From 87913b73dba9b0e7d88d94298d303c7c7db3bcb3 Mon Sep 17 00:00:00 2001
From: Franck Galpin <franck.galpin@interdigital.com>
Date: Fri, 22 Nov 2024 08:16:34 +0000
Subject: [PATCH] fix python scripts, model names, sadl version
---
Makefile | 2 +-
cfg/nn-based/vlop3.cfg | 3 +
fixsadl.patch | 15 -
sadl | 2 +-
source/App/DecoderApp/DecAppCfg.cpp | 2 +-
source/App/EncoderApp/EncAppCfg.cpp | 2 +-
source/Lib/CommonLib/NNFilterSet1.cpp | 4 -
.../NN_Filtering/LOP4/model/model.py | 285 +++++++++++++++---
.../LOP4/quantize/quantize_lop4.py | 227 +++++++-------
.../NN_Filtering/VLOP3/model/model.py | 285 +++++++++++++++---
.../NN_Filtering/VLOP3/quantize/quantize.py | 149 ++++-----
11 files changed, 691 insertions(+), 285 deletions(-)
create mode 100644 cfg/nn-based/vlop3.cfg
delete mode 100644 fixsadl.patch
diff --git a/Makefile b/Makefile
index 0ed4aa783..83fb371cf 100644
--- a/Makefile
+++ b/Makefile
@@ -18,7 +18,7 @@ BUILD_SCRIPT := $(CURDIR)/cmake/CMakeBuild/bin/cmake.py
TARGETS := CommonLib DecoderAnalyserApp DecoderAnalyserLib DecoderApp DecoderLib
TARGETS += EncoderApp EncoderLib Utilities SEIRemovalApp StreamMergeApp
-SADL_HASH := "v10" # hard coded because of windows $(shell git submodule status | grep sadl | cut -d' ' -f2)
+SADL_HASH := "2088b26" # hard coded because of windows $(shell git submodule status | grep sadl | cut -d' ' -f2)
ifeq ($(OS),Windows_NT)
ifneq ($(MSYSTEM),)
diff --git a/cfg/nn-based/vlop3.cfg b/cfg/nn-based/vlop3.cfg
new file mode 100644
index 000000000..61457e8cf
--- /dev/null
+++ b/cfg/nn-based/vlop3.cfg
@@ -0,0 +1,3 @@
+NnlfOption : 3
+NnlfModelName : models/nnlf_vlop3_model_int16.sadl
+
diff --git a/fixsadl.patch b/fixsadl.patch
deleted file mode 100644
index 285e68819..000000000
--- a/fixsadl.patch
+++ /dev/null
@@ -1,15 +0,0 @@
-diff --git a/sadl/tensor.h b/sadl/tensor.h
-index 98d7157..db524f4 100644
---- a/sadl/tensor.h
-+++ b/sadl/tensor.h
-@@ -651,8 +651,8 @@ template<typename T> void Tensor<T>::resize(Dimensions d)
- template<typename T> void Tensor<T>::resizeSparse(Dimensions d, int32_t sizeSparse, int32_t packedSparsitySize)
- {
- m_dims = d;
-- int64_t m = m_dims.nbElements();
-- assert(m < kMaxSize);
-+ //int64_t m = m_dims.nbElements();
-+ //assert(m < kMaxSize);
- assert(sizeSparse > 0);
- assert(packedSparsitySize > 0);
- m_data_sparse.clear();
diff --git a/sadl b/sadl
index 0854e53ac..ff7e50a0f 160000
--- a/sadl
+++ b/sadl
@@ -1 +1 @@
-Subproject commit 0854e53ac91cedc26fe0469a2ed930d62ba1fc4c
+Subproject commit ff7e50a0f0432843b989dfd87cb359c1b540467b
diff --git a/source/App/DecoderApp/DecAppCfg.cpp b/source/App/DecoderApp/DecAppCfg.cpp
index 757ccc9ae..c30d5dbd1 100644
--- a/source/App/DecoderApp/DecAppCfg.cpp
+++ b/source/App/DecoderApp/DecAppCfg.cpp
@@ -81,7 +81,7 @@ bool DecAppCfg::parseCfg( int argc, char* argv[] )
("DumpBasename", m_dumpBasename, string(""), "basename for data dumping\n")
#endif
#if NN_LF_UNIFIED
- ("NnlfModelName", m_nnModel[NnModel::LOP_UNIFIED_FILTER], string("models/nnlf_lop3_model_int16.sadl"), "loop filter model name\n")
+ ("NnlfModelName", m_nnModel[NnModel::LOP_UNIFIED_FILTER], string("models/nnlf_lop4_model_int16.sadl"), "loop filter model name\n")
#if NN_LF_FORCE_USE
( "NnlfUnifiedDebugOption", m_nnlfDebugOption, 0, "Option used to debug stage 1 model. 0: default, 1: apply only on I slice, 2: apply on all slices using I type as input" )
#endif
diff --git a/source/App/EncoderApp/EncAppCfg.cpp b/source/App/EncoderApp/EncAppCfg.cpp
index 0c683c0e0..5fd696d07 100644
--- a/source/App/EncoderApp/EncAppCfg.cpp
+++ b/source/App/EncoderApp/EncAppCfg.cpp
@@ -1188,7 +1188,7 @@ bool EncAppCfg::parseCfg( int argc, char* argv[] )
#else
("NnlfOption", m_nnlfOption, 1, "NN-based in-loop filter option (0:disable nnlf, 1: enable unified nnlf, [legacy: 10: enable nnlf-0, 11: enable nnlf-1, 12: enable nnlf-LC])")
#endif
- ("NnlfModelName", m_nnModel[NnModel::LOP_UNIFIED_FILTER], string("models/nnlf_lop3_model_int16.sadl"), "unified nnlf model name." )
+ ("NnlfModelName", m_nnModel[NnModel::LOP_UNIFIED_FILTER], string("models/nnlf_lop4_model_int16.sadl"), "unified nnlf model name." )
#endif
#if NN_LF_UNIFIED
( "NnlfBlockSize", m_nnlfBlockSize, 128u, "Base inference size of NN-based in-loop filter")
diff --git a/source/Lib/CommonLib/NNFilterSet1.cpp b/source/Lib/CommonLib/NNFilterSet1.cpp
index fcb3d9d70..b6b23c866 100644
--- a/source/Lib/CommonLib/NNFilterSet1.cpp
+++ b/source/Lib/CommonLib/NNFilterSet1.cpp
@@ -424,8 +424,6 @@ void NNFilterSet1::cnnFilterLumaBlock(Picture* pic, UnitArea inferArea, int extL
int log2InputScale = 10;
int log2QpScale = 6;
- const int border_to_skip = 0;
- if (border_to_skip>0) sadl::Tensor<float>::skip_border = true;
// get model
#if JVET_AC0177_MULTI_FRAME
@@ -570,8 +568,6 @@ void NNFilterSet1::cnnFilterChromaBlock(Picture* pic, UnitArea inferArea, int ex
int log2InputScale = 10;
int log2QpScale = 6;
- const int border_to_skip = 0;
- if (border_to_skip>0) sadl::Tensor<float>::skip_border = true;
// get model
ModelData<T> &m = getModel<T>(inferArea.lheight(), inferArea.lwidth(), false, inter, inter ? m_interChroma : m_intraChroma);
diff --git a/training/training_scripts/NN_Filtering/LOP4/model/model.py b/training/training_scripts/NN_Filtering/LOP4/model/model.py
index 853f307e0..e4dee907a 100644
--- a/training/training_scripts/NN_Filtering/LOP4/model/model.py
+++ b/training/training_scripts/NN_Filtering/LOP4/model/model.py
@@ -122,7 +122,7 @@ class Conv(nn.Sequential):
super(Conv, self).__init__(*modules)
-
+
class MultiBranchModule(nn.Module):
"""A module representing multple, parallel branches. If the input is a list, each element in the list is fed into the corresponding branch,
otherwise the input is fed into every branch. The outputs of each branch are then merged."""
