From eebf94d99507d30fc81d55a0d4e2d1d5fde690af Mon Sep 17 00:00:00 2001
From: Birdylx <29754889+Birdylx@users.noreply.github.com>
Date: Fri, 19 Aug 2022 14:41:57 +0800
Subject: [PATCH] [TIPC] add tipc benchmark for msvsr (#672)
* add tipc benchmark for msvsr
* update tipc readme img
---
ppgan/engine/trainer.py | 25 ++++++++-
ppgan/models/msvsr_model.py | 43 ++++++++++++++
ppgan/utils/options.py | 34 +++++++----
ppgan/utils/setup.py | 10 +++-
test_tipc/{readme.md => README.md} | 7 +--
test_tipc/benchmark_train.sh | 17 +++---
.../configs/msvsr/train_amp_infer_python.txt | 53 ++++++++++++++++++
.../configs/msvsr/train_infer_python.txt | 14 +++--
test_tipc/docs/benchmark_train.md | 16 +++---
test_tipc/docs/test.png | Bin 0 -> 229200 bytes
test_tipc/prepare.sh | 7 ++-
test_tipc/test_train_inference_python.sh | 51 +++++++++--------
12 files changed, 211 insertions(+), 66 deletions(-)
rename test_tipc/{readme.md => README.md} (93%)
create mode 100644 test_tipc/configs/msvsr/train_amp_infer_python.txt
create mode 100644 test_tipc/docs/test.png
diff --git a/ppgan/engine/trainer.py b/ppgan/engine/trainer.py
index fa2dfa0..eed6f89 100755
--- a/ppgan/engine/trainer.py
+++ b/ppgan/engine/trainer.py
@@ -32,6 +32,7 @@ from ..utils.profiler import add_profiler_step
class IterLoader:
+
def __init__(self, dataloader):
self._dataloader = dataloader
self.iter_loader = iter(self._dataloader)
@@ -79,6 +80,7 @@ class Trainer:
# | ||
# save checkpoint (model.nets) \/
"""
+
def __init__(self, cfg):
# base config
self.logger = logging.getLogger(__name__)
@@ -181,6 +183,22 @@ class Trainer:
iter_loader = IterLoader(self.train_dataloader)
+ # use amp
+ if self.cfg.amp:
+ self.logger.info('use AMP to train. AMP level = {}'.format(
+ self.cfg.amp_level))
+ assert self.cfg.model.name == 'MultiStageVSRModel', "AMP only support msvsr model"
+ scaler = paddle.amp.GradScaler(init_loss_scaling=1024)
+ # need to decorate model and optim if amp_level == 'O2'
+ if self.cfg.amp_level == 'O2':
+ # msvsr has only one generator and one optimizer
+ self.model.nets['generator'], self.optimizers[
+ 'optim'] = paddle.amp.decorate(
+ models=self.model.nets['generator'],
+ optimizers=self.optimizers['optim'],
+ level='O2',
+ save_dtype='float32')
+
# set model.is_train = True
self.model.setup_train_mode(is_train=True)
while self.current_iter < (self.total_iters + 1):
@@ -195,7 +213,12 @@ class Trainer:
# unpack data from dataset and apply preprocessing
# data input should be dict
self.model.setup_input(data)
- self.model.train_iter(self.optimizers)
+
+ if self.cfg.amp:
+ self.model.train_iter_amp(self.optimizers, scaler,
+ self.cfg.amp_level) # amp train
+ else:
+ self.model.train_iter(self.optimizers) # norm train
batch_cost_averager.record(
time.time() - step_start_time,
diff --git a/ppgan/models/msvsr_model.py b/ppgan/models/msvsr_model.py
index 3ee6fbd..4642dde 100644
--- a/ppgan/models/msvsr_model.py
+++ b/ppgan/models/msvsr_model.py
@@ -30,6 +30,7 @@ class MultiStageVSRModel(BaseSRModel):
Paper:
PP-MSVSR: Multi-Stage Video Super-Resolution, 2021
"""
+
def __init__(self, generator, fix_iter, pixel_criterion=None):
"""Initialize the PP-MSVSR class.
@@ -96,6 +97,48 @@ class MultiStageVSRModel(BaseSRModel):
self.current_iter += 1
+ # amp train with brute force implementation, maybe decorator can simplify this
+ def train_iter_amp(self, optims=None, scaler=None, amp_level='O1'):
+ optims['optim'].clear_grad()
+ if self.fix_iter:
+ if self.current_iter == 1:
+ print('Train MSVSR with fixed spynet for', self.fix_iter,
+ 'iters.')
+ for name, param in self.nets['generator'].named_parameters():
+ if 'spynet' in name:
+ param.trainable = False
+ elif self.current_iter >= self.fix_iter + 1 and self.flag:
+ print('Train all the parameters.')
