Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,70 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
Basic TinyCNN PyTorch model trained on Sklearn Digits dataset.
|
2 |
+
|
3 |
+
```python
|
4 |
+
"""
|
5 |
+
Credits to Zama.ai - https://github.com/zama-ai/concrete-ml/blob/main/docs/user/advanced_examples/ConvolutionalNeuralNetwork.ipynb
|
6 |
+
"""
|
7 |
+
import numpy as np
|
8 |
+
|
9 |
+
import torch
|
10 |
+
from torch import nn
|
11 |
+
from torch.nn.utils import prune
|
12 |
+
|
13 |
+
class TinyCNN(nn.Module):
|
14 |
+
"""A very small CNN to classify the sklearn digits dataset.
|
15 |
+
|
16 |
+
This class also allows pruning to a maximum of 10 active neurons, which
|
17 |
+
should help keep the accumulator bit width low.
|
18 |
+
"""
|
19 |
+
|
20 |
+
def __init__(self, n_classes) -> None:
|
21 |
+
"""Construct the CNN with a configurable number of classes."""
|
22 |
+
super().__init__()
|
23 |
+
|
24 |
+
# This network has a total complexity of 1216 MAC
|
25 |
+
self.conv1 = nn.Conv2d(1, 2, 3, stride=1, padding=0)
|
26 |
+
self.conv2 = nn.Conv2d(2, 3, 3, stride=2, padding=0)
|
27 |
+
self.conv3 = nn.Conv2d(3, 16, 2, stride=1, padding=0)
|
28 |
+
self.fc1 = nn.Linear(16, n_classes)
|
29 |
+
|
30 |
+
# Enable pruning, prepared for training
|
31 |
+
self.toggle_pruning(True)
|
32 |
+
|
33 |
+
def toggle_pruning(self, enable):
|
34 |
+
"""Enables or removes pruning."""
|
35 |
+
|
36 |
+
# Maximum number of active neurons (i.e. corresponding weight != 0)
|
37 |
+
n_active = 10
|
38 |
+
|
39 |
+
# Go through all the convolution layers
|
40 |
+
for layer in (self.conv1, self.conv2, self.conv3):
|
41 |
+
s = layer.weight.shape
|
42 |
+
|
43 |
+
# Compute fan-in (number of inputs to a neuron)
|
44 |
+
# and fan-out (number of neurons in the layer)
|
45 |
+
st = [s[0], np.prod(s[1:])]
|
46 |
+
|
47 |
+
# The number of input neurons (fan-in) is the product of
|
48 |
+
# the kernel width x height x inChannels.
|
49 |
+
if st[1] > n_active:
|
50 |
+
if enable:
|
51 |
+
# This will create a forward hook to create a mask tensor that is multiplied
|
52 |
+
# with the weights during forward. The mask will contain 0s or 1s
|
53 |
+
prune.l1_unstructured(layer, "weight", (st[1] - n_active) * st[0])
|
54 |
+
else:
|
55 |
+
# When disabling pruning, the mask is multiplied with the weights
|
56 |
+
# and the result is stored in the weights member
|
57 |
+
prune.remove(layer, "weight")
|
58 |
+
|
59 |
+
def forward(self, x):
|
60 |
+
"""Run inference on the tiny CNN, apply the decision layer on the reshaped conv output."""
|
61 |
+
x = self.conv1(x)
|
62 |
+
x = torch.relu(x)
|
63 |
+
x = self.conv2(x)
|
64 |
+
x = torch.relu(x)
|
65 |
+
x = self.conv3(x)
|
66 |
+
x = torch.relu(x)
|
67 |
+
x = x.view(-1, 16)
|
68 |
+
x = self.fc1(x)
|
69 |
+
return x
|
70 |
+
```
|