diff --git a/examples/export_LeNet/lenet.py b/examples/export_LeNet/lenet.py
index 63411ab0dd5df262f7556135227342e8d59efbe1..c28265561e61495d7d19cee0c7e140ebd2da68f9 100644
--- a/examples/export_LeNet/lenet.py
+++ b/examples/export_LeNet/lenet.py
@@ -22,7 +22,7 @@ from aidge_export_cpp.export_utils import (
set_nodes_datatypes,
exclude_unwanted_producers)
-from aidge_core.export_utils import remove_optional_inputs
+from aidge_core.export_utils import remove_optional_inputs, get_node_from_metaop
# Torch (Dataset)
import torch
@@ -89,8 +89,7 @@ Export configuration details :
- OPTIM_SIGN : Quantization optional optimization based on data sign.
- SINGLE_SHIFT : Quantization option specifying if inserted scaling nodes should be
single shift or floating point.
-- ROUNDING : Apply rounding on the data after the single shift step.
-- NO_QUANTIZATION : Skip the quantization step.
+- NO_QUANT : Skip the quantization step.
- CLIPPING : Clipping method during quantization.
- FOLD_GRAPH : The quantization step adds cast nodes to cast the graph into the given TARGET_TYPE.
Enabling the FOLD_GRAPH will automatically fold these nodes into the following
@@ -136,7 +135,7 @@ DO_EXAMPLES = True
OPTIM_SIGN = False
SINGLE_SHIFT = True
ROUNDING = True
-NO_QUANTIZATION = False
+NO_QUANT = False
CLIPPING = aidge_quantization.Clipping.MSE # 'MAX'
FOLD_GRAPH = True
@@ -150,12 +149,11 @@ print(' NB_TEST = ', NB_TEST)
print(' NB_CALIB = ', NB_CALIB)
print(' NB_BITS = ', NB_BITS)
print(' OPTIM_SIGN = ', OPTIM_SIGN)
-print(' NO_QUANTIZATION = ', NO_QUANTIZATION)
+print(' NO_QUANT = ', NO_QUANT)
print(' CLIPPING = ', CLIPPING)
print(' SINGLE_SHIFT = ', SINGLE_SHIFT)
print(' USE_CUDA = ', USE_CUDA)
print(' DEV_MODE = ', DEV_MODE)
-print(' ROUNDING = ', ROUNDING)
torch.manual_seed(RNG_SEED)
random.seed(RNG_SEED)
@@ -207,11 +205,13 @@ for input, label in test_set:
Load the .onnx model and perform some usual graph modifications :
- Remove the flatten nodes;
- Fuse the batchnorm nodes into the biases producers.
+ - Expand the metaOperators to perform the desired fusions.
"""
model = aidge_onnx.load_onnx(MODEL_NAME + ".onnx", verbose=False)
aidge_core.remove_flatten(model)
aidge_core.fuse_batchnorm(model)
+aidge_core.expand_metaops(model)
model.save("imported_model")
# --------------------------------------------------------------
@@ -269,15 +269,24 @@ if quantize_model:
aidge_quantization.quantize_network(
network = model,
nb_bits = NB_BITS,
- input_dataset = tensors[0:NB_CALIB],
+ calibration_set = tensors[0:NB_CALIB],
clipping_mode = CLIPPING,
target_type = TARGET_TYPE,
- no_quantization = NO_QUANTIZATION,
+ no_quant = NO_QUANT,
optimize_signs = OPTIM_SIGN,
single_shift = SINGLE_SHIFT,
use_cuda = USE_CUDA,
- fold_graph = FOLD_GRAPH,
- bitshift_rounding = ROUNDING)
+ fold_graph = FOLD_GRAPH)
+
+# Tag the scaling producers
+for node in model.get_nodes():
+ if node.type() == "Quantizer":
+ for SNode in get_node_from_metaop(node, "BitShift"):
+ SNode.get_parent(1).attributes().shift_prod = True
+ for CNode in get_node_from_metaop(node, "Mul"):
+ CNode.get_parent(1).attributes().coef_prod = True
+
+model.save("post_ptq_model")
# --------------------------------------------------------------
# RESCALE THE INPUT SAMPLES
@@ -358,13 +367,6 @@ In this step, we use graph regex techniques to find the desired patterns
within the graph in order to match the export implementation of the kernels.
"""
-# Expand meta ops
-"""
-We first need to expand the graph to break all the metaops that may already
-exist. For instance, PaddedConv will become Pad -> Conv.
