Feat (examples/a2q+): new super resolution models (Xilinx#811)

* Feat (a2q+): adding to super_res example

* Updating links to pre-trained checkpoints

src/brevitas_examples/super_resolution/README.md

@@ -20,10 +20,12 @@ Note that this is a difference from many academic works that train only on the Y
 | [quant_espcn_x2_w8a8_base](https://github.com/Xilinx/brevitas/releases/download/super_res_r1/quant_espcn_x2_w8a8_base-f761e4a1.pth) | x2 | int8 | (u)int8 | 30.96 |
 | [quant_espcn_x2_w8a8_a2q_32b](https://github.com/Xilinx/brevitas/releases/download/super_res_r1/quant_espcn_x2_w8a8_a2q_32b-85470d9b.pth) | x2 | int8 | (u)int8 | 30.79 |
 | [quant_espcn_x2_w8a8_a2q_16b](https://github.com/Xilinx/brevitas/releases/download/super_res_r1/quant_espcn_x2_w8a8_a2q_16b-f9e1da66.pth) | x2 | int8 | (u)int8 | 30.56 |
+| [quant_espcn_x2_w8a8_a2q_plus_16b](https://github.com/Xilinx/brevitas/releases/download/super_res_r2/quant_espcn_x2_w8a8_a2q_plus_16b-0ddf46f1.pth) | x2 | int8 | (u)int8 | 31.24 |
 ||
 | [quant_espcn_x2_w4a4_base](https://github.com/Xilinx/brevitas/releases/download/super_res_r1/quant_espcn_x2_w4a4_base-80658e6d.pth) | x2 | int4 | (u)int4 | 30.30 |
 | [quant_espcn_x2_w4a4_a2q_32b](https://github.com/Xilinx/brevitas/releases/download/super_res_r1/quant_espcn_x2_w4a4_a2q_32b-8702a412.pth) | x2 | int4 | (u)int4 | 30.27 |
 | [quant_espcn_x2_w4a4_a2q_13b](https://github.com/Xilinx/brevitas/releases/download/super_res_r1/quant_espcn_x2_w4a4_a2q_13b-9fff234e.pth) | x2 | int4 | (u)int4 | 30.24 |
+| [quant_espcn_x2_w4a4_a2q_plus_13b](https://github.com/Xilinx/brevitas/releases/download/super_res_r2/quant_espcn_x2_w4a4_a2q_plus_13b-6e6d55f0.pth) | x2 | int4 | (u)int4 | 30.95 |
 
 
 ## Train

src/brevitas_examples/super_resolution/models/__init__.py

@@ -7,6 +7,7 @@
 from torch import hub
 import torch.nn as nn
 
+from .common import CommonIntAccumulatorAwareZeroCenterWeightQuant
 from .espcn import *
 
 model_impl = {
@@ -43,18 +44,45 @@
             upscale_factor=2,
             weight_bit_width=4,
             act_bit_width=4,
-            acc_bit_width=13)}
+            acc_bit_width=13),
+    'quant_espcn_x2_w4a4_a2q_plus_13b':
+        partial(
+            quant_espcn,
+            upscale_factor=2,
+            weight_bit_width=4,
+            act_bit_width=4,
+            acc_bit_width=13,
+            weight_quant=CommonIntAccumulatorAwareZeroCenterWeightQuant),
+    'quant_espcn_x2_w8a8_a2q_plus_16b':
+        partial(
+            quant_espcn,
+            upscale_factor=2,
+            weight_bit_width=8,
+            act_bit_width=8,
+            acc_bit_width=16,
+            weight_quant=CommonIntAccumulatorAwareZeroCenterWeightQuant)}
 
-root_url = 'https://github.com/Xilinx/brevitas/releases/download/super_res_r1'
+root_url = 'https://github.com/Xilinx/brevitas/releases/download/'
 
