fix: fixing the call to train.py directly in python

else we have issues with the versions

use_case_examples/deployment/breast_cancer/README.md

@@ -8,14 +8,57 @@ To run this example on AWS you will also need to have the AWS CLI properly setup
 To do so please refer to [AWS documentation](https://docs.aws.amazon.com/cli/latest/userguide/cli-configure-quickstart.html).
 One can also run this example locally using Docker, or just by running the scripts locally.
 
-1. To train your model you can use `train.py`, or `train_with_docker.sh` to use Docker (recommended way).
+#### On the developer machine:
+
+1. To train your model you can
+      - use `train_with_docker.sh` to use Docker (recommended way),
+      - or, only if you know what you're doing and will manage synchronisation between versions, use `python train.py`
+
    This will train a model and [serialize the FHE circuit](../../../docs/guides/client_server.md) in a new folder called `./dev`.
-1. Once that's done you can use the script provided in Concrete ML in `use_case_examples/deployment/server/`, use `deploy_to_docker.py`.
 
-- `python use_case_examples/deployment/server/deploy_to_docker.py --path-to-model ./dev`
+#### On the server machine:
+
+1. Copy the './dev' directory from the developer machine.
+1. If you need to delete existing Dockers: `docker rm -f $(docker ps -a -q)`
+1. Launch the server via:
+
+```
+python ../server/deploy_to_docker.py --path-to-model ./dev
+```
+
+You will finally see some
+
+> INFO:     Uvicorn running on http://0.0.0.0:5000 (Press CTRL+C to quit)
+
+which means the server is ready to server, on Port 5000.
+
+#### On the client machine:
+
+##### If you go for a Docker part on the client side:
+
+1. Launch the `build_docker_client_image.py` to build a client Docker image.
+1. Run the client with `client.sh` script. This will run the container in interactive mode.
+1. Then, in this Docker, you can launch the client script to interact with the server:
+
+```
+URL="<my_url>" python client.py
+```
+
+where `<my_url>` is the content of the `url.txt` file (if you don't set URL, the default is `0.0.0.0`; this defines the IP to use when running server in Docker on localhost).
+
+#### If you go for client side done in Python:
+
+1. Prepare the client side:
+
+```
+python3.8 -m venv .venvclient
+source .venvclient/bin/activate
+pip install -r client_requirements.txt
+```
+1. Run the client script:
 
-3. Once that's done you can launch the `build_docker_client_image.py` script to build a client Docker image.
-1. You can then run the client by using the `client.sh` script. This will run the container in interactive mode.
-   To interact with the server you can launch the `client.py` script using `URL="<my_url>" python client.py` where `<my_url>` is the content of the `url.txt` file (default is `0.0.0.0`, ip to use when running server in Docker on localhost).
+```
+URL="http://localhost:8888" python client.py
+```
 
-And here it is you deployed a Concrete ML model and ran an inference using Fully Homormophic Encryption.
+And here it is! Whether you use Docker or Python for the client side, you deployed a Concrete ML model and ran an inference using Fully Homormophic Encryption. In particular, you can see that the FHE predictions are correct.

use_case_examples/deployment/breast_cancer/client.py

@@ -21,7 +21,7 @@
 
 from concrete.ml.deployment import FHEModelClient
 
-URL = os.environ.get("URL", f"http://localhost:5000")
+URL = os.environ.get("URL", f"http://localhost:8888")
 STATUS_OK = 200
 ROOT = Path(__file__).parent / "client"
 ROOT.mkdir(exist_ok=True)
@@ -105,4 +105,13 @@
         encrypted_result = result.content
         decrypted_prediction = client.deserialize_decrypt_dequantize(encrypted_result)[0]
         decrypted_predictions.append(decrypted_prediction)
-    print(decrypted_predictions)
+    print(f"Decrypted predictions are: {decrypted_predictions}")
+
+    decrypted_predictions_classes = numpy.array(decrypted_predictions).argmax(axis=1)
+    print(f"Decrypted prediction classes are: {decrypted_predictions_classes}")
+
+    # Let's check the results and compare them against the clear model
+    clear_prediction_classes = y[0:10]
+    accuracy = (clear_prediction_classes == decrypted_predictions_classes).mean()
+    print(f"Accuracy between FHE prediction and expected results is: {accuracy*100:.0f}%")
+

