Informer

https://github.com/zhouhaoyi/informer2020
Admin
Informer is an efficient Transformer-based model for long sequence time-series forecasting that...
Tokens:8,217
Snippets:51
Trust Score:7.5
License:Apache-2.0
Update:6 months ago
Show doc for...
Tokens:
Raw
### Install Project Dependencies

Source: https://github.com/zhouhaoyi/informer2020/blob/main/README.md

Installs the required Python libraries for the Informer project using pip and the provided requirements file.

```bash
pip install -r requirements.txt
```

--------------------------------

### Manage Experiments with Docker and Make

Source: https://context7.com/zhouhaoyi/informer2020/llms.txt

Provides Makefile commands for environment setup, dataset downloading, and executing experiments in a containerized environment.

```makefile
make init
make dataset
make run_module module="bash scripts/ETTh1.sh"
make jupyter
make bash_docker
```

--------------------------------

### Build Encoder with Self-Attention Distilling

Source: https://context7.com/zhouhaoyi/informer2020/llms.txt

Illustrates the construction of an Informer encoder using distillation layers to reduce sequence length. Includes examples for both standard Encoder and EncoderStack configurations.

```python
import torch
import torch.nn as nn
from models.encoder import Encoder, EncoderLayer, ConvLayer, EncoderStack
from models.attn import ProbAttention, AttentionLayer

d_model = 512
n_heads = 8
d_ff = 2048
e_layers = 2

def create_attn_layer():
    return AttentionLayer(
        ProbAttention(False, factor=5, attention_dropout=0.05, output_attention=False),
        d_model, n_heads, mix=False
    )

encoder_layers = [
    EncoderLayer(
        attention=create_attn_layer(),
        d_model=d_model,
        d_ff=d_ff,
        dropout=0.05,
        activation='gelu'
    ) for _ in range(e_layers)
]

conv_layers = [ConvLayer(d_model) for _ in range(e_layers - 1)]

encoder = Encoder(
    attn_layers=encoder_layers,
    conv_layers=conv_layers,
    norm_layer=nn.LayerNorm(d_model)
)

batch_size, seq_len = 32, 96
x = torch.randn(batch_size, seq_len, d_model)
enc_out, attentions = encoder(x, attn_mask=None)
```

--------------------------------

### Execute Informer Model Training and Inference

Source: https://github.com/zhouhaoyi/informer2020/blob/main/README.md

Commands to initiate the training or prediction process for the Informer model. The first example demonstrates a specific run on the ETTm1 dataset, while the second provides a template for full parameter configuration.

```bash
python -u main_informer.py --model informer --data ETTm1 --attn prob --freq t
```

```bash
python -u main_informer.py --model <model> --data <data> --root_path <root_path> --data_path <data_path> --features <features> --target <target> --freq <freq> --checkpoints <checkpoints> --seq_len <seq_len> --label_len <label_len> --pred_len <pred_len> --enc_in <enc_in> --dec_in <dec_in> --c_out <c_out> --d_model <d_model> --n_heads <n_heads> --e_layers <e_layers> --d_layers <d_layers> --s_layers <s_layers> --d_ff <d_ff> --factor <factor> --padding <padding> --distil --dropout <dropout> --attn <attn> --embed <embed> --activation <activation> --output_attention --do_predict --mix --cols <cols> --itr <itr> --num_workers <num_workers> --train_epochs <train_epochs> --batch_size <batch_size> --patience <patience> --des <des> --learning_rate <learning_rate> --loss <loss> --lradj <lradj> --use_amp --inverse --use_gpu <use_gpu> --gpu <gpu> --use_multi_gpu --devices <devices>
```

--------------------------------

### Decoder with Cross-Attention

Source: https://context7.com/zhouhaoyi/informer2020/llms.txt

This section details the decoder component of the Informer model, which utilizes both self-attention and cross-attention mechanisms for sequence generation. It includes a Python code example demonstrating its initialization and a forward pass.

```APIDOC
## Decoder with Cross-Attention

### Description
The decoder processes the label sequence and attends to encoder output using both self-attention and cross-attention for generating predictions.

### Method
N/A (This is a model component description, not an API endpoint)

### Endpoint
N/A

### Parameters
N/A

### Request Example
```python
import torch
import torch.nn as nn
from models.decoder import Decoder, DecoderLayer
from models.attn import ProbAttention, FullAttention, AttentionLayer

