Test (quantize): preliminary unit tests to the CNNs 'quantize_model' (#…

…927)

* Added some prelininary unit tests to the CNNs 'quantize_model'. We test some aspects of 'layerwise' and 'fx' quantization, as well as some invalid inputs, e.g. invalid strings and zero and negative valued bit widths.

* We test po2 scale values, and have the negative bit widths fail.

* Updated to lastest dev

* Added tests for symm/asym quantization, per_chan quantization, and started work on minifloat quantization

* Remove unecessary inputs for minifloat

* Removed minifloat tests

* Add tests for checking calibration percentile performance, and testing MSE for qparam calibration.

* Fix default and condition

* Use Brevitas FX

---------

Co-authored-by: Giuseppe Franco <giuseppefranco4@gmail.com>

src/brevitas_examples/imagenet_classification/ptq/ptq_common.py

@@ -174,7 +174,7 @@ def quantize_model(
         act_param_method='stats',
         weight_quant_type='sym',
         act_quant_granularity='per_tensor',
-        act_scale_computation_type='dynamic',
+        act_scale_computation_type='static',
         uint_sym_act_for_unsigned_values=True,
         dtype=torch.float32,
         device='cpu'):
@@ -196,7 +196,7 @@ def quantize_model(
         act_mantissa_bit_width,
         act_exponent_bit_width)
 
-    if act_scale_computation_type == 'dynamic':
+    if act_scale_computation_type == 'dynamic' and backend != 'layerwise':
         assert bias_bit_width is None, "Bias quantization is not supported with dynamic activation quantization"
 
     weight_quant_format = quant_format
@@ -602,4 +602,5 @@ def check_positive_int(*args):
         if arg is None:
             continue
         assert arg > 0.0
+        assert not math.isclose(arg, 0.0)
         assert math.isclose(arg % 1, 0.0)