SuperLinear / modeling_super_linear.py
lirannoc's picture
Upload 7 files
5a08ee9 verified
from typing import Optional, Tuple, Dict, List, Union
import copy
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.nn.functional import interpolate
from transformers import PreTrainedModel
from transformers.modeling_outputs import CausalLMOutputWithCrossAttentions
from .configuration_super_linear import SuperLinearConfig
"-------------------------------------------------------------------------------------------------------------------"
class RevIN(nn.Module):
def __init__(self, num_features: int, eps=1e-5, affine=True, norm_type=None, subtract_last=False):
"""
:param num_features: the number of features or channels
:param eps: a value added for numerical stability
:param affine: if True, RevIN has learnable affine parameters
"""
super(RevIN, self).__init__()
self.num_features = num_features
self.eps = eps
self.affine = affine
self.subtract_last = subtract_last
self.norm_type = norm_type
if self.affine:
self._init_params()
def forward(self, x, mode: str):
if mode == 'norm':
self._get_statistics(x)
x = self._normalize(x)
elif mode == 'denorm':
x = self._denormalize(x)
else:
raise NotImplementedError
return x
def _init_params(self):
# initialize RevIN params: (C,)
self.affine_weight = nn.Parameter(torch.ones(self.num_features))
self.affine_bias = nn.Parameter(torch.zeros(self.num_features))
def _get_statistics(self, x):
dim2reduce = tuple(range(1, x.ndim-1))
if self.subtract_last:
self.last = x[:, -1:, :].detach()
self.mean = torch.mean(x[:, :-1, :], dim=dim2reduce, keepdim=True).detach()
else:
self.mean = torch.mean(x, dim=dim2reduce, keepdim=True).detach()
self.stdev = torch.sqrt(torch.var(x, dim=dim2reduce, keepdim=True, unbiased=False) + self.eps).detach()
if self.norm_type == "l1":
self.stdev = torch.mean(torch.abs(x - self.mean), dim=dim2reduce, keepdim=True).detach()
elif self.norm_type == "l2":
self.stdev = torch.sqrt(torch.mean((x - self.mean) ** 2, dim=dim2reduce, keepdim=True) + self.eps).detach()
def _normalize(self, x):
if self.subtract_last:
x = x - self.last
else:
x = x - self.mean
x = x / self.stdev
if self.norm_type in ["l1", "l2"]:
x = x / self.stdev
if self.affine:
x = x * self.affine_weight
x = x + self.affine_bias
return x
def _denormalize(self, x):
if self.affine:
x = x - self.affine_bias
x = x / (self.affine_weight + self.eps*self.eps)
if self.norm_type in ["l1", "l2"]:
x = x * self.stdev
x = x * self.stdev
if self.subtract_last:
x = x + self.last
else:
x = x + self.mean
return x
"-------------------------------------------------------------------------------------------------------------------"
class Linear(nn.Module):
"""Simple linear layer expert."""
def __init__(self, input_len, output_len):
super(Linear, self).__init__()
self.Linear = nn.Linear(input_len, output_len)
def forward(self, x):
# x: [Batch*Channel, Input length]
x = x.clone()
x = self.Linear(x).clone()
return x # to [Batch, Output length, Channel]
class Naive(nn.Module):
"""Naive forecasting expert - repeats last value."""
def __init__(self, input_len, output_len):
super(Naive, self).__init__()
self.output_len = output_len
def forward(self, x):
# x: [Batch*Channel, Input length]
x = x[:,-1].unsqueeze(1).repeat(1, self.output_len)
return x # to [Batch, Output length, Channel]
class Mean(nn.Module):
"""Mean forecasting expert - repeats mean value."""
def __init__(self, input_len, output_len):
super(Mean, self).__init__()
self.output_len = output_len
def forward(self, x):
# x: [Batch*Channel, Input length]
x = x.mean(dim=1).unsqueeze(1).repeat(1, self.output_len)
return x # to [Batch, Output length, Channel]
class RLinear(nn.Module):
"""Reversible Instance Normalization Linear layer expert."""
def __init__(self, input_len, output_len):
super(RLinear, self).__init__()
self.Linear = nn.Linear(input_len, output_len)
self.revin_layer = RevIN(num_features = None, affine=False, norm_type = None, subtract_last = False)
def forward(self, x):
# x: [Batch, Input length,Channel]
x_shape = x.shape
if len(x_shape) == 2:
x = x.unsqueeze(-1)
x = x.clone()
x = self.revin_layer(x, 'norm')
x = self.Linear(x.permute(0,2,1)).permute(0,2,1).clone()
x = self.revin_layer(x, 'denorm')
if len(x_shape) == 2:
x = x.squeeze(-1)
return x # to [Batch, Output length, Channel]
"-------------------------------------------------------------------------------------------------------------------"
class SparseMoE(nn.Module):
"""
Sparse Mixture of Experts (MoE) module that routes inputs to the most relevant experts.
