PoliBa-DeepLearning/NanoSocrates
2
0
Fork 0
Code Issues 3 Pull Requests 1 Actions Packages Projects Releases Wiki Activity

Activated Dropout to avoid overfitting

Browse Source
This commit is contained in:
Christian Risi
2025-10-12 12:28:06 +02:00
parent f463f699cf
commit 4ca1d0a189
3 changed files with 22 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
Project_Model/Libs/Transformer/Classes/Decoder.py
View File

@@ -19,7 +19,7 @@ class Decoder(nn.Module):
self.__attention_heads = number_of_attention_heads
super().__init__()
self.__masked_attention = MultiHeadAttention(
embedding_dimension, number_of_attention_heads, dropout=0.1
@@ -68,12 +68,12 @@ class Decoder(nn.Module):
)
# 2) Dropout
# DROPPED_MASKED_ATTENTION = self.__dropout(MASKED_ATTENTION)
# del MASKED_ATTENTION
DROPPED_MASKED_ATTENTION = self.__dropout(MASKED_ATTENTION)
del MASKED_ATTENTION
# 3) Residual Connection
x = x + MASKED_ATTENTION
del MASKED_ATTENTION
x = x + DROPPED_MASKED_ATTENTION
del DROPPED_MASKED_ATTENTION
# 4) Layer Normalization
x = self.__layer_norm_1(x)
@@ -86,12 +86,12 @@ class Decoder(nn.Module):
)
# 6) Dropout
# DROPPED_CROSS_ATTENTION = self.__dropout(CROSS_ATTENTION)
# del CROSS_ATTENTION
DROPPED_CROSS_ATTENTION = self.__dropout(CROSS_ATTENTION)
del CROSS_ATTENTION
# 7) Residual Connection
x = x + CROSS_ATTENTION
del CROSS_ATTENTION
x = x + DROPPED_CROSS_ATTENTION
del DROPPED_CROSS_ATTENTION
# 8) Layer Normalization
x = self.__layer_norm_2(x)
@@ -100,12 +100,12 @@ class Decoder(nn.Module):
FEED_FORWARD = self.__feed_forward_network(x)
# 10) Dropout
# DROPPED_FEED_FORWARD = self.__dropout(FEED_FORWARD)
# del FEED_FORWARD
DROPPED_FEED_FORWARD = self.__dropout(FEED_FORWARD)
del FEED_FORWARD
# 11) Residual Connection
x = x + FEED_FORWARD
del FEED_FORWARD
x = x + DROPPED_FEED_FORWARD
del DROPPED_FEED_FORWARD
# 12) Layer Normalization
x = self.__layer_norm_3(x)