coreml-complexity-analyzer

FLOPS estimation, memory analysis, and per-layer profiling for Core ML models.

Why This Exists

Before deploying Core ML models to Apple devices, developers need to understand:

• Compute cost: How many FLOPS/MACs does the model require?
• Memory footprint: How much memory for parameters and activations?
• Bottleneck layers: Which layers dominate compute and memory?

coremltools does not provide these analysis capabilities out of the box. This package fills that gap.

Installation

pip install coreml-complexity-analyzer

Requirements: coremltools >= 7.0, numpy

Quick Start

import coremltools as ct
from coreml_complexity_analyzer import generate_report

model = ct.models.MLModel("my_model.mlpackage")
report = generate_report(model, "ResNet50")

print(f"Total GFLOPS: {report.total_gflops:.2f}")
print(f"Parameters: {report.parameters_millions:.2f}M")
print(f"Memory: {report.memory_breakdown.total_mb:.2f} MB")

# Full markdown report
print(report.to_markdown())

API

FLOPSAnalyzer

Compute floating-point operations for each layer in a model.

from coreml_complexity_analyzer import FLOPSAnalyzer

analyzer = FLOPSAnalyzer(model)
total = analyzer.get_total_flops()
breakdown = analyzer.get_layer_breakdown()
by_type = analyzer.get_flops_by_op_type()

MemoryEstimator

Estimate memory requirements (parameters + activations + overhead).

from coreml_complexity_analyzer import MemoryEstimator

estimator = MemoryEstimator(model)
breakdown = estimator.estimate()
param_count = estimator.get_parameter_count()

print(f"Total: {breakdown.total_mb:.2f} MB")
print(f"Parameters: {breakdown.parameter_mb:.2f} MB")

LayerProfiler

Detailed per-layer analysis combining FLOPS, memory, and shape information.

from coreml_complexity_analyzer import LayerProfiler

profiler = LayerProfiler(model)
profiles = profiler.profile()

top_layers = profiler.get_top_layers(n=5, by="flops")
for layer in top_layers:
    print(f"{layer.name}: {layer.mflops:.2f} MFLOPS")

generate_report

Generate comprehensive reports combining all analyses.

from coreml_complexity_analyzer import generate_report

report = generate_report(model, "ModelName")

# Output formats
print(report.to_markdown())   # Markdown table
print(report.to_text())       # Plain text
data = report.to_dict()       # Dictionary

Supported Operations

Category	Operations
Convolutions	conv, conv_transpose
Linear	linear, matmul, einsum
Activations	relu, sigmoid, tanh, gelu, silu, softplus
Normalization	batch_norm, layer_norm, instance_norm
Pooling	max_pool, avg_pool
Element-wise	add, sub, mul, div
Reductions	reduce_sum, reduce_mean, reduce_max, reduce_min
Other	softmax

Example Output

============================================================
Model Complexity Report: ResNet50
============================================================
SUMMARY
----------------------------------------
  Total FLOPS:            4,089,184,256
  Total GFLOPS:                    4.09
  Parameters:                25,557,032
  Parameters (M):                 25.56

MEMORY ANALYSIS
----------------------------------------
  Total Memory:              106.42 MB
  Parameters:                 97.52 MB

TOP OPERATIONS BY FLOPS
----------------------------------------
  conv                 3,891,200,000 ( 95.2%)
  linear                 102,400,000 (  2.5%)
  batch_norm              51,200,000 (  1.3%)
============================================================

Use Cases

• Model Optimization: Identify bottleneck layers before applying compression
• Deployment Planning: Estimate if a model fits device constraints
• Model Comparison: Compare efficiency across architectures
• Research: Analyze compute/memory trade-offs

Development

git clone https://github.com/yaswanth169/coreml-complexity-analyzer.git
cd coreml-complexity-analyzer
pip install -e ".[dev]"
pytest

License

BSD-3-Clause

Authors

• Devavarapu Yashwanth
• Ireddi Rakshitha