@@ -148,8 +148,24 @@ class NewResBlock_separate_prelu(nn.Sequential):
super().__init__()
self.prelu = nn.PReLU()
self.conv1_11 = Conv(C1, C, kernel_size=1, post_activation=None)
- self.conv2_13 = Conv(C, C, kernel_size=3, post_activation=None, is_separable=True, is_horizontal=True, hidden_separable_channels=C21)
- self.conv3_31 = Conv(C, C, kernel_size=3, post_activation=None, is_separable=True, is_horizontal=False, hidden_separable_channels=C21)
+ self.conv2_13 = Conv(
+ C,
+ C,
+ kernel_size=3,
+ post_activation=None,
+ is_separable=True,
+ is_horizontal=True,
+ hidden_separable_channels=C21,
+ )
+ self.conv3_31 = Conv(
+ C,
+ C,
+ kernel_size=3,
+ post_activation=None,
+ is_separable=True,
+ is_horizontal=False,
+ hidden_separable_channels=C21,
+ )
self.conv4_11 = Conv(C, C1, kernel_size=1, post_activation=None)
def forward(self, x: torch.Tensor) -> torch.Tensor:
@@ -167,8 +183,26 @@ class NewResBlock_separate_prelu_crop(nn.Sequential):
super().__init__()
self.prelu = nn.PReLU()
self.conv1_11 = Conv(C1, C, kernel_size=1, post_activation=None)
- self.conv2_13 = Conv(C, C, kernel_size=3, padding=(0,0), post_activation=None, is_separable=True, is_horizontal=True, hidden_separable_channels=C21)
- self.conv3_31 = Conv(C, C, kernel_size=3, padding=(0,0), post_activation=None, is_separable=True, is_horizontal=False, hidden_separable_channels=C21)
+ self.conv2_13 = Conv(
+ C,
+ C,
+ kernel_size=3,
+ padding=(0, 0),
+ post_activation=None,
+ is_separable=True,
+ is_horizontal=True,
+ hidden_separable_channels=C21,
+ )
+ self.conv3_31 = Conv(
+ C,
+ C,
+ kernel_size=3,
+ padding=(0, 0),
+ post_activation=None,
+ is_separable=True,
+ is_horizontal=False,
+ hidden_separable_channels=C21,
+ )
self.conv4_11 = Conv(C, C1, kernel_size=1, post_activation=None)
def forward(self, x: torch.Tensor) -> torch.Tensor:
@@ -180,6 +214,7 @@ class NewResBlock_separate_prelu_crop(nn.Sequential):
x5 = self.conv4_11(x4)
return x5 + temp
+
class Slice_only_y(nn.Sequential):
def __init__(self):
super().__init__()
@@ -194,8 +229,8 @@ class Slice_only_x(nn.Sequential):
def forward(self, x: torch.Tensor) -> torch.Tensor:
return x[:, :, :, 1:-1]
-
-
+
+
class SplitLumaChromaBlocks(nn.Sequential):
def __init__(
self,
@@ -218,26 +253,91 @@ class SplitLumaChromaBlocks(nn.Sequential):
self.Y_no_lower_extra_crop = N_Y[1]
self.split_y_path = nn.Sequential(
Conv(C_Y, C1_Y, kernel_size=1, post_activation=None),
- *[NewResBlock_separate_prelu(C_Y, C1_Y, C21_Y) for _ in range(N_Y[0]) ], # Start with full margin 36x36
- *[NewResBlock_separate_prelu_crop(C_Y, C1_Y, C21_Y) for _ in range(N_Y[1]) ], # Go down once to 34x34
- *[NewResBlock_separate_prelu(C_Y, C1_Y, C21_Y) for _ in range(N_Y[2]) ], # continue once at 34x34
- *[NewResBlock_separate_prelu_crop(C_Y, C1_Y, C21_Y) for _ in range(N_Y[3]) ], # go down once again to 32x32
- *[NewResBlock_separate_prelu(C_Y, C1_Y, C21_Y) for _ in range(N_Y[4]) ], # continue with 32x32 for the rest
+ *[
+ NewResBlock_separate_prelu(C_Y, C1_Y, C21_Y) for _ in range(N_Y[0])
+ ], # Start with full margin 36x36
+ *[
+ NewResBlock_separate_prelu_crop(C_Y, C1_Y, C21_Y) for _ in range(N_Y[1])
+ ], # Go down once to 34x34
+ *[
+ NewResBlock_separate_prelu(C_Y, C1_Y, C21_Y) for _ in range(N_Y[2])
+ ], # continue once at 34x34
+ *[
+ NewResBlock_separate_prelu_crop(C_Y, C1_Y, C21_Y) for _ in range(N_Y[3])
+ ], # go down once again to 32x32
+ *[
+ NewResBlock_separate_prelu(C_Y, C1_Y, C21_Y) for _ in range(N_Y[4])
+ ], # continue with 32x32 for the rest
Conv(C1_Y, C_Y, kernel_size=1, post_activation=None),
-
- # If we need extra cropping, we will do padding = (0,0) here.
- *[Conv(C_Y, C_Y, kernel_size=3, padding=(0,0), is_separable=True, is_horizontal=True, hidden_separable_channels=C21_Y, post_activation=None) for _ in range(self.Y_upper_extra_crop)],
- *[Conv(C_Y, C_Y, kernel_size=3, padding=(0,0), is_separable=True, is_horizontal=False, hidden_separable_channels=C21_Y, post_activation=None) for _ in range(self.Y_upper_extra_crop)],
+ # If we need extra cropping, we will do padding = (0,0) here.
+ *[
+ Conv(
+ C_Y,
+ C_Y,
+ kernel_size=3,
+ padding=(0, 0),
+ is_separable=True,
+ is_horizontal=True,
+ hidden_separable_channels=C21_Y,
+ post_activation=None,
+ )
+ for _ in range(self.Y_upper_extra_crop)
+ ],
+ *[
+ Conv(
+ C_Y,
+ C_Y,
+ kernel_size=3,
+ padding=(0, 0),
+ is_separable=True,
+ is_horizontal=False,
+ hidden_separable_channels=C21_Y,
+ post_activation=None,
+ )
+ for _ in range(self.Y_upper_extra_crop)
+ ],
# Otherwise, we will do padding = (1,1) (which is implicit if not specified).
- *[Conv(C_Y, C_Y, kernel_size=3, is_separable=True, is_horizontal=True, hidden_separable_channels=C21_Y, post_activation=None) for _ in range(self.Y_no_upper_extra_crop)],
- *[Conv(C_Y, C_Y, kernel_size=3, is_separable=True, is_horizontal=False, hidden_separable_channels=C21_Y, post_activation=None) for _ in range(self.Y_no_upper_extra_crop)],
-
+ *[
+ Conv(
+ C_Y,
+ C_Y,
+ kernel_size=3,
+ is_separable=True,
+ is_horizontal=True,
+ hidden_separable_channels=C21_Y,
+ post_activation=None,
+ )
+ for _ in range(self.Y_no_upper_extra_crop)
+ ],
+ *[
+ Conv(
+ C_Y,
+ C_Y,
+ kernel_size=3,
+ is_separable=True,
+ is_horizontal=False,
+ hidden_separable_channels=C21_Y,
+ post_activation=None,
+ )
+ for _ in range(self.Y_no_upper_extra_crop)
+ ],
Conv(C_Y, C_Y, kernel_size=1),
-
# If we need extra cropping, we will do padding = (0,0) here.
- *[Conv(C_Y, output_channels_y, kernel_size=3, padding=(0,0), post_activation=None) for _ in range(self.Y_lower_extra_crop)],
+ *[
+ Conv(
+ C_Y,
+ output_channels_y,
+ kernel_size=3,
+ padding=(0, 0),
+ post_activation=None,
+ )
+ for _ in range(self.Y_lower_extra_crop)
+ ],
# Otherwise, we will do padding = (1,1) (which is implicit if not specified).
- *[Conv(C_Y, output_channels_y, kernel_size=3, post_activation=None) for _ in range(self.Y_no_lower_extra_crop)],
+ *[
+ Conv(C_Y, output_channels_y, kernel_size=3, post_activation=None)
+ for _ in range(self.Y_no_lower_extra_crop)
+ ],
)
self.UV_upper_extra_crop = 1 - N_UV[3]
@@ -246,26 +346,93 @@ class SplitLumaChromaBlocks(nn.Sequential):
self.UV_no_lower_extra_crop = N_UV[1]
self.split_uv_path = nn.Sequential(
Conv(C_UV, C1_UV, kernel_size=1, post_activation=None),
- *[NewResBlock_separate_prelu(C_UV, C1_UV, C21_UV) for _ in range(N_UV[0])], # Start with full margin 36x36
- *[NewResBlock_separate_prelu_crop(C_UV, C1_UV, C21_UV) for _ in range(N_UV[1])], # Go down once to 34x34
- *[NewResBlock_separate_prelu(C_UV, C1_UV, C21_UV) for _ in range(N_UV[2])], # continue once at 34x34
- *[NewResBlock_separate_prelu_crop(C_UV, C1_UV, C21_UV) for _ in range(N_UV[3])], # go down once again to 32x32
- *[NewResBlock_separate_prelu(C_UV, C1_UV, C21_UV) for _ in range(N_UV[4])], # continue with 32x32 for the rest
+ *[
+ NewResBlock_separate_prelu(C_UV, C1_UV, C21_UV) for _ in range(N_UV[0])
+ ], # Start with full margin 36x36
+ *[
+ NewResBlock_separate_prelu_crop(C_UV, C1_UV, C21_UV)
+ for _ in range(N_UV[1])
+ ], # Go down once to 34x34
+ *[
+ NewResBlock_separate_prelu(C_UV, C1_UV, C21_UV) for _ in range(N_UV[2])
+ ], # continue once at 34x34
+ *[
+ NewResBlock_separate_prelu_crop(C_UV, C1_UV, C21_UV)
+ for _ in range(N_UV[3])
+ ], # go down once again to 32x32
+ *[
+ NewResBlock_separate_prelu(C_UV, C1_UV, C21_UV) for _ in range(N_UV[4])
+ ], # continue with 32x32 for the rest
Conv(C1_UV, C_UV, kernel_size=1, post_activation=None),
-
# If we need extra cropping, we will do padding = (0,0) here.
- *[Conv(C_UV, C_UV, kernel_size=3, padding=(0,0), is_separable=True, is_horizontal=True, hidden_separable_channels=C21_UV, post_activation=None) for _ in range(self.UV_upper_extra_crop)],
- *[Conv(C_UV, C_UV, kernel_size=3, padding=(0,0), is_separable=True, is_horizontal=False, hidden_separable_channels=C21_UV, post_activation=None) for _ in range(self.UV_upper_extra_crop)],
+ *[
+ Conv(
+ C_UV,
+ C_UV,
+ kernel_size=3,
+ padding=(0, 0),
+ is_separable=True,
+ is_horizontal=True,
+ hidden_separable_channels=C21_UV,
+ post_activation=None,
+ )
+ for _ in range(self.UV_upper_extra_crop)
+ ],
+ *[
+ Conv(
+ C_UV,
+ C_UV,
+ kernel_size=3,
+ padding=(0, 0),
+ is_separable=True,
+ is_horizontal=False,
+ hidden_separable_channels=C21_UV,
+ post_activation=None,
+ )
+ for _ in range(self.UV_upper_extra_crop)
