Skip to content

Commit 756cc19

Browse files
mhjensenmhjensen
committed
update week 7
1 parent f0c88bb commit 756cc19

File tree

8 files changed

+1452
-21
lines changed

8 files changed

+1452
-21
lines changed

‎doc/pub/week7/ipynb/ipynb-week7-src.tar.gz‎

2.23 MB
Binary file not shown.

‎doc/pub/week7/ipynb/week7.ipynb‎

Lines changed: 725 additions & 21 deletions
Large diffs are not rendered by default.

‎doc/pub/week7/pdf/week7.pdf‎

1.6 MB
Binary file not shown.

‎doc/src/week7/_minted/45BFBE4480AD7F905AB70F217399FB65.highlight.minted‎

Lines changed: 132 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,132 @@
1+
\begin{MintedVerbatim}[commandchars=\\\{\},codes={\catcode`\$=3\catcode`\^=7\catcode`\_=8\relax}]
2+
\PYG{l+s+sd}{\PYGZdq{}\PYGZdq{}\PYGZdq{}}
3+
\PYG{l+s+sd}{Key components:}
4+
\PYG{l+s+sd}{1. **Data Handling**: Uses PyTorch DataLoader with MNIST dataset}
5+
\PYG{l+s+sd}{2. **LSTM Architecture**:}
6+
\PYG{l+s+sd}{ \PYGZhy{} Input sequence of 28 timesteps (image rows)}
7+
\PYG{l+s+sd}{ \PYGZhy{} 128 hidden units in LSTM layer}
8+
\PYG{l+s+sd}{ \PYGZhy{} Fully connected layer for classification}
9+
\PYG{l+s+sd}{3. **Training**:}
10+
\PYG{l+s+sd}{ \PYGZhy{} Cross\PYGZhy{}entropy loss}
11+
\PYG{l+s+sd}{ \PYGZhy{} Adam optimizer}
12+
\PYG{l+s+sd}{ \PYGZhy{} Automatic GPU utilization if available}
13+
14+
\PYG{l+s+sd}{This implementation typically achieves **97\PYGZhy{}98\PYGZpc{} accuracy** after 10 epochs. The main differences from the TensorFlow/Keras version:}
15+
\PYG{l+s+sd}{\PYGZhy{} Explicit device management (CPU/GPU)}
16+
\PYG{l+s+sd}{\PYGZhy{} Manual training loop}
17+
\PYG{l+s+sd}{\PYGZhy{} Different data loading pipeline}
18+
\PYG{l+s+sd}{\PYGZhy{} More explicit tensor reshaping}
19+
20+
\PYG{l+s+sd}{To improve performance, you could:}
21+
\PYG{l+s+sd}{1. Add dropout regularization}
22+
\PYG{l+s+sd}{2. Use bidirectional LSTM}
23+
\PYG{l+s+sd}{3. Implement learning rate scheduling}
24+
\PYG{l+s+sd}{4. Add batch normalization}
25+
\PYG{l+s+sd}{5. Increase model capacity (more layers/units)}
26+
\PYG{l+s+sd}{\PYGZdq{}\PYGZdq{}\PYGZdq{}}
27+
28+
\PYG{k+kn}{import}\PYG{+w}{ }\PYG{n+nn}{torch}
29+
\PYG{k+kn}{import}\PYG{+w}{ }\PYG{n+nn}{torch}\PYG{n+nn}{.}\PYG{n+nn}{nn}\PYG{+w}{ }\PYG{k}{as}\PYG{+w}{ }\PYG{n+nn}{nn}
30+
\PYG{k+kn}{import}\PYG{+w}{ }\PYG{n+nn}{torch}\PYG{n+nn}{.}\PYG{n+nn}{optim}\PYG{+w}{ }\PYG{k}{as}\PYG{+w}{ }\PYG{n+nn}{optim}
