Model Analysis

Neural-Scope provides comprehensive model analysis capabilities to help you understand your models better.

Basic Analysis

The basic analysis provides information about the model’s architecture, parameters, memory usage, and inference time.

from neural_scope import NeuralScope
import torch

# Load a model
model = torch.hub.load('pytorch/vision:v0.10.0', 'resnet18', pretrained=True)

# Create Neural-Scope instance
neural_scope = NeuralScope()

# Analyze the model
results = neural_scope.analyze_model(
    model=model,
    model_name="resnet18",
    framework="pytorch"
)

# Print results
print(f"Parameters: {results['parameters']}")
print(f"Layers: {results['layers']}")
print(f"Memory usage: {results['memory_usage_mb']} MB")
print(f"Inference time: {results['inference_time_ms']} ms")

Analysis Results

The analysis results include:

Parameters: The total number of trainable parameters in the model
Layers: The total number of layers in the model
Architecture: The architecture of the model
Memory Usage: The amount of memory required to store the model weights
Inference Time: The average time taken to perform a forward pass on a single input

Analyzing Pre-trained Models

Neural-Scope supports analyzing pre-trained models from various sources:

PyTorch Hub

from neural_scope import NeuralScope
from model_sources.pytorch_hub import PyTorchHubSource

# Create source
source = PyTorchHubSource()

# Fetch model
model_path = source.fetch_model("resnet18", "models")

# Load model
import torch
model = torch.load(model_path)

# Analyze model
neural_scope = NeuralScope()
results = neural_scope.analyze_model(
    model=model,
    model_name="resnet18",
    framework="pytorch"
)

TensorFlow Hub

from neural_scope import NeuralScope
from model_sources.tensorflow_hub import TensorFlowHubSource

# Create source
source = TensorFlowHubSource()

# Fetch model
model_path = source.fetch_model("efficientnet_b0", "models")

# Load model
import tensorflow as tf
model = tf.saved_model.load(model_path)

# Analyze model
neural_scope = NeuralScope()
results = neural_scope.analyze_model(
    model=model,
    model_name="efficientnet_b0",
    framework="tensorflow"
)

Hugging Face

from neural_scope import NeuralScope
from model_sources.huggingface import HuggingFaceSource

# Create source
source = HuggingFaceSource()

# Fetch model
model_path = source.fetch_model("bert-base-uncased", "models")

# Load model
from transformers import AutoModel
model = AutoModel.from_pretrained(model_path)

# Analyze model
neural_scope = NeuralScope()
results = neural_scope.analyze_model(
    model=model,
    model_name="bert-base-uncased",
    framework="pytorch"
)

Generating Reports

Neural-Scope can generate comprehensive reports in JSON and HTML formats:

from neural_scope import NeuralScope
import torch

# Load a model
model = torch.hub.load('pytorch/vision:v0.10.0', 'resnet18', pretrained=True)

# Create Neural-Scope instance
neural_scope = NeuralScope()

# Analyze the model and generate reports
results = neural_scope.analyze_model(
    model=model,
    model_name="resnet18",
    framework="pytorch",
    output_dir="results",
    generate_reports=True
)

# Print report paths
print(f"JSON report: {results['report_path']}")
print(f"HTML report: {results['html_report_path']}")

Command Line Interface

Neural-Scope provides a command-line interface for model analysis:

neural-scope analyze \
    --model-path models/model.pt \
    --framework pytorch \
    --output-dir results \
    --generate-reports