+ ],
# Otherwise, we will do padding = (1,1) (which is implicit if not specified).
- *[Conv(C_UV, C_UV, kernel_size=3, is_separable=True, is_horizontal=True, hidden_separable_channels=C21_UV, post_activation=None) for _ in range(self.UV_no_upper_extra_crop)],
- *[Conv(C_UV, C_UV, kernel_size=3, is_separable=True, is_horizontal=False, hidden_separable_channels=C21_UV, post_activation=None) for _ in range(self.UV_no_upper_extra_crop)],
-
+ *[
+ Conv(
+ C_UV,
+ C_UV,
+ kernel_size=3,
+ is_separable=True,
+ is_horizontal=True,
+ hidden_separable_channels=C21_UV,
+ post_activation=None,
+ )
+ for _ in range(self.UV_no_upper_extra_crop)
+ ],
+ *[
+ Conv(
+ C_UV,
+ C_UV,
+ kernel_size=3,
+ is_separable=True,
+ is_horizontal=False,
+ hidden_separable_channels=C21_UV,
+ post_activation=None,
+ )
+ for _ in range(self.UV_no_upper_extra_crop)
+ ],
Conv(C_UV, C_UV, kernel_size=1),
-
# If we need extra cropping, we will do padding = (0,0) here.
- *[Conv(C_UV, output_channels_uv, padding=(0,0), kernel_size=3, post_activation=None) for _ in range(self.UV_lower_extra_crop)],
+ *[
+ Conv(
+ C_UV,
+ output_channels_uv,
+ padding=(0, 0),
+ kernel_size=3,
+ post_activation=None,
+ )
+ for _ in range(self.UV_lower_extra_crop)
+ ],
# Otherwise, we will do padding = (1,1) (which is implicit if not specified).
- *[Conv(C_UV, output_channels_uv, kernel_size=3, post_activation=None) for _ in range(self.UV_no_lower_extra_crop)],
+ *[
+ Conv(C_UV, output_channels_uv, kernel_size=3, post_activation=None)
+ for _ in range(self.UV_no_lower_extra_crop)
+ ],
)
self.Cy = C_Y
@@ -324,10 +491,39 @@ class SADLNet(nn.Sequential):
]
),
Conv(sum(self.input_features), D6, kernel_size=1),
- Conv(D6, D6, kernel_size=3, stride=2, post_activation=None, is_separable=True, is_horizontal=True, hidden_separable_channels=D6),
- Conv(D6, D6, kernel_size=3, stride=2, post_activation=None, is_separable=True, is_horizontal=False, hidden_separable_channels=D6),
+ Conv(
+ D6,
+ D6,
+ kernel_size=3,
+ stride=2,
+ post_activation=None,
+ is_separable=True,
+ is_horizontal=True,
+ hidden_separable_channels=D6,
+ ),
+ Conv(
+ D6,
+ D6,
+ kernel_size=3,
+ stride=2,
+ post_activation=None,
+ is_separable=True,
+ is_horizontal=False,
+ hidden_separable_channels=D6,
+ ),
Conv(D6, C_Y + C_UV, kernel_size=1),
- SplitLumaChromaBlocks(N_Y, N_UV, C_Y, C_UV, C1_Y, C1_UV, C21_Y, C21_UV, output_channels_y, output_channels_uv),
+ SplitLumaChromaBlocks(
+ N_Y,
+ N_UV,
+ C_Y,
+ C_UV,
+ C1_Y,
+ C1_UV,
+ C21_Y,
+ C21_UV,
+ output_channels_y,
+ output_channels_uv,
+ ),
)
def get_example_inputs(
@@ -399,7 +595,14 @@ class Net(nn.Module):
input_kernels
), "[ERROR] input size and kernels size not equal"
self.input_channels = input_channels
- sizes = [dct_ch + dct_ch // 2, dct_ch + dct_ch // 2, dct_ch + dct_ch // 2, 1, 1, dct_ch]
+ sizes = [
+ dct_ch + dct_ch // 2,
+ dct_ch + dct_ch // 2,
+ dct_ch + dct_ch // 2,
+ 1,
+ 1,
+ dct_ch,
+ ]
self.SADL_model = SADLNet(
sizes,
input_kernels,
@@ -418,7 +621,7 @@ class Net(nn.Module):
C21_Y,
C21_UV,
4 * dct_ch,
- 2 * dct_ch
+ 2 * dct_ch,
)
self.chroma_upsampler = nn.Upsample(scale_factor=2, mode="nearest")
self.dct_ch = dct_ch
diff --git a/training/training_scripts/NN_Filtering/LOP4/quantize/quantize_lop4.py b/training/training_scripts/NN_Filtering/LOP4/quantize/quantize_lop4.py
index a937bbe21..106e7718b 100644
--- a/training/training_scripts/NN_Filtering/LOP4/quantize/quantize_lop4.py
+++ b/training/training_scripts/NN_Filtering/LOP4/quantize/quantize_lop4.py
@@ -45,9 +45,7 @@ import json
def parse_arguments():
parser = argparse.ArgumentParser()
- parser.add_argument(
- "--config", "-c", type=str, help="config file"
- )
+ parser.add_argument("--config", "-c", type=str, help="config file")
parser.add_argument(
"--stage",
type=int,
@@ -59,37 +57,37 @@ def parse_arguments():
def qfactor_gen(content):
- output_q_str = ''
+ output_q_str = ""
linenumber = 0
- q0 = '13'
+ q0 = "13"
q_prev = q0
q_save_prev = q_prev
for line in content:
linenumber += 1
- if line[:11] == '[INFO] id: ':
- layer_str = line.split(' ')
+ if line[:11] == "[INFO] id: ":
+ layer_str = line.split(" ")
layer_id = layer_str[2]
layer_name = layer_str[-1][:-1]
- if layer_name == 'Placeholder':
+ if layer_name == "Placeholder":
qq = q0
- print('Setting qq = q0 for', layer_name)
- elif layer_name == 'Const':
+ print("Setting qq = q0 for", layer_name)
+ elif layer_name == "Const":
next_layer = content[linenumber + 5]
- layer_str2 = next_layer.split(' ')
+ layer_str2 = next_layer.split(" ")
next_layer_name = layer_str2[-1][:-1]
- if next_layer_name == 'Concat':
- qq = '0'
+ if next_layer_name == "Concat":
+ qq = "0"
else:
node_name_str0 = content[linenumber]
- if 'split_y_path.' in node_name_str0:
- node_name_str = node_name_str0.split('split_y_path')
- node_module_str = node_name_str[1].split('.')[1]
+ if "split_y_path." in node_name_str0:
+ node_name_str = node_name_str0.split("split_y_path")
+ node_module_str = node_name_str[1].split(".")[1]
node_module_int = int(node_module_str)
- if node_module_int ==0 :
+ if node_module_int == 0:
# 35 SADL_model.5.split_y_path.0.0 len 45056 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.00 [64] 0.14 [32] 0.95 [16] 5.76 [8]20.39 exceed 64 bit 10
- qq = '11'
- elif node_module_int ==1 :
+ qq = "11"
+ elif node_module_int == 1:
# 36 SADL_model.5.split_y_path.1.size
# 37 SADL_model.5.split_y_path.1.sub
# 38 SADL_model.5.split_y_path.1.size_1
@@ -101,8 +99,8 @@ def qfactor_gen(content):
# 44 SADL_model.5.split_y_path.1.conv3_31.0 len 6272 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.00 [64] 0.00 [32] 0.00 [16] 4.18 [8]19.89 exceed 16 bit 12
# 45 SADL_model.5.split_y_path.1.conv4_11.0 len 34496 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.00 [64] 0.00 [32] 0.00 [16] 0.59 [8]11.18 exceed 16 bit 12
# 46 SADL_model.5.split_y_path.1.add len 34496 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.00 [64] 0.12 [32] 1.57 [16] 9.91 [8]34.72 exceed 64 bit 10
- qq = '11' #or maybe 12. Also try prelu=11 or 10 and rest 12 or 11.
- elif node_module_int ==2 :
+ qq = "11" # or maybe 12. Also try prelu=11 or 10 and rest 12 or 11.