d_model = 512
n_heads = 8
d_ff = 2048
d_layers = 1

# Decoder uses ProbAttention for self-attention, FullAttention for cross-attention
decoder_layers = [
    DecoderLayer(
        self_attention=AttentionLayer(ProbAttention(True, factor=5, attention_dropout=0.05), d_model, n_heads, mix=True),
        cross_attention=AttentionLayer(FullAttention(False, factor=5, attention_dropout=0.05), d_model, n_heads, mix=False),
        d_model=d_model,
        d_ff=d_ff,
        dropout=0.05,
        activation='gelu'
    ) for _ in range(d_layers)
]

decoder = Decoder(layers=decoder_layers, norm_layer=nn.LayerNorm(d_model))

# Forward pass
batch_size = 32
label_len, pred_len = 48, 24
dec_seq_len = label_len + pred_len

# Decoder input and encoder output
x = torch.randn(batch_size, dec_seq_len, d_model)      # Decoder input
cross = torch.randn(batch_size, 48, d_model)           # Encoder output (after distillation)

dec_out = decoder(x, cross, x_mask=None, cross_mask=None)
print(f"Decoder output shape: {dec_out.shape}")  # [32, 72, 512]
```

### Response
#### Success Response (200)
N/A

#### Response Example
N/A
```

--------------------------------

### Configure and Run Informer Experiment

Source: https://context7.com/zhouhaoyi/informer2020/llms.txt

Demonstrates how to initialize experiment arguments using a dotdict, instantiate the Informer model, and execute training, testing, and prediction workflows.

```python
args = dotdict({'data': 'ETTh1', 'seq_len': 96, 'label_len': 48, 'pred_len': 24, 'model': 'informer', 'train_epochs': 6, 'batch_size': 32, 'learning_rate': 0.0001, 'use_gpu': True})
exp = Exp_Informer(args)
setting = f"{args.model}_{args.data}_ft{args.features}_sl{args.seq_len}_pl{args.pred_len}"
model = exp.train(setting)
exp.test(setting)
exp.predict(setting, load=True)
```

--------------------------------

### Initialize and Run Informer Model in Python

Source: https://context7.com/zhouhaoyi/informer2020/llms.txt

Shows how to import and instantiate the Informer and InformerStack classes. It includes setting up input tensors and performing a forward pass to generate predictions.

```python
import torch
from models.model import Informer, InformerStack

# Initialize Informer model
model = Informer(
    enc_in=7,
    dec_in=7,
    c_out=7,
    seq_len=96,
    label_len=48,
    out_len=24,
    factor=5,
    d_model=512,
    n_heads=8,
    e_layers=2,
    d_layers=1,
    d_ff=2048,
    dropout=0.05,
    attn='prob',
    embed='timeF',
    freq='h',
    activation='gelu',
    output_attention=False,
    distil=True,
    mix=True,
    device=torch.device('cuda:0')
)

# Forward pass
batch_size = 32
x_enc = torch.randn(batch_size, 96, 7).cuda()
x_mark_enc = torch.randn(batch_size, 96, 4).cuda()
x_dec = torch.randn(batch_size, 72, 7).cuda()
x_mark_dec = torch.randn(batch_size, 72, 4).cuda()

# Get predictions
output = model(x_enc, x_mark_enc, x_dec, x_mark_dec)
print(f"Output shape: {output.shape}")

# InformerStack variant
model_stack = InformerStack(
    enc_in=7, dec_in=7, c_out=7,
    seq_len=96, label_len=48, out_len=24,
    e_layers=[3, 2, 1],
    d_layers=1,
    d_model=512, n_heads=8, d_ff=2048,
    attn='prob', distil=True,
    device=torch.device('cuda:0')
)
```

--------------------------------

### Configure Multi-GPU Training

Source: https://context7.com/zhouhaoyi/informer2020/llms.txt

Demonstrates how to enable multi-GPU training via command-line arguments or using PyTorch DataParallel in the model initialization script.

```bash
python -u main_informer.py --model informer --data ETTh1 --use_gpu True --use_multi_gpu --devices '0,1,2,3'
```

```python
args_use_multi_gpu = True
args_device_ids = [0, 1, 2, 3]
model = Informer(...)
if args_use_multi_gpu:
    model = nn.DataParallel(model, device_ids=args_device_ids)
model = model.cuda()
```