use_case_examples/deployment/breast_cancer/client_requirements.txt

@@ -1,3 +1,6 @@
 grequests
 requests
 tqdm
+numpy
+scikit-learn
+concrete-ml

use_case_examples/deployment/breast_cancer/client_via_tfhe-rs.py

@@ -0,0 +1,208 @@
+"""Client script.
+
+This script does the following:
+    - Query crypto-parameters and pre/post-processing parameters
+    - Quantize the inputs using the parameters
+    - Encrypt data using the crypto-parameters
+    - Send the encrypted data to the server (async using grequests)
+    - Collect the data and decrypt it
+    - De-quantize the decrypted results
+"""
+
+import io
+import os
+from pathlib import Path
+
+import grequests
+import numpy
+import requests
+from sklearn.datasets import load_breast_cancer
+from tqdm import tqdm
+
+from concrete import fhe
+from concrete.ml.deployment import FHEModelClient
+
+URL = os.environ.get("URL", f"http://localhost:8888")
+STATUS_OK = 200
+ROOT = Path(__file__).parent / "client"
+ROOT.mkdir(exist_ok=True)
+
+encrypt_with_tfhe = False
+nb_samples = 10
+
+def to_tuple(x) -> tuple:
+    """Make the input a tuple if it is not already the case.
+
+    Args:
+        x (Any): The input to consider. It can already be an input.
+
+    Returns:
+        tuple: The input as a tuple.
+    """
+    # If the input is not a tuple, return a tuple of a single element
+    if not isinstance(x, tuple):
+        return (x,)
+
+    return x
+
+def serialize_encrypted_values(
+    *values_enc,
+):
+    """Serialize encrypted values.
+
+    If a value is None, None is returned.
+
+    Args:
+        values_enc (Optional[fhe.Value]): The values to serialize.
+
+    Returns:
+        Union[Optional[bytes], Optional[Tuple[bytes]]]: The serialized values.
+    """
+    values_enc_serialized = tuple(
+        value_enc.serialize() if value_enc is not None else None for value_enc in values_enc
+    )
+
+    if len(values_enc_serialized) == 1:
+        return values_enc_serialized[0]
+
+    return values_enc_serialized
+
+def deserialize_encrypted_values(
+    *values_serialized,
+):
+    """Deserialize encrypted values.
+
+    If a value is None, None is returned.
+
+    Args:
+        values_serialized (Optional[bytes]): The values to deserialize.
+
+    Returns:
+        Union[Optional[fhe.Value], Optional[Tuple[fhe.Value]]]: The deserialized values.
+    """
+    values_enc = tuple(
+        fhe.Value.deserialize(value_serialized) if value_serialized is not None else None
+        for value_serialized in values_serialized
+    )
+
+    if len(values_enc) == 1:
+        return values_enc[0]
+
+    return values_enc
+
+
+if __name__ == "__main__":
+    # Get the necessary data for the client
+    # client.zip
+    zip_response = requests.get(f"{URL}/get_client")
+    assert zip_response.status_code == STATUS_OK
+    with open(ROOT / "client.zip", "wb") as file:
+        file.write(zip_response.content)
+
+    # Get the data to infer
+    X, y = load_breast_cancer(return_X_y=True)
+    assert isinstance(X, numpy.ndarray)
+    assert isinstance(y, numpy.ndarray)
+    X = X[-nb_samples:]
+    y = y[-nb_samples:]
+
+    assert isinstance(X, numpy.ndarray)
+    assert isinstance(y, numpy.ndarray)
+
+    # Create the client
+    client = FHEModelClient(path_dir=str(ROOT.resolve()), key_dir=str((ROOT / "keys").resolve()))
+
+    # The client first need to create the private and evaluation keys.
+    serialized_evaluation_keys = client.get_serialized_evaluation_keys()
+
+    assert isinstance(serialized_evaluation_keys, bytes)
+
+    # Evaluation keys can be quite large files but only have to be shared once with the server.
+
+    # Check the size of the evaluation keys (in MB)
+    print(f"Evaluation keys size: {len(serialized_evaluation_keys) / (10**6):.2f} MB")
+
+    # Send this evaluation key to the server (this has to be done only once)
+    # send_evaluation_key_to_server(serialized_evaluation_keys)