+ elif node_module_int == 2:
# 47 SADL_model.5.split_y_path.2.size
# 48 SADL_model.5.split_y_path.2.sub
# 49 SADL_model.5.split_y_path.2.size_1
@@ -114,177 +112,183 @@ def qfactor_gen(content):
# 55 SADL_model.5.split_y_path.2.conv3_31.0 len 4608 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.00 [64] 0.00 [32] 0.00 [16] 2.93 [8]32.10 exceed 16 bit 12
# 56 SADL_model.5.split_y_path.2.conv4_11.0 len 25344 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.00 [64] 0.00 [32] 0.00 [16] 0.88 [8]14.56 exceed 32 bit 11
# 57 SADL_model.5.split_y_path.2.add len 25344 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.00 [64] 0.15 [32] 1.82 [16]12.17 [8]36.14 exceed 64 bit 10
- qq = '11' #or maybe 12. Also try prelu=11 or 10 and rest 12 or 11
- elif node_module_int ==3 :
+ qq = "11" # or maybe 12. Also try prelu=11 or 10 and rest 12 or 11
+ elif node_module_int == 3:
# 58 SADL_model.5.split_y_path.3.prelu len 25344 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.00 [64] 0.13 [32] 1.71 [16]11.62 [8]34.97 exceed 64 bit 10
# 59 SADL_model.5.split_y_path.3.conv1_11.0 len 4608 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.00 [64] 0.07 [32] 3.41 [16] 8.99 [8]29.35 exceed 64 bit 10
# 60 SADL_model.5.split_y_path.3.conv2_13.0 len 4608 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.00 [64] 0.00 [32] 1.65 [16] 7.92 [8]26.85 exceed 32 bit 11
# 61 SADL_model.5.split_y_path.3.conv3_31.0 len 4608 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.00 [64] 0.00 [32] 0.07 [16] 5.47 [8]22.64 exceed 32 bit 11
# 62 SADL_model.5.split_y_path.3.conv4_11.0 len 25344 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.00 [64] 0.00 [32] 0.01 [16] 0.90 [8] 8.56 exceed 32 bit 11
# 63 SADL_model.5.split_y_path.3.add len 25344 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.00 [64] 0.13 [32] 3.11 [16]15.77 [8]39.54 exceed 64 bit 10
- qq = '11'
- elif node_module_int ==4 :
+ qq = "11"
+ elif node_module_int == 4:
# 64 SADL_model.5.split_y_path.4.prelu len 25344 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.00 [64] 0.16 [32] 3.20 [16]16.14 [8]40.24 exceed 64 bit 10
# 65 SADL_model.5.split_y_path.4.conv1_11.0 len 4608 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.00 [64] 2.54 [32] 4.82 [16]19.49 [8]46.19 exceed 64 bit 10
# 66 SADL_model.5.split_y_path.4.conv2_13.0 len 4608 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.00 [64] 0.59 [32] 4.51 [16]18.25 [8]45.58 exceed 64 bit 10
# 67 SADL_model.5.split_y_path.4.conv3_31.0 len 4608 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.00 [64] 0.00 [32] 5.14 [16]15.89 [8]44.87 exceed 32 bit 11
# 68 SADL_model.5.split_y_path.4.conv4_11.0 len 25344 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.00 [64] 0.00 [32] 0.19 [16] 5.64 [8]31.95 exceed 32 bit 11
# 69 SADL_model.5.split_y_path.4.add len 25344 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.00 [64] 0.11 [32] 3.71 [16]22.80 [8]52.47 exceed 64 bit 10
- qq = '10' #minskad
- elif node_module_int ==5 :
+ qq = "10" # minskad
+ elif node_module_int == 5:
# 70 SADL_model.5.split_y_path.5.prelu len 25344 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.00 [64] 0.01 [32] 2.28 [16]13.31 [8]26.94 exceed 64 bit 10
# 71 SADL_model.5.split_y_path.5.conv1_11.0 len 4608 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.00 [64] 0.07 [32] 5.71 [16]24.90 [8]44.69 exceed 64 bit 10
# 72 SADL_model.5.split_y_path.5.conv2_13.0 len 4608 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.00 [64] 0.00 [32] 8.38 [16]26.46 [8]47.67 exceed 32 bit 11
# 73 SADL_model.5.split_y_path.5.conv3_31.0 len 4608 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.00 [64] 0.00 [32]11.37 [16]26.76 [8]43.85 exceed 32 bit 11
# 74 SADL_model.5.split_y_path.5.conv4_11.0 len 25344 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.00 [64] 0.00 [32] 0.73 [16]13.54 [8]39.33 exceed 32 bit 11
# 75 SADL_model.5.split_y_path.5.add len 25344 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.00 [64] 0.23 [32] 6.98 [16]30.09 [8]56.88 exceed 64 bit 10
- qq = '11'
- elif node_module_int ==6 :
+ qq = "11"
+ elif node_module_int == 6:
# 76 SADL_model.5.split_y_path.6.prelu len 25344 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.00 [64] 0.06 [32] 4.49 [16]18.49 [8]31.93 exceed 64 bit 10
# 77 SADL_model.5.split_y_path.6.conv1_11.0 len 4608 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.00 [64] 0.00 [32] 5.10 [16]26.57 [8]49.19 exceed 32 bit 11
# 78 SADL_model.5.split_y_path.6.conv2_13.0 len 4608 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.00 [64] 0.04 [32] 5.77 [16]30.24 [8]51.14 exceed 64 bit 10
# 79 SADL_model.5.split_y_path.6.conv3_31.0 len 4608 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.00 [64] 0.04 [32] 7.58 [16]31.52 [8]49.34 exceed 64 bit 10
# 80 SADL_model.5.split_y_path.6.conv4_11.0 len 25344 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.00 [64] 0.00 [32] 1.51 [16]13.79 [8]42.65 exceed 32 bit 11
# 81 SADL_model.5.split_y_path.6.add len 25344 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.00 [64] 0.75 [32]12.88 [16]40.20 [8]65.95 exceed 64 bit 10
- qq = '11'
- elif node_module_int ==7 :
+ qq = "11"
+ elif node_module_int == 7:
# 82 SADL_model.5.split_y_path.7.prelu len 25344 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.00 [64] 0.29 [32] 7.79 [16]22.67 [8]35.08 exceed 64 bit 10
# 83 SADL_model.5.split_y_path.7.conv1_11.0 len 4608 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.00 [64] 3.56 [32]10.33 [16]29.52 [8]54.61 exceed 64 bit 10
# 84 SADL_model.5.split_y_path.7.conv2_13.0 len 4608 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.00 [64] 4.19 [32]10.92 [16]33.15 [8]57.76 exceed 64 bit 10
# 85 SADL_model.5.split_y_path.7.conv3_31.0 len 4608 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.00 [64] 3.34 [32]12.31 [16]35.84 [8]59.67 exceed 64 bit 10
# 86 SADL_model.5.split_y_path.7.conv4_11.0 len 25344 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.00 [64] 0.34 [32] 9.21 [16]35.11 [8]62.59 exceed 64 bit 10
# 87 SADL_model.5.split_y_path.7.add len 25344 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.02 [64] 3.73 [32]23.88 [16]51.23 [8]72.27 exceed 128 bit 9
- qq = '10'
- elif node_module_int ==8 :
+ qq = "10"
+ elif node_module_int == 8:
# 88 SADL_model.5.split_y_path.8.prelu len 25344 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.02 [64] 3.68 [32]23.66 [16]51.01 [8]72.15 exceed 128 bit 9
# 89 SADL_model.5.split_y_path.8.conv1_11.0 len 4608 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 3.08 [64] 5.30 [32]21.01 [16]45.37 [8]69.26 exceed 128 bit 9
# 90 SADL_model.5.split_y_path.8.conv2_13.0 len 4608 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 1.93 [64] 4.97 [32]20.16 [16]42.54 [8]63.71 exceed 128 bit 9
# 91 SADL_model.5.split_y_path.8.conv3_31.0 len 4608 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.11 [64] 4.95 [32]19.90 [16]39.64 [8]60.30 exceed 128 bit 9
# 92 SADL_model.5.split_y_path.8.conv4_11.0 len 25344 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.00 [64] 1.68 [32]15.37 [16]40.81 [8]64.95 exceed 64 bit 10
# 93 SADL_model.5.split_y_path.8.add len 25344 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.51 [64] 9.45 [32]34.87 [16]60.19 [8]78.80 exceed 128 bit 9
- qq = '9'
- elif node_module_int ==9 :
+ qq = "9"
+ elif node_module_int == 9:
# 94 SADL_model.5.split_y_path.9.prelu len 25344 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 2.65 [64]24.04 [32]52.62 [16]73.08 [8]85.48 exceed 128 bit 9
# 95 SADL_model.5.split_y_path.9.conv1_11.0 len 4608 [1024] 0.00 [512] 0.00 [256] 1.28 [128] 3.32 [64]12.03 [32]31.60 [16]54.83 [8]73.15 exceed 256 bit 8
# 96 SADL_model.5.split_y_path.9.conv2_13.0 len 4608 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 2.91 [64] 8.40 [32]25.70 [16]50.38 [8]69.76 exceed 128 bit 9
# 97 SADL_model.5.split_y_path.9.conv3_31.0 len 4608 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.22 [64] 5.06 [32]19.82 [16]42.74 [8]63.84 exceed 128 bit 9
# 98 SADL_model.5.split_y_path.9.conv4_11.0 len 25344 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.00 [64] 0.70 [32]11.60 [16]39.87 [8]66.43 exceed 64 bit 10
# 99 SADL_model.5.split_y_path.9.add len 25344 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.61 [64]12.96 [32]42.79 [16]69.78 [8]85.22 exceed 128 bit 9
- qq = '9' # minskad