--------------------------------

### Implement ProbSparse and Full Attention Mechanisms

Source: https://context7.com/zhouhaoyi/informer2020/llms.txt

Demonstrates how to initialize and use ProbSparse Attention for O(L log L) complexity compared to standard Full Attention. It includes wrapping these mechanisms into an AttentionLayer with projection heads.

```python
import torch
from models.attn import ProbAttention, FullAttention, AttentionLayer

prob_attention = ProbAttention(
    mask_flag=True,
    factor=5,
    scale=None,
    attention_dropout=0.1,
    output_attention=False
)

full_attention = FullAttention(
    mask_flag=True,
    factor=5,
    attention_dropout=0.1,
    output_attention=False
)

d_model = 512
n_heads = 8
attention_layer = AttentionLayer(
    attention=prob_attention,
    d_model=d_model,
    n_heads=n_heads,
    d_keys=None,
    d_values=None,
    mix=False
)

batch_size, seq_len = 32, 96
queries = torch.randn(batch_size, seq_len, d_model)
keys = torch.randn(batch_size, seq_len, d_model)
values = torch.randn(batch_size, seq_len, d_model)

output, attn_weights = attention_layer(queries, keys, values, attn_mask=None)
print(f"Attention output shape: {output.shape}")
```

--------------------------------

### Train Informer Model via CLI

Source: https://context7.com/zhouhaoyi/informer2020/llms.txt

Demonstrates how to execute training, evaluation, and prediction tasks using the main_informer.py script. It covers basic configurations, multivariate forecasting, univariate long-term forecasting, and enabling prediction mode.

```bash
# Basic training on ETTh1 dataset with ProbSparse attention
python -u main_informer.py --model informer --data ETTh1 --attn prob --freq h

# Full multivariate training with custom parameters
python -u main_informer.py \
    --model informer \
    --data ETTh1 \
    --features M \
    --seq_len 96 \
    --label_len 48 \
    --pred_len 24 \
    --enc_in 7 \
    --dec_in 7 \
    --c_out 7 \
    --d_model 512 \
    --n_heads 8 \
    --e_layers 2 \
    --d_layers 1 \
    --d_ff 2048 \
    --factor 5 \
    --dropout 0.05 \
    --attn prob \
    --embed timeF \
    --activation gelu \
    --train_epochs 6 \
    --batch_size 32 \
    --learning_rate 0.0001 \
    --patience 3 \
    --des 'Exp' \
    --itr 5

# Univariate long-term forecasting (720 hours ahead)
python -u main_informer.py \
    --model informer \
    --data ETTh1 \
    --features S \
    --seq_len 720 \
    --label_len 336 \
    --pred_len 720 \
    --e_layers 2 \
    --d_layers 1 \
    --attn prob \
    --des 'Exp' \
    --itr 5

# Enable prediction on unseen future data
python -u main_informer.py \
    --model informer \
    --data ETTh1 \
    --features M \
    --seq_len 96 \
    --label_len 48 \
    --pred_len 24 \
    --attn prob \
    --do_predict
```

--------------------------------

### Implement Training Loop with Early Stopping and LR Adjustment

Source: https://context7.com/zhouhaoyi/informer2020/llms.txt

Demonstrates how to integrate early stopping and dynamic learning rate scheduling into a PyTorch training loop. It assumes the existence of early_stopping and adjust_learning_rate utility functions.

```python
model = torch.nn.Linear(10, 1)
optimizer = torch.optim.Adam(model.parameters(), lr=0.0001)
checkpoint_path = './checkpoints/experiment/'
for epoch in range(100):
    val_loss = 0.05
    early_stopping(val_loss, model, checkpoint_path)
    if early_stopping.early_stop:
        break

class Args:
    learning_rate = 0.0001
    lradj = 'type1'
args = Args()
for epoch in range(1, 20):
    adjust_learning_rate(optimizer, epoch, args)
```

--------------------------------

### Execute Training Scripts

Source: https://github.com/zhouhaoyi/informer2020/blob/main/README.md

Iterates through all scripts located in the scripts directory and executes them using the project's makefile system.

```bash
for file in `ls scripts`; do make run_module module="bash scripts/$script"; done
```