 model_url = {
-    'float_espcn_x2': f'{root_url}/float_espcn_x2-2f85a454.pth',
-    'quant_espcn_x2_w4a4_a2q_13b': f'{root_url}/quant_espcn_x2_w4a4_a2q_13b-9fff234e.pth',
-    'quant_espcn_x2_w4a4_a2q_32b': f'{root_url}/quant_espcn_x2_w4a4_a2q_32b-8702a412.pth',
-    'quant_espcn_x2_w4a4_base': f'{root_url}/quant_espcn_x2_w4a4_base-80658e6d.pth',
-    'quant_espcn_x2_w8a8_a2q_16b': f'{root_url}/quant_espcn_x2_w8a8_a2q_16b-f9e1da66.pth',
-    'quant_espcn_x2_w8a8_a2q_32b': f'{root_url}/quant_espcn_x2_w8a8_a2q_32b-85470d9b.pth',
-    'quant_espcn_x2_w8a8_base': f'{root_url}/quant_espcn_x2_w8a8_base-f761e4a1.pth'}
+    'float_espcn_x2':
+        f'{root_url}/super_res_r1/float_espcn_x2-2f85a454.pth',
+    'quant_espcn_x2_w4a4_a2q_13b':
+        f'{root_url}/super_res_r1/quant_espcn_x2_w4a4_a2q_13b-9fff234e.pth',
+    'quant_espcn_x2_w4a4_a2q_32b':
+        f'{root_url}/super_res_r1/quant_espcn_x2_w4a4_a2q_32b-8702a412.pth',
+    'quant_espcn_x2_w4a4_base':
+        f'{root_url}/super_res_r1/quant_espcn_x2_w4a4_base-80658e6d.pth',
+    'quant_espcn_x2_w8a8_a2q_16b':
+        f'{root_url}/super_res_r1/quant_espcn_x2_w8a8_a2q_16b-f9e1da66.pth',
+    'quant_espcn_x2_w8a8_a2q_32b':
+        f'{root_url}/super_res_r1/quant_espcn_x2_w8a8_a2q_32b-85470d9b.pth',
+    'quant_espcn_x2_w8a8_base':
+        f'{root_url}/super_res_r1/quant_espcn_x2_w8a8_base-f761e4a1.pth',
+    'quant_espcn_x2_w4a4_a2q_plus_13b':
+        f'{root_url}/super_res_r2/quant_espcn_x2_w4a4_a2q_plus_13b-6e6d55f0.pth',
+    'quant_espcn_x2_w8a8_a2q_plus_16b':
+        f'{root_url}/super_res_r2/quant_espcn_x2_w8a8_a2q_plus_16b-0ddf46f1.pth'}
 
 
 def get_model_by_name(name: str, pretrained: bool = False) -> Union[FloatESPCN, QuantESPCN]:

src/brevitas_examples/super_resolution/models/common.py

@@ -12,6 +12,7 @@
 import brevitas.nn as qnn
 from brevitas.nn.quant_layer import WeightQuantType
 from brevitas.quant import Int8AccumulatorAwareWeightQuant
+from brevitas.quant import Int8AccumulatorAwareZeroCenterWeightQuant
 from brevitas.quant import Int8ActPerTensorFloat
 from brevitas.quant import Int8WeightPerTensorFloat
 from brevitas.quant import Uint8ActPerTensorFloat
@@ -26,9 +27,14 @@ class CommonIntWeightPerChannelQuant(Int8WeightPerTensorFloat):
 
 
 class CommonIntAccumulatorAwareWeightQuant(Int8AccumulatorAwareWeightQuant):
+    """A2Q: Accumulator-Aware Quantization with Guaranteed Overflow Avoidance"""
     restrict_scaling_impl = FloatRestrictValue  # backwards compatibility
-    pre_scaling_min_val = 1e-10
-    scaling_min_val = 1e-10
+    bit_width = None
+
+
+class CommonIntAccumulatorAwareZeroCenterWeightQuant(Int8AccumulatorAwareZeroCenterWeightQuant):
+    """A2Q+: Improving Accumulator-Aware Weight Quantization"""
+    bit_width = None
 
 
 class CommonIntActQuant(Int8ActPerTensorFloat):

src/brevitas_examples/super_resolution/models/espcn.py

@@ -164,15 +164,15 @@ def float_espcn(upscale_factor: int, num_channels: int = 3) -> FloatESPCN:
 
 
 def quant_espcn(
-        upcsale_factor: int,
+        upscale_factor: int,
         num_channels: int = 3,
         weight_bit_width: int = 8,
         act_bit_width: int = 8,
         acc_bit_width: int = 32,
         weight_quant: WeightQuantType = CommonIntWeightPerChannelQuant) -> QuantESPCN:
     """ """
     return QuantESPCN(
-        upscale_factor=upcsale_factor,
+        upscale_factor=upscale_factor,
         num_channels=num_channels,
         act_bit_width=act_bit_width,
         acc_bit_width=acc_bit_width,