31+
\PYG{k+kn}{from}\PYG{+w}{ }\PYG{n+nn}{torchvision}\PYG{+w}{ }\PYG{k+kn}{import} \PYG{n}{datasets}\PYG{p}{,} \PYG{n}{transforms}
32+
\PYG{k+kn}{from}\PYG{+w}{ }\PYG{n+nn}{torch}\PYG{n+nn}{.}\PYG{n+nn}{utils}\PYG{n+nn}{.}\PYG{n+nn}{data}\PYG{+w}{ }\PYG{k+kn}{import} \PYG{n}{DataLoader}
33+
34+
\PYG{c+c1}{\PYGZsh{} Hyperparameters}
35+
\PYG{n}{input\PYGZus{}size} \PYG{o}{=} \PYG{l+m+mi}{28} \PYG{c+c1}{\PYGZsh{} Number of features (pixels per row)}
36+
\PYG{n}{hidden\PYGZus{}size} \PYG{o}{=} \PYG{l+m+mi}{128} \PYG{c+c1}{\PYGZsh{} LSTM hidden state size}
37+
\PYG{n}{num\PYGZus{}classes} \PYG{o}{=} \PYG{l+m+mi}{10} \PYG{c+c1}{\PYGZsh{} Digits 0\PYGZhy{}9}
38+
\PYG{n}{num\PYGZus{}epochs} \PYG{o}{=} \PYG{l+m+mi}{10} \PYG{c+c1}{\PYGZsh{} Training iterations}
39+
\PYG{n}{batch\PYGZus{}size} \PYG{o}{=} \PYG{l+m+mi}{64} \PYG{c+c1}{\PYGZsh{} Batch size}
40+
\PYG{n}{learning\PYGZus{}rate} \PYG{o}{=} \PYG{l+m+mf}{0.001}
41+
42+
\PYG{c+c1}{\PYGZsh{} Device configuration}
43+
\PYG{n}{device} \PYG{o}{=} \PYG{n}{torch}\PYG{o}{.}\PYG{n}{device}\PYG{p}{(}\PYG{l+s+s1}{\PYGZsq{}}\PYG{l+s+s1}{cuda}\PYG{l+s+s1}{\PYGZsq{}} \PYG{k}{if} \PYG{n}{torch}\PYG{o}{.}\PYG{n}{cuda}\PYG{o}{.}\PYG{n}{is\PYGZus{}available}\PYG{p}{(}\PYG{p}{)} \PYG{k}{else} \PYG{l+s+s1}{\PYGZsq{}}\PYG{l+s+s1}{cpu}\PYG{l+s+s1}{\PYGZsq{}}\PYG{p}{)}
44+
45+
\PYG{c+c1}{\PYGZsh{} MNIST dataset}
46+
\PYG{n}{transform} \PYG{o}{=} \PYG{n}{transforms}\PYG{o}{.}\PYG{n}{Compose}\PYG{p}{(}\PYG{p}{[}
47+
\PYG{n}{transforms}\PYG{o}{.}\PYG{n}{ToTensor}\PYG{p}{(}\PYG{p}{)}\PYG{p}{,}
48+
\PYG{n}{transforms}\PYG{o}{.}\PYG{n}{Normalize}\PYG{p}{(}\PYG{p}{(}\PYG{l+m+mf}{0.1307}\PYG{p}{,}\PYG{p}{)}\PYG{p}{,} \PYG{p}{(}\PYG{l+m+mf}{0.3081}\PYG{p}{,}\PYG{p}{)}\PYG{p}{)} \PYG{c+c1}{\PYGZsh{} MNIST mean and std}
49+
\PYG{p}{]}\PYG{p}{)}
50+
51+
\PYG{n}{train\PYGZus{}dataset} \PYG{o}{=} \PYG{n}{datasets}\PYG{o}{.}\PYG{n}{MNIST}\PYG{p}{(}\PYG{n}{root}\PYG{o}{=}\PYG{l+s+s1}{\PYGZsq{}}\PYG{l+s+s1}{./data}\PYG{l+s+s1}{\PYGZsq{}}\PYG{p}{,}
52+
\PYG{n}{train}\PYG{o}{=}\PYG{k+kc}{True}\PYG{p}{,}
53+
\PYG{n}{transform}\PYG{o}{=}\PYG{n}{transform}\PYG{p}{,}
54+
\PYG{n}{download}\PYG{o}{=}\PYG{k+kc}{True}\PYG{p}{)}
55+
56+
\PYG{n}{test\PYGZus{}dataset} \PYG{o}{=} \PYG{n}{datasets}\PYG{o}{.}\PYG{n}{MNIST}\PYG{p}{(}\PYG{n}{root}\PYG{o}{=}\PYG{l+s+s1}{\PYGZsq{}}\PYG{l+s+s1}{./data}\PYG{l+s+s1}{\PYGZsq{}}\PYG{p}{,}
57+
