File size: 10,764 Bytes
68b32f4 80dd9c4 68b32f4 |
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 |
import torch
import torch.nn as nn
import torch.nn.functional as F
def compute_ctc_loss(predictions, targets, blank_label=0):
"""
Computes the Connectionist Temporal Classification (CTC) loss.
Args:
predictions: A tensor of shape [B, C, L] representing the logits of the
predicted sequences. B is the batch size, C is the number
of classes (including the blank label), and L is the sequence
length of the predictions.
targets: A tensor of shape [B, T] representing the target sequences.
B is the batch size and T is the target sequence length.
Note that T can vary within the batch.
blank_label: The index of the blank label. Defaults to 0.
Returns:
The CTC loss (a scalar tensor).
"""
batch_size, num_classes, prediction_length = predictions.shape
_, target_length = targets.shape
# 1. Log softmax on predictions: Crucially, CTC loss requires log probabilities.
log_probs = F.log_softmax(predictions, dim=1) # Shape: [B, C, L]
# 2. Prepare inputs for torch.nn.CTCLoss:
# a. Convert log_probs to shape (L, B, C): CTCLoss expects time first.
log_probs = log_probs.permute(2, 0, 1) # Shape: [L, B, C]
# b. Get lengths of the predicted sequences (all L in this case).
input_lengths = torch.full(size=(batch_size,), fill_value=prediction_length, dtype=torch.long)
# c. Get lengths of the target sequences.
target_lengths = torch.tensor([t.shape[0] for t in targets], dtype=torch.long) # Handle variable target lengths
# 3. Create the CTCLoss criterion. `blank=blank_label` is essential!
ctc_loss = torch.nn.CTCLoss(blank=blank_label, reduction='mean') # 'mean' for averaging over the batch
# 4. Calculate the loss. `targets` needs to be a concatenated tensor.
# We handle padding by only passing the valid lengths to CTCLoss.
concatenated_targets = torch.cat(list(targets)) # Concatenate targets
loss = ctc_loss(log_probs, concatenated_targets, input_lengths, target_lengths)
return loss
def sort_loss(predictions, targets):
"""
The sort task was used partly to show that ctc loss can work.
"""
loss = compute_ctc_loss(predictions, targets, blank_label=predictions.shape[1]-1)
return loss
def image_classification_loss(predictions, certainties, targets, use_most_certain=True):
"""
Computes the maze loss with auto-extending cirriculum.
Predictions are of shape: (B, class, internal_ticks),
Certainties are of shape: (B, 2, internal_ticks),
where the inside dimension (2) is [normalised_entropy, 1-normalised_entropy]
Targets are of shape: [B]
use_most_certain will select either the most certain point or the final point.
"""
targets_expanded = torch.repeat_interleave(targets.unsqueeze(-1), predictions.size(-1), -1)
# Losses are of shape [B, internal_ticks]
losses = nn.CrossEntropyLoss(reduction='none')(predictions, targets_expanded)
loss_index_1 = losses.argmin(dim=1)
loss_index_2 = certainties[:,1].argmax(-1)
if not use_most_certain: # Revert to final loss if set
loss_index_2[:] = -1
batch_indexer = torch.arange(predictions.size(0), device=predictions.device)
loss_minimum_ce = losses[batch_indexer, loss_index_1].mean()
loss_selected = losses[batch_indexer, loss_index_2].mean()
loss = (loss_minimum_ce + loss_selected)/2
return loss, loss_index_2
def maze_loss(predictions, certainties, targets, cirriculum_lookahead=5, use_most_certain=True):
"""
Computes the maze loss with auto-extending cirriculum.
Predictions are of shape: (B, route_length, class, internal_ticks),
where classes are in [0,1,2,3,4] for [Up, Down, Left, Right, Wait]
Certainties are of shape: (B, 2, internal_ticks),
where the inside dimension (2) is [normalised_entropy, 1-normalised_entropy]
Targets are of shape: [B, route_length]
cirriculum_lookahead: how far to look ahead in the auto-cirriculum
use_most_certain will select either the most certain point or the final point. For baselines,
the final point proved the only usable option.
"""
# Predictions reshaped to: [B*route_length, 5, internal_ticks]
predictions_reshaped = predictions.flatten(0,1)
# Targets reshaped to: [B*route_length, internal_ticks]
targets_reshaped = torch.repeat_interleave(targets.unsqueeze(-1),
predictions.size(-1), -1).flatten(0,1).long()
# Losses are of shape [B, route_length, internal_ticks]
losses = nn.CrossEntropyLoss(reduction='none')(predictions_reshaped, targets_reshaped)
losses = losses.reshape(predictions[:,:,0].shape)
# Below is the code for auto-cirriculum
# Find where correct, and make sure to always push +5 beyond that
iscorrects = (predictions.argmax(2) == targets.unsqueeze(-1)).cumsum(1)
correct_mask = (iscorrects == torch.arange(1, iscorrects.size(1)+1, device=iscorrects.device).reshape(1, -1, 1))
correct_mask[:,0,:] = 1
upto_where = correct_mask.cumsum(1).argmax(1).max(-1)[0]+cirriculum_lookahead
loss_mask = torch.zeros_like(losses)
for bi in range(predictions.size(0)):
loss_mask[bi, :upto_where[bi]] = 1
# Reduce losses along route dimension
# Will now be of shape [B, internal_ticks]
losses = (losses * loss_mask).sum(1)/(loss_mask.sum(1))
loss_index_1 = losses.argmin(dim=1)
loss_index_2 = certainties[:,1].argmax(-1)
if not use_most_certain:
loss_index_2[:] = -1
batch_indexer = torch.arange(predictions.size(0), device=predictions.device)
loss_minimum_ce = losses[batch_indexer, loss_index_1]
loss_selected = losses[batch_indexer, loss_index_2]
loss = ((loss_minimum_ce + loss_selected)/2).mean()
return loss, loss_index_2, upto_where.detach().cpu().numpy()
def parity_loss(predictions, certainties, targets, use_most_certain=True):
"""
Computes the parity loss.