+ for name, param in self.nets['generator'].named_parameters():
+ param.trainable = True
+ if 'spynet' in name:
+ param.optimize_attr['learning_rate'] = 0.25
+ self.flag = False
+ for net in self.nets.values():
+ net.find_unused_parameters = False
+
+ # put loss computation in amp context
+ with paddle.amp.auto_cast(enable=True, level=amp_level):
+ output = self.nets['generator'](self.lq)
+ if isinstance(output, (list, tuple)):
+ out_stage2, output = output
+ loss_pix_stage2 = self.pixel_criterion(out_stage2, self.gt)
+ self.losses['loss_pix_stage2'] = loss_pix_stage2
+ self.visual_items['output'] = output[:, 0, :, :, :]
+ # pixel loss
+ loss_pix = self.pixel_criterion(output, self.gt)
+ self.losses['loss_pix'] = loss_pix
+
+ self.loss = sum(_value for _key, _value in self.losses.items()
+ if 'loss_pix' in _key)
+ scaled_loss = scaler.scale(self.loss)
+ self.losses['loss'] = scaled_loss
+
+ scaled_loss.backward()
+ scaler.minimize(optims['optim'], scaled_loss)
+
+ self.current_iter += 1
+
def test_iter(self, metrics=None):
self.gt = self.gt.cpu()
self.nets['generator'].eval()
diff --git a/ppgan/utils/options.py b/ppgan/utils/options.py
index 2a0abc8..d371d63 100644
--- a/ppgan/utils/options.py
+++ b/ppgan/utils/options.py
@@ -45,9 +45,9 @@ def parse_args():
default=False,
help='skip validation during training')
# config options
- parser.add_argument("-o",
- "--opt",
- nargs='+',
+ parser.add_argument("-o",
+ "--opt",
+ nargs='+',
help="set configuration options")
#for inference
@@ -60,19 +60,31 @@ def parse_args():
help="path to reference images")
parser.add_argument("--model_path", default=None, help="model for loading")
- # for profiler
- parser.add_argument('-p',
- '--profiler_options',
- type=str,
- default=None,
- help='The option of profiler, which should be in format \"key1=value1;key2=value2;key3=value3\".'
+ # for profiler
+ parser.add_argument(
+ '-p',
+ '--profiler_options',
+ type=str,
+ default=None,
+ help=
+ 'The option of profiler, which should be in format \"key1=value1;key2=value2;key3=value3\".'
)
# fix random numbers by setting seed
parser.add_argument('--seed',
type=int,
default=None,
- help='fix random numbers by setting seed\".'
- )
+ help='fix random numbers by setting seed\".')
+
+ # add for amp training
+ parser.add_argument('--amp',
+ action='store_true',
+ default=False,
+ help='whether to enable amp training')
+ parser.add_argument('--amp_level',
+ type=str,
+ default='O1',
+ choices=['O1', 'O2'],
+ help='level of amp training; O2 represent pure fp16')
args = parser.parse_args()
return args
diff --git a/ppgan/utils/setup.py b/ppgan/utils/setup.py
index 78df109..5531076 100644
--- a/ppgan/utils/setup.py
+++ b/ppgan/utils/setup.py
@@ -19,6 +19,7 @@ import numpy as np
import random
from .logger import setup_logger
+
def setup(args, cfg):
if args.evaluate_only:
cfg.is_train = False
@@ -44,10 +45,13 @@ def setup(args, cfg):
paddle.set_device('gpu')
else:
paddle.set_device('cpu')
-
+
if args.seed:
paddle.seed(args.seed)
random.seed(args.seed)
- np.random.seed(args.seed)
+ np.random.seed(args.seed)
paddle.framework.random._manual_program_seed(args.seed)
-
+
+ # add amp and amp_level args into cfg
+ cfg['amp'] = args.amp
+ cfg['amp_level'] = args.amp_level
diff --git a/test_tipc/readme.md b/test_tipc/README.md
similarity index 93%
rename from test_tipc/readme.md
rename to test_tipc/README.md
index 3d4778c..4ffe955 100644
--- a/test_tipc/readme.md
+++ b/test_tipc/README.md
@@ -57,9 +57,8 @@ test_tipc/
### 测试流程
使用本工具，可以测试不同功能的支持情况，以及预测结果是否对齐，测试流程如下：
-
-
-
u?JLrIl~GD${BE5`cJ)Ka}>AMguU0i6W*ZR_NPbBpeOb2CTVLZ;Pfz
z=#h|!@(&Wj#hV{ESUrK}@0ZUj(BdNuN5qK}7Z9POpMxCr8#gkqkRvhJdZ2j*hl!-1
zu=@o(ea>LBwfn788s5YEB;zW{@L1%sqdt=Ta-lI%orHb4!H$dynf0b7!88Kz
zJ5hLRIA9yC3YUFx%MnQ34T*NV&|WoZhpOew;9;v0;!6JIB!~gzGqEFM{XlCU?V3K3
z3R&xbOW%miX_aAPffvgR)tDHI@VCREy4AJ7y_6%qTuKy0@zKytV&eC9O-`Q)sbue-
z)6O*!5r|50`yxR=U5yY=C{u;Q!!n*&K-t>Q&)kAEQ!