\PYG{n}{train}\PYG{o}{=}\PYG{k+kc}{False}\PYG{p}{,}
58+
\PYG{n}{transform}\PYG{o}{=}\PYG{n}{transform}\PYG{p}{)}
59+
60+
\PYG{n}{train\PYGZus{}loader} \PYG{o}{=} \PYG{n}{DataLoader}\PYG{p}{(}\PYG{n}{dataset}\PYG{o}{=}\PYG{n}{train\PYGZus{}dataset}\PYG{p}{,}
61+
\PYG{n}{batch\PYGZus{}size}\PYG{o}{=}\PYG{n}{batch\PYGZus{}size}\PYG{p}{,}
62+
\PYG{n}{shuffle}\PYG{o}{=}\PYG{k+kc}{True}\PYG{p}{)}
63+
64+
\PYG{n}{test\PYGZus{}loader} \PYG{o}{=} \PYG{n}{DataLoader}\PYG{p}{(}\PYG{n}{dataset}\PYG{o}{=}\PYG{n}{test\PYGZus{}dataset}\PYG{p}{,}
65+
\PYG{n}{batch\PYGZus{}size}\PYG{o}{=}\PYG{n}{batch\PYGZus{}size}\PYG{p}{,}
66+
\PYG{n}{shuffle}\PYG{o}{=}\PYG{k+kc}{False}\PYG{p}{)}
67+
68+
\PYG{c+c1}{\PYGZsh{} LSTM model}
69+
\PYG{k}{class}\PYG{+w}{ }\PYG{n+nc}{LSTMModel}\PYG{p}{(}\PYG{n}{nn}\PYG{o}{.}\PYG{n}{Module}\PYG{p}{)}\PYG{p}{:}
70+
\PYG{k}{def}\PYG{+w}{ }\PYG{n+nf+fm}{\PYGZus{}\PYGZus{}init\PYGZus{}\PYGZus{}}\PYG{p}{(}\PYG{n+nb+bp}{self}\PYG{p}{,} \PYG{n}{input\PYGZus{}size}\PYG{p}{,} \PYG{n}{hidden\PYGZus{}size}\PYG{p}{,} \PYG{n}{num\PYGZus{}classes}\PYG{p}{)}\PYG{p}{:}
71+
\PYG{n+nb}{super}\PYG{p}{(}\PYG{n}{LSTMModel}\PYG{p}{,} \PYG{n+nb+bp}{self}\PYG{p}{)}\PYG{o}{.}\PYG{n+nf+fm}{\PYGZus{}\PYGZus{}init\PYGZus{}\PYGZus{}}\PYG{p}{(}\PYG{p}{)}
72+
\PYG{n+nb+bp}{self}\PYG{o}{.}\PYG{n}{hidden\PYGZus{}size} \PYG{o}{=} \PYG{n}{hidden\PYGZus{}size}
73+
\PYG{n+nb+bp}{self}\PYG{o}{.}\PYG{n}{lstm} \PYG{o}{=} \PYG{n}{nn}\PYG{o}{.}\PYG{n}{LSTM}\PYG{p}{(}\PYG{n}{input\PYGZus{}size}\PYG{p}{,} \PYG{n}{hidden\PYGZus{}size}\PYG{p}{,} \PYG{n}{batch\PYGZus{}first}\PYG{o}{=}\PYG{k+kc}{True}\PYG{p}{)}
74+
\PYG{n+nb+bp}{self}\PYG{o}{.}\PYG{n}{fc} \PYG{o}{=} \PYG{n}{nn}\PYG{o}{.}\PYG{n}{Linear}\PYG{p}{(}\PYG{n}{hidden\PYGZus{}size}\PYG{p}{,} \PYG{n}{num\PYGZus{}classes}\PYG{p}{)}
75+
76+
\PYG{k}{def}\PYG{+w}{ }\PYG{n+nf}{forward}\PYG{p}{(}\PYG{n+nb+bp}{self}\PYG{p}{,} \PYG{n}{x}\PYG{p}{)}\PYG{p}{:}
77+
\PYG{c+c1}{\PYGZsh{} Reshape input to (batch\PYGZus{}size, sequence\PYGZus{}length, input\PYGZus{}size)}
78+
\PYG{n}{x} \PYG{o}{=} \PYG{n}{x}\PYG{o}{.}\PYG{n}{reshape}\PYG{p}{(}\PYG{o}{\PYGZhy{}}\PYG{l+m+mi}{1}\PYG{p}{,} \PYG{l+m+mi}{28}\PYG{p}{,} \PYG{l+m+mi}{28}\PYG{p}{)}
79+
80+
\PYG{c+c1}{\PYGZsh{} Forward propagate LSTM}
81+
\PYG{n}{out}\PYG{p}{,} \PYG{n}{\PYGZus{}} \PYG{o}{=} \PYG{n+nb+bp}{self}\PYG{o}{.}\PYG{n}{lstm}\PYG{p}{(}\PYG{n}{x}\PYG{p}{)} \PYG{c+c1}{\PYGZsh{} out: (batch\PYGZus{}size, seq\PYGZus{}length, hidden\PYGZus{}size)}
