mirror of
https://github.com/zama-ai/concrete.git
synced 2026-02-09 12:15:09 -05:00
d35a183a1e
- add Compiling a Torch Model to the Getting Started - rename to Compiling and Executing a Numpy Function for more clarity - changed the FAQ - removed a .md which is no more needed
110 lines
3.3 KiB
Markdown
110 lines
3.3 KiB
Markdown
# Compiling and Executing a Numpy Function
|
|
|
|
## Importing necessary components
|
|
|
|
Everything you need to compile and execute homomorphic functions is included in a single module. You can import it like so:
|
|
|
|
```python
|
|
import concrete.numpy as hnp
|
|
```
|
|
|
|
## Defining a function to compile
|
|
|
|
You need to have a python function that follows the [limits](../explanation/fhe_and_framework_limits.md) of the **Concrete Framework**. Here is a simple example:
|
|
|
|
<!--python-test:cont-->
|
|
```python
|
|
def f(x, y):
|
|
return x + y
|
|
```
|
|
|
|
## Compiling the function
|
|
|
|
To compile the function, you need to identify the inputs that it is expecting. In the example function above, `x` and `y` could be scalars or tensors (though, for now, only dot between tensors are supported), they can be encrypted or clear, they can be signed or unsigned, they can have different bit-widths. So, we need to know what they are beforehand. We can do that like so:
|
|
|
|
<!--python-test:cont-->
|
|
```python
|
|
x = "encrypted"
|
|
y = "encrypted"
|
|
```
|
|
|
|
In this configuration, both `x` and `y` will be encrypted values.
|
|
|
|
We also need an inputset. It is to determine the bit-widths of the intermediate results. It should be an iterable yielding tuples in the same order as the inputs of the function to compile. There should be at least 10 inputs in the input set to avoid warnings (except for functions with less than 10 possible inputs). The warning is there because the bigger the input set, the better the bounds will be.
|
|
|
|
<!--python-test:cont-->
|
|
```python
|
|
inputset = [(2, 3), (0, 0), (1, 6), (7, 7), (7, 1), (3, 2), (6, 1), (1, 7), (4, 5), (5, 4)]
|
|
```
|
|
|
|
Finally, we can compile our function to its homomorphic equivalent.
|
|
|
|
<!--python-test:cont-->
|
|
```python
|
|
compiler = hnp.NPFHECompiler(
|
|
f, {"x": x, "y": y},
|
|
)
|
|
circuit = compiler.compile_on_inputset(inputset)
|
|
|
|
# If you want, you can separate tracing and compilation steps like so:
|
|
|
|
# You can either evaluate in one go:
|
|
compiler.eval_on_inputset(inputset)
|
|
|
|
# Or progressively:
|
|
for input_values in inputset:
|
|
compiler(*input_values)
|
|
|
|
# You can print the traced graph:
|
|
print(str(compiler))
|
|
|
|
# Outputs
|
|
|
|
# %0 = x # EncryptedScalar<uint3>
|
|
# %1 = y # EncryptedScalar<uint3>
|
|
# %2 = add(%0, %1) # EncryptedScalar<uint4>
|
|
# return %2
|
|
|
|
# Or draw it
|
|
compiler.draw_graph(show=True)
|
|
|
|
circuit = compiler.get_compiled_fhe_circuit()
|
|
|
|
```
|
|
|
|
Here is the graph from the previous code block drawn with `draw_graph`:
|
|
|
|
|
|
|
|
## Performing homomorphic evaluation
|
|
|
|
You can use `.run(...)` method of `FHECircuit` returned by `hnp.compile_numpy_function(...)` to perform fully homomorphic evaluation. Here are some examples:
|
|
|
|
<!--python-test:cont-->
|
|
```python
|
|
circuit.run(3, 4)
|
|
# 7
|
|
circuit.run(1, 2)
|
|
# 3
|
|
circuit.run(7, 7)
|
|
# 14
|
|
circuit.run(0, 0)
|
|
# 0
|
|
```
|
|
|
|
```{caution}
|
|
Be careful about the inputs, though.
|
|
If you were to run with values outside the range of the inputset, the result might not be correct.
|
|
```
|
|
|
|
```{warning}
|
|
FIXME(benoit): explain the API to encrypt, run_inference, decrypt, keygen etc when they are available
|
|
|
|
```
|
|
|
|
## Further reading
|
|
|
|
- [Working With Floating Points Tutorial](../tutorial/working_with_floating_points.md)
|
|
- [Table Lookup Tutorial](../tutorial/table_lookup.md)
|
|
- [Compiling a torch model](../howto/compiling_torch_model.md)
|