src/brevitas_examples/super_resolution/utils/evaluate.py

@@ -5,12 +5,52 @@
 from torch import Tensor
 import torch.nn as nn
 
+from brevitas.core.scaling import AccumulatorAwareParameterPreScaling
+from brevitas.core.scaling import AccumulatorAwareZeroCenterParameterPreScaling
 import brevitas.nn as qnn
 from brevitas.nn.quant_layer import QuantWeightBiasInputOutputLayer as QuantWBIOL
 
 EPS = 1e-10
 
 
+def _get_a2q_module(module: nn.Module):
+    for submod in module.modules():
+        if isinstance(submod, AccumulatorAwareParameterPreScaling):
+            return submod
+    return None
+
+
+def _calc_a2q_acc_bit_width(
+        weight_max_l1_norm: Tensor, input_bit_width: Tensor, input_is_signed: bool):
+    """Using the closed-form bounds on accumulator bit-width as derived in
+    `A2Q: Accumulator-Aware Quantization with Guaranteed Overflow Avoidance`.
+    This function returns the minimum accumulator bit-width that can be used
+    without risk of overflow."""
+    assert weight_max_l1_norm.numel() == 1
+    input_is_signed = float(input_is_signed)
+    weight_max_l1_norm = torch.clamp_min(weight_max_l1_norm, EPS)
+    alpha = torch.log2(weight_max_l1_norm) + input_bit_width - input_is_signed
+    phi = lambda x: torch.log2(1. + pow(2., -x))
+    min_bit_width = alpha + phi(alpha) + 1.
+    min_bit_width = torch.ceil(min_bit_width)
+    return min_bit_width
+
+
+def _calc_a2q_plus_acc_bit_width(
+        weight_max_l1_norm: Tensor, input_bit_width: Tensor, input_is_signed: bool):
+    """Using the closed-form bounds on accumulator bit-width as derived in `A2Q+:
+    Improving Accumulator-Aware Weight Quantization`. This function returns the
+    minimum accumulator bit-width that can be used without risk of overflow,
+    assuming that the floating-point weights are zero-centered."""
+    input_is_signed = float(input_is_signed)
+    assert weight_max_l1_norm.numel() == 1
+    weight_max_l1_norm = torch.clamp_min(weight_max_l1_norm, EPS)
+    input_range = pow(2., input_bit_width) - 1.  # 2^N - 1.
+    min_bit_width = torch.log2(weight_max_l1_norm * input_range + 2.)
+    min_bit_width = torch.ceil(min_bit_width)
+    return min_bit_width
+
+
 def _calc_min_acc_bit_width(module: QuantWBIOL) -> Tensor:
     assert isinstance(module, qnn.QuantConv2d), "Error: function only support QuantConv2d."
 
@@ -24,12 +64,15 @@ def _calc_min_acc_bit_width(module: QuantWBIOL) -> Tensor:
     quant_weight_per_channel_l1_norm = quant_weight.norm(p=1, dim=(1, 2, 3))
 
     # using the closed-form bounds on accumulator bit-width
-    weight_max_l1_norm = quant_weight_per_channel_l1_norm.max()
-    weight_max_l1_norm = torch.clamp_min(weight_max_l1_norm, EPS)
-    alpha = torch.log2(weight_max_l1_norm) + input_bit_width - input_is_signed
-    phi = lambda x: torch.log2(1. + pow(2., -x))
-    min_bit_width = alpha + phi(alpha) + 1.
-    min_bit_width = torch.ceil(min_bit_width)
+    min_bit_width = _calc_a2q_acc_bit_width(
+        quant_weight_per_channel_l1_norm.max(),
+        input_bit_width=input_bit_width,
+        input_is_signed=input_is_signed)
+    if isinstance(_get_a2q_module(module), AccumulatorAwareZeroCenterParameterPreScaling):
+        min_bit_width = _calc_a2q_plus_acc_bit_width(
+            quant_weight_per_channel_l1_norm.max(),
+            input_bit_width=input_bit_width,
+            input_is_signed=input_is_signed)
     return min_bit_width