82+
83+
\PYG{c+c1}{\PYGZsh{} Decode the hidden state of the last time step}
84+
\PYG{n}{out} \PYG{o}{=} \PYG{n}{out}\PYG{p}{[}\PYG{p}{:}\PYG{p}{,} \PYG{o}{\PYGZhy{}}\PYG{l+m+mi}{1}\PYG{p}{,} \PYG{p}{:}\PYG{p}{]}
85+
\PYG{n}{out} \PYG{o}{=} \PYG{n+nb+bp}{self}\PYG{o}{.}\PYG{n}{fc}\PYG{p}{(}\PYG{n}{out}\PYG{p}{)}
86+
\PYG{k}{return} \PYG{n}{out}
87+
88+
\PYG{c+c1}{\PYGZsh{} Initialize model}
89+
\PYG{n}{model} \PYG{o}{=} \PYG{n}{LSTMModel}\PYG{p}{(}\PYG{n}{input\PYGZus{}size}\PYG{p}{,} \PYG{n}{hidden\PYGZus{}size}\PYG{p}{,} \PYG{n}{num\PYGZus{}classes}\PYG{p}{)}\PYG{o}{.}\PYG{n}{to}\PYG{p}{(}\PYG{n}{device}\PYG{p}{)}
90+
91+
\PYG{c+c1}{\PYGZsh{} Loss and optimizer}
92+
\PYG{n}{criterion} \PYG{o}{=} \PYG{n}{nn}\PYG{o}{.}\PYG{n}{CrossEntropyLoss}\PYG{p}{(}\PYG{p}{)}
93+
\PYG{n}{optimizer} \PYG{o}{=} \PYG{n}{optim}\PYG{o}{.}\PYG{n}{Adam}\PYG{p}{(}\PYG{n}{model}\PYG{o}{.}\PYG{n}{parameters}\PYG{p}{(}\PYG{p}{)}\PYG{p}{,} \PYG{n}{lr}\PYG{o}{=}\PYG{n}{learning\PYGZus{}rate}\PYG{p}{)}
94+
95+
\PYG{c+c1}{\PYGZsh{} Training loop}
96+
\PYG{n}{total\PYGZus{}step} \PYG{o}{=} \PYG{n+nb}{len}\PYG{p}{(}\PYG{n}{train\PYGZus{}loader}\PYG{p}{)}
97+
\PYG{k}{for} \PYG{n}{epoch} \PYG{o+ow}{in} \PYG{n+nb}{range}\PYG{p}{(}\PYG{n}{num\PYGZus{}epochs}\PYG{p}{)}\PYG{p}{:}
98+
\PYG{n}{model}\PYG{o}{.}\PYG{n}{train}\PYG{p}{(}\PYG{p}{)}
99+
\PYG{k}{for} \PYG{n}{i}\PYG{p}{,} \PYG{p}{(}\PYG{n}{images}\PYG{p}{,} \PYG{n}{labels}\PYG{p}{)} \PYG{o+ow}{in} \PYG{n+nb}{enumerate}\PYG{p}{(}\PYG{n}{train\PYGZus{}loader}\PYG{p}{)}\PYG{p}{:}
100+
\PYG{n}{images} \PYG{o}{=} \PYG{n}{images}\PYG{o}{.}\PYG{n}{to}\PYG{p}{(}\PYG{n}{device}\PYG{p}{)}
101+
\PYG{n}{labels} \PYG{o}{=} \PYG{n}{labels}\PYG{o}{.}\PYG{n}{to}\PYG{p}{(}\PYG{n}{device}\PYG{p}{)}
102+
103+
\PYG{c+c1}{\PYGZsh{} Forward pass}
104+
\PYG{n}{outputs} \PYG{o}{=} \PYG{n}{model}\PYG{p}{(}\PYG{n}{images}\PYG{p}{)}
105+
\PYG{n}{loss} \PYG{o}{=} \PYG{n}{criterion}\PYG{p}{(}\PYG{n}{outputs}\PYG{p}{,} \PYG{n}{labels}\PYG{p}{)}
106+
107+
\PYG{c+c1}{\PYGZsh{} Backward and optimize}
108+
\PYG{n}{optimizer}\PYG{o}{.}\PYG{n}{zero\PYGZus{}grad}\PYG{p}{(}\PYG{p}{)}
109+
\PYG{n}{loss}\PYG{o}{.}\PYG{n}{backward}\PYG{p}{(}\PYG{p}{)}
110+
\PYG{n}{optimizer}\PYG{o}{.}\PYG{n}{step}\PYG{p}{(}\PYG{p}{)}
111+
112+
\PYG{k}{if} \PYG{p}{(}\PYG{n}{i}\PYG{o}{+}\PYG{l+m+mi}{1}\PYG{p}{)} \PYG{o}{\PYGZpc{}} \PYG{l+m+mi}{100} \PYG{o}{==} \PYG{l+m+mi}{0}\PYG{p}{:}