- elif node_module_int ==10 :
+ qq = "9" # minskad
+ elif node_module_int == 10:
# 100 SADL_model.5.split_y_path.10.prelu len 25344 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.60 [64] 9.31 [32]24.48 [16]35.04 [8]51.07 exceed 128 bit 9
# 101 SADL_model.5.split_y_path.10.conv1_11.0 len 4608 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 3.00 [64] 5.80 [32]21.34 [16]45.39 [8]67.61 exceed 128 bit 9
# 102 SADL_model.5.split_y_path.10.conv2_13.0 len 4608 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 2.93 [64] 7.38 [32]25.31 [16]49.49 [8]70.44 exceed 128 bit 9
# 103 SADL_model.5.split_y_path.10.conv3_31.0 len 4608 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 2.76 [64] 9.68 [32]28.80 [16]53.83 [8]73.61 exceed 128 bit 9
# 104 SADL_model.5.split_y_path.10.conv4_11.0 len 25344 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.02 [64] 3.86 [32]23.43 [16]52.58 [8]73.89 exceed 128 bit 9
# 105 SADL_model.5.split_y_path.10.add len 25344 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 1.59 [64]22.03 [32]51.45 [16]73.37 [8]86.02 exceed 128 bit 9
- qq = '9' # minskad
- elif node_module_int ==11 :
+ qq = "9" # minskad
+ elif node_module_int == 11:
# 106 SADL_model.5.split_y_path.11.prelu len 25344 [1024] 0.00 [512] 0.00 [256] 0.07 [128] 7.94 [64]36.46 [32]63.57 [16]80.01 [8]89.45 exceed 256 bit 8
# 107 SADL_model.5.split_y_path.11.conv1_11.0 len 4608 [1024] 0.00 [512] 0.00 [256] 0.07 [128] 9.59 [64]24.01 [32]41.96 [16]60.69 [8]73.41 exceed 256 bit 8
# 108 SADL_model.5.split_y_path.11.conv2_13.0 len 4608 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 4.73 [64]21.79 [32]38.64 [16]58.02 [8]72.17 exceed 128 bit 9
# 109 SADL_model.5.split_y_path.11.conv3_31.0 len 4608 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 1.80 [64]18.84 [32]35.45 [16]55.35 [8]69.70 exceed 128 bit 9
# 110 SADL_model.5.split_y_path.11.conv4_11.0 len 25344 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.69 [64] 7.70 [32]29.72 [16]60.00 [8]80.87 exceed 128 bit 9
# 111 SADL_model.5.split_y_path.11.add len 25344 [1024] 0.00 [512] 0.00 [256] 0.25 [128] 6.65 [64]27.63 [32]52.13 [16]73.17 [8]86.38 exceed 256 bit 8
- qq = '9'
- elif node_module_int ==12 :
+ qq = "9"
+ elif node_module_int == 12:
# 112 SADL_model.5.split_y_path.12.prelu len 25344 [1024] 0.00 [512] 0.00 [256] 0.25 [128] 5.20 [64]23.71 [32]46.21 [16]68.53 [8]83.83 exceed 256 bit 8
# 113 SADL_model.5.split_y_path.12.conv1_11.0 len 4608 [1024] 0.00 [512] 0.00 [256] 2.87 [128] 5.93 [64]24.40 [32]49.92 [16]69.83 [8]81.64 exceed 256 bit 8
# 114 SADL_model.5.split_y_path.12.conv2_13.0 len 4608 [1024] 0.00 [512] 0.00 [256] 1.76 [128] 6.71 [64]26.52 [32]52.07 [16]70.68 [8]81.61 exceed 256 bit 8
# 115 SADL_model.5.split_y_path.12.conv3_31.0 len 4608 [1024] 0.00 [512] 0.00 [256] 0.41 [128] 8.20 [64]28.85 [32]53.29 [16]71.48 [8]81.75 exceed 256 bit 8
# 116 SADL_model.5.split_y_path.12.conv4_11.0 len 25344 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 2.49 [64]20.81 [32]48.03 [16]71.36 [8]85.46 exceed 128 bit 9
# 117 SADL_model.5.split_y_path.12.add len 25344 [1024] 0.00 [512] 0.00 [256] 0.16 [128]10.86 [64]39.23 [32]65.59 [16]82.41 [8]90.97 exceed 256 bit 8
- qq = '8' # minskad
- elif node_module_int ==13 :
+ qq = "8" # minskad
+ elif node_module_int == 13:
# 118 SADL_model.5.split_y_path.13.prelu len 25344 [1024] 0.00 [512] 0.00 [256] 0.14 [128] 6.01 [64]19.81 [32]33.71 [16]54.24 [8]73.76 exceed 256 bit 8
# 119 SADL_model.5.split_y_path.13.conv1_11.0 len 4608 [1024] 0.00 [512] 0.00 [256] 0.37 [128] 8.07 [64]17.04 [32]41.70 [16]66.25 [8]80.68 exceed 256 bit 8
# 120 SADL_model.5.split_y_path.13.conv2_13.0 len 4608 [1024] 0.00 [512] 0.00 [256] 0.24 [128] 9.29 [64]21.10 [32]46.73 [16]68.81 [8]81.16 exceed 256 bit 8
# 121 SADL_model.5.split_y_path.13.conv3_31.0 len 4608 [1024] 0.00 [512] 0.00 [256] 1.22 [128]10.40 [64]27.22 [32]52.25 [16]71.52 [8]82.79 exceed 256 bit 8
# 122 SADL_model.5.split_y_path.13.conv4_11.0 len 25344 [1024] 0.00 [512] 0.00 [256] 0.06 [128] 4.83 [64]24.14 [32]55.16 [16]77.26 [8]88.55 exceed 256 bit 8
# 123 SADL_model.5.split_y_path.13.add len 25344 [1024] 0.00 [512] 0.00 [256] 0.83 [128]18.66 [64]52.06 [32]75.17 [16]87.59 [8]93.60 exceed 256 bit 8
- qq = '9'
- elif node_module_int ==14 :
+ qq = "9"
+ elif node_module_int == 14:
# 124 SADL_model.5.split_y_path.14.prelu len 25344 [1024] 0.00 [512] 0.00 [256] 0.59 [128]16.81 [64]49.17 [32]73.73 [16]86.81 [8]93.29 exceed 256 bit 8
# 125 SADL_model.5.split_y_path.14.conv1_11.0 len 4608 [1024] 0.00 [512] 0.04 [256] 6.01 [128] 9.59 [64]20.79 [32]47.30 [16]70.57 [8]84.00 exceed 512 bit 7
# 126 SADL_model.5.split_y_path.14.conv2_13.0 len 4608 [1024] 0.00 [512] 0.00 [256] 6.19 [128]10.31 [64]24.66 [32]51.23 [16]73.37 [8]85.04 exceed 256 bit 8
# 127 SADL_model.5.split_y_path.14.conv3_31.0 len 4608 [1024] 0.00 [512] 0.00 [256] 6.03 [128]11.33 [64]29.02 [32]55.81 [16]75.52 [8]86.15 exceed 256 bit 8
# 128 SADL_model.5.split_y_path.14.conv4_11.0 len 25344 [1024] 0.00 [512] 0.00 [256] 0.49 [128] 5.56 [64]35.73 [32]70.33 [16]86.02 [8]93.41 exceed 256 bit 8
# 129 SADL_model.5.split_y_path.14.add len 25344 [1024] 0.00 [512] 0.00 [256] 1.30 [128]20.90 [64]54.87 [32]76.70 [16]88.52 [8]94.26 exceed 256 bit 8
- qq = '8'
- elif node_module_int ==15 :
+ qq = "8"
+ elif node_module_int == 15:
# 130 SADL_model.5.split_y_path.15.prelu len 25344 [1024] 0.00 [512] 0.62 [256]12.04 [128]41.69 [64]69.22 [32]83.67 [16]91.73 [8]95.89 exceed 512 bit 7
# 131 SADL_model.5.split_y_path.15.conv1_11.0 len 4608 [1024] 1.56 [512] 3.13 [256] 3.45 [128]10.61 [64]29.43 [32]55.20 [16]74.56 [8]86.46 exceed 1024 bit 6
# 132 SADL_model.5.split_y_path.15.conv2_13.0 len 4608 [1024] 0.15 [512] 3.13 [256] 3.13 [128] 6.51 [64]15.35 [32]36.31 [16]58.69 [8]76.97 exceed 1024 bit 6
# 133 SADL_model.5.split_y_path.15.conv3_31.0 len 4608 [1024] 0.00 [512] 2.37 [256] 3.13 [128] 5.19 [64]10.27 [32]20.12 [16]40.76 [8]61.41 exceed 512 bit 7
# 134 SADL_model.5.split_y_path.15.conv4_11.0 len 25344 [1024] 0.00 [512] 0.00 [256] 0.01 [128]14.34 [64]48.19 [32]76.03 [16]89.79 [8]95.26 exceed 256 bit 8
# 135 SADL_model.5.split_y_path.15.add len 25344 [1024] 0.00 [512] 0.00 [256] 0.19 [128] 2.15 [64]18.03 [32]49.42 [16]73.32 [8]86.87 exceed 256 bit 8
- qq = '7'
- elif node_module_int ==16 :
+ qq = "7"
+ elif node_module_int == 16:
# 136 SADL_model.5.split_y_path.16.0 len 4608 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.00 [64] 1.76 [32]16.04 [16]37.42 [8]61.82 exceed 64 bit 10
- qq = '10'
- elif node_module_int ==17 :
+ qq = "10"
+ elif node_module_int == 17:
# 137 SADL_model.5.split_y_path.17.0 len 4608 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.00 [64] 0.00 [32] 0.00 [16] 0.43 [8] 9.94 exceed 16 bit 12
- qq = '13'
- elif node_module_int ==18 :
+ qq = "13"
+ elif node_module_int == 18:
# 138 SADL_model.5.split_y_path.18.0 len 4608 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.00 [64] 0.00 [32] 0.00 [16] 0.00 [8] 0.00
- qq = '13'
- elif node_module_int ==19 :
+ qq = "13"
+ elif node_module_int == 19:
# 139 SADL_model.5.split_y_path.19.0 len 4608 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.00 [64] 0.00 [32] 0.00 [16] 0.00 [8] 0.00
- qq = '13'
- elif node_module_int ==20 :
+ qq = "13"
+ elif node_module_int == 20:
# 141 SADL_model.5.split_y_path.20.0 len 2304 [1024] 0.00 [512] 0.00 [256] 0.00 [128] 0.00 [64] 0.00 [32] 0.00 [16] 0.00 [8] 0.00
- qq = '13'
- else :
+ qq = "13"
+ else:
qq = q0
print(layer_id, node_name_str0[:-1], node_module_str, qq)
- elif 'branches' in node_name_str0 or \
- 'split_uv_path.5' in node_name_str0 or \
- 'split_uv_path.6' in node_name_str0 or \
- 'split_uv_path.7' in node_name_str0 or \
- 'split_uv_path.8' in node_name_str0 or \
- 'split_uv_path.9' in node_name_str0 :
+ elif (
+ "branches" in node_name_str0
+ or "split_uv_path.5" in node_name_str0
+ or "split_uv_path.6" in node_name_str0
+ or "split_uv_path.7" in node_name_str0
+ or "split_uv_path.8" in node_name_str0
+ or "split_uv_path.9" in node_name_str0
+ ):
qq = q0
- print(layer_id, node_name_str0[:-1], qq)