--------------------------------

### Initialize Experiment Runner

Source: https://context7.com/zhouhaoyi/informer2020/llms.txt

This snippet initializes the Exp_Informer class, which serves as the main entry point for managing the training, validation, and testing lifecycle of the Informer model.

```python
import argparse
from exp.exp_informer import Exp_Informer
from utils.tools import dotdict
```

--------------------------------

### Load and Preprocess Time-Series Datasets

Source: https://context7.com/zhouhaoyi/informer2020/llms.txt

This snippet shows how to use the provided PyTorch Dataset classes to load ETT or custom time-series data. It covers configuration for multivariate/univariate features, normalization, and creating a DataLoader for training.

```python
import torch
from torch.utils.data import DataLoader
from data.data_loader import Dataset_ETT_hour, Dataset_ETT_minute, Dataset_Custom

train_dataset = Dataset_ETT_hour(
    root_path='./data/ETT/',
    data_path='ETTh1.csv',
    flag='train',
    size=[96, 48, 24],
    features='M',
    target='OT',
    scale=True,
    inverse=False,
    timeenc=1,
    freq='h'
)

train_loader = DataLoader(
    train_dataset,
    batch_size=32,
    shuffle=True,
    num_workers=0,
    drop_last=True
)

for batch_x, batch_y, batch_x_mark, batch_y_mark in train_loader:
    print(f"batch_x (encoder input): {batch_x.shape}")
    break

predictions = torch.randn(32, 24, 7)
original_scale = train_dataset.inverse_transform(predictions)
```

--------------------------------

### Implement Early Stopping

Source: https://context7.com/zhouhaoyi/informer2020/llms.txt

Configures an EarlyStopping object to monitor model performance during training and halt execution when validation metrics stop improving.

```python
from utils.tools import EarlyStopping
early_stopping = EarlyStopping(patience=3, verbose=True, delta=0)
```

--------------------------------

### Implement Informer Decoder with Cross-Attention

Source: https://context7.com/zhouhaoyi/informer2020/llms.txt

This snippet demonstrates how to configure and instantiate the Informer decoder. It utilizes ProbAttention for self-attention and FullAttention for cross-attention mechanisms to process encoder outputs and generate sequence predictions.

```python
import torch
import torch.nn as nn
from models.decoder import Decoder, DecoderLayer
from models.attn import ProbAttention, FullAttention, AttentionLayer

d_model = 512
n_heads = 8
d_ff = 2048
d_layers = 1

decoder_layers = [
    DecoderLayer(
        self_attention=AttentionLayer(
            ProbAttention(True, factor=5, attention_dropout=0.05),
            d_model, n_heads, mix=True
        ),
        cross_attention=AttentionLayer(
            FullAttention(False, factor=5, attention_dropout=0.05),
            d_model, n_heads, mix=False
        ),
        d_model=d_model,
        d_ff=d_ff,
        dropout=0.05,
        activation='gelu'
    ) for _ in range(d_layers)
]

decoder = Decoder(
    layers=decoder_layers,
    norm_layer=nn.LayerNorm(d_model)
)

batch_size = 32
label_len, pred_len = 48, 24
dec_seq_len = label_len + pred_len

x = torch.randn(batch_size, dec_seq_len, d_model)
cross = torch.randn(batch_size, 48, d_model)

dec_out = decoder(x, cross, x_mask=None, cross_mask=None)
print(f"Decoder output shape: {dec_out.shape}")
```

--------------------------------

### Train Informer Model

Source: https://github.com/zhouhaoyi/informer2020/blob/main/README.md

Command to initiate training of the Informer model on the ETTh1 dataset using ProbSparse attention.

```bash
python -u main_informer.py --model informer --data ETTh1 --attn prob --freq h
```

--------------------------------

### Dataset Loading

Source: https://context7.com/zhouhaoyi/informer2020/llms.txt

This section describes the PyTorch Dataset classes provided for loading and preprocessing time-series data. It covers automatic train/validation/test splitting, normalization, and handling different data frequencies (hourly, minutely, custom).

```APIDOC
## Dataset Loading

### Description
PyTorch Dataset classes for loading and preprocessing time-series data with automatic train/val/test splitting and normalization.

### Method
N/A (This is a utility description, not an API endpoint)

### Endpoint
N/A

### Parameters
N/A

### Request Example
```python
import torch
from torch.utils.data import DataLoader
from data.data_loader import Dataset_ETT_hour, Dataset_ETT_minute, Dataset_Custom