113+
\PYG{n+nb}{print}\PYG{p}{(}\PYG{l+s+sa}{f}\PYG{l+s+s1}{\PYGZsq{}}\PYG{l+s+s1}{Epoch [}\PYG{l+s+si}{\PYGZob{}}\PYG{n}{epoch}\PYG{o}{+}\PYG{l+m+mi}{1}\PYG{l+s+si}{\PYGZcb{}}\PYG{l+s+s1}{/}\PYG{l+s+si}{\PYGZob{}}\PYG{n}{num\PYGZus{}epochs}\PYG{l+s+si}{\PYGZcb{}}\PYG{l+s+s1}{], Step [}\PYG{l+s+si}{\PYGZob{}}\PYG{n}{i}\PYG{o}{+}\PYG{l+m+mi}{1}\PYG{l+s+si}{\PYGZcb{}}\PYG{l+s+s1}{/}\PYG{l+s+si}{\PYGZob{}}\PYG{n}{total\PYGZus{}step}\PYG{l+s+si}{\PYGZcb{}}\PYG{l+s+s1}{], Loss: }\PYG{l+s+si}{\PYGZob{}}\PYG{n}{loss}\PYG{o}{.}\PYG{n}{item}\PYG{p}{(}\PYG{p}{)}\PYG{l+s+si}{:}\PYG{l+s+s1}{.4f}\PYG{l+s+si}{\PYGZcb{}}\PYG{l+s+s1}{\PYGZsq{}}\PYG{p}{)}
114+
115+
\PYG{c+c1}{\PYGZsh{} Test the model}
116+
\PYG{n}{model}\PYG{o}{.}\PYG{n}{eval}\PYG{p}{(}\PYG{p}{)}
117+
\PYG{k}{with} \PYG{n}{torch}\PYG{o}{.}\PYG{n}{no\PYGZus{}grad}\PYG{p}{(}\PYG{p}{)}\PYG{p}{:}
118+
\PYG{n}{correct} \PYG{o}{=} \PYG{l+m+mi}{0}
119+
\PYG{n}{total} \PYG{o}{=} \PYG{l+m+mi}{0}
120+
\PYG{k}{for} \PYG{n}{images}\PYG{p}{,} \PYG{n}{labels} \PYG{o+ow}{in} \PYG{n}{test\PYGZus{}loader}\PYG{p}{:}
121+
\PYG{n}{images} \PYG{o}{=} \PYG{n}{images}\PYG{o}{.}\PYG{n}{to}\PYG{p}{(}\PYG{n}{device}\PYG{p}{)}
122+
\PYG{n}{labels} \PYG{o}{=} \PYG{n}{labels}\PYG{o}{.}\PYG{n}{to}\PYG{p}{(}\PYG{n}{device}\PYG{p}{)}
123+
\PYG{n}{outputs} \PYG{o}{=} \PYG{n}{model}\PYG{p}{(}\PYG{n}{images}\PYG{p}{)}
124+
\PYG{n}{\PYGZus{}}\PYG{p}{,} \PYG{n}{predicted} \PYG{o}{=} \PYG{n}{torch}\PYG{o}{.}\PYG{n}{max}\PYG{p}{(}\PYG{n}{outputs}\PYG{o}{.}\PYG{n}{data}\PYG{p}{,} \PYG{l+m+mi}{1}\PYG{p}{)}
125+
\PYG{n}{total} \PYG{o}{+}\PYG{o}{=} \PYG{n}{labels}\PYG{o}{.}\PYG{n}{size}\PYG{p}{(}\PYG{l+m+mi}{0}\PYG{p}{)}
126+
\PYG{n}{correct} \PYG{o}{+}\PYG{o}{=} \PYG{p}{(}\PYG{n}{predicted} \PYG{o}{==} \PYG{n}{labels}\PYG{p}{)}\PYG{o}{.}\PYG{n}{sum}\PYG{p}{(}\PYG{p}{)}\PYG{o}{.}\PYG{n}{item}\PYG{p}{(}\PYG{p}{)}
127+
128+
\PYG{n+nb}{print}\PYG{p}{(}\PYG{l+s+sa}{f}\PYG{l+s+s1}{\PYGZsq{}}\PYG{l+s+s1}{Test Accuracy: }\PYG{l+s+si}{\PYGZob{}}\PYG{l+m+mi}{100}\PYG{+w}{ }\PYG{o}{*}\PYG{+w}{ }\PYG{n}{correct}\PYG{+w}{ }\PYG{o}{/}\PYG{+w}{ }\PYG{n}{total}\PYG{l+s+si}{:}\PYG{l+s+s1}{.2f}\PYG{l+s+si}{\PYGZcb{}}\PYG{l+s+s1}{\PYGZpc{}}\PYG{l+s+s1}{\PYGZsq{}}\PYG{p}{)}
129+
130+
\PYG{n+nb}{print}\PYG{p}{(}\PYG{l+s+s1}{\PYGZsq{}}\PYG{l+s+s1}{Training finished.}\PYG{l+s+s1}{\PYGZsq{}}\PYG{p}{)}
131+
132+
\end{MintedVerbatim}

