mathdatech/Python
1
0
Fork 0
mirror of https://github.com/TheAlgorithms/Python.git synced 2025-01-30 22:23:42 +00:00
Code Issues Packages Projects Releases Wiki Activity

[pre-commit.ci] auto fixes from pre-commit.com hooks

Browse Source 
for more information, see https://pre-commit.ci
This commit is contained in:
pre-commit-ci[bot] 2024-10-20 16:52:44 +00:00
parent 3b8848430c
commit 4f573e0d8d
1 changed files with 9 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

18
neural_network/sliding_window_attention.py
View File

@ -1,7 +1,7 @@
""" """
- - - - - -- - - - - - - - - - - - - - - - - - - - - - - - - - - - -- - - - - - - - - - - - - - - - - - - - - - -
Name - - sliding_window_attention.py Name - - sliding_window_attention.py
Goal - - Implement a neural network architecture using sliding Goal - - Implement a neural network architecture using sliding
window attention for sequence modeling tasks. window attention for sequence modeling tasks.
Detail: Total 5 layers neural network Detail: Total 5 layers neural network
* Input layer * Input layer
@ -12,11 +12,11 @@ Author: Stephen Lee
Github: 245885195@qq.com Github: 245885195@qq.com
Date: 2024.10.20 Date: 2024.10.20
References: References:
1. Choromanska, A., et al. (2020). "On the Importance of 1. Choromanska, A., et al. (2020). "On the Importance of
Initialization and Momentum in Deep Learning." *Proceedings Initialization and Momentum in Deep Learning." *Proceedings
of the 37th International Conference on Machine Learning*. of the 37th International Conference on Machine Learning*.
2. Dai, Z., et al. (2020). "Transformers are RNNs: Fast 2. Dai, Z., et al. (2020). "Transformers are RNNs: Fast
Autoregressive Transformers with Linear Attention." Autoregressive Transformers with Linear Attention."
*arXiv preprint arXiv:2006.16236*. *arXiv preprint arXiv:2006.16236*.
3. [Attention Mechanisms in Neural Networks](https://en.wikipedia.org/wiki/Attention_(machine_learning)) 3. [Attention Mechanisms in Neural Networks](https://en.wikipedia.org/wiki/Attention_(machine_learning))
- - - - - -- - - - - - - - - - - - - - - - - - - - - - - - - - - - -- - - - - - - - - - - - - - - - - - - - - - -
@ -28,7 +28,7 @@ import numpy as np
class SlidingWindowAttention: class SlidingWindowAttention:
"""Sliding Window Attention Module. """Sliding Window Attention Module.
This class implements a sliding window attention mechanism where This class implements a sliding window attention mechanism where
the model attends to a fixed-size window of context around each token. the model attends to a fixed-size window of context around each token.
Attributes: Attributes:
@ -54,13 +54,13 @@ class SlidingWindowAttention:
Forward pass for the sliding window attention. Forward pass for the sliding window attention.
Args: Args:
input_tensor (np.ndarray): Input tensor of shape (batch_size, input_tensor (np.ndarray): Input tensor of shape (batch_size,
seq_length, embed_dim). seq_length, embed_dim).
Returns: Returns:
np.ndarray: Output tensor of shape (batch_size, seq_length, embed_dim). np.ndarray: Output tensor of shape (batch_size, seq_length, embed_dim).
>>> x = np.random.randn(2, 10, 4) # Batch size 2, sequence >>> x = np.random.randn(2, 10, 4) # Batch size 2, sequence
>>> attention = SlidingWindowAttention(embed_dim=4, window_size=3) >>> attention = SlidingWindowAttention(embed_dim=4, window_size=3)
>>> output = attention.forward(x) >>> output = attention.forward(x)
>>> output.shape >>> output.shape
@ -95,7 +95,7 @@ if __name__ == "__main__":
# usage # usage
rng = np.random.default_rng() rng = np.random.default_rng()
x = rng.standard_normal((2, 10, 4)) # Batch size 2, x = rng.standard_normal((2, 10, 4)) # Batch size 2,
attention = SlidingWindowAttention(embed_dim=4, window_size=3) attention = SlidingWindowAttention(embed_dim=4, window_size=3)
output = attention.forward(x) output = attention.forward(x)
print(output) print(output)