6ed69a3d48
Major changes: - Organized Python v1 implementation into v1/ subdirectory - Created Rust workspace with 9 modular crates: - wifi-densepose-core: Core types, traits, errors - wifi-densepose-signal: CSI processing, phase sanitization, FFT - wifi-densepose-nn: Neural network inference (ONNX/Candle/tch) - wifi-densepose-api: Axum-based REST/WebSocket API - wifi-densepose-db: SQLx database layer - wifi-densepose-config: Configuration management - wifi-densepose-hardware: Hardware abstraction - wifi-densepose-wasm: WebAssembly bindings - wifi-densepose-cli: Command-line interface Documentation: - ADR-001: Workspace structure - ADR-002: Signal processing library selection - ADR-003: Neural network inference strategy - DDD domain model with bounded contexts Testing: - 69 tests passing across all crates - Signal processing: 45 tests - Neural networks: 21 tests - Core: 3 doc tests Performance targets: - 10x faster CSI processing (~0.5ms vs ~5ms) - 5x lower memory usage (~100MB vs ~500MB) - WASM support for browser deployment
367 lines
14 KiB
Python
367 lines
14 KiB
Python
import pytest
|
|
import torch
|
|
import torch.nn as nn
|
|
import numpy as np
|
|
from unittest.mock import Mock, patch
|
|
from src.models.densepose_head import DensePoseHead, DensePoseError
|
|
|
|
|
|
class TestDensePoseHead:
|
|
"""Test suite for DensePose Head following London School TDD principles"""
|
|
|
|
@pytest.fixture
|
|
def mock_config(self):
|
|
"""Configuration for DensePose head"""
|
|
return {
|
|
'input_channels': 256,
|
|
'num_body_parts': 24,
|
|
'num_uv_coordinates': 2,
|
|
'hidden_channels': [128, 64],
|
|
'kernel_size': 3,
|
|
'padding': 1,
|
|
'dropout_rate': 0.1,
|
|
'use_deformable_conv': False,
|
|
'use_fpn': True,
|
|
'fpn_levels': [2, 3, 4, 5],
|
|
'output_stride': 4
|
|
}
|
|
|
|
@pytest.fixture
|
|
def densepose_head(self, mock_config):
|
|
"""Create DensePose head instance for testing"""
|
|
return DensePoseHead(mock_config)
|
|
|
|
@pytest.fixture
|
|
def mock_feature_input(self):
|
|
"""Generate mock feature input tensor"""
|
|
batch_size = 2
|
|
channels = 256
|
|
height = 56
|
|
width = 56
|
|
return torch.randn(batch_size, channels, height, width)
|
|
|
|
@pytest.fixture
|
|
def mock_target_masks(self):
|
|
"""Generate mock target segmentation masks"""
|
|
batch_size = 2
|
|
num_parts = 24
|
|
height = 224
|
|
width = 224
|
|
return torch.randint(0, num_parts + 1, (batch_size, height, width))
|
|
|
|
@pytest.fixture
|
|
def mock_target_uv(self):
|
|
"""Generate mock target UV coordinates"""
|
|
batch_size = 2
|
|
num_coords = 2
|
|
height = 224
|
|
width = 224
|
|
return torch.randn(batch_size, num_coords, height, width)
|
|
|
|
def test_head_initialization_creates_correct_architecture(self, mock_config):
|
|
"""Test that DensePose head initializes with correct architecture"""
|
|
# Act
|
|
head = DensePoseHead(mock_config)
|
|
|
|
# Assert
|
|
assert head is not None
|
|
assert isinstance(head, nn.Module)
|
|
assert head.input_channels == mock_config['input_channels']
|
|
assert head.num_body_parts == mock_config['num_body_parts']
|
|
assert head.num_uv_coordinates == mock_config['num_uv_coordinates']
|
|
assert head.use_fpn == mock_config['use_fpn']
|
|
assert hasattr(head, 'segmentation_head')
|
|
assert hasattr(head, 'uv_regression_head')
|
|
if mock_config['use_fpn']:
|
|
assert hasattr(head, 'fpn')
|
|
|
|
def test_forward_pass_produces_correct_output_format(self, densepose_head, mock_feature_input):
|
|
"""Test that forward pass produces correctly formatted output"""
|
|
# Act
|
|
output = densepose_head(mock_feature_input)
|
|
|