This implementation uses a gating network to determine which experts should process each input.
Only the top-k experts are used for each input, creating a sparse computation pattern.
Args:
configs: Configuration object containing MoE parameters
experts: Collection of expert modules (neural networks)
"""
def __init__(self, configs, experts=None):
super(SparseMoE, self).__init__()
self.noise_std = configs.noisy_gating_std
self.experts = nn.ModuleList(experts) # Store experts in ModuleList for proper registration
self.num_experts = len(experts)
self.k = configs.top_k_experts
if self.k > self.num_experts:
self.k = self.num_experts
self.moe_temp = configs.moe_temp
self.use_fft = configs.use_fft
self.fft_len = configs.fft_len
self.moe_norm = configs.moe_norm
# Initialize gating network based on configuration
if self.use_fft:
self.gating_network = nn.Linear(self.fft_len//2, self.num_experts, bias=True)
else:
self.gating_network = nn.Linear(configs.train_seq_len, self.num_experts, bias=True)
if self.moe_norm:
self.batch_norm = nn.BatchNorm1d(self.num_experts)
def get_periodogram(self, inputs, n=10000):
"""
Calculate the periodogram (power spectral density) of input time series.
The periodogram is used as a frequency-domain representation of the signal
to help the gating network identify periodic patterns.
Args:
inputs: Input time series tensor of shape [batch_size, sequence_length] or [batch_size, sequence_length, features]
n: Number of points in FFT computation
Returns:
Normalized periodogram of the input signals
"""
x_0 = inputs - torch.mean(inputs, dim=1, keepdim=True) # Remove mean (DC component)
# Compute FFT and normalize
dft = torch.fft.fft(x_0, dim=1, n=n) / np.sqrt(n)
dft = dft[:, :n//2] # Keep only positive frequencies
I = torch.abs(dft) ** 2 # Power spectral density
# Normalize periodogram
I_sum = torch.sum(I, dim=1, keepdim=True)
I_sum[I_sum == 0] = 1 # Avoid division by zero
I = I / I_sum
return I
def forward(self, x, get_prob=False, get_prob_only=False):
"""
Forward pass through the Mixture of Experts.
Args:
x: Input tensor of shape [batch_size, sequence_length]
get_prob: Whether to return expert selection probabilities
get_prob_only: Whether to return only probabilities without computation
Returns:
- Output tensor from the selected experts
- (Optional) Expert selection probabilities if get_prob is True
"""
# Preprocess input if using FFT-based gating
if self.use_fft:
x_0 = self.get_periodogram(x, n=self.fft_len)
else:
x_0 = x
# Get gating logits
gate_outputs = self.gating_network(x_0) # Raw gating scores
if self.moe_norm:
gate_outputs = self.batch_norm(gate_outputs)
# Apply temperature scaling during inference
if not self.training:
gate_outputs = gate_outputs / self.moe_temp
if get_prob_only:
expert_probs = F.softmax(gate_outputs, dim=1)
return expert_probs
# Add noise to gating logits during training (for exploration)
if self.training:
noise = torch.randn_like(gate_outputs).to(x.device) * self.noise_std
noisy_gate_outputs = gate_outputs + noise
topk_values, topk_indices = torch.topk(noisy_gate_outputs, self.k, dim=1)
else:
topk_values, topk_indices = torch.topk(gate_outputs, self.k, dim=1)
# Normalize the gate values with softmax
topk_gates = F.softmax(topk_values, dim=1)
# Get outputs from all experts
expert_outputs = torch.stack([self.experts[i](x) for i in range(self.num_experts)], dim=1)
# Select only the outputs from the top-k experts
topk_indices_expanded = topk_indices.unsqueeze(-1).expand(-1, -1, expert_outputs.size(2))
sparse_expert_outputs = torch.gather(expert_outputs, 1, topk_indices_expanded)
# Combine expert outputs using the gate values
output = torch.sum(topk_gates.unsqueeze(2) * sparse_expert_outputs, dim=1)
if get_prob:
expert_probs = F.softmax(gate_outputs, dim=1)
return output, expert_probs
return output
class Model(nn.Module):
"""
Main model class that employs a Mixture of Experts for time series forecasting.
This model can work with various types of linear layers as experts and supports
both standard prediction and auto-regressive prediction for longer horizons.