# Load hourly ETT dataset
train_dataset = Dataset_ETT_hour(
    root_path='./data/ETT/',
    data_path='ETTh1.csv',
    flag='train',                    # 'train', 'val', or 'test'
    size=[96, 48, 24],               # [seq_len, label_len, pred_len]
    features='M',                    # 'M': multivariate, 'S': univariate, 'MS': multi-to-single
    target='OT',                     # Target column for S/MS tasks
    scale=True,                      # Enable StandardScaler normalization
    inverse=False,                   # Inverse transform outputs
    timeenc=1,                       # 0: categorical, 1: continuous time features
    freq='h'                         # Time frequency
)

# Create DataLoader
train_loader = DataLoader(
    train_dataset,
    batch_size=32,
    shuffle=True,
    num_workers=0,
    drop_last=True
)

# Iterate through batches
for batch_x, batch_y, batch_x_mark, batch_y_mark in train_loader:
    print(f"batch_x (encoder input): {batch_x.shape}")       # [32, 96, 7]
    print(f"batch_y (decoder target): {batch_y.shape}")      # [32, 72, 7]
    print(f"batch_x_mark (time features): {batch_x_mark.shape}")  # [32, 96, 4]
    print(f"batch_y_mark (time features): {batch_y_mark.shape}")  # [32, 72, 4]
    break

# Load minutely data
minute_dataset = Dataset_ETT_minute(
    root_path='./data/ETT/',
    data_path='ETTm1.csv',
    flag='train',
    size=[96, 48, 24],
    features='M',
    timeenc=1,
    freq='t'  # minutely
)

# Custom dataset for your own data
# CSV format: ['date', feature1, feature2, ..., target]
custom_dataset = Dataset_Custom(
    root_path='./data/',
    data_path='your_data.csv',
    flag='train',
    size=[96, 48, 24],
    features='M',
    target='your_target_column',
    cols=['feature1', 'feature2', 'target'],  # Optional: select specific columns
    timeenc=1,
    freq='h'
)

# Inverse transform predictions back to original scale
predictions = torch.randn(32, 24, 7)
original_scale = train_dataset.inverse_transform(predictions)
```

### Response
#### Success Response (200)
N/A

#### Response Example
N/A
```

--------------------------------

### Normalize Time-Series Data with StandardScaler

Source: https://context7.com/zhouhaoyi/informer2020/llms.txt

Shows how to fit a scaler on training data and transform it, as well as how to inverse transform predictions back to the original scale.

```python
scaler = StandardScaler()
train_data = np.random.randn(1000, 7)
scaler.fit(train_data)
normalized = scaler.transform(train_data)
predictions = torch.randn(32, 24, 7)
original_scale = scaler.inverse_transform(predictions)
```

--------------------------------

### Configure Data Embedding Layers

Source: https://context7.com/zhouhaoyi/informer2020/llms.txt

Shows how to combine value, positional, and temporal embeddings to create rich input representations. It covers the DataEmbedding wrapper and individual component usage.

```python
import torch
from models.embed import DataEmbedding, TokenEmbedding, PositionalEmbedding

embedding = DataEmbedding(
    c_in=7,
    d_model=512,
    embed_type='timeF',
    freq='h',
    dropout=0.1
)

batch_size, seq_len = 32, 96
x = torch.randn(batch_size, seq_len, 7)
x_mark = torch.randn(batch_size, seq_len, 4)

embedded = embedding(x, x_mark)
print(f"Embedded shape: {embedded.shape}")

token_embed = TokenEmbedding(c_in=7, d_model=512)
pos_embed = PositionalEmbedding(d_model=512, max_len=5000)

token_out = token_embed(x)
pos_out = pos_embed(x)
```

--------------------------------

### Experiment Runner

Source: https://context7.com/zhouhaoyi/informer2020/llms.txt

This section introduces the `Exp_Informer` class, which serves as the central orchestrator for the entire Informer model pipeline, including training, validation, testing, and prediction phases. It also manages functionalities like early stopping and learning rate scheduling.

```APIDOC
## Experiment Runner

### Description
The `Exp_Informer` class orchestrates the complete training, validation, testing, and prediction pipeline with early stopping and learning rate scheduling.

### Method
N/A (This is a class description, not an API endpoint)

### Endpoint
N/A

### Parameters
N/A

### Request Example
```python
import argparse
from exp.exp_informer import Exp_Informer
from utils.tools import dotdict