- elif 'split_uv_path.0' in node_name_str0 or \
- 'split_uv_path.1' in node_name_str0 :
- qq = '12'
- print(layer_id, node_name_str0[:-1], qq)
- elif 'split_uv' in node_name_str0 :
- qq = '11'
- print(layer_id, node_name_str0[:-1], qq)
- elif next_layer_name == 'PReLU':
+ print(layer_id, node_name_str0[:-1], qq)
+ elif (
+ "split_uv_path.0" in node_name_str0
+ or "split_uv_path.1" in node_name_str0
+ ):
+ qq = "12"
+ print(layer_id, node_name_str0[:-1], qq)
+ elif "split_uv" in node_name_str0:
+ qq = "11"
+ print(layer_id, node_name_str0[:-1], qq)
+ elif next_layer_name == "PReLU":
qq = q_prev
- if qq == '0':
+ if qq == "0":
qq = q_save_prev
- print('Overriding q_prev with', qq)
- print(layer_id,'PReLU', qq)
+ print("Overriding q_prev with", qq)
+ print(layer_id, "PReLU", qq)
else:
- print('Other: node name =', node_name_str0)
+ print("Other: node name =", node_name_str0)
# fuse layers
- qq = '12'
+ qq = "12"
print(layer_id, node_name_str0[:-1], qq)
- elif layer_name == 'Conv2D':
+ elif layer_name == "Conv2D":
q_save_prev = q_prev
- qq = '0'
- elif layer_name == 'Mul':
- qq = '0'
- elif layer_name == 'BiasAdd' or \
- layer_name == 'LeakyRelu' or \
- layer_name == 'Concat' or \
- layer_name == 'Shape' or \
- layer_name == 'Expand' or \
- layer_name == 'Add' or \
- layer_name == 'Transpose' or \
- layer_name == 'Reshape' or \
- layer_name == 'PReLU' or \
- layer_name == 'Slice' :
+ qq = "0"
+ elif layer_name == "Mul":
+ qq = "0"
+ elif (
+ layer_name == "BiasAdd"
+ or layer_name == "LeakyRelu"
+ or layer_name == "Concat"
+ or layer_name == "Shape"
+ or layer_name == "Expand"
+ or layer_name == "Add"
+ or layer_name == "Transpose"
+ or layer_name == "Reshape"
+ or layer_name == "PReLU"
+ or layer_name == "Slice"
+ ):
continue
else:
- print('Undecided layer name', layer_name)
+ print("Undecided layer name", layer_name)
continue
q_prev = qq
- output_q_str = output_q_str + layer_id + ' ' + qq + ' '
+ output_q_str = output_q_str + layer_id + " " + qq + " "
return output_q_str
@@ -299,20 +303,20 @@ if __name__ == "__main__":
except Exception:
quit("[ERROR] unable to open json config")
- main_dir = config['binaries']['sadl_path']
- sadl_build_dir = os.path.join(main_dir, 'sample_test')
- debug_cpp_dir = os.path.join(sadl_build_dir, 'debug_model')
- float_sadl_dir = config[f"stage{stage}"]['conversion']['full_path_filename']
- int_sadl_dir = config[f"stage{stage}"]['quantize']['full_path_filename']
-
- debug_log_dir = float_sadl_dir[:-5] + '_debug.txt'
+ main_dir = config["binaries"]["sadl_path"]
+ sadl_build_dir = os.path.join(main_dir, "sample_test")
+ debug_cpp_dir = os.path.join(sadl_build_dir, "debug_model")
+ float_sadl_dir = config[f"stage{stage}"]["conversion"]["full_path_filename"]
+ int_sadl_dir = config[f"stage{stage}"]["quantize"]["full_path_filename"]
+
+ debug_log_dir = float_sadl_dir[:-5] + "_debug.txt"
print(f"[INFO] SADL build dir {sadl_build_dir}")
print(f"[INFO] input model {float_sadl_dir}")
print(f"[INFO] debug_log_dir {debug_log_dir}")
# debug float SADL model and get the log
- cmd_debug = [debug_cpp_dir, float_sadl_dir, '>', debug_log_dir]
+ cmd_debug = [debug_cpp_dir, float_sadl_dir, ">", debug_log_dir]
cmd_debug = " ".join(cmd_debug)
print("Fore")
os.system(cmd_debug)
@@ -325,12 +329,17 @@ if __name__ == "__main__":
file.close()
# apply quantization
- naive_quantizer_cpp_dir = os.path.join(sadl_build_dir, 'naive_quantization')
- cmd_quantize = ["echo '", output_q_str , " ' | ",
- naive_quantizer_cpp_dir,
- float_sadl_dir,
- int_sadl_dir, ";"]
+ naive_quantizer_cpp_dir = os.path.join(sadl_build_dir, "naive_quantization")
+ cmd_quantize = [
+ "echo '",
+ output_q_str,
+ " ' | ",
+ naive_quantizer_cpp_dir,
+ float_sadl_dir,
+ int_sadl_dir,
+ ";",
+ ]
cmd_quantize = " ".join(cmd_quantize)
- print('Executing the following command:')
+ print("Executing the following command:")
print(cmd_quantize)
os.system(cmd_quantize)
diff --git a/training/training_scripts/NN_Filtering/VLOP3/model/model.py b/training/training_scripts/NN_Filtering/VLOP3/model/model.py
index 853f307e0..e4dee907a 100644
--- a/training/training_scripts/NN_Filtering/VLOP3/model/model.py
+++ b/training/training_scripts/NN_Filtering/VLOP3/model/model.py
@@ -122,7 +122,7 @@ class Conv(nn.Sequential):
super(Conv, self).__init__(*modules)
-
+
class MultiBranchModule(nn.Module):
"""A module representing multple, parallel branches. If the input is a list, each element in the list is fed into the corresponding branch,
otherwise the input is fed into every branch. The outputs of each branch are then merged."""
@@ -148,8 +148,24 @@ class NewResBlock_separate_prelu(nn.Sequential):
super().__init__()
self.prelu = nn.PReLU()
self.conv1_11 = Conv(C1, C, kernel_size=1, post_activation=None)
- self.conv2_13 = Conv(C, C, kernel_size=3, post_activation=None, is_separable=True, is_horizontal=True, hidden_separable_channels=C21)
- self.conv3_31 = Conv(C, C, kernel_size=3, post_activation=None, is_separable=True, is_horizontal=False, hidden_separable_channels=C21)
+ self.conv2_13 = Conv(
+ C,
+ C,
+ kernel_size=3,
+ post_activation=None,
+ is_separable=True,
+ is_horizontal=True,
+ hidden_separable_channels=C21,
+ )
+ self.conv3_31 = Conv(
+ C,
+ C,
+ kernel_size=3,
+ post_activation=None,
+ is_separable=True,
+ is_horizontal=False,
+ hidden_separable_channels=C21,
+ )
self.conv4_11 = Conv(C, C1, kernel_size=1, post_activation=None)
def forward(self, x: torch.Tensor) -> torch.Tensor:
@@ -167,8 +183,26 @@ class NewResBlock_separate_prelu_crop(nn.Sequential):
super().__init__()
self.prelu = nn.PReLU()
self.conv1_11 = Conv(C1, C, kernel_size=1, post_activation=None)
- self.conv2_13 = Conv(C, C, kernel_size=3, padding=(0,0), post_activation=None, is_separable=True, is_horizontal=True, hidden_separable_channels=C21)
- self.conv3_31 = Conv(C, C, kernel_size=3, padding=(0,0), post_activation=None, is_separable=True, is_horizontal=False, hidden_separable_channels=C21)
+ self.conv2_13 = Conv(
+ C,
+ C,
+ kernel_size=3,
+ padding=(0, 0),
+ post_activation=None,
+ is_separable=True,
+ is_horizontal=True,
+ hidden_separable_channels=C21,
+ )
+ self.conv3_31 = Conv(
+ C,
+ C,
+ kernel_size=3,
+ padding=(0, 0),
+ post_activation=None,
+ is_separable=True,
+ is_horizontal=False,
+ hidden_separable_channels=C21,
+ )
self.conv4_11 = Conv(C, C1, kernel_size=1, post_activation=None)
def forward(self, x: torch.Tensor) -> torch.Tensor:
@@ -180,6 +214,7 @@ class NewResBlock_separate_prelu_crop(nn.Sequential):
x5 = self.conv4_11(x4)
return x5 + temp
+
class Slice_only_y(nn.Sequential):
def __init__(self):
super().__init__()
@@ -194,8 +229,8 @@ class Slice_only_x(nn.Sequential):
def forward(self, x: torch.Tensor) -> torch.Tensor:
return x[:, :, :, 1:-1]
-
-
+
+
class SplitLumaChromaBlocks(nn.Sequential):
def __init__(
self,
@@ -218,26 +253,91 @@ class SplitLumaChromaBlocks(nn.Sequential):
self.Y_no_lower_extra_crop = N_Y[1]
self.split_y_path = nn.Sequential(
Conv(C_Y, C1_Y, kernel_size=1, post_activation=None),
- *[NewResBlock_separate_prelu(C_Y, C1_Y, C21_Y) for _ in range(N_Y[0]) ], # Start with full margin 36x36
- *[NewResBlock_separate_prelu_crop(C_Y, C1_Y, C21_Y) for _ in range(N_Y[1]) ], # Go down once to 34x34
- *[NewResBlock_separate_prelu(C_Y, C1_Y, C21_Y) for _ in range(N_Y[2]) ], # continue once at 34x34
- *[NewResBlock_separate_prelu_crop(C_Y, C1_Y, C21_Y) for _ in range(N_Y[3]) ], # go down once again to 32x32
- *[NewResBlock_separate_prelu(C_Y, C1_Y, C21_Y) for _ in range(N_Y[4]) ], # continue with 32x32 for the rest
+ *[
+ NewResBlock_separate_prelu(C_Y, C1_Y, C21_Y) for _ in range(N_Y[0])
+ ], # Start with full margin 36x36
+ *[
+ NewResBlock_separate_prelu_crop(C_Y, C1_Y, C21_Y) for _ in range(N_Y[1])
+ ], # Go down once to 34x34
+ *[
+ NewResBlock_separate_prelu(C_Y, C1_Y, C21_Y) for _ in range(N_Y[2])
+ ], # continue once at 34x34
+ *[
+ NewResBlock_separate_prelu_crop(C_Y, C1_Y, C21_Y) for _ in range(N_Y[3])
+ ], # go down once again to 32x32
+ *[
+ NewResBlock_separate_prelu(C_Y, C1_Y, C21_Y) for _ in range(N_Y[4])
+ ], # continue with 32x32 for the rest
Conv(C1_Y, C_Y, kernel_size=1, post_activation=None),
-
- # If we need extra cropping, we will do padding = (0,0) here.
