File size: 10,287 Bytes
a10ba7f | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 | import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader
import os
import time
import numpy as np
from tqdm import tqdm
import sys
import argparse
from pathlib import Path
# Add project root to path
sys.path.append(str(Path(__file__).parent.parent))
from src.data.dataset import GazeDataset
# --- 1. Loss Functions ---
class SmoothAWLoss(nn.Module):
def __init__(self, omega=8.0, alpha=1.5, theta=0.5, epsilon=1.0):
super(SmoothAWLoss, self).__init__()
self.omega = omega
self.alpha = alpha
self.theta = theta
self.epsilon = epsilon
def forward(self, y_pred, y_true):
delta_y = (y_true - y_pred).abs()
device = y_pred.device
mask = delta_y < self.theta
loss = torch.zeros_like(delta_y)
loss[mask] = self.omega * torch.log(1 + torch.pow(delta_y[mask] / self.epsilon, self.alpha))
theta_eps = torch.tensor(self.theta / self.epsilon, device=device)
A = self.omega * (1.0 / (1.0 + torch.pow(theta_eps, self.alpha))) * \
(self.alpha * torch.pow(theta_eps, self.alpha - 1.0) * (1.0 / self.epsilon))
B = A * self.theta - self.omega * torch.log(1.0 + torch.pow(theta_eps, self.alpha))
loss[~mask] = A * delta_y[~mask] - B
return loss.mean()
class DistillationLoss(nn.Module):
def __init__(self, temperature=2.0, alpha=0.5, base_criterion=nn.MSELoss()):
super(DistillationLoss, self).__init__()
self.temp = temperature
self.alpha = alpha
self.base_criterion = base_criterion
self.kl_div = nn.KLDivLoss(reduction='batchmean')
def forward(self, student_logits, teacher_logits, labels):
# student_logits/teacher_logits are tuples (pitch, yaw)
p_s, y_s = student_logits
p_t, y_t = teacher_logits
# Soft targets loss
loss_p_kd = self.kl_div(torch.log_softmax(p_s / self.temp, dim=1),
torch.softmax(p_t / self.temp, dim=1)) * (self.temp ** 2)
loss_y_kd = self.kl_div(torch.log_softmax(y_s / self.temp, dim=1),
torch.softmax(y_t / self.temp, dim=1)) * (self.temp ** 2)
# Hard targets loss (GT)
loss_gt = self.base_criterion(student_logits, labels) # This assumes base_criterion handles the logits/labels correctly
return self.alpha * (loss_p_kd + loss_y_kd) + (1 - self.alpha) * loss_gt
# --- 2. Model Architecture with Ablation Toggles ---
class FlexibleMiniConv(nn.Module):
def __init__(self, pool_type='avg'):
super(FlexibleMiniConv, self).__init__()
self.conv = nn.Sequential(
nn.Conv2d(1, 16, kernel_size=3, padding=0), nn.ReLU(inplace=True),
nn.Conv2d(16, 32, kernel_size=3, padding=0), nn.ReLU(inplace=True),
nn.Conv2d(32, 64, kernel_size=3, padding=0), nn.ReLU(inplace=True)
)
self.pool_type = pool_type
if pool_type == 'avg':
self.pool = nn.AdaptiveAvgPool2d(1)
self.out_channels = 64
else: # Dual Pool
self.avg_pool = nn.AdaptiveAvgPool2d(1)
self.max_pool = nn.AdaptiveMaxPool2d(1)
self.out_channels = 128
def forward(self, x):
x = self.conv(x)
if self.pool_type == 'avg':
return self.pool(x).flatten(1)
else:
return torch.cat([self.avg_pool(x), self.max_pool(x)], dim=1).flatten(1)
class LIPEV2AblationModel(nn.Module):
def __init__(self, pool_type='avg'):
super(LIPEV2AblationModel, self).__init__()
self.app_net = FlexibleMiniConv(pool_type=pool_type)
# app_net output: out_channels * 4 patches
app_out_dim = self.app_net.out_channels * 4
self.geo_net = nn.Sequential(
nn.Linear(956, 256),
nn.LayerNorm(256),
nn.ReLU(inplace=True),
nn.Linear(256, 256),
nn.ReLU(inplace=True)
)
self.post_concat_bn = nn.BatchNorm1d(app_out_dim + 256)
self.fusion = nn.Sequential(
nn.Linear(app_out_dim + 256, 256),
nn.ReLU(inplace=True),
nn.Dropout(0.1),
nn.Linear(256, 128),
nn.ReLU(inplace=True)
)
self.pitch_head = nn.Linear(128, 90)
self.yaw_head = nn.Linear(128, 90)
def forward(self, patches, landmarks):
batch_size = patches.shape[0]
p_h, p_w = patches.shape[2], patches.shape[3]
app_feat = self.app_net(patches.view(-1, 1, p_h, p_w)).view(batch_size, -1)
geo_feat = self.geo_net(landmarks)
combined = torch.cat([app_feat, geo_feat], dim=1)
combined = self.post_concat_bn(combined)
fused = self.fusion(combined)
return self.pitch_head(fused), self.yaw_head(fused)
# --- 3. Training Function ---
def train_ablation(args):
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print(f"\n>>> ABLATION STUDY ID: #{args.id} | Participant: {args.participant}")