‎doc/src/week7/_minted/DECAD7444034D9AA0E1812FE8926F6AB.highlight.minted‎

Lines changed: 63 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,63 @@
1+
\begin{MintedVerbatim}[commandchars=\\\{\},codes={\catcode`\$=3\catcode`\^=7\catcode`\_=8\relax}]
2+
\PYG{l+s+sd}{\PYGZdq{}\PYGZdq{}\PYGZdq{}}
3+
\PYG{l+s+sd}{Key points:}
4+
\PYG{l+s+sd}{1. The input images (28x28 pixels) are treated as sequences of 28 timesteps with 28 features each}
5+
\PYG{l+s+sd}{2. The LSTM layer processes this sequential data}
6+
\PYG{l+s+sd}{3. A final dense layer with softmax activation handles the classification}
7+
\PYG{l+s+sd}{4. Typical accuracy ranges between 95\PYGZhy{}98\PYGZpc{} (lower than CNNs but reasonable for demonstration)}
8+
9+
\PYG{l+s+sd}{Note: LSTMs are not typically used for image classification (CNNs are more efficient), but this demonstrates how to adapt them for such tasks. Training might take longer compared to CNN architectures.}
10+
11+
\PYG{l+s+sd}{To improve performance, you could:}
12+
\PYG{l+s+sd}{1. Add more LSTM layers}
13+
\PYG{l+s+sd}{2. Use Bidirectional LSTMs}
14+
\PYG{l+s+sd}{3. Increase the number of units}
15+
\PYG{l+s+sd}{4. Add dropout for regularization}
16+
\PYG{l+s+sd}{5. Use learning rate scheduling}
17+
\PYG{l+s+sd}{\PYGZdq{}\PYGZdq{}\PYGZdq{}}
18+
19+
\PYG{k+kn}{import}\PYG{+w}{ }\PYG{n+nn}{tensorflow}\PYG{+w}{ }\PYG{k}{as}\PYG{+w}{ }\PYG{n+nn}{tf}
20+
\PYG{k+kn}{from}\PYG{+w}{ }\PYG{n+nn}{tensorflow}\PYG{n+nn}{.}\PYG{n+nn}{keras}\PYG{n+nn}{.}\PYG{n+nn}{models}\PYG{+w}{ }\PYG{k+kn}{import} \PYG{n}{Sequential}
21+
\PYG{k+kn}{from}\PYG{+w}{ }\PYG{n+nn}{tensorflow}\PYG{n+nn}{.}\PYG{n+nn}{keras}\PYG{n+nn}{.}\PYG{n+nn}{layers}\PYG{+w}{ }\PYG{k+kn}{import} \PYG{n}{LSTM}\PYG{p}{,} \PYG{n}{Dense}
22+
\PYG{k+kn}{from}\PYG{+w}{ }\PYG{n+nn}{tensorflow}\PYG{n+nn}{.}\PYG{n+nn}{keras}\PYG{n+nn}{.}\PYG{n+nn}{utils}\PYG{+w}{ }\PYG{k+kn}{import} \PYG{n}{to\PYGZus{}categorical}
23+
24+
\PYG{c+c1}{\PYGZsh{} Load and preprocess data}
25+
\PYG{p}{(}\PYG{n}{x\PYGZus{}train}\PYG{p}{,} \PYG{n}{y\PYGZus{}train}\PYG{p}{)}\PYG{p}{,} \PYG{p}{(}\PYG{n}{x\PYGZus{}test}\PYG{p}{,} \PYG{n}{y\PYGZus{}test}\PYG{p}{)} \PYG{o}{=} \PYG{n}{tf}\PYG{o}{.}\PYG{n}{keras}\PYG{o}{.}\PYG{n}{datasets}\PYG{o}{.}\PYG{n}{mnist}\PYG{o}{.}\PYG{n}{load\PYGZus{}data}\PYG{p}{(}\PYG{p}{)}