Args:
configs: Configuration object containing model parameters
"""
def __init__(self, configs):
super(Model, self).__init__()
self.configs = copy.deepcopy(configs)
# Core model configuration
self.train_pred_len = configs.train_pred_len
self.train_seq_len = configs.train_seq_len
self.layer_type = configs.layer_type
# Initialize additional configuration attributes with defaults
self.long_horizon_scaling = configs.long_horizon_scaling
self.lookback_resampling = configs.lookback_resampling
lookback_scale_str = configs.scale_list
if isinstance(lookback_scale_str, str):
self.scale_list = [float(x.strip()) for x in lookback_scale_str.split(',')]
else:
self.scale_list = lookback_scale_str # Already a list
self.threshold = configs.threshold
self.freq_bound = configs.freq_bound
self.penalty_scale = configs.penalty_scale
self.fft_len = configs.fft_len
# Parse frequency experts from configuration
freq_experts_str = configs.freq_experts
if freq_experts_str == "":
self.freq_experts = None
else:
self.freq_experts = freq_experts_str.split('_')
# Expert configuration
self.top_k_experts = configs.top_k_experts
self.freeze_experts = configs.freeze_experts
# Initialize experts based on frequency specification or create generic experts
self.experts = {}
if self.freq_experts is not None:
for expert_freq in self.freq_experts:
if expert_freq.lower() == "naive":
self.experts[expert_freq] = Naive(self.train_seq_len, self.train_pred_len)
elif expert_freq.lower() == "mean":
self.experts[expert_freq] = Mean(self.train_seq_len, self.train_pred_len)
else:
self.experts[expert_freq] = RLinear(self.train_seq_len, self.train_pred_len)
self.n_experts = len(self.experts)
else:
raise ValueError("Please specify experts in the configuration.")
# Create additional complementary experts if specified
comp_moe = configs.comp_moe
if comp_moe > 0:
if comp_moe == 1:
print("Creating complementary expert")
self.experts["comp"] = RLinear(self.train_seq_len, self.train_pred_len)
else:
for i in range(comp_moe):
print(f"Creating complementary expert {i}")
self.experts["comp_"+str(i)] = RLinear(self.train_seq_len, self.train_pred_len)
# Initialize the MoE layer and dropout
self.moe = SparseMoE(configs, experts=self.experts.values())
print("Experts:", self.experts.keys())
def add_experts(self, experts: Dict[str, nn.Module]) -> nn.Module:
"""
Add new experts to the model.
Args:
experts: Dictionary of expert instances to add
Returns:
Updated MoE layer
"""
for name, expert in experts.items():
if name not in self.experts:
self.experts[name] = expert
print(f"Added expert: {name}")
else:
print(f"Expert {name} already exists. Skipping addition.")
# Reinitialize the MoE layer with the updated experts
self.moe = SparseMoE(self.configs, experts=self.experts.values())
return self.moe
def apply_long_horizon_scaling(self, ar_out: torch.Tensor, ar_x: torch.Tensor) -> torch.Tensor:
"""
Apply scaling to auto-regressive outputs to maintain statistical properties during long horizon prediction.
This function identifies cases where the variance of the new predictions exceeds the variance
of the input sequence and applies scaling to maintain consistent statistical properties.
Args:
ar_out: Auto-regressive output tensor of shape [batch_size * features, pred_len]
ar_x: Input sequence tensor of shape [batch_size * features, seq_len]
Returns:
Scaled auto-regressive output tensor
"""
if not (self.long_horizon_scaling and not self.training):
return ar_out
# Calculate statistics for scaling
std_new = torch.std(ar_out, dim=1, keepdim=True)
mean_new = torch.mean(ar_out, dim=1, keepdim=True)
std_old = torch.std(ar_x, dim=1, keepdim=True)
# Find indices where new variance exceeds old variance
inds = torch.where(std_new / std_old > 1)[0]
if len(inds) > 0:
# Center the outputs around their mean
ar_out_centered = ar_out[inds] - mean_new[inds]
# Calculate scaling factor to match old variance
scaling = std_old[inds] / (std_new[inds] + 1e-8)
# Scale and shift back to mean_new
ar_out_adjusted = ar_out_centered * scaling + mean_new[inds]
ar_out[inds] = ar_out_adjusted
return ar_out
def lookback_resample_search(self, x, scale_list=[2,4,6], min_lookback=512):
"""
Search for optimal resampling scale based on lookback analysis of expert selection.
This function analyzes the frequency content and expert selection lookback to determine
the best resampling scale for each input sequence, potentially improving model performance
by matching input characteristics to expert capabilities.