z|0aBDnqbkWTYM}wrKe~0nGjw`jDPt6AL5r5?>|Zkag&d#I
z@ntPP)LWV?`kD1u=JqRZdp?l)o?msFv7NxJASh`X2ud6=P9e^Wl%`bwc=+SJmvrmG
zOX6QpNM+cr$ljgR%R&`2&w={Y?OM|FJ%3^i3AJJLi;2Cg<)FLcuU)kx0*(&`j$)3f
zQ&5ke+S81NxsxV-lE{;n6VFIj273oCmO@{K{=Fd=^`a=;9V*3PUB1EeCz%1Hb!Dpi
z6sm>gO+mG%Nxs*=BxzZ|aQGl~4ZA-&Kb1IOg);eh!>?d$*!-u|%t}{lrXq36@r2H=
zlo>E`G#^-|DFL{m#ke+ZBsgIw;)ULsP_Ui!wE0QeP+M*7B&HyTae^x4
z;9$(9Mo~T}q-K)ak&r`p@6H_Id`mgSL`E_6vOeR|`5$66%6YJI^=$^Lk`Z37cNlZV
z3$3}BYN_WEszWmfPz8ZgVvy|DvmpnDot;cJ0wgK$f@%^oQ6}%(n+rO8pn~b9>x~tv
zEa6MN&fS>5PB_s@ZzgVkL6Ss}z@6~(6|p7$Dya@kdN$)mgH28SLR?J`3H$e#9HbqU
z2|_dkwp|rr;)I#sQ8+~Ev-L-estl@spnt<7n_9^QZVHl|4by1z80hql%vQkdbz
zsUeZfRu&`c3t0VEiYtZpSof`%6o2>y*60KCklL;7S>WK}C9cA|gD!UAty}1ydwpV7
zBKZ$|eoLwbDdg(Q`A#?(@sGWm^k%JR<*gG~Jtq3dVFS`G5(HIC_#nFLU)Zd=YO2v^
z`m2^|_$pewy9S)}bBWEXR?G2}x?Ga)h=lK&%;Ca{!|UgZ>Ks#Z9&bPmasz>c;Pk0b?%
zrJ-j72kl1dT~O-?Cf9TT1VTE%+{K
zU+}E1a(i0DJX$Y@2pGCN7-sc3%K12I0le8Z8k?a7|{T)F@-(zNi78>@`v$3ARlT5hN
zyg#7q1Zce8vOYr9zSnJPzO6S_zH+PhGxFYow08z`uVno+rxu+gmd?9twA=0>g+;7k
zyO?D<*`woGiG6jadc`DMtrn7MY8Cg?p#7z|?OHy1^g=C)x`#zm-0af|s{8v
zwz}%AG^u<|L~}IDnbalsD%!X-gLM@x!o`*lC)?u;W}J~ztbS%$?sXCcfqG)#TBq_O
zuTrW*2`1iIQ@am+?0|^&IQ=z_vA5R118Sdgn1^CBA+#3*T=>B0iofrJx!An-p{=vT
zkU{wbScAV`!RkpJ!MK6=ZR*Yi415f+IfKNJJ9C?bFE_+6eP+u<4w?QOM$0Ln=idT;
zW~shIqut_ez}2kVQYAriQ@hLZ#9Y>0-j7!n_G6tMV8(`Q-e9$t7BP+gf&@TL`t`ZJ
zqJY0n@4(;0AZe^C137aE@)MbdpEMDdnms)0?N0LF6cMg~zx&?KU}xlh=)Z{>-Gf}2
zKJ792Ja9?ey{w4;tzDPN{lvD8mskbC3X^U24=2k7Bqt4p{bPwdtA9A-x1D!_SP;u=
z%Vfnq5HBCZWThQ@x4|8ZK6WZ#I!;^`dCk$%)}R?<$2U;q_Ufm_YyDwJqd`HwVlinW
zz5qCnfk^qZ62_0K^JlV#`E#-b+|bXo-J6ecGrlaS$$
z9|)_7Zz6hQHVQgH?4G^_j2UNWV{C|xrCiyIqd1=K(wQEKiI~ERhT$%MUTydvwYQMI
z!mh!1(#B|n^%B0?U_emi=Z&0SR%UX~;k?T?EI0WrWk3-zTQ(KHTQw4)P9(&r*7gQX
zixi&SHx5}5AVi|4=cc*x{?c|o{SAnXJ4xg^<0KJfpe?uV_0W23dic}I^ujT2Sf?vB
zH%yeujUDyv;TrM7b)+c2s_5_Nba;$u6r0SIiuRw3%y3kw0G8u_9dE0IAY2H*
