Skip to content

Fix TextRNN and centernet_hg104 tuning issue #171

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 4 commits into from
Dec 5, 2022
Merged

Fix TextRNN and centernet_hg104 tuning issue #171

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
f4270f9
Fix centernet_hg104 tuning issue
sys-lpot-val Nov 26, 2022
ba8e78b
Fix TextRNN tuning issue
sys-lpot-val Nov 26, 2022
db71b5d
Merge branch 'master' of https://github.com/intel/neural-compressor i…
lvliang-intel Nov 26, 2022
b0ef2e9
Merge branch 'master' of https://github.com/intel/neural-compressor i…
lvliang-intel Dec 4, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 6 additions & 0 deletions examples/tensorflow/oob_models/quantization/ptq/model_detail.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -385,5 +385,11 @@
'low': -1.0,
'high': 1.0
},
# centernet_hg104
{
'model_name': 'centernet_hg104',
'input': {'input_tensor': generate_data([224, 224, 3]),},
'output': ['Identity'],
},
]

3 changes: 3 additions & 0 deletions examples/tensorflow/oob_models/quantization/ptq/run_benchmark.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -101,6 +101,9 @@ function set_args {
NeuMF
PRNet
DIEN_Deep-Interest-Evolution-Network
EfficientDet-D2-768x768
EfficientDet-D4-1024x1024
centernet_hg104
--------
)

Expand Down
1 change: 1 addition & 0 deletions examples/tensorflow/oob_models/quantization/ptq/run_tuning.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -83,6 +83,7 @@ function set_args {
DIEN_Deep-Interest-Evolution-Network
EfficientDet-D2-768x768
EfficientDet-D4-1024x1024
centernet_hg104
--------
)

Expand Down
41 changes: 24 additions & 17 deletions neural_compressor/adaptor/tf_utils/graph_rewriter/generic/fuse_pad_with_conv.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -45,6 +45,7 @@ def do_transformation(self):
target_nodes = cur_graph.query_fusion_pattern_nodes(
[["Pad"], ["Conv2D", "Conv3D", "DepthwiseConv2dNative"], ('BiasAdd', 'Add', 'AddV2')])

padding_tensor_dict = {}
for node_combination in target_nodes:
conv_name = node_combination[1]

Expand All @@ -70,21 +71,26 @@ def do_transformation(self):
continue

padding_tensor = None
pad_node = graph_info[node_combination[0]].node
if graph_info[pad_node.input[1]].node.op != 'Const':
input_node = graph_info[pad_node.input[1]].node
if input_node.op == 'DataFormatVecPermute':
parent_input_node = graph_info[input_node.input[0]].node
if parent_input_node.op == 'Const':
padding_tensor = tensor_util.MakeNdarray( \
parent_input_node.attr["value"].tensor).flatten()
pad_node = None
if node_combination[0] not in padding_tensor_dict:
pad_node = graph_info[node_combination[0]].node
if graph_info[pad_node.input[1]].node.op != 'Const':
input_node = graph_info[pad_node.input[1]].node
if input_node.op == 'DataFormatVecPermute':
parent_input_node = graph_info[input_node.input[0]].node
if parent_input_node.op == 'Const':
padding_tensor = tensor_util.MakeNdarray( \
parent_input_node.attr["value"].tensor).flatten()
else:
continue
else:
continue
else:
continue
padding_tensor = tensor_util.MakeNdarray(
graph_info[pad_node.input[1]].node.attr["value"].tensor).flatten()
padding_tensor_dict[node_combination[0]] = padding_tensor
else:
padding_tensor = tensor_util.MakeNdarray(
graph_info[pad_node.input[1]].node.attr["value"].tensor).flatten()
padding_tensor = padding_tensor_dict[node_combination[0]]

if self.itex_qdq_mode:
enabled_pad_conv2d = bool(tf.version.VERSION == '1.15.0-up3' or \
Expand All @@ -95,12 +101,13 @@ def do_transformation(self):
if any(padding_tensor) and not enabled_pad_conv2d: # pragma: no cover
continue

if graph_info[pad_node.input[1]].node.op != 'Const':
cur_graph.node_name_details[pad_node.name].node.input.remove(pad_node.input[1])
cur_graph.remove_node_with_single_input_output(pad_node.name)
else:
cur_graph.remove_node_with_single_input_output(pad_node.name)
cur_graph.remove_node(pad_node.input[1])
if pad_node:
if graph_info[pad_node.input[1]].node.op != 'Const':
cur_graph.node_name_details[pad_node.name].node.input.remove(pad_node.input[1])
cur_graph.remove_node_with_single_input_output(pad_node.name)
else:
cur_graph.remove_node_with_single_input_output(pad_node.name)
cur_graph.remove_node(pad_node.input[1])
conv_node = graph_info[node_combination[1]].node
if self.itex_qdq_mode:
if any(padding_tensor) and enabled_pad_conv2d: # pragma: no cover
Expand Down
42 changes: 25 additions & 17 deletions neural_compressor/adaptor/tf_utils/graph_rewriter/generic/fuse_pad_with_fp32_conv.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -46,6 +46,7 @@ def do_transformation(self):
target_nodes = cur_graph.query_fusion_pattern_nodes(
[["Pad"], ["Conv2D", "DepthwiseConv2dNative"], ('BiasAdd', 'Add', 'AddV2')])

padding_tensor_dict = {}
for node_combination in target_nodes:
conv_name = node_combination[1]

