a Pull Request.
License
This project is licensed under the Apache 2.0 License - see the LICENSE file for details.