Args:
x: Input tensor of shape [batch_size, features, sequence_length]
scale_list: List of potential downsampling scales to evaluate
min_lookback: Minimum sequence length required after resampling
Returns:
Tuple of (resampled_input, final_scales) where:
- resampled_input: Optimally resampled input tensor
- final_scales: Scale factors used for each sample
"""
B, V, L = x.shape
lookback = self.train_seq_len
x_0 = x.reshape(B*V, L)[:, -lookback:]
output_x = x_0.clone()[:, -lookback:]
x_reshape = x.reshape(B*V, L)
x_fft_init = self.moe.get_periodogram(x_reshape, n=self.fft_len)
right_cumsum = torch.cumsum(x_fft_init, dim=-1)
mask = right_cumsum > 1-self.threshold
j_threshold = mask.float().argmax(dim=-1)
freqs = np.array([np.linspace(0, 0.5, self.fft_len//2)])
threshhold_freqs = np.take_along_axis(freqs, j_threshold.unsqueeze(-1).detach().cpu().numpy(), axis=1)
# where threshhold_freqs is 0, set to a small value to avoid division by zero
threshhold_freqs[threshhold_freqs == 0] = self.freq_bound
max_scale_factor = (self.freq_bound/ threshhold_freqs).astype(int).flatten()
if self.threshold==0:
max_scale_factor = np.inf * np.ones(B*V, dtype=int)
# Compute energy loss penalty for each potential scale
energy_loss_penalties = {}
total_energy = torch.sum(x_fft_init, dim=-1) # Total energy per sample
for scale in scale_list:
if scale <= 1:
continue # No penalty for upsampling or no scaling
# Calculate Nyquist frequency after downsampling
nyquist_after_downsample = 0.5 / scale
# Find frequency bins that will be lost (above new Nyquist)
freq_bins = torch.linspace(0, 0.5, self.fft_len//2, device=x_fft_init.device)
lost_freq_mask = freq_bins > nyquist_after_downsample
# Calculate energy that will be lost
lost_energy = torch.sum(x_fft_init[:, lost_freq_mask], dim=-1)
# Energy loss fraction (0 = no loss, 1 = all energy lost)
energy_loss_fraction = lost_energy / (total_energy + 1e-10)
energy_loss_penalties[scale] = energy_loss_fraction
# Get initial entropy
prob = self.moe(x_0, get_prob_only=True)
best_scores = -torch.sum(prob * torch.log(prob + 1e-10), dim=-1)
final_scales = torch.ones(B*V, device=x.device)
for scale in scale_list:
x_interp = torch.nn.functional.interpolate(
x, scale_factor=1/scale, mode='linear', align_corners=True
)
if x_interp.shape[2] >= min_lookback:
x_interp_reshaped = x_interp.reshape(B*V, x_interp.shape[-1])
x_interp_reshaped = x_interp_reshaped[:, -lookback:]
prob = self.moe(x_interp_reshaped, get_prob_only=True)
scores = -torch.sum(prob * torch.log(prob + 1e-10), dim=-1)
# Add energy loss penalty
if scale in energy_loss_penalties:
energy_penalty = energy_loss_penalties[scale]
scores = scores + energy_penalty*self.penalty_scale
idx = np.where((scores < best_scores).cpu() & torch.tensor(max_scale_factor >= scale))[0]
if len(idx) > 0:
output_x[idx] = x_interp_reshaped[idx]
best_scores[idx] = scores[idx]
final_scales[idx] = scale
return output_x.reshape(B, V, output_x.shape[-1]), final_scales
def lookback_resample_reverse(self, y, final_scales, inf_pred_len=None):
"""
Reverse the resampling operation on the output.
This function upsamples the model outputs back to the original scale
based on the resampling factors used during input processing.
Args:
y: Output tensor from model of shape [batch_size, features, pred_len]
final_scales: Scale factors used during input resampling
inf_pred_len: Target prediction length
Returns:
Upsampled output tensor of shape [batch_size, features, inf_pred_len]
"""
B, V, L = y.shape
y_reshaped = y.view(B*V, L)
y_out = y_reshaped[:, :inf_pred_len]
unique_scales = torch.unique(final_scales)
for scale in unique_scales:
scale_val = scale.item() # Convert tensor to scalar
if scale_val > 1:
idx = torch.where(final_scales == scale)[0]
if len(idx) > 0:
y_interp = torch.nn.functional.interpolate(
y_reshaped[idx].unsqueeze(1), scale_factor=scale_val, mode='linear', align_corners=True
)
y_out[idx] = y_interp.reshape(len(idx), y_interp.shape[-1])[:, :inf_pred_len]
return y_out.reshape(B, V, inf_pred_len)
def forward(self, x_in: torch.Tensor, get_prob: bool = False, pred_len: Optional[int] = None) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
"""
Forward pass through the model.
