Agent Conversation: deepseek-reasoner

Model: deepseek-reasoner
Total steps: 9
Messages: 20
UserStep 0

Task: cv-multitask-loss

CV Multi-Task Loss Combination Strategy Design

Research Question

Design a novel multi-task loss combination strategy for jointly training fine-grained (100-class) and coarse (20-superclass) classification on CIFAR-100 that maximizes fine-class test accuracy.

Background

CIFAR-100 contains 100 fine classes organized into 20 coarse superclasses. Training a model with two classification heads (fine + coarse) provides a natural multi-task learning setup where the coarse task acts as an auxiliary signal. The key challenge is how to combine the two losses effectively.

Classic approaches include:

  • Equal weighting: Simply sum the losses (baseline default)
  • Uncertainty weighting (Kendall et al., 2018): Learn task-specific uncertainty as log-variance parameters
  • Dynamic Weight Average (Liu et al., 2019): Weight tasks by their relative loss change rate
  • PCGrad (Yu et al., NeurIPS 2020): Project conflicting task gradients onto each other's normal plane to reduce gradient interference

The coarse labels encode semantic hierarchy. The task is to balance the auxiliary coarse signal against the primary fine-class objective across different architectures and training stages.

What You Can Modify

The MultiTaskLoss class (lines 195-216) in custom_mtl.py. This class receives individual task losses and must combine them into a single scalar loss.

You can modify:

  • The __init__ method: add learnable parameters (log-variances, weights, etc.)
  • The forward method: implement any combination strategy
  • Use epoch and total_epochs for curriculum/scheduling approaches
  • Add any auxiliary state (e.g., loss history buffers)

The forward method receives:

  • fine_loss: scalar tensor, cross-entropy for 100-class fine prediction
  • coarse_loss: scalar tensor, cross-entropy for 20-class coarse prediction
  • epoch: int, current epoch (0-indexed)
  • total_epochs: int, total number of training epochs

Note: The MultiTaskLoss parameters are included in the optimizer, so learnable parameters will be trained.

Evaluation

  • Metric: Best fine-class test accuracy (%, higher is better)
  • Architectures (all on CIFAR-100 with fine+coarse heads):
    • ResNet-20 (shallow residual network)
    • ResNet-56 (deeper residual network)
    • VGG-16-BN (deep non-residual with BatchNorm) — hidden, evaluated on final submission only
  • Training: SGD (lr=0.1, momentum=0.9, wd=5e-4), cosine annealing, 200 epochs
  • Data augmentation: RandomCrop(32, pad=4) + RandomHorizontalFlip

pytorch-vision/custom_mtl.py [EDITABLE — lines 195–216 only]

     1: """CV Multi-Task Loss Benchmark.
     2: 
     3: Train vision models (ResNet, VGG) on CIFAR-100 with TWO classification heads
     4: (fine: 100 classes, coarse: 20 superclasses) to evaluate multi-task loss
     5: combination strategies.
     6: 
     7: FIXED: Model architectures, data pipeline, training loop.
     8: EDITABLE: MultiTaskLoss class.
AssistantStep 1edit
Tool Input
1{
2  "op": "replace",
3  "filename": "pytorch-vision/custom_mtl.py",
4  "start_line": 195,
5  "end_line": 216,
6  "content": "class MultiTaskLoss(nn.Module):\n    \"\"\"Hierarchical Adaptive Weighting (HAW): Uncertainty weighting with curriculum decay on coarse task.\n\n    Learns per-task log-variances, and reduces coarse weight over epochs.\n    Total = exp(-log_var_fine) * L_fine + exp(-log_var_coarse) * (1 - epoch/total_epochs)^β * L_coarse + log_var_fine + log_var_coarse\n    \"\"\"\n\n    def __init__(self, num_tasks=2):\n        super().__init__()\n        self.log_vars = nn.Parameter(torch.zeros(num_tasks))\n        # beta controls decay strength; hardcoded\n        self.beta = 2.0\n\n    def forward(self, fine_loss, coarse_loss, epoch, total_epochs):\n        fine_weight = torch.exp(-self.log_vars[0])\n        coarse_weight = torch.exp(-self.log_vars[1])\n        # curriculum decay factor for coarse task\n        progress = epoch / total_epochs\n        decay = (1.0 - progress) ** self.beta\n        coarse_weight = coarse_weight * decay\n        \n        total = fine_weight * fine_loss + coarse_weight * coarse_loss + self.log_vars[0] + self.log_vars[1]\n        return total"
7}
Tool ResultStep 1
OK: Replaced lines 195..216 with 23 line(s) in pytorch-vision/custom_mtl.py. Editable range: 195–217.