Expand All @@ -71,21 +72,26 @@ def do_transformation(self):
continue

padding_tensor = None
pad_node = graph_info[node_combination[0]].node
if graph_info[pad_node.input[1]].node.op != 'Const':
input_node = graph_info[pad_node.input[1]].node
if input_node.op == 'DataFormatVecPermute':
parent_input_node = graph_info[input_node.input[0]].node
if parent_input_node.op == 'Const':
padding_tensor = tensor_util.MakeNdarray( \
parent_input_node.attr["value"].tensor).flatten()
pad_node = None
if node_combination[0] not in padding_tensor_dict:
pad_node = graph_info[node_combination[0]].node
if graph_info[pad_node.input[1]].node.op != 'Const':
input_node = graph_info[pad_node.input[1]].node
if input_node.op == 'DataFormatVecPermute':
parent_input_node = graph_info[input_node.input[0]].node
if parent_input_node.op == 'Const':
padding_tensor = tensor_util.MakeNdarray( \
parent_input_node.attr["value"].tensor).flatten()
else:
continue
else:
continue
else:
continue
padding_tensor = tensor_util.MakeNdarray(
graph_info[pad_node.input[1]].node.attr["value"].tensor).flatten()
padding_tensor_dict[node_combination[0]] = padding_tensor
else:
padding_tensor = tensor_util.MakeNdarray(
graph_info[pad_node.input[1]].node.attr["value"].tensor).flatten()
padding_tensor = padding_tensor_dict[node_combination[0]]

if self.itex_qdq_mode:
enabled_pad_conv2d = bool(tf.version.VERSION == '1.15.0-up3' or \
Expand All @@ -95,12 +101,14 @@ def do_transformation(self):

if any(padding_tensor) and not enabled_pad_conv2d: # pragma: no cover
continue
if graph_info[pad_node.input[1]].node.op != 'Const':
cur_graph.node_name_details[pad_node.name].node.input.remove(pad_node.input[1])
cur_graph.remove_node_with_single_input_output(pad_node.name)
else:
cur_graph.remove_node_with_single_input_output(pad_node.name)
cur_graph.remove_node(pad_node.input[1])

if pad_node:
if graph_info[pad_node.input[1]].node.op != 'Const':
cur_graph.node_name_details[pad_node.name].node.input.remove(pad_node.input[1])
cur_graph.remove_node_with_single_input_output(pad_node.name)
else:
cur_graph.remove_node_with_single_input_output(pad_node.name)
cur_graph.remove_node(pad_node.input[1])
conv_node = graph_info[node_combination[1]].node
# Helper.set_attr_int_list(conv_node, "padding_list", padding_tensor)
# only when padding attr is explicit, the explicit_paddings is not empty
Expand Down
4 changes: 2 additions & 2 deletions neural_compressor/adaptor/tf_utils/graph_rewriter/qdq/insert_qdq_pattern.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -182,7 +182,7 @@ def _check_op_list(self, node_type):
"MaxPool", "MaxPool3D", "FusedBatchNormV3", "Requantize", "RequantizePerChannel", "AvgPool", "Pad",
"CropAndResize", "Dequantize", "Mean", "MatMul", "BatchMatMul", "BatchMatMulV2",
"FakeQuantWithMinMaxVars", "_MklFusedInstanceNorm",
"Conv2DBackpropInput", "Conv3DBackpropInputV2")
"Conv2DBackpropInput", "Conv3DBackpropInputV2", "Sigmoid", "BiasAdd")
return any([node_type.find(i) != -1 for i in op_list])

def _find_relu_node(self, node):
Expand All @@ -198,7 +198,7 @@ def _find_relu_node(self, node):
or len(self.node_name_mapping \
[Helper.node_name_from_input(node.input[0])].output) > 1):
return True
elif 'T' in node.attr and node.attr['T'].type in (dtypes.quint8, dtypes.uint8):
elif 'T' in node.attr and dtypes.DType(node.attr['T'].type) in (dtypes.quint8, dtypes.uint8):
return True
elif (node.op.find("QuantizedConv") != -1
or node.op.find("QuantizedDepthwiseConv") != -1 or
Expand Down
4 changes: 3 additions & 1 deletion neural_compressor/adaptor/tf_utils/graph_util.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -918,11 +918,13 @@ def gen_per_iter(data):
if i.startswith(first_line):
iterations += 1

step = len(valid_data) / iterations
step = int(len(valid_data) / iterations)
final_res = []

for i in range(iterations):
final_res.extend(gen_per_iter(valid_data[int(i*step): int(step*( i+ 1))]))
if i + 1 == iterations and int(step*( i+ 1)) < len(valid_data):
final_res.extend(gen_per_iter(valid_data[int(step*( i+ 1)): len(valid_data)]))

return final_res

Expand Down