-"""
-aidge_core.expand_metaops(model)
-
# Exclude unwanted producers
"""
Before fusing the nodes, we set a tag on the Producers in order to exclude
diff --git a/examples/export_ResNet18/resnet18.py b/examples/export_ResNet18/resnet18.py
index 354e4740b809364be0571a6c6b0ee2fc244281ff..3ba1ff4eef95c9ce89b55fee27fb79bd5af5b7ca 100644
--- a/examples/export_ResNet18/resnet18.py
+++ b/examples/export_ResNet18/resnet18.py
@@ -27,7 +27,7 @@ from aidge_export_cpp.export_utils import (
set_nodes_datatypes,
normalize)
-from aidge_core.export_utils import remove_optional_inputs
+from aidge_core.export_utils import remove_optional_inputs, get_node_from_metaop
# Torch (Dataset)
import torch
@@ -94,8 +94,7 @@ Export configuration details :
- OPTIM_SIGN : Quantization optional optimization based on data sign.
- SINGLE_SHIFT : Quantization option specifying if inserted scaling nodes should be
single shift or floating point.
-- ROUNDING : Apply rounding on the data after the single shift step.
-- NO_QUANTIZATION : Skip the quantization step. Should be set to False.
+- NO_QUANT : Skip the quantization step. Should be set to False.
- CLIPPING : Clipping method during quantization.
- FOLD_GRAPH : The quantization step adds cast nodes to cast the graph into the given TARGET_TYPE.
Enabling the FOLD_GRAPH will automatically fold these nodes into the following
@@ -161,14 +160,13 @@ def print_cfg():
print(' NB_CALIB = ', NB_CALIB)
print(' NB_BITS = ', NB_BITS)
print(' OPTIM_SIGN = ', OPTIM_SIGN)
- print(' NO_QUANTIZATION = ', NO_QUANTIZATION)
+ print(' NO_QUANT = ', NO_QUANT)
print(' CLIPPING = ', CLIPPING)
print(' SINGLE_SHIFT = ', SINGLE_SHIFT)
print(' TARGET_TYPE = ', TARGET_TYPE)
print(' FOLD_GRAPH = ', FOLD_GRAPH)
print(' USE_CUDA = ', USE_CUDA)
print(' DEV_MODE = ', DEV_MODE)
- print(' ROUNDING = ', ROUNDING)
print_cfg()
@@ -239,12 +237,14 @@ for tensor in tensors:
Load the .onnx model and perform some usual graph modifications :
- Remove the flatten nodes;
- Fuse the batchnorm nodes into the biases producers.
+ - Expand the metaOperators to perform the desired fusions.
"""
model = aidge_onnx.load_onnx(MODEL_NAME + ".onnx", verbose=False)
model.save("imported_model")
aidge_core.remove_flatten(model)
aidge_core.fuse_batchnorm(model)
+aidge_core.expand_metaops(model)
model.save("imported_model_fused_bn")
# --------------------------------------------------------------
@@ -301,15 +301,24 @@ if quantize_model:
aidge_quantization.quantize_network(
network = model,
nb_bits = NB_BITS,
- input_dataset = aidge_tensors[0:NB_CALIB],
+ calibration_set = aidge_tensors[0:NB_CALIB],
clipping_mode = CLIPPING,
target_type = TARGET_TYPE,
- no_quantization = NO_QUANTIZATION,
+ no_quant = NO_QUANT,
optimize_signs = OPTIM_SIGN,
single_shift = SINGLE_SHIFT,
use_cuda = USE_CUDA,
- fold_graph = FOLD_GRAPH,
- bitshift_rounding = ROUNDING)
+ fold_graph = FOLD_GRAPH)
+
+# Tag the scaling producers
+for node in model.get_nodes():
+ if node.type() == "Quantizer":
+ for SNode in get_node_from_metaop(node, "BitShift"):
+ SNode.get_parent(1).attributes().shift_prod = True
+ for CNode in get_node_from_metaop(node, "Mul"):
+ CNode.get_parent(1).attributes().coef_prod = True
+
+model.save("post_ptq_model")
# --------------------------------------------------------------
# RESCALE THE INPUT SAMPLES
@@ -382,16 +391,6 @@ In this step, we use graph regex techniques to find the desired patterns
within the graph in order to match the export implementation of the kernels.
"""
-# Expand meta ops
-"""
-We first need to expand the graph to break all the metaops that may already
-exist. For instance, PaddedConv will become Pad -> Conv.
-"""
-aidge_core.expand_metaops(model)
-
-
-model.save("after_expand")
-
# Exclude unwanted producers
"""
Before fusing the nodes, we set a tag on the Producers in order to exclude