feat: use signed weights in compile_torch_model

concrete/quantization/quantized_array.py

@@ -41,6 +41,11 @@ class QuantizedArray:
 
         if rmax - rmin < STABILITY_CONST:
             # In this case there is  a single unique value to quantize
+
+            # is is_signed is True, we need to set the offset back to 0.
+            # Signed quantization does not make sense for a single value.
+            self.offset = 0
+
             # This value could be multiplied with inputs at some point in the model
             # Since zero points need to be integers, if this value is a small float (ex: 0.01)
             # it will be quantized to 0 with a 0 zero-point, thus becoming useless in multiplication

concrete/torch/compile.py

@@ -78,7 +78,7 @@ def compile_torch_model(
     )
 
     # Quantize with post-training static method, to have a model with integer weights
-    post_training_quant = PostTrainingAffineQuantization(n_bits, numpy_model)
+    post_training_quant = PostTrainingAffineQuantization(n_bits, numpy_model, is_signed=True)
     quantized_module = post_training_quant.quantize_module(numpy_inputset_as_single_array)
 
     # Quantize input

docs/user/howto/compiling_torch_model.md

@@ -9,7 +9,6 @@ A simple command can compile a torch model to its FHE counterpart. This process
 ```python
 from torch import nn
 import torch
-torch.manual_seed(0)
 class LogisticRegression(nn.Module):
     """LogisticRegression with Torch"""
 

tests/quantization/test_quantized_module.py

@@ -80,7 +80,11 @@ N_BITS_ATOL_TUPLE_LIST = [
         pytest.param(FC, (100, 32 * 32 * 3)),
     ],
 )
-def test_quantized_linear(model, input_shape, n_bits, atol, seed_torch):
+@pytest.mark.parametrize(
+    "is_signed",
+    [pytest.param([False, True])],
+)
+def test_quantized_linear(model, input_shape, n_bits, atol, is_signed, seed_torch):
     """Test the quantized module with a post-training static quantization.
 
     With n_bits>>0 we expect the results of the quantized module
@@ -97,7 +101,9 @@ def test_quantized_linear(model, input_shape, n_bits, atol, seed_torch):
     # Predict with real model
     numpy_prediction = numpy_fc_model(numpy_input)
     # Quantize with post-training static method
-    post_training_quant = PostTrainingAffineQuantization(n_bits, numpy_fc_model)
+    post_training_quant = PostTrainingAffineQuantization(
+        n_bits, numpy_fc_model, is_signed=is_signed
+    )
     quantized_model = post_training_quant.quantize_module(numpy_input)
     # Quantize input
     q_input = QuantizedArray(n_bits, numpy_input)