26+
27+
\PYG{c+c1}{\PYGZsh{} Normalize pixel values to [0, 1]}
28+
\PYG{n}{x\PYGZus{}train} \PYG{o}{=} \PYG{n}{x\PYGZus{}train}\PYG{o}{.}\PYG{n}{astype}\PYG{p}{(}\PYG{l+s+s1}{\PYGZsq{}}\PYG{l+s+s1}{float32}\PYG{l+s+s1}{\PYGZsq{}}\PYG{p}{)} \PYG{o}{/} \PYG{l+m+mf}{255.0}
29+
\PYG{n}{x\PYGZus{}test} \PYG{o}{=} \PYG{n}{x\PYGZus{}test}\PYG{o}{.}\PYG{n}{astype}\PYG{p}{(}\PYG{l+s+s1}{\PYGZsq{}}\PYG{l+s+s1}{float32}\PYG{l+s+s1}{\PYGZsq{}}\PYG{p}{)} \PYG{o}{/} \PYG{l+m+mf}{255.0}
30+
31+
\PYG{c+c1}{\PYGZsh{} Reshape data for LSTM (samples, timesteps, features)}
32+
\PYG{c+c1}{\PYGZsh{} MNIST images are 28x28, so we treat each image as 28 timesteps of 28 features}
33+
\PYG{n}{x\PYGZus{}train} \PYG{o}{=} \PYG{n}{x\PYGZus{}train}\PYG{o}{.}\PYG{n}{reshape}\PYG{p}{(}\PYG{p}{(}\PYG{o}{\PYGZhy{}}\PYG{l+m+mi}{1}\PYG{p}{,} \PYG{l+m+mi}{28}\PYG{p}{,} \PYG{l+m+mi}{28}\PYG{p}{)}\PYG{p}{)}
34+
\PYG{n}{x\PYGZus{}test} \PYG{o}{=} \PYG{n}{x\PYGZus{}test}\PYG{o}{.}\PYG{n}{reshape}\PYG{p}{(}\PYG{p}{(}\PYG{o}{\PYGZhy{}}\PYG{l+m+mi}{1}\PYG{p}{,} \PYG{l+m+mi}{28}\PYG{p}{,} \PYG{l+m+mi}{28}\PYG{p}{)}\PYG{p}{)}
35+
36+
\PYG{c+c1}{\PYGZsh{} Convert labels to one\PYGZhy{}hot encoding}
37+
\PYG{n}{y\PYGZus{}train} \PYG{o}{=} \PYG{n}{to\PYGZus{}categorical}\PYG{p}{(}\PYG{n}{y\PYGZus{}train}\PYG{p}{,} \PYG{l+m+mi}{10}\PYG{p}{)}
38+
\PYG{n}{y\PYGZus{}test} \PYG{o}{=} \PYG{n}{to\PYGZus{}categorical}\PYG{p}{(}\PYG{n}{y\PYGZus{}test}\PYG{p}{,} \PYG{l+m+mi}{10}\PYG{p}{)}
39+
40+
\PYG{c+c1}{\PYGZsh{} Build LSTM model}
41+
\PYG{n}{model} \PYG{o}{=} \PYG{n}{Sequential}\PYG{p}{(}\PYG{p}{)}
42+
\PYG{n}{model}\PYG{o}{.}\PYG{n}{add}\PYG{p}{(}\PYG{n}{LSTM}\PYG{p}{(}\PYG{l+m+mi}{128}\PYG{p}{,} \PYG{n}{input\PYGZus{}shape}\PYG{o}{=}\PYG{p}{(}\PYG{l+m+mi}{28}\PYG{p}{,} \PYG{l+m+mi}{28}\PYG{p}{)}\PYG{p}{)}\PYG{p}{)} \PYG{c+c1}{\PYGZsh{} 128 LSTM units}
43+
\PYG{n}{model}\PYG{o}{.}\PYG{n}{add}\PYG{p}{(}\PYG{n}{Dense}\PYG{p}{(}\PYG{l+m+mi}{10}\PYG{p}{,} \PYG{n}{activation}\PYG{o}{=}\PYG{l+s+s1}{\PYGZsq{}}\PYG{l+s+s1}{softmax}\PYG{l+s+s1}{\PYGZsq{}}\PYG{p}{)}\PYG{p}{)}
