Profiling Tools

The profiling API provides a comprehensive suite of tools for measuring various aspects of model performance, including memory usage, computational complexity, and runtime. This documentation covers the profiling functions available in the Metinor package for deep learning models.

To begin profiling your model, first import the necessary functions from the profiler:

from nervai_optim.profiler.functional import profile_metrics, profile_macs, profile_flops, profile_runtime, profile_memory, profile_madds, profile_params
from nervai_optim.profiler.metrics import MetricType
from nervai_optim.profiler import ProfilerReport, Profiler
from nervai_optim.profiler.analysis import ReportAnalyzer

• Profiling Metrics
• Profiler Reports
• Analyzing Profiling Results
• Support Matrix
  • Linear Layers
  • Pooling Layers
  • Normalization Layers
  • Activation Layers
  • Recurrent Layers
  • Transformer Layers
  • Miscellaneous Layers
• Tutorial: Profiling a PyTorch Model
  • Create a model
  • Import the profiler
    • Profiling a model using the functional API
    • Profiling a model using the profiler class
  • Analyzing the results
    • Find layers with the highest values of a metric
    • Determine relative impact of each layer on a metric
    • Filter layers by type
    • Analyze memory usage
  • Plotting the results