File size: 5,823 Bytes
84e9a1f | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 | import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
# MediaPipe skeletal anatomical connections
MEDIAPIPE_EDGES = [
(0, 1), (1, 2), (2, 3), (0, 4), (4, 5), (5, 6), (3, 7), (6, 8), (9, 10),
(11, 12), (11, 13), (13, 15), (12, 14), (14, 16),
(15, 17), (15, 19), (15, 21), (17, 19),
(16, 18), (16, 20), (16, 22), (18, 20),
(11, 23), (12, 24), (23, 24),
(23, 25), (25, 27), (27, 29), (29, 31), (27, 31),
(24, 26), (26, 28), (28, 30), (30, 32), (28, 32)
]
def get_normalized_adjacency():
"""
Builds the normalized skeletal adjacency matrix A_norm for the 33 MediaPipe joints.
A_norm = D^(-1/2) * (A + I) * D^(-1/2)
"""
num_nodes = 33
A = np.zeros((num_nodes, num_nodes), dtype=np.float32)
# Add skeletal edges
for i, j in MEDIAPIPE_EDGES:
A[i, j] = 1.0
A[j, i] = 1.0
# Add self loops
A_loop = A + np.eye(num_nodes, dtype=np.float32)
# Calculate degree matrix D
deg = np.sum(A_loop, axis=1)
deg_inv_sqrt = np.power(deg, -0.5, where=deg>0)
deg_inv_sqrt[deg == 0] = 0.0
D_inv_sqrt = np.diag(deg_inv_sqrt)
# Symmetric normalization
A_norm = D_inv_sqrt.dot(A_loop).dot(D_inv_sqrt)
return torch.tensor(A_norm, dtype=torch.float32)
class SpatialGraphConv(nn.Module):
"""
True Spatial Graph Convolutional Layer.
Aggregates joint features along the normalized skeletal adjacency matrix A.
"""
def __init__(self, in_channels, out_channels, A_norm):
super(SpatialGraphConv, self).__init__()
self.register_buffer('A', A_norm) # Keep adjacency matrix on the correct device
self.conv = nn.Conv2d(in_channels, out_channels, kernel_size=1)
def forward(self, x):
# x shape: [batch_size, in_channels, num_nodes, seq_len] -> [N, C, 33, T]
# 1. Project features: W * X
x = self.conv(x) # [N, C_out, 33, T]
# 2. Graph aggregation: A * (W * X)
# We multiply adjacency matrix A (shape [33, 33]) across the node dimension (dim=2)
out = torch.einsum('vw,ncwt->ncvt', self.A, x)
return out
class TemporalConv(nn.Module):
"""
Temporal Convolutional Layer.
Convolves features along the temporal dimension for each node independently.
"""
def __init__(self, in_channels, out_channels, kernel_size=9, stride=1, dropout=0.3):
super(TemporalConv, self).__init__()
padding = (kernel_size - 1) // 2
self.conv = nn.Conv2d(
in_channels,
out_channels,
kernel_size=(kernel_size, 1),
stride=(stride, 1),
padding=(padding, 0)
)
self.bn = nn.BatchNorm2d(out_channels)
self.dropout = nn.Dropout(dropout)
def forward(self, x):
# x shape: [batch_size, in_channels, num_nodes, seq_len] -> [N, C, 33, T]
# Swap node and temporal dimensions to convolve over time: [N, C, T, 33]
x = x.transpose(2, 3)
x = self.conv(x)
x = self.bn(x)
x = self.dropout(x)
# Swap back: [N, C, 33, T]
return x.transpose(2, 3)
class STGCNBlock(nn.Module):
"""
A single Spatio-Temporal Graph Convolutional Block containing:
1. Spatial Graph Convolution (GCN)
2. Temporal Graph Convolution (TCN)
"""
def __init__(self, in_channels, out_channels, A_norm, stride=1, dropout=0.3):
super(STGCNBlock, self).__init__()
self.gcn = SpatialGraphConv(in_channels, out_channels, A_norm)
self.tcn = TemporalConv(out_channels, out_channels, kernel_size=9, stride=stride, dropout=dropout)
# Residual skip connection
if in_channels != out_channels or stride != 1:
self.residual = nn.Sequential(
nn.Conv2d(in_channels, out_channels, kernel_size=1),
nn.BatchNorm2d(out_channels)
)
else:
self.residual = nn.Identity()
def forward(self, x):
# x shape: [N, C_in, 33, T]
res = self.residual(x)
x = self.gcn(x)
x = F.gelu(x)
x = self.tcn(x)
return F.gelu(x + res)
# True ST-GCN Sequence Classifier (maintaining class name YogaSequenceLSTM for backend loading compatibility)
class YogaSequenceLSTM(nn.Module):
def __init__(self, input_dim, hidden_dim, num_layers, num_classes):
super(YogaSequenceLSTM, self).__init__()
A_norm = get_normalized_adjacency()
# Input channels: 3 (x, y, z joint coordinates)
self.block1 = STGCNBlock(3, 64, A_norm, stride=1, dropout=0.2)
self.block2 = STGCNBlock(64, 128, A_norm, stride=1, dropout=0.3)
self.block3 = STGCNBlock(128, 256, A_norm, stride=1, dropout=0.3)
self.fc = nn.Sequential(
nn.Linear(256, 128),
nn.LayerNorm(128),
nn.GELU(),
nn.Dropout(0.4),
nn.Linear(128, num_classes)
)
def forward(self, x):
# Input x shape: [batch_size, seq_len, input_dim] -> [batch_size, 60, 99]
# Reshape to [batch_size, seq_len, 33, 3] and permute to [batch_size, channels=3, nodes=33, seq_len=60]
batch_size = x.size(0)
x_reshaped = x.view(batch_size, 60, 33, 3).permute(0, 3, 2, 1) # [batch_size, 3, 33, 60]
# ST-GCN convolutions
out = self.block1(x_reshaped) # [batch_size, 64, 33, 60]
out = self.block2(out) # [batch_size, 128, 33, 60]
out = self.block3(out) # [batch_size, 256, 33, 60]
# Global Pooling over both Spatial nodes (33) and Temporal frames (60)
# out shape: [batch_size, 256]
out = F.adaptive_avg_pool2d(out, (1, 1)).view(out.size(0), -1)
return self.fc(out)
|