[Current file: pytorch-vision/custom_mtl.py | editable: 195–217 | total: 388 lines]
   192| # ============================================================================
   193| 
   194| # -- EDITABLE REGION START (lines 195-216) ------------------------------------
   195  class MultiTaskLoss(nn.Module):
   196      """Hierarchical Adaptive Weighting (HAW): Uncertainty weighting with curriculum decay on coarse task.
   197  
   198      Learns per-task log-variances, and reduces coarse weight over epochs.
   199      Total = exp(-log_var_fine) * L_fine + exp(-log_var_coarse) * (1 - epoch/total_epochs)^β * L_coarse + log_var_fine + log_var_coarse
   200      """
   201  
   202      def __init__(self, num_tasks=2):
   203          super().__init__()
   204          self.log_vars = nn.Parameter(torch.zeros(num_tasks))
   205          # beta controls decay strength; hardcoded
   206
AssistantStep 2test
Tool ResultStep 2
[Test #1] (2 tests remaining; call submit(n=N) to choose which test result to submit as final)

### resnet20-cifar100mt (scripts/resnet20_cifar100mt.sh)
Training progress (last epochs):
TRAIN_METRICS: epoch=160 train_loss=-2.0740 train_acc=71.45 test_acc=61.86 lr=0.009549
TRAIN_METRICS: epoch=170 train_loss=-2.9604 train_acc=74.15 test_acc=63.35 lr=0.005450
TRAIN_METRICS: epoch=180 train_loss=-4.1847 train_acc=77.10 test_acc=64.40 lr=0.002447
TRAIN_METRICS: epoch=190 train_loss=-6.0121 train_acc=79.91 test_acc=65.63 lr=0.000616
TRAIN_METRICS: epoch=200 train_loss=-8.9071 train_acc=81.40 test_acc=66.54 lr=0.000000
Final metrics (resnet20-cifar100mt): test_acc_resnet20-cifar100mt=66.55

### resnet56-cifar100mt (scripts/resnet56_cifar100mt.sh)
Training progress (last epochs):
TRAIN_METRICS: epoch=160 train_loss=-2.8977 train_acc=78.32 test_acc=65.40 lr=0.009549
TRAIN_METRICS: epoch=170 train_loss=-4.0024 train_acc=81.83 test_acc=67.64 lr=0.005450
TRAIN_METRICS: epoch=180 train_loss=-5.513
AssistantStep 3edit
Tool Input
1{
2  "op": "replace",
3  "filename": "pytorch-vision/custom_mtl.py",
4  "start_line": 195,
5  "end_line": 217,
6  "content": "class MultiTaskLoss(nn.Module):\n    \"\"\"Hierarchical Adaptive Weighting with regularization (HAW-Reg).\n    \n    Uncertainty weighting with curriculum decay on coarse task, plus L2 penalty on log-variances.\n    Total = exp(-log_var_fine) * L_fine + exp(-log_var_coarse) * (1 - epoch/total_epochs)^β * L_coarse \n            + log_var_fine + log_var_coarse + λ * (log_var_fine^2 + log_var_coarse^2)\n    \"\"\"\n\n    def __init__(self, num_tasks=2):\n        super().__init__()\n        self.log_vars = nn.Parameter(torch.zeros(num_tasks))\n        # beta controls decay strength; hardcoded\n        self.beta = 2.0\n        # regularization strength\n        self.lambda_reg = 0.01\n\n    def forward(self, fine_loss, coarse_loss, epoch, total_epochs):\n        fine_weight = torch.exp(-self.log_vars[0])\n        coarse_weight = torch.exp(-self.log_vars[1])\n        # curriculum decay factor for coarse task\n        progress = epoch / total_epochs\n        decay = (1.0 - progress) ** self.beta\n        coarse_weight = coarse_weight * decay\n        \n        total = (fine_weight * fine_loss + coarse_weight * coarse_loss + \n                 self.log_vars[0] + self.log_vars[1] +\n                 self.lambda_reg * (self.log_vars[0]**2 + self.log_vars[1]**2))\n        return total"
7}
Tool ResultStep 3
OK: Replaced lines 195..217 with 28 line(s) in pytorch-vision/custom_mtl.py. Editable range: 195–222.