|
# Assert
|
|
assert output is not None
|
|
assert isinstance(output, dict)
|
|
assert 'segmentation' in output
|
|
assert 'uv_coordinates' in output
|
|
|
|
seg_output = output['segmentation']
|
|
uv_output = output['uv_coordinates']
|
|
|
|
assert isinstance(seg_output, torch.Tensor)
|
|
assert isinstance(uv_output, torch.Tensor)
|
|
assert seg_output.shape[0] == mock_feature_input.shape[0] # Batch size preserved
|
|
assert uv_output.shape[0] == mock_feature_input.shape[0] # Batch size preserved
|
|
|
|
def test_segmentation_head_produces_correct_shape(self, densepose_head, mock_feature_input):
|
|
"""Test that segmentation head produces correct output shape"""
|
|
# Act
|
|
output = densepose_head(mock_feature_input)
|
|
seg_output = output['segmentation']
|
|
|
|
# Assert
|
|
expected_channels = densepose_head.num_body_parts + 1 # +1 for background
|
|
assert seg_output.shape[1] == expected_channels
|
|
assert seg_output.shape[2] >= mock_feature_input.shape[2] # Height upsampled
|
|
assert seg_output.shape[3] >= mock_feature_input.shape[3] # Width upsampled
|
|
|
|
def test_uv_regression_head_produces_correct_shape(self, densepose_head, mock_feature_input):
|
|
"""Test that UV regression head produces correct output shape"""
|
|
# Act
|
|
output = densepose_head(mock_feature_input)
|
|
uv_output = output['uv_coordinates']
|
|
|
|
# Assert
|
|
assert uv_output.shape[1] == densepose_head.num_uv_coordinates
|
|
assert uv_output.shape[2] >= mock_feature_input.shape[2] # Height upsampled
|
|
assert uv_output.shape[3] >= mock_feature_input.shape[3] # Width upsampled
|
|
|
|
def test_compute_segmentation_loss_measures_pixel_classification(self, densepose_head, mock_feature_input, mock_target_masks):
|
|
"""Test that compute_segmentation_loss measures pixel classification accuracy"""
|
|
# Arrange
|
|
output = densepose_head(mock_feature_input)
|
|
seg_logits = output['segmentation']
|
|
|
|
# Resize target to match output
|
|
target_resized = torch.nn.functional.interpolate(
|
|
mock_target_masks.float().unsqueeze(1),
|
|
size=seg_logits.shape[2:],
|
|
mode='nearest'
|
|
).squeeze(1).long()
|
|
|
|
# Act
|
|
loss = densepose_head.compute_segmentation_loss(seg_logits, target_resized)
|
|
|
|
# Assert
|
|
assert loss is not None
|
|
assert isinstance(loss, torch.Tensor)
|
|
assert loss.dim() == 0 # Scalar loss
|
|
assert loss.item() >= 0 # Loss should be non-negative
|
|
|
|
def test_compute_uv_loss_measures_coordinate_regression(self, densepose_head, mock_feature_input, mock_target_uv):
|
|
"""Test that compute_uv_loss measures UV coordinate regression accuracy"""
|
|
# Arrange
|
|
output = densepose_head(mock_feature_input)
|
|
uv_pred = output['uv_coordinates']
|
|
|
|
# Resize target to match output
|
|
target_resized = torch.nn.functional.interpolate(
|
|
mock_target_uv,
|
|
size=uv_pred.shape[2:],
|
|
mode='bilinear',
|
|
align_corners=False
|
|
)
|
|
|
|
# Act
|
|
loss = densepose_head.compute_uv_loss(uv_pred, target_resized)
|
|
|
|
# Assert
|
|
assert loss is not None
|
|
assert isinstance(loss, torch.Tensor)
|
|
assert loss.dim() == 0 # Scalar loss
|
|
assert loss.item() >= 0 # Loss should be non-negative
|
|
|
|
def test_compute_total_loss_combines_segmentation_and_uv_losses(self, densepose_head, mock_feature_input, mock_target_masks, mock_target_uv):
|
|
"""Test that compute_total_loss combines segmentation and UV losses"""
|
|
# Arrange
|
|
output = densepose_head(mock_feature_input)
|
|
|
|
# Resize targets to match outputs
|
|
seg_target = torch.nn.functional.interpolate(
|
|
mock_target_masks.float().unsqueeze(1),
|
|