44+
45+
\PYG{c+c1}{\PYGZsh{} Compile the model}
46+
\PYG{n}{model}\PYG{o}{.}\PYG{n}{compile}\PYG{p}{(}\PYG{n}{loss}\PYG{o}{=}\PYG{l+s+s1}{\PYGZsq{}}\PYG{l+s+s1}{categorical\PYGZus{}crossentropy}\PYG{l+s+s1}{\PYGZsq{}}\PYG{p}{,}
47+
\PYG{n}{optimizer}\PYG{o}{=}\PYG{l+s+s1}{\PYGZsq{}}\PYG{l+s+s1}{adam}\PYG{l+s+s1}{\PYGZsq{}}\PYG{p}{,}
48+
\PYG{n}{metrics}\PYG{o}{=}\PYG{p}{[}\PYG{l+s+s1}{\PYGZsq{}}\PYG{l+s+s1}{accuracy}\PYG{l+s+s1}{\PYGZsq{}}\PYG{p}{]}\PYG{p}{)}
49+
50+
\PYG{c+c1}{\PYGZsh{} Display model summary}
51+
\PYG{n}{model}\PYG{o}{.}\PYG{n}{summary}\PYG{p}{(}\PYG{p}{)}
52+
53+
\PYG{c+c1}{\PYGZsh{} Train the model}
54+
\PYG{n}{history} \PYG{o}{=} \PYG{n}{model}\PYG{o}{.}\PYG{n}{fit}\PYG{p}{(}\PYG{n}{x\PYGZus{}train}\PYG{p}{,} \PYG{n}{y\PYGZus{}train}\PYG{p}{,}
55+
\PYG{n}{batch\PYGZus{}size}\PYG{o}{=}\PYG{l+m+mi}{64}\PYG{p}{,}
56+
\PYG{n}{epochs}\PYG{o}{=}\PYG{l+m+mi}{10}\PYG{p}{,}
57+
\PYG{n}{validation\PYGZus{}split}\PYG{o}{=}\PYG{l+m+mf}{0.2}\PYG{p}{)}
58+
59+
\PYG{c+c1}{\PYGZsh{} Evaluate on test data}
60+
\PYG{n}{test\PYGZus{}loss}\PYG{p}{,} \PYG{n}{test\PYGZus{}acc} \PYG{o}{=} \PYG{n}{model}\PYG{o}{.}\PYG{n}{evaluate}\PYG{p}{(}\PYG{n}{x\PYGZus{}test}\PYG{p}{,} \PYG{n}{y\PYGZus{}test}\PYG{p}{,} \PYG{n}{verbose}\PYG{o}{=}\PYG{l+m+mi}{2}\PYG{p}{)}
61+
\PYG{n+nb}{print}\PYG{p}{(}\PYG{l+s+sa}{f}\PYG{l+s+s1}{\PYGZsq{}}\PYG{l+s+se}{\PYGZbs{}n}\PYG{l+s+s1}{Test accuracy: }\PYG{l+s+si}{\PYGZob{}}\PYG{n}{test\PYGZus{}acc}\PYG{l+s+si}{:}\PYG{l+s+s1}{.4f}\PYG{l+s+si}{\PYGZcb{}}\PYG{l+s+s1}{\PYGZsq{}}\PYG{p}{)}
62+
63+
\end{MintedVerbatim}

‎doc/src/week7/_minted/_306ADBB672267A7DCBC3A4933F3A8570.index.minted‎

Lines changed: 11 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,11 @@
1+
{
2+
"jobname": "week7",
3+
"md5": "306ADBB672267A7DCBC3A4933F3A8570",
4+
"timestamp": "20260304125736",
5+
"cachefiles": [
6+
"45BFBE4480AD7F905AB70F217399FB65.highlight.minted",
7+
"DECAD7444034D9AA0E1812FE8926F6AB.highlight.minted",
8+
"_306ADBB672267A7DCBC3A4933F3A8570.index.minted",
9+
"default.style.minted"
10+
]
11+
}

0 commit comments

Comments
 (0)