+
+    # Now we have everything for the client to interact with the server
+
+    # We create a loop to send the input to the server and receive the encrypted prediction
+    execution_time = []
+    encrypted_input = None
+    clear_input = None
+
+    # Update all base64 queries encodings with UploadFile
+    response = requests.post(
+        f"{URL}/add_key", files={"key": io.BytesIO(initial_bytes=serialized_evaluation_keys)}
+    )
+    assert response.status_code == STATUS_OK
+    uid = response.json()["uid"]
+
+    inferences = []
+    # Launch the queries
+    for i in tqdm(range(len(X))):
+        clear_input = X[[i], :]
+
+        assert isinstance(clear_input, numpy.ndarray)
+
+        quantized_input = to_tuple(client.model.quantize_input(clear_input))
+
+        # Here, we can encrypt with TFHE-rs instead of Concrete
+        if encrypt_with_tfhe:
+            pass
+        else:
+            encrypted_input = to_tuple(client.client.encrypt(*quantized_input))
+
+        encrypted_input = serialize_encrypted_values(*encrypted_input)
+
+        # Debugging
+        if False:
+            print(f"Clear input: {clear_input}")
+            print(f"Quantized input: {quantized_input}")
+            print(f"Quantized input: {encrypted_input}")
+
+        assert isinstance(encrypted_input, bytes)
+
+        inferences.append(
+            grequests.post(
+                f"{URL}/compute",
+                files={
+                    "model_input": io.BytesIO(encrypted_input),
+                },
+                data={
+                    "uid": uid,
+                },
+            )
+        )
+
+    # Unpack the results
+    decrypted_predictions = []
+    for result in grequests.map(inferences):
+        if result is None:
+            raise ValueError("Result is None, probably due to a crash on the server side.")
+        assert result.status_code == STATUS_OK
+
+        encrypted_result = result.content
+
+        # Decrypt and deserialize the values
+        result_quant_encrypted = to_tuple(
+            deserialize_encrypted_values(encrypted_result)
+        )
+
+        result_quant = to_tuple(client.client.decrypt(*result_quant_encrypted))
+
+        result = to_tuple(client.model.dequantize_output(*result_quant))
+        decrypted_prediction = client.model.post_processing(*result)[0]
+
+        decrypted_predictions.append(decrypted_prediction)
+    print(f"Decrypted predictions are: {decrypted_predictions}")
+
+    decrypted_predictions_classes = numpy.array(decrypted_predictions).argmax(axis=1)
+    print(f"Decrypted prediction classes are: {decrypted_predictions_classes}")
+
+    # Let's check the results and compare them against the clear model
+    clear_prediction_classes = y[0:nb_samples]
+    accuracy = (clear_prediction_classes == decrypted_predictions_classes).mean()
+    print(f"Accuracy between FHE prediction and expected results is: {accuracy*100:.0f}%")
+

use_case_examples/deployment/breast_cancer/train.py

@@ -20,4 +20,4 @@
     model.fit(X_train, y_train)
     model.compile(X_train)
     dev = FHEModelDev("./dev", model)
-    dev.save()
+    dev.save(via_mlir=True)

use_case_examples/deployment/server/deploy_to_docker.py

@@ -97,11 +97,13 @@ def main(path_to_model: Path, image_name: str):
     if args.only_build:
         return
 
+    PORT_TO_CHOOSE=8888
+
     # Run newly created Docker server
     try:
         with open("./url.txt", mode="w", encoding="utf-8") as file:
-            file.write("http://localhost:5000")
-        subprocess.check_output(f"docker run -p 5000:5000 {image_name}", shell=True)
+            file.write(f"http://localhost:{PORT_TO_CHOOSE}")
+        subprocess.check_output(f"docker run -p {PORT_TO_CHOOSE}:5000 {image_name}", shell=True)
     except KeyboardInterrupt:
         message = "Terminate container? (y/n) "
         shutdown_instance = input(message).lower()