- *[Conv(C_Y, C_Y, kernel_size=3, padding=(0,0), is_separable=True, is_horizontal=True, hidden_separable_channels=C21_Y, post_activation=None) for _ in range(self.Y_upper_extra_crop)],
- *[Conv(C_Y, C_Y, kernel_size=3, padding=(0,0), is_separable=True, is_horizontal=False, hidden_separable_channels=C21_Y, post_activation=None) for _ in range(self.Y_upper_extra_crop)],
+ # If we need extra cropping, we will do padding = (0,0) here.
+ *[
+ Conv(
+ C_Y,
+ C_Y,
+ kernel_size=3,
+ padding=(0, 0),
+ is_separable=True,
+ is_horizontal=True,
+ hidden_separable_channels=C21_Y,
+ post_activation=None,
+ )
+ for _ in range(self.Y_upper_extra_crop)
+ ],
+ *[
+ Conv(
+ C_Y,
+ C_Y,
+ kernel_size=3,
+ padding=(0, 0),
+ is_separable=True,
+ is_horizontal=False,
+ hidden_separable_channels=C21_Y,
+ post_activation=None,
+ )
+ for _ in range(self.Y_upper_extra_crop)
+ ],
# Otherwise, we will do padding = (1,1) (which is implicit if not specified).
- *[Conv(C_Y, C_Y, kernel_size=3, is_separable=True, is_horizontal=True, hidden_separable_channels=C21_Y, post_activation=None) for _ in range(self.Y_no_upper_extra_crop)],
- *[Conv(C_Y, C_Y, kernel_size=3, is_separable=True, is_horizontal=False, hidden_separable_channels=C21_Y, post_activation=None) for _ in range(self.Y_no_upper_extra_crop)],
-
+ *[
+ Conv(
+ C_Y,
+ C_Y,
+ kernel_size=3,
+ is_separable=True,
+ is_horizontal=True,
+ hidden_separable_channels=C21_Y,
+ post_activation=None,
+ )
+ for _ in range(self.Y_no_upper_extra_crop)
+ ],
+ *[
+ Conv(
+ C_Y,
+ C_Y,
+ kernel_size=3,
+ is_separable=True,
+ is_horizontal=False,
+ hidden_separable_channels=C21_Y,
+ post_activation=None,
+ )
+ for _ in range(self.Y_no_upper_extra_crop)
+ ],
Conv(C_Y, C_Y, kernel_size=1),
-
# If we need extra cropping, we will do padding = (0,0) here.
- *[Conv(C_Y, output_channels_y, kernel_size=3, padding=(0,0), post_activation=None) for _ in range(self.Y_lower_extra_crop)],
+ *[
+ Conv(
+ C_Y,
+ output_channels_y,
+ kernel_size=3,
+ padding=(0, 0),
+ post_activation=None,
+ )
+ for _ in range(self.Y_lower_extra_crop)
+ ],
# Otherwise, we will do padding = (1,1) (which is implicit if not specified).
- *[Conv(C_Y, output_channels_y, kernel_size=3, post_activation=None) for _ in range(self.Y_no_lower_extra_crop)],
+ *[
+ Conv(C_Y, output_channels_y, kernel_size=3, post_activation=None)
+ for _ in range(self.Y_no_lower_extra_crop)
+ ],
)
self.UV_upper_extra_crop = 1 - N_UV[3]
@@ -246,26 +346,93 @@ class SplitLumaChromaBlocks(nn.Sequential):
self.UV_no_lower_extra_crop = N_UV[1]
self.split_uv_path = nn.Sequential(
Conv(C_UV, C1_UV, kernel_size=1, post_activation=None),
- *[NewResBlock_separate_prelu(C_UV, C1_UV, C21_UV) for _ in range(N_UV[0])], # Start with full margin 36x36
- *[NewResBlock_separate_prelu_crop(C_UV, C1_UV, C21_UV) for _ in range(N_UV[1])], # Go down once to 34x34
- *[NewResBlock_separate_prelu(C_UV, C1_UV, C21_UV) for _ in range(N_UV[2])], # continue once at 34x34
- *[NewResBlock_separate_prelu_crop(C_UV, C1_UV, C21_UV) for _ in range(N_UV[3])], # go down once again to 32x32
- *[NewResBlock_separate_prelu(C_UV, C1_UV, C21_UV) for _ in range(N_UV[4])], # continue with 32x32 for the rest
+ *[
+ NewResBlock_separate_prelu(C_UV, C1_UV, C21_UV) for _ in range(N_UV[0])
+ ], # Start with full margin 36x36
+ *[
+ NewResBlock_separate_prelu_crop(C_UV, C1_UV, C21_UV)
+ for _ in range(N_UV[1])
+ ], # Go down once to 34x34
+ *[
+ NewResBlock_separate_prelu(C_UV, C1_UV, C21_UV) for _ in range(N_UV[2])
+ ], # continue once at 34x34
+ *[
+ NewResBlock_separate_prelu_crop(C_UV, C1_UV, C21_UV)
+ for _ in range(N_UV[3])
+ ], # go down once again to 32x32
+ *[
+ NewResBlock_separate_prelu(C_UV, C1_UV, C21_UV) for _ in range(N_UV[4])
+ ], # continue with 32x32 for the rest
Conv(C1_UV, C_UV, kernel_size=1, post_activation=None),
-
# If we need extra cropping, we will do padding = (0,0) here.
- *[Conv(C_UV, C_UV, kernel_size=3, padding=(0,0), is_separable=True, is_horizontal=True, hidden_separable_channels=C21_UV, post_activation=None) for _ in range(self.UV_upper_extra_crop)],
- *[Conv(C_UV, C_UV, kernel_size=3, padding=(0,0), is_separable=True, is_horizontal=False, hidden_separable_channels=C21_UV, post_activation=None) for _ in range(self.UV_upper_extra_crop)],
+ *[
+ Conv(
+ C_UV,
+ C_UV,
+ kernel_size=3,
+ padding=(0, 0),
+ is_separable=True,
+ is_horizontal=True,
+ hidden_separable_channels=C21_UV,
+ post_activation=None,
+ )
+ for _ in range(self.UV_upper_extra_crop)
+ ],
+ *[
+ Conv(
+ C_UV,
+ C_UV,
+ kernel_size=3,
+ padding=(0, 0),
+ is_separable=True,
+ is_horizontal=False,
+ hidden_separable_channels=C21_UV,
+ post_activation=None,
+ )
+ for _ in range(self.UV_upper_extra_crop)
+ ],
# Otherwise, we will do padding = (1,1) (which is implicit if not specified).
- *[Conv(C_UV, C_UV, kernel_size=3, is_separable=True, is_horizontal=True, hidden_separable_channels=C21_UV, post_activation=None) for _ in range(self.UV_no_upper_extra_crop)],
- *[Conv(C_UV, C_UV, kernel_size=3, is_separable=True, is_horizontal=False, hidden_separable_channels=C21_UV, post_activation=None) for _ in range(self.UV_no_upper_extra_crop)],
-
+ *[
+ Conv(
+ C_UV,
+ C_UV,
+ kernel_size=3,
+ is_separable=True,
+ is_horizontal=True,
+ hidden_separable_channels=C21_UV,
+ post_activation=None,
+ )
+ for _ in range(self.UV_no_upper_extra_crop)
+ ],
+ *[
+ Conv(
+ C_UV,
+ C_UV,
+ kernel_size=3,
+ is_separable=True,
+ is_horizontal=False,
+ hidden_separable_channels=C21_UV,
+ post_activation=None,
+ )
+ for _ in range(self.UV_no_upper_extra_crop)
+ ],
Conv(C_UV, C_UV, kernel_size=1),
-
# If we need extra cropping, we will do padding = (0,0) here.
- *[Conv(C_UV, output_channels_uv, padding=(0,0), kernel_size=3, post_activation=None) for _ in range(self.UV_lower_extra_crop)],
+ *[
+ Conv(
+ C_UV,
+ output_channels_uv,
+ padding=(0, 0),
+ kernel_size=3,
+ post_activation=None,
+ )
+ for _ in range(self.UV_lower_extra_crop)