Predictions are of shape: (B, parity_sequence_length, class, internal_ticks),
where classes are in [0,1,2,3,4] for [Up, Down, Left, Right, Wait]
Certainties are of shape: (B, 2, internal_ticks),
where the inside dimension (2) is [normalised_entropy, 1-normalised_entropy]
Targets are of shape: [B, parity_sequence_length]
use_most_certain will select either the most certain point or the final point. For baselines,
the final point proved the only usable option.
"""
# Losses are of shape [B, parity_sequence_length, internal_ticks]
losses = nn.CrossEntropyLoss(reduction='none')(predictions.flatten(0,1),
torch.repeat_interleave(targets.unsqueeze(-1),
predictions.size(-1), -1).flatten(0,1).long()).reshape(predictions[:,:,0].shape)
# Average the loss over the parity sequenece dimension
losses = losses.mean(1)
loss_index_1 = losses.argmin(dim=1)
loss_index_2 = certainties[:,1].argmax(-1)
if not use_most_certain:
loss_index_2[:] = -1
batch_indexer = torch.arange(predictions.size(0), device=predictions.device)
loss_minimum_ce = losses[batch_indexer, loss_index_1].mean()
loss_selected = losses[batch_indexer, loss_index_2].mean()
loss = (loss_minimum_ce + loss_selected)/2
return loss, loss_index_2
class EnergyContrastiveLoss(nn.Module):
def __init__(self, margin=10.0, energy_scale=0.1):
super().__init__()
self.margin = margin
self.energy_scale = energy_scale
self.ce_loss = nn.CrossEntropyLoss(reduction='none')
def forward(self, logits_history, energy_history, targets):
"""
logits_history: [B, Class, T]
energy_history: [B, 1, T]
targets: [B]
"""
B, C, T = logits_history.shape
# Flatten for easy computation
logits_flat = logits_history.permute(0, 2, 1).reshape(B * T, C)
energy_flat = energy_history.permute(0, 2, 1).reshape(B * T)
targets_expanded = targets.unsqueeze(1).repeat(1, T).reshape(B * T)
# 1. Standard Classification Loss (Cross Entropy)
ce_vals = self.ce_loss(logits_flat, targets_expanded)
# 2. Determine "Correctness" for Contrastive Divergence
# We treat a step as "positive" if the prediction matches the target
predictions = logits_flat.argmax(dim=1)
is_correct = (predictions == targets_expanded).float() # 1.0 if correct, 0.0 if wrong
# 3. Energy Loss Logic
# If Correct: Minimize Energy (Pull down to 0)
# If Incorrect: Maximize Energy (Push up to margin)
# L_pos = ||E(x)||^2 (Push correct states to 0 energy)
loss_pos = energy_flat ** 2
# L_neg = max(0, m - E(x))^2 (Push incorrect states above margin m)
loss_neg = F.relu(self.margin - energy_flat) ** 2
# Combine: correct samples use loss_pos, incorrect use loss_neg
energy_objective = (is_correct * loss_pos) + ((1 - is_correct) * loss_neg)
# Total Loss
total_loss = ce_vals.mean() + (self.energy_scale * energy_objective.mean())
return total_loss, {
"ce_loss": ce_vals.mean().item(),
"energy_loss": energy_objective.mean().item(),
"avg_energy": energy_flat.mean().item()
}
def qamnist_loss(predictions, certainties, targets, use_most_certain=True):
"""
Computes the qamnist loss over the last num_answer_steps steps.
Predictions are of shape: (B, class, internal_ticks),
Certainties are of shape: (B, 2, internal_ticks),
where the inside dimension (2) is [normalised_entropy, 1-normalised_entropy]
Targets are of shape: [B]
num_answer_steps: number of steps to consider for the loss
use_most_certain will select either the most certain point or the final point.
"""
losses = nn.CrossEntropyLoss(reduction='none')(predictions,
torch.repeat_interleave(targets.unsqueeze(-1), predictions.size(-1), -1))
loss_index_1 = losses.argmin(dim=1)
loss_index_2 = certainties[:,1].argmax(-1)
if not use_most_certain:
loss_index_2[:] = -1
batch_indexer = torch.arange(predictions.size(0), device=predictions.device)
loss_minimum_ce = losses[batch_indexer, loss_index_1].mean()
loss_selected = losses[batch_indexer, loss_index_2].mean()
loss = (loss_minimum_ce + loss_selected)/2
return loss, loss_index_2 |