[Current file: pytorch-vision/custom_mtl.py | editable: 195–222 | total: 393 lines]
   192| # ============================================================================
   193| 
   194| # -- EDITABLE REGION START (lines 195-216) ------------------------------------
   195  class MultiTaskLoss(nn.Module):
   196      """Hierarchical Adaptive Weighting with regularization (HAW-Reg).
   197      
   198      Uncertainty weighting with curriculum decay on coarse task, plus L2 penalty on log-variances.
   199      Total = exp(-log_var_fine) * L_fine + exp(-log_var_coarse) * (1 - epoch/total_epochs)^β * L_coarse 
   200              + log_var_fine + log_var_coarse + λ * (log_var_fine^2 + log_var_coarse^2)
   201      """
   202  
   203      def __init__(self, num_tasks=2):
   204          super().__init__()
   205          self.log_vars = nn.Parameter(torch.zeros(num_tasks))
   206
AssistantStep 4test
Tool ResultStep 4
[Test #2] (1 test remaining; call submit(n=N) to choose which test result to submit as final)

### resnet20-cifar100mt (scripts/resnet20_cifar100mt.sh)
Training progress (last epochs):
TRAIN_METRICS: epoch=160 train_loss=-1.9509 train_acc=71.88 test_acc=61.86 lr=0.009549
TRAIN_METRICS: epoch=170 train_loss=-2.7809 train_acc=74.47 test_acc=64.64 lr=0.005450
TRAIN_METRICS: epoch=180 train_loss=-3.9269 train_acc=78.00 test_acc=64.98 lr=0.002447
TRAIN_METRICS: epoch=190 train_loss=-5.5384 train_acc=80.75 test_acc=66.63 lr=0.000616
TRAIN_METRICS: epoch=200 train_loss=-7.8515 train_acc=82.03 test_acc=67.05 lr=0.000000
Final metrics (resnet20-cifar100mt): test_acc_resnet20-cifar100mt=67.10