# Example usage (conceptual):
# parser = argparse.ArgumentParser(description='Informer Experiment Runner')
# # Add arguments for configuration, data paths, hyperparameters, etc.
# args = parser.parse_args()

# # Convert args to dotdict for easier access
# config = dotdict(vars(args))

# # Initialize the experiment runner
# exp = Exp_Informer(config)

# # Run the experiment (e.g., training)
# exp.train()

# # Evaluate the model
# exp.test()
```

### Response
#### Success Response (200)
N/A

#### Response Example
N/A
```

--------------------------------

### Compute Time-Series Forecasting Metrics

Source: https://context7.com/zhouhaoyi/informer2020/llms.txt

Provides utility functions to calculate standard regression metrics such as MAE, MSE, RMSE, MAPE, and MSPE for time-series predictions against ground truth data.

```python
import numpy as np
from utils.metrics import metric, MAE, MSE, RMSE, MAPE, MSPE
pred = np.random.randn(1000, 24, 7)
true = np.random.randn(1000, 24, 7)
mae, mse, rmse, mape, mspe = metric(pred, true)
mae_value = MAE(pred, true)
```

--------------------------------

### Encoder with Distilling

Source: https://context7.com/zhouhaoyi/informer2020/llms.txt

Constructs an encoder stack with optional distillation layers to reduce sequence length and highlight dominant features.

```APIDOC
## Encoder with Self-Attention Distilling

### Description
Builds a deep encoder structure using multiple EncoderLayers and ConvLayer-based distillation.

### Parameters
- **attn_layers** (list) - Required - List of EncoderLayer instances.
- **conv_layers** (list) - Optional - List of ConvLayer instances for distillation.
- **norm_layer** (nn.Module) - Required - Normalization layer.

### Request Example
encoder = Encoder(attn_layers=encoder_layers, conv_layers=conv_layers, norm_layer=nn.LayerNorm(d_model))

### Response
- **enc_out** (Tensor) - The distilled output representation from the encoder.
```

--------------------------------

### Extract Temporal Features

Source: https://context7.com/zhouhaoyi/informer2020/llms.txt

Shows how to extract time-based features from datetime indices, supporting both continuous normalized encoding and categorical encoding for various time granularities.

```python
import pandas as pd
from utils.timefeatures import time_features
dates = pd.DataFrame({'date': pd.date_range('2020-01-01', periods=100, freq='H')})
features = time_features(dates, timeenc=1, freq='h')
features_cat = time_features(dates.copy(), timeenc=0, freq='h')
```

--------------------------------

### ProbSparse Attention Mechanism

Source: https://context7.com/zhouhaoyi/informer2020/llms.txt

Configures and executes the ProbSparse Attention mechanism, which reduces complexity to O(L log L) by selecting dominant queries.

```APIDOC
## ProbSparse Attention Mechanism

### Description
Initializes and runs the ProbSparse Attention layer, designed for long sequence processing by sampling dominant keys.

### Method
N/A (Class Initialization)

### Parameters
- **mask_flag** (bool) - Required - Enable causal masking.
- **factor** (int) - Required - Sampling factor for key selection.
- **attention_dropout** (float) - Optional - Dropout rate.

### Request Example
prob_attention = ProbAttention(mask_flag=True, factor=5, attention_dropout=0.1)

### Response
- **output** (Tensor) - The attention-weighted representation of the input sequence.
```

--------------------------------

### Data Embedding Module

Source: https://context7.com/zhouhaoyi/informer2020/llms.txt

Combines value, positional, and temporal embeddings to create a rich input representation for time-series forecasting.

```APIDOC
## Data Embedding

### Description
Processes raw input features and time markers into high-dimensional embeddings.

### Parameters
- **c_in** (int) - Required - Number of input features.
- **d_model** (int) - Required - Embedding dimension.
- **embed_type** (string) - Optional - Type of embedding ('timeF', 'fixed', 'learned').
- **freq** (string) - Optional - Time frequency code.

### Request Example
embedding = DataEmbedding(c_in=7, d_model=512, embed_type='timeF', freq='h')

### Response
- **embedded** (Tensor) - Combined embedding tensor of shape [batch_size, seq_len, d_model].
```
Tokens:
Raw