size=output['segmentation'].shape[2:],
|
|
mode='nearest'
|
|
).squeeze(1).long()
|
|
|
|
uv_target = torch.nn.functional.interpolate(
|
|
mock_target_uv,
|
|
size=output['uv_coordinates'].shape[2:],
|
|
mode='bilinear',
|
|
align_corners=False
|
|
)
|
|
|
|
# Act
|
|
total_loss = densepose_head.compute_total_loss(output, seg_target, uv_target)
|
|
seg_loss = densepose_head.compute_segmentation_loss(output['segmentation'], seg_target)
|
|
uv_loss = densepose_head.compute_uv_loss(output['uv_coordinates'], uv_target)
|
|
|
|
# Assert
|
|
assert total_loss is not None
|
|
assert isinstance(total_loss, torch.Tensor)
|
|
assert total_loss.item() > 0
|
|
# Total loss should be combination of individual losses
|
|
expected_total = seg_loss + uv_loss
|
|
assert torch.allclose(total_loss, expected_total, atol=1e-6)
|
|
|
|
def test_fpn_integration_enhances_multi_scale_features(self, mock_config, mock_feature_input):
|
|
"""Test that FPN integration enhances multi-scale feature processing"""
|
|
# Arrange
|
|
config_with_fpn = mock_config.copy()
|
|
config_with_fpn['use_fpn'] = True
|
|
|
|
config_without_fpn = mock_config.copy()
|
|
config_without_fpn['use_fpn'] = False
|
|
|
|
head_with_fpn = DensePoseHead(config_with_fpn)
|
|
head_without_fpn = DensePoseHead(config_without_fpn)
|
|
|
|
# Act
|
|
output_with_fpn = head_with_fpn(mock_feature_input)
|
|
output_without_fpn = head_without_fpn(mock_feature_input)
|
|
|
|
# Assert
|
|
assert output_with_fpn['segmentation'].shape == output_without_fpn['segmentation'].shape
|
|
assert output_with_fpn['uv_coordinates'].shape == output_without_fpn['uv_coordinates'].shape
|
|
# Outputs should be different due to FPN
|
|
assert not torch.allclose(output_with_fpn['segmentation'], output_without_fpn['segmentation'], atol=1e-6)
|
|
|
|
def test_get_prediction_confidence_provides_uncertainty_estimates(self, densepose_head, mock_feature_input):
|
|
"""Test that get_prediction_confidence provides uncertainty estimates"""
|
|
# Arrange
|
|
output = densepose_head(mock_feature_input)
|
|
|
|
# Act
|
|
confidence = densepose_head.get_prediction_confidence(output)
|
|
|
|
# Assert
|
|
assert confidence is not None
|
|
assert isinstance(confidence, dict)
|
|
assert 'segmentation_confidence' in confidence
|
|
assert 'uv_confidence' in confidence
|
|
|
|
seg_conf = confidence['segmentation_confidence']
|
|
uv_conf = confidence['uv_confidence']
|
|
|
|
assert isinstance(seg_conf, torch.Tensor)
|
|
assert isinstance(uv_conf, torch.Tensor)
|
|
assert seg_conf.shape[0] == mock_feature_input.shape[0]
|
|
assert uv_conf.shape[0] == mock_feature_input.shape[0]
|
|
|
|
def test_post_process_predictions_formats_output(self, densepose_head, mock_feature_input):
|
|
"""Test that post_process_predictions formats output correctly"""
|
|
# Arrange
|
|
raw_output = densepose_head(mock_feature_input)
|
|
|
|
# Act
|
|
processed = densepose_head.post_process_predictions(raw_output)
|
|
|
|
# Assert
|
|
assert processed is not None
|
|
assert isinstance(processed, dict)
|
|
assert 'body_parts' in processed
|
|
assert 'uv_coordinates' in processed
|
|
assert 'confidence_scores' in processed
|
|
|
|
def test_training_mode_enables_dropout(self, densepose_head, mock_feature_input):
|
|
"""Test that training mode enables dropout for regularization"""
|
|
# Arrange
|
|
densepose_head.train()
|
|
|
|
# Act
|
|
output1 = densepose_head(mock_feature_input)
|
|
output2 = densepose_head(mock_feature_input)
|
|
|
|
# Assert - outputs should be different due to dropout
|
|
assert not torch.allclose(output1['segmentation'], output2['segmentation'], atol=1e-6)
|
|