print(f">>> Config: Pool={args.pooling}, Loss={args.loss}, KD={args.kd}")
# Load Data
h5_dir = 'data/processed'
train_files = [os.path.join(h5_dir, f) for f in os.listdir(h5_dir) if f.endswith('_v16.h5') and not f.startswith(args.participant)]
val_files = [os.path.join(h5_dir, f) for f in os.listdir(h5_dir) if f.startswith(args.participant) and f.endswith('_v16.h5')]
train_loader = DataLoader(GazeDataset(train_files, transform=True), batch_size=32, shuffle=True, num_workers=args.workers)
val_loader = DataLoader(GazeDataset(val_files, transform=False), batch_size=32, shuffle=False, num_workers=args.workers)
# Model
model = LIPEV2AblationModel(pool_type=args.pooling).to(device)
# KD Setup
teacher_model = None
if args.kd:
from src.models.loss import GazeDistillationLoss
# We assume the GazeDataset provides teacher predictions or we need a teacher model
# For simplicity in this ablation, if GazeDataset has t_p, t_y, we use them
kd_criterion = GazeDistillationLoss(w_aw=1.0, w_kd=0.5)
print(">>> KD Enabled: Using GazeDistillationLoss")
# Base Loss
if args.loss == 'smoothaw':
criterion = SmoothAWLoss()
else:
criterion = nn.MSELoss()
optimizer = optim.AdamW(model.parameters(), lr=args.lr, weight_decay=1e-2)
scheduler = optim.lr_scheduler.OneCycleLR(optimizer, max_lr=args.lr, steps_per_epoch=len(train_loader), epochs=args.epochs)
idx_tensor = torch.arange(90).float().to(device)
best_mae = float('inf')
# Checkpoints dir
ckpt_dir = f"checkpoints/ablation/id_{args.id}"
os.makedirs(ckpt_dir, exist_ok=True)
for epoch in range(1, args.epochs + 1):
model.train()
running_loss = 0.0
pbar = tqdm(train_loader, desc=f"Epoch {epoch}")
for batch in pbar:
# batch: [patches, landmarks, gaze, t_p, t_y, domain]
batch = [b.to(device) for b in batch]
optimizer.zero_grad()
p_logits, y_logits = model(batch[0], batch[1])
s_p = (torch.softmax(p_logits, dim=1) * idx_tensor).sum(1) * 2 - 90
s_y = (torch.softmax(y_logits, dim=1) * idx_tensor).sum(1) * 2 - 90
if args.kd:
# GazeDistillationLoss takes (p_logits, y_logits), gaze_gt, (t_p, t_y)
loss, _, _ = kd_criterion((p_logits, y_logits), batch[2], (batch[3], batch[4]))
else:
# Simple loss for this ablation
loss = criterion(torch.stack([s_p, s_y], dim=1), batch[2] * (180.0 / np.pi))
loss.backward()
optimizer.step()
scheduler.step()
running_loss += loss.item()
# Validation
model.eval()
total_error, count = 0.0, 0
with torch.no_grad():
for batch in val_loader:
batch = [b.to(device) for b in batch]
p_l, y_l = model(batch[0], batch[1])
p_d = (torch.softmax(p_l, dim=1) * idx_tensor).sum(1) * 2 - 90
y_d = (torch.softmax(y_l, dim=1) * idx_tensor).sum(1) * 2 - 90
gt_d = batch[2] * (180.0 / np.pi)
total_error += (torch.abs(p_d - gt_d[:,0]) + torch.abs(y_d - gt_d[:,1])).sum().item()
count += batch[2].shape[0]
val_mae = total_error / (count * 2) if count > 0 else 0
print(f"Epoch {epoch}: Val MAE {val_mae:.4f}")
if val_mae < best_mae:
best_mae = val_mae
torch.save(model.state_dict(), f"{ckpt_dir}/best_{args.participant}.pt")
print(f"--- New Best: {best_mae:.4f} ---")
# Final log
with open('logs/ablation_results.log', 'a') as f:
f.write(f"ID: {args.id}, Participant: {args.participant}, MAE: {best_mae:.4f}, Config: {args.pooling}/{args.loss}/KD={args.kd}\n")
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--id', type=int, required=True, help="Ablation ID (1-8)")
parser.add_argument('--participant', type=str, default='p11')
parser.add_argument('--pooling', choices=['avg', 'dual'], default='avg')
parser.add_argument('--loss', choices=['mse', 'smoothaw'], default='mse')
parser.add_argument('--kd', action='store_true')
parser.add_argument('--epochs', type=int, default=50)
parser.add_argument('--lr', type=float, default=1e-4)
parser.add_argument('--workers', type=int, default=4)
args = parser.parse_args()
# Auto-config based on ID if provided (optional override)
# 1: avg/mse/no, 2: dual/mse/no, 3: avg/saw/no, 4: avg/mse/yes
# 5: dual/saw/no, 6: dual/mse/yes, 7: avg/saw/yes, 8: dual/saw/yes
configs = {
1: ('avg', 'mse', False),
2: ('dual', 'mse', False),
3: ('avg', 'smoothaw', False),
4: ('avg', 'mse', True),
5: ('dual', 'smoothaw', False),
6: ('dual', 'mse', True),
7: ('avg', 'smoothaw', True),
8: ('dual', 'smoothaw', True),
}
if args.id in configs:
args.pooling, args.loss, args.kd = configs[args.id]
train_ablation(args)
|