### resnet56-cifar100mt (scripts/resnet56_cifar100mt.sh)
Training progress (last epochs):
TRAIN_METRICS: epoch=160 train_loss=-2.7788 train_acc=79.14 test_acc=66.01 lr=0.009549
TRAIN_METRICS: epoch=170 train_loss=-3.7797 train_acc=82.21 test_acc=67.59 lr=0.005450
TRAIN_METRICS: epoch=180 train_loss=-5.1890
AssistantStep 5edit
Tool Input
1{
2  "op": "replace",
3  "filename": "pytorch-vision/custom_mtl.py",
4  "start_line": 195,
5  "end_line": 222,
6  "content": "class MultiTaskLoss(nn.Module):\n    \"\"\"Adaptive Task Weighting with Hierarchical Curriculum (ATWHC).\n    \n    Learn task weights via softmax over learned logits, apply curriculum decay to coarse weight.\n    Total = w_fine * L_fine + w_coarse * (1 - epoch/total_epochs)^β * L_coarse.\n    \"\"\"\n    def __init__(self, num_tasks=2):\n        super().__init__()\n        self.logits = nn.Parameter(torch.zeros(num_tasks))\n        self.beta = 2.0\n        self.tau = 0.5  # temperature\n    def forward(self, fine_loss, coarse_loss, epoch, total_epochs):\n        weights = F.softmax(self.logits / self.tau, dim=0)\n        fine_weight = weights[0]\n        coarse_weight = weights[1] * (1.0 - epoch / total_epochs) ** self.beta\n        return fine_weight * fine_loss + coarse_weight * coarse_loss"
7}
Tool ResultStep 5
OK: Replaced lines 195..222 with 16 line(s) in pytorch-vision/custom_mtl.py. Editable range: 195–210.

[Current file: pytorch-vision/custom_mtl.py | editable: 195–210 | total: 381 lines]
   192| # ============================================================================
   193| 
   194| # -- EDITABLE REGION START (lines 195-216) ------------------------------------
   195  class MultiTaskLoss(nn.Module):
   196      """Adaptive Task Weighting with Hierarchical Curriculum (ATWHC).
   197      
   198      Learn task weights via softmax over learned logits, apply curriculum decay to coarse weight.
   199      Total = w_fine * L_fine + w_coarse * (1 - epoch/total_epochs)^β * L_coarse.
   200      """
   201      def __init__(self, num_tasks=2):
   202          super().__init__()
   203          self.logits = nn.Parameter(torch.zeros(num_tasks))
   204          self.beta = 2.0
   205          self.tau = 0.5  # temperature
   206      def forward(self, fine_loss, coarse_loss, epoch, tota
AssistantStep 6test
Tool ResultStep 6
[Test #3] (0 tests remaining; call submit(n=N) to choose which test result to submit as final)

[NOTE] This was your last test. You MUST now call submit(n=X) to choose which test result to submit as your final answer.

### resnet20-cifar100mt (scripts/resnet20_cifar100mt.sh)
Training progress (last epochs):
TRAIN_METRICS: epoch=160 train_loss=0.0195 train_acc=26.51 test_acc=23.59 lr=0.009549
TRAIN_METRICS: epoch=170 train_loss=0.0108 train_acc=26.67 test_acc=23.86 lr=0.005450
TRAIN_METRICS: epoch=180 train_loss=0.0053 train_acc=26.63 test_acc=24.11 lr=0.002447
TRAIN_METRICS: epoch=190 train_loss=0.0019 train_acc=26.63 test_acc=24.04 lr=0.000616
TRAIN_METRICS: epoch=200 train_loss=0.0006 train_acc=26.63 test_acc=24.15 lr=0.000000
Final metrics (resnet20-cifar100mt): test_acc_resnet20-cifar100mt=24.69