+ ],
# Otherwise, we will do padding = (1,1) (which is implicit if not specified).
- *[Conv(C_UV, output_channels_uv, kernel_size=3, post_activation=None) for _ in range(self.UV_no_lower_extra_crop)],
+ *[
+ Conv(C_UV, output_channels_uv, kernel_size=3, post_activation=None)
+ for _ in range(self.UV_no_lower_extra_crop)
+ ],
)
self.Cy = C_Y
@@ -324,10 +491,39 @@ class SADLNet(nn.Sequential):
]
),
Conv(sum(self.input_features), D6, kernel_size=1),
- Conv(D6, D6, kernel_size=3, stride=2, post_activation=None, is_separable=True, is_horizontal=True, hidden_separable_channels=D6),
- Conv(D6, D6, kernel_size=3, stride=2, post_activation=None, is_separable=True, is_horizontal=False, hidden_separable_channels=D6),
+ Conv(
+ D6,
+ D6,
+ kernel_size=3,
+ stride=2,
+ post_activation=None,
+ is_separable=True,
+ is_horizontal=True,
+ hidden_separable_channels=D6,
+ ),
+ Conv(
+ D6,
+ D6,
+ kernel_size=3,
+ stride=2,
+ post_activation=None,
+ is_separable=True,
+ is_horizontal=False,
+ hidden_separable_channels=D6,
+ ),
Conv(D6, C_Y + C_UV, kernel_size=1),
- SplitLumaChromaBlocks(N_Y, N_UV, C_Y, C_UV, C1_Y, C1_UV, C21_Y, C21_UV, output_channels_y, output_channels_uv),
+ SplitLumaChromaBlocks(
+ N_Y,
+ N_UV,
+ C_Y,
+ C_UV,
+ C1_Y,
+ C1_UV,
+ C21_Y,
+ C21_UV,
+ output_channels_y,
+ output_channels_uv,
+ ),
)
def get_example_inputs(
@@ -399,7 +595,14 @@ class Net(nn.Module):
input_kernels
), "[ERROR] input size and kernels size not equal"
self.input_channels = input_channels
- sizes = [dct_ch + dct_ch // 2, dct_ch + dct_ch // 2, dct_ch + dct_ch // 2, 1, 1, dct_ch]
+ sizes = [
+ dct_ch + dct_ch // 2,
+ dct_ch + dct_ch // 2,
+ dct_ch + dct_ch // 2,
+ 1,
+ 1,
+ dct_ch,
+ ]
self.SADL_model = SADLNet(
sizes,
input_kernels,
@@ -418,7 +621,7 @@ class Net(nn.Module):
C21_Y,
C21_UV,
4 * dct_ch,
- 2 * dct_ch
+ 2 * dct_ch,
)
self.chroma_upsampler = nn.Upsample(scale_factor=2, mode="nearest")
self.dct_ch = dct_ch
diff --git a/training/training_scripts/NN_Filtering/VLOP3/quantize/quantize.py b/training/training_scripts/NN_Filtering/VLOP3/quantize/quantize.py
index d2782acae..835214607 100755
--- a/training/training_scripts/NN_Filtering/VLOP3/quantize/quantize.py
+++ b/training/training_scripts/NN_Filtering/VLOP3/quantize/quantize.py
@@ -46,9 +46,7 @@ import json
def parse_arguments():
parser = argparse.ArgumentParser()
- parser.add_argument(
- "--config", "-c", type=str, help="config file"
- )
+ parser.add_argument("--config", "-c", type=str, help="config file")
parser.add_argument(
"--stage",
type=int,
@@ -68,47 +66,47 @@ def parse_arguments():
def qfactor_gen(content):
- output_q_str = ''
+ output_q_str = ""
linenumber = 0
- q0 = '13'
+ q0 = "13"
q_prev = q0
for line in content:
linenumber += 1
- if line[:11] == '[INFO] id: ':
- layer_str = line.split(' ')
+ if line[:11] == "[INFO] id: ":
+ layer_str = line.split(" ")
layer_id = layer_str[2]
layer_name = layer_str[-1][:-1]
- if layer_name == 'Placeholder':
+ if layer_name == "Placeholder":
qq = q0
- elif layer_name == 'Const':
+ elif layer_name == "Const":
next_layer = content[linenumber + 5]
- layer_str2 = next_layer.split(' ')
+ layer_str2 = next_layer.split(" ")
next_layer_name = layer_str2[-1][:-1]
- if next_layer_name == 'Concat':
- qq = '0'
+ if next_layer_name == "Concat":
+ qq = "0"
else:
node_name_str0 = content[linenumber]
- if 'split_y_path.' in node_name_str0:
- node_name_str = node_name_str0.split('split_y_path')
- node_module_str = node_name_str[1].split('.')[1]
+ if "split_y_path." in node_name_str0:
+ node_name_str = node_name_str0.split("split_y_path")
+ node_module_str = node_name_str[1].split(".")[1]
node_module_int = int(node_module_str)
- if node_module_int <= 2 :
- qq = '11'
+ if node_module_int <= 2:
+ qq = "11"
elif node_module_int > 2 and node_module_int <= 4:
- qq = '10'
- #elif node_module_int > 4 and node_module_int <= 6:
+ qq = "10"
+ # elif node_module_int > 4 and node_module_int <= 6:
# qq = '9'
- #elif node_module_int == 7:
+ # elif node_module_int == 7:
# qq = '11'
elif node_module_int > 4 and node_module_int <= 7:
- qq = '9'
+ qq = "9"
elif node_module_int == 8:
- qq = '11'
- else :
+ qq = "11"
+ else:
qq = q0
print(layer_id, node_name_str0[:-1], node_module_str, qq)
- #elif 'branches' in node_name_str0 or \
+ # elif 'branches' in node_name_str0 or \
# 'split_uv_path.2' in node_name_str0 or \
# 'split_uv_path.3' in node_name_str0 or \
# 'split_uv_path.4' in node_name_str0 or \
@@ -116,56 +114,60 @@ def qfactor_gen(content):
# 'split_uv_path.6' in node_name_str0 :
# qq = q0
# print(layer_id, node_name_str0[:-1], qq)
- #elif 'split_uv_path.0' in node_name_str0:
+ # elif 'split_uv_path.0' in node_name_str0:
# qq = '12'
# print(layer_id, node_name_str0[:-1], qq)
- #elif 'split_uv' in node_name_str0 :
+ # elif 'split_uv' in node_name_str0 :
# qq = '11'
# print(layer_id, node_name_str0[:-1], qq)
- elif 'branches' in node_name_str0 or \
- 'split_uv_path.0' in node_name_str0 or \
- 'split_uv_path.1' in node_name_str0 or \
- 'split_uv_path.2' in node_name_str0 or \
- 'split_uv_path.3' in node_name_str0 or \
- 'split_uv_path.4' in node_name_str0 or \
- 'split_uv_path.5' in node_name_str0 or \
- 'split_uv_path.6' in node_name_str0 or \
- 'split_uv_path.7' in node_name_str0 or \
- 'split_uv_path.8' in node_name_str0 or \
- 'split_uv_path.9' in node_name_str0 or \
- 'split_uv_path.10' in node_name_str0 :
+ elif (
+ "branches" in node_name_str0
+ or "split_uv_path.0" in node_name_str0
+ or "split_uv_path.1" in node_name_str0
+ or "split_uv_path.2" in node_name_str0
+ or "split_uv_path.3" in node_name_str0
+ or "split_uv_path.4" in node_name_str0
+ or "split_uv_path.5" in node_name_str0
+ or "split_uv_path.6" in node_name_str0
+ or "split_uv_path.7" in node_name_str0
+ or "split_uv_path.8" in node_name_str0
+ or "split_uv_path.9" in node_name_str0
+ or "split_uv_path.10" in node_name_str0
+ ):
qq = q0
print(layer_id, node_name_str0[:-1], qq)
- elif next_layer_name == 'PReLU':
+ elif next_layer_name == "PReLU":
qq = q_prev
- if qq == '0':
- print('PReLU quantization is wrong. ', qq)
- print(layer_id, 'PReLU', qq)
+ if qq == "0":
+ print("PReLU quantization is wrong. ", qq)
+ print(layer_id, "PReLU", qq)
else:
# fuse layers
- qq = '12'
+ qq = "12"
print(layer_id, node_name_str0[:-1], qq)
- elif layer_name == 'Conv2D':
- qq = '0'
- elif layer_name == 'Mul':
- qq = '0'
- elif layer_name == 'BiasAdd' or \
- layer_name == 'LeakyRelu' or \
- layer_name == 'Concat' or \
- layer_name == 'Shape' or \
- layer_name == 'Expand' or \
- layer_name == 'Add' or \
- layer_name == 'Transpose' or \
- layer_name == 'Reshape' or \
- layer_name == 'PReLU' or \
- layer_name == 'Slice' :
+ elif layer_name == "Conv2D":
+ qq = "0"
+ elif layer_name == "Mul":
+ qq = "0"
+ elif (
+ layer_name == "BiasAdd"
+ or layer_name == "LeakyRelu"
+ or layer_name == "Concat"
+ or layer_name == "Shape"
+ or layer_name == "Expand"
+ or layer_name == "Add"
+ or layer_name == "Transpose"
+ or layer_name == "Reshape"
+ or layer_name == "PReLU"
+ or layer_name == "Slice"
+ ):
continue
else:
- print('Undecided layer name', layer_name)
+ print("Undecided layer name", layer_name)
continue
q_prev = qq
- output_q_str = output_q_str + layer_id + ' ' + qq + ' '
+ output_q_str = output_q_str + layer_id + " " + qq + " "
return output_q_str
@@ -181,24 +183,24 @@ if __name__ == "__main__":
except Exception:
quit("[ERROR] unable to open json config")
- main_dir = config['binaries']['sadl_path']
- sadl_build_dir = os.path.join(main_dir, 'sample_test')
- debug_cpp_dir = os.path.join(sadl_build_dir, 'debug_model')
+ main_dir = config["binaries"]["sadl_path"]
+ sadl_build_dir = os.path.join(main_dir, "sample_test")
+ debug_cpp_dir = os.path.join(sadl_build_dir, "debug_model")
if args["float_model"]:
float_sadl_dir = args["float_model"]
else:
- float_sadl_dir = config[f"stage{stage}"]['conversion']['full_path_filename']
+ float_sadl_dir = config[f"stage{stage}"]["conversion"]["full_path_filename"]
if args["int_model"]:
int_sadl_dir = args["int_model"]
else:
- int_sadl_dir = config[f"stage{stage}"]['quantize']['full_path_filename']
- debug_log_dir = float_sadl_dir[:-5] + '_debug.txt'
+ int_sadl_dir = config[f"stage{stage}"]["quantize"]["full_path_filename"]
+ debug_log_dir = float_sadl_dir[:-5] + "_debug.txt"
print(f"[INFO] SADL build dir {sadl_build_dir}")
print(f"[INFO] input model {float_sadl_dir}")
# debug float SADL model and get the log
- cmd_debug = [debug_cpp_dir, float_sadl_dir, '>', debug_log_dir]
+ cmd_debug = [debug_cpp_dir, float_sadl_dir, ">", debug_log_dir]
cmd_debug = " ".join(cmd_debug)
os.system(cmd_debug)
@@ -209,10 +211,15 @@ if __name__ == "__main__":
file.close()
# apply quantization
- naive_quantizer_cpp_dir = os.path.join(sadl_build_dir, 'naive_quantization')
- cmd_quantize = ["echo '", output_q_str , " ' | ",
- naive_quantizer_cpp_dir,
- float_sadl_dir,
- int_sadl_dir, ";"]
+ naive_quantizer_cpp_dir = os.path.join(sadl_build_dir, "naive_quantization")
+ cmd_quantize = [
+ "echo '",
+ output_q_str,
+ " ' | ",
+ naive_quantizer_cpp_dir,
+ float_sadl_dir,
+ int_sadl_dir,
+ ";",
+ ]
cmd_quantize = " ".join(cmd_quantize)
os.system(cmd_quantize)
--
GitLab