zZ&-q;ahInnAZsN1*wUsba$we&_tuUX$#cV1U|$5L*OIym8tz1^J-cXN^?gnVU4l<7
zKi4&_X%H~D~t>;`KMkX}u}7{_m;L8^ItZu_Hc6Q|2J;RBZi{sx1l5H)6N
zcg;#K7$WkD9>@jWI5j}8uVVtD!&M@#4ntjzOTJ(DX)r=?X9WH8)omzl$!pl@0Jc)*cf|BHucJ8{5a-5)=*3mc}J45Gda#1)Y5$1vr5T<&Jt
z4vo)n^xwWo5TooyAX1V$)RQW;#9;AtXtdjtCN|Vpf_Z#`^w1PW9b=Av{N=$ZVkiHF
zjN#nf^M*BZ7eWuz8rZaI8Gr!?44y%!7k-8!`##pUy?>^_WuH0)t0-Ks8f8Mm1Ff)=
zD#4k%DixDW8#Bm!F%`StU#ydn(;>B&YmkVq%eVo9Ble-MzGmlsu@2KZyc>$DIcw%n
zJg;Wr^7^RWfQFFc#BG_Rn%p|yI5>wq(*#{UW9Vw%3~Ak7{ueKhcomzj=iFfwWKT`X
z1vKJOBl>N%bzwXFe+L6qO+SJJ;bXgRu8;U_(be`xfp#{}%`NQ0t!)kaexqC$Y1Ab$|5lIGZ44zCg!M~Sj7PUp#TA|+-fD;g=r?~BQ2#*F9h
ziBuiPpuKl^+)&SU8O6l?SNAkSk0rw5S{D;ED#*G%)(`18^ZkU$OAVEvlM&)Kdw0@E
z53$BC&_zhk3;9m~!(E82GDx$Dg;5EOuV%!#To+?8-eJ7pvYjp3Msgbzxy8PzJ>G~M
z_5X539JHrwPD`#sjj=6%&JI6GpmX9+{p6l=SA=tZ4G+5C)RX5qVi|XH+gR4D%
zWEZZ_5xd1U;GZb+eARBlPM4q0?wC!zCBQuR_aP@}*dLaoc*H<^HPi8OJtHIXBf+xz
zg*D@F87du5k)q9iKnj%^_!4N)v%2jH)i}h*fM&xeoD%6^3s9sUH?k|!bC7JsQtf}i
z%wL&5zemD7GZ^UE&DnOo?W&Gn3D1NDWYvYtwY_f2J}Y-z#JgTu`_=a1^Nxh2;|675
z>DP~ri&g~L6u6z3$d;vPFCg?NJr@Yjw#rEtQ8{NBg*S$7#H%W&0A<-=X>gODP3R%N
z93m9y`PlCFZ7Ke$RDfN6p^YE(L|@MoQxc)1MY_B3t{+{6Jm
zy~1o8&EHm4s;)0bcsL}uiG^%`vSO|3;oH;uX0d?NV~~C$4%}r6zTemUraLj84I@J(R!Kb
z;#M1^>UB=p_5C)^sevw-X_gn^$prVS72tiqv5*e;y^p`nECno)9vi=Uqz;?>FbS9D
zciE?5@+JdH4WsVO+aXr25ulzNX&f&mtLO(&6(y@^{l3QH?N9k*^!2Zua@nQ18vIj^Gwzr4#AD9&Jan@U9;!<4CT$gHRsjcr)?Bd9bb-l+r
z#&!7p^$pnXKm%-Du$Pn%F!g!Ee=ly+v8l~08VBwi@s99Wr`^krm>b?5^a$fxkm)0j
z#=&aB89_fHA*rc{=@Q!-B*u0nLzk)F$H}-*}+V-P@nw{A&mQ<
zT=5*-x1hF*>N#p;DQa8BbQ~Pt!lN3md~qJbAaW>A9Ecs&dV~!#LhA=xAdc0xJc|9Y
z$3k%y-S)ql2;Dj~PTvAe((#!-IyX01y2ll9O8_5It>!8Ayy@CEy=1*^;U2f!xU2S0
zgIJrnrOZ!WXOp*H&cFYBvSq_AkjfdIUSEFkwUX1$&mDeGE_|jb?xZ(r)&>-vLEa%%
zQ?GS}I`*W5^UnMUU3jOc?fIvl)na@}(N38SDw;<4yd0frM%@pO7eU2Se8!WTp=3fg
ztC!on-7A45;2WpfB2i!P(FnlZiP!I_