### resnet56-cifar100mt (scripts/resnet56_cifar100mt.sh)
Training progress (last epochs):
TRAIN_METRICS: epoch=160 train_loss=0.0077 train_acc=24.20 test_acc=21.72 lr=0.009549
TRAIN_METRIC
AssistantStep 7undo
Tool Input
1{
2  "n": 1
3}
Tool ResultStep 7
Undo complete:
Restored: pytorch-vision/custom_mtl.py
AssistantStep 8edit
Tool Input
1{
2  "op": "replace",
3  "filename": "pytorch-vision/custom_mtl.py",
4  "start_line": 195,
5  "end_line": 222,
6  "content": "class MultiTaskLoss(nn.Module):\n    \"\"\"Hierarchical Adaptive Weighting with regularization (HAW-Reg).\n    \n    Uncertainty weighting with curriculum decay on coarse task, plus L2 penalty on log-variances.\n    Total = exp(-log_var_fine) * L_fine + exp(-log_var_coarse) * (1 - epoch/total_epochs)^β * L_coarse \n            + log_var_fine + log_var_coarse + λ * (log_var_fine^2 + log_var_coarse^2)\n    \"\"\"\n\n    def __init__(self, num_tasks=2):\n        super().__init__()\n        self.log_vars = nn.Parameter(torch.zeros(num_tasks))\n        # beta controls decay strength; hardcoded\n        self.beta = 2.0\n        # regularization strength\n        self.lambda_reg = 0.01\n\n    def forward(self, fine_loss, coarse_loss, epoch, total_epochs):\n        fine_weight = torch.exp(-self.log_vars[0])\n        coarse_weight = torch.exp(-self.log_vars[1])\n        # curriculum decay factor for coarse task\n        progress = epoch / total_epochs\n        decay = (1.0 - progress) ** self.beta\n        coarse_weight = coarse_weight * decay\n        \n        total = (fine_weight * fine_loss + coarse_weight * coarse_loss + \n                 self.log_vars[0] + self.log_vars[1] +\n                 self.lambda_reg * (self.log_vars[0]**2 + self.log_vars[1]**2))\n        return total"
7}
Tool ResultStep 8
OK: Replaced lines 195..222 with 28 line(s) in pytorch-vision/custom_mtl.py. Editable range: 195–222.

[Current file: pytorch-vision/custom_mtl.py | editable: 195–222 | total: 393 lines]
   192| # ============================================================================
   193| 
   194| # -- EDITABLE REGION START (lines 195-216) ------------------------------------
   195  class MultiTaskLoss(nn.Module):
   196      """Hierarchical Adaptive Weighting with regularization (HAW-Reg).
   197      
   198      Uncertainty weighting with curriculum decay on coarse task, plus L2 penalty on log-variances.
   199      Total = exp(-log_var_fine) * L_fine + exp(-log_var_coarse) * (1 - epoch/total_epochs)^β * L_coarse 
   200              + log_var_fine + log_var_coarse + λ * (log_var_fine^2 + log_var_coarse^2)
   201      """
   202  
   203      def __init__(self, num_tasks=2):
   204          super().__init__()
   205          self.log_vars = nn.Parameter(torch.zeros(num_tasks))
   206
AssistantStep 9submit
Tool Input
1{
2  "n": 2
3}
Tool ResultStep 9
[submit] Submitting result from test #2 as final.

### resnet20-cifar100mt (scripts/resnet20_cifar100mt.sh)
Training progress (last epochs):
TRAIN_METRICS: epoch=160 train_loss=-1.9509 train_acc=71.88 test_acc=61.86 lr=0.009549
TRAIN_METRICS: epoch=170 train_loss=-2.7809 train_acc=74.47 test_acc=64.64 lr=0.005450
TRAIN_METRICS: epoch=180 train_loss=-3.9269 train_acc=78.00 test_acc=64.98 lr=0.002447
TRAIN_METRICS: epoch=190 train_loss=-5.5384 train_acc=80.75 test_acc=66.63 lr=0.000616
TRAIN_METRICS: epoch=200 train_loss=-7.8515 train_acc=82.03 test_acc=67.05 lr=0.000000
Final metrics (resnet20-cifar100mt): test_acc_resnet20-cifar100mt=67.10

### resnet56-cifar100mt (scripts/resnet56_cifar100mt.sh)
Training progress (last epochs):
TRAIN_METRICS: epoch=160 train_loss=-2.7788 train_acc=79.14 test_acc=66.01 lr=0.009549
TRAIN_METRICS: epoch=170 train_loss=-3.7797 train_acc=82.21 test_acc=67.59 lr=0.005450
TRAIN_METRICS: epoch=180 train_loss=-5.1890 train_acc=85.41 test_acc=68.50 lr=0.002447