Args:
x_in: Encoder input tensor of shape [batch_size, sequence_length] or [batch_size, features, sequence_length]
get_prob: Whether to return expert selection probabilities
pred_len: Override for prediction length
Returns:
- Prediction tensor
- (Optional) Expert selection probabilities if get_prob is True
"""
if pred_len is None:
pred_len = self.train_pred_len
x = x_in
# If input is 2D, add a channel dimension
if x_in.dim() == 2:
x = x.unsqueeze(1)
B, V, L = x.shape
short_lookback = False
orig_pred_len = pred_len
if L < self.train_seq_len:
# Handle case where input sequence is shorter than expected
# by interpolating to the required length
scale_factor = self.train_seq_len / L
scale_factor = int(np.ceil(scale_factor))
pred_len = pred_len * scale_factor
x = interpolate(x, scale_factor=scale_factor, mode='linear')
x = x[:, :, -self.train_seq_len:]
L = self.train_seq_len
short_lookback = True
# lookback resampling logic
final_scales = None
if self.lookback_resampling and L > self.train_seq_len:
x_resampled, final_scales = self.lookback_resample_search(
x, self.scale_list, self.train_seq_len
)
# Update x and L for the resampled input
x = x_resampled
L = x.shape[-1]
# Reshape to process each feature independently
x = x.reshape(B * V, L)
expert_probs = None
# Forward pass through MoE
if get_prob:
out, expert_probs = self.moe(x, get_prob=True)
else:
out = self.moe(x)
# Auto-regressive prediction for long horizons
if self.train_pred_len < pred_len:
outputs = [out]
ar_x = torch.cat([x, out], dim=1)[:, -self.train_seq_len:]
for i in range(0, pred_len, self.train_pred_len):
ar_out = self.moe(ar_x)
ar_out = self.apply_long_horizon_scaling(ar_out, ar_x)
outputs.append(ar_out)
ar_x = torch.cat([ar_x, ar_out], dim=1)[:, -self.train_seq_len:]
out = torch.cat(outputs, dim=1)[:, :pred_len]
# Reshape back to batch format
out = out.reshape(B, V, out.shape[-1])
# Apply lookback resampling reverse if it was used
if self.lookback_resampling and final_scales is not None and not short_lookback:
out = self.lookback_resample_reverse(out, final_scales, orig_pred_len)
# If we used interpolation earlier, now downsample back to original scale
if short_lookback:
out = interpolate(out, scale_factor=1/scale_factor, mode='linear')
out = out[:, :, :orig_pred_len]
if x_in.dim() == 2:
out = out.squeeze(1)
if get_prob:
expert_probs = expert_probs.reshape(B, V, expert_probs.shape[-1])
# expert_probs = expert_probs.permute(0, 2, 1) # [batch_size, num_experts, sequence_length]
if x_in.dim() == 2:
expert_probs = expert_probs.squeeze(-1)
return out, expert_probs
return out
def map_to_cycle(self, freq: str) -> int:
"""
Map frequency string notation to cycle length (number of periods).
Args:
freq: String representing a time frequency (e.g., "h" for hourly, "D" for daily)
Returns:
Integer representing the number of periods in the cycle
"""
cycle = int(freq.split("/")[1])
return cycle
"-------------------------------------------------------------------------------------------------------------------"
class SuperLinearForCausalLM(PreTrainedModel):
config_class = SuperLinearConfig
def __init__(self, config: SuperLinearConfig):
super().__init__(config)
# the backbone keeps its own Config dataclass, so build one on-the-fly:
backbone_cfg = type("Cfg", (), config.to_dict())()
self.args = backbone_cfg
self.backbone = Model(backbone_cfg)
self.post_init()
# ------------------------------------------------------------------
# Forward pass expected by AutoModelForCausalLM
# ------------------------------------------------------------------
def forward(self,
inputs_embeds: torch.Tensor = None,
pred_len: Optional[int] = None,
get_prob: bool = False,
**kwargs) -> CausalLMOutputWithCrossAttentions:
if inputs_embeds is None:
raise ValueError("inputs_embeds must be provided")
# backbone expects (B, C, L) or (B, L)
x_enc = inputs_embeds
# backbone returns (B, pred_len, C)
if get_prob:
preds, probs = self.backbone(x_enc, pred_len=pred_len, get_prob=True)
else:
preds = self.backbone(x_enc, pred_len=pred_len, get_prob=False)
probs = None
return CausalLMOutputWithCrossAttentions(
logits=preds,
hidden_states=None,
attentions=probs
)