assert not torch.allclose(output1['uv_coordinates'], output2['uv_coordinates'], atol=1e-6)
|
|
|
|
def test_evaluation_mode_disables_dropout(self, densepose_head, mock_feature_input):
|
|
"""Test that evaluation mode disables dropout for consistent inference"""
|
|
# Arrange
|
|
densepose_head.eval()
|
|
|
|
# Act
|
|
output1 = densepose_head(mock_feature_input)
|
|
output2 = densepose_head(mock_feature_input)
|
|
|
|
# Assert - outputs should be identical in eval mode
|
|
assert torch.allclose(output1['segmentation'], output2['segmentation'], atol=1e-6)
|
|
assert torch.allclose(output1['uv_coordinates'], output2['uv_coordinates'], atol=1e-6)
|
|
|
|
def test_head_validates_input_dimensions(self, densepose_head):
|
|
"""Test that head validates input dimensions"""
|
|
# Arrange
|
|
invalid_input = torch.randn(2, 128, 56, 56) # Wrong number of channels
|
|
|
|
# Act & Assert
|
|
with pytest.raises(DensePoseError):
|
|
densepose_head(invalid_input)
|
|
|
|
def test_head_handles_different_input_sizes(self, densepose_head):
|
|
"""Test that head handles different input sizes"""
|
|
# Arrange
|
|
small_input = torch.randn(1, 256, 28, 28)
|
|
large_input = torch.randn(1, 256, 112, 112)
|
|
|
|
# Act
|
|
small_output = densepose_head(small_input)
|
|
large_output = densepose_head(large_input)
|
|
|
|
# Assert
|
|
assert small_output['segmentation'].shape[2:] != large_output['segmentation'].shape[2:]
|
|
assert small_output['uv_coordinates'].shape[2:] != large_output['uv_coordinates'].shape[2:]
|
|
|
|
def test_head_supports_gradient_computation(self, densepose_head, mock_feature_input, mock_target_masks, mock_target_uv):
|
|
"""Test that head supports gradient computation for training"""
|
|
# Arrange
|
|
densepose_head.train()
|
|
optimizer = torch.optim.Adam(densepose_head.parameters(), lr=0.001)
|
|
|
|
output = densepose_head(mock_feature_input)
|
|
|
|
# Resize targets
|
|
seg_target = torch.nn.functional.interpolate(
|
|
mock_target_masks.float().unsqueeze(1),
|
|
size=output['segmentation'].shape[2:],
|
|
mode='nearest'
|
|
).squeeze(1).long()
|
|
|
|
uv_target = torch.nn.functional.interpolate(
|
|
mock_target_uv,
|
|
size=output['uv_coordinates'].shape[2:],
|
|
mode='bilinear',
|
|
align_corners=False
|
|
)
|
|
|
|
# Act
|
|
loss = densepose_head.compute_total_loss(output, seg_target, uv_target)
|
|
|
|
optimizer.zero_grad()
|
|
loss.backward()
|
|
|
|
# Assert
|
|
for param in densepose_head.parameters():
|
|
if param.requires_grad:
|
|
assert param.grad is not None
|
|
assert not torch.allclose(param.grad, torch.zeros_like(param.grad))
|
|
|
|
def test_head_configuration_validation(self):
|
|
"""Test that head validates configuration parameters"""
|
|
# Arrange
|
|
invalid_config = {
|
|
'input_channels': 0, # Invalid
|
|
'num_body_parts': -1, # Invalid
|
|
'num_uv_coordinates': 2
|
|
}
|
|
|
|
# Act & Assert
|
|
with pytest.raises(ValueError):
|
|
DensePoseHead(invalid_config)
|
|
|
|
def test_save_and_load_model_state(self, densepose_head, mock_feature_input):
|
|
"""Test that model state can be saved and loaded"""
|
|
# Arrange
|
|
original_output = densepose_head(mock_feature_input)
|
|
|
|
# Act - Save state
|
|
state_dict = densepose_head.state_dict()
|
|
|
|
# Create new head and load state
|
|
new_head = DensePoseHead(densepose_head.config)
|
|
new_head.load_state_dict(state_dict)
|
|
new_output = new_head(mock_feature_input)
|
|
|
|
# Assert
|
|
assert torch.allclose(original_output['segmentation'], new_output['segmentation'], atol=1e-6)
|
|
assert torch.allclose(original_output['uv_coordinates'], new_output['uv_coordinates'], atol=1e-6) |