Star7
6
代码
代码
Issues17
Pull Requests9
项目讨论
Wiki
分析
Star7
6
已开启
【众智】【清华大学深圳国际研究生院】【ID0953】【MUNIT】NPU精度达标，提交精度验收 #5
AtomGit-Bot创建于 2022年6月13日
【众智】【清华大学深圳国际研究生院】【ID0953】【MUNIT】NPU精度达标，提交精度验收 #5
已开启
AtomGit-Bot创建于 2022年6月13日
从refs/pull/5/head合入到master
共 11 个文件变更+2297-0
Acontrib/Tensorflow/MUNIT_ID0953_for_TensorFlow/LICENSE+251-0
@@ -0,0 +1,251 @@
1+ Apache License
2+ Version 2.0, January 2004
3+ http://www.apache.org/licenses/
4+ 
5+ TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
6+ 
7+ 1. Definitions.
8+ 
9+ "License" shall mean the terms and conditions for use, reproduction,
10+ and distribution as defined by Sections 1 through 9 of this document.
11+ 
12+ "Licensor" shall mean the copyright owner or entity authorized by
13+ the copyright owner that is granting the License.
14+ 
15+ "Legal Entity" shall mean the union of the acting entity and all
16+ other entities that control, are controlled by, or are under common
17+ control with that entity. For the purposes of this definition,
18+ "control" means (i) the power, direct or indirect, to cause the
19+ direction or management of such entity, whether by contract or
20+ otherwise, or (ii) ownership of fifty percent (50%) or more of the
21+ outstanding shares, or (iii) beneficial ownership of such entity.
22+ 
23+ "You" (or "Your") shall mean an individual or Legal Entity
24+ exercising permissions granted by this License.
25+ 
26+ "Source" form shall mean the preferred form for making modifications,
27+ including but not limited to software source code, documentation
28+ source, and configuration files.
29+ 
30+ "Object" form shall mean any form resulting from mechanical
31+ transformation or translation of a Source form, including but
32+ not limited to compiled object code, generated documentation,
33+ and conversions to other media types.
34+ 
35+ "Work" shall mean the work of authorship, whether in Source or
36+ Object form, made available under the License, as indicated by a
37+ copyright notice that is included in or attached to the work
38+ (an example is provided in the Appendix below).
39+ 
40+ "Derivative Works" shall mean any work, whether in Source or Object
41+ form, that is based on (or derived from) the Work and for which the
42+ editorial revisions, annotations, elaborations, or other modifications
43+ represent, as a whole, an original work of authorship. For the purposes
44+ of this License, Derivative Works shall not include works that remain
45+ separable from, or merely link (or bind by name) to the interfaces of,
46+ the Work and Derivative Works thereof.
47+ 
48+ "Contribution" shall mean any work of authorship, including
49+ the original version of the Work and any modifications or additions
50+ to that Work or Derivative Works thereof, that is intentionally
51+ submitted to Licensor for inclusion in the Work by the copyright owner
52+ or by an individual or Legal Entity authorized to submit on behalf of
53+ the copyright owner. For the purposes of this definition, "submitted"
54+ means any form of electronic, verbal, or written communication sent
55+ to the Licensor or its representatives, including but not limited to
56+ communication on electronic mailing lists, source code control systems,
57+ and issue tracking systems that are managed by, or on behalf of, the
58+ Licensor for the purpose of discussing and improving the Work, but
59+ excluding communication that is conspicuously marked or otherwise
60+ designated in writing by the copyright owner as "Not a Contribution."
61+ 
62+ "Contributor" shall mean Licensor and any individual or Legal Entity
63+ on behalf of whom a Contribution has been received by Licensor and
64+ subsequently incorporated within the Work.
65+ 
66+ 2. Grant of Copyright License. Subject to the terms and conditions of
67+ this License, each Contributor hereby grants to You a perpetual,
68+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
69+ copyright license to reproduce, prepare Derivative Works of,
70+ publicly display, publicly perform, sublicense, and distribute the
71+ Work and such Derivative Works in Source or Object form.
72+ 
73+ 3. Grant of Patent License. Subject to the terms and conditions of
74+ this License, each Contributor hereby grants to You a perpetual,
75+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
76+ (except as stated in this section) patent license to make, have made,
77+ use, offer to sell, sell, import, and otherwise transfer the Work,
78+ where such license applies only to those patent claims licensable
79+ by such Contributor that are necessarily infringed by their
80+ Contribution(s) alone or by combination of their Contribution(s)
81+ with the Work to which such Contribution(s) was submitted. If You
82+ institute patent litigation against any entity (including a
83+ cross-claim or counterclaim in a lawsuit) alleging that the Work
84+ or a Contribution incorporated within the Work constitutes direct
85+ or contributory patent infringement, then any patent licenses
86+ granted to You under this License for that Work shall terminate
87+ as of the date such litigation is filed.
88+ 
89+ 4. Redistribution. You may reproduce and distribute copies of the
90+ Work or Derivative Works thereof in any medium, with or without
91+ modifications, and in Source or Object form, provided that You
92+ meet the following conditions:
93+ 
94+ (a) You must give any other recipients of the Work or
95+ Derivative Works a copy of this License; and
96+ 
97+ (b) You must cause any modified files to carry prominent notices
98+ stating that You changed the files; and
99+ 
100+ (c) You must retain, in the Source form of any Derivative Works
101+ that You distribute, all copyright, patent, trademark, and
102+ attribution notices from the Source form of the Work,
103+ excluding those notices that do not pertain to any part of
104+ the Derivative Works; and
105+ 
106+ (d) If the Work includes a "NOTICE" text file as part of its
107+ distribution, then any Derivative Works that You distribute must
108+ include a readable copy of the attribution notices contained
109+ within such NOTICE file, excluding those notices that do not
110+ pertain to any part of the Derivative Works, in at least one
111+ of the following places: within a NOTICE text file distributed
112+ as part of the Derivative Works; within the Source form or
113+ documentation, if provided along with the Derivative Works; or,
114+ within a display generated by the Derivative Works, if and
115+ wherever such third-party notices normally appear. The contents
116+ of the NOTICE file are for informational purposes only and
117+ do not modify the License. You may add Your own attribution
118+ notices within Derivative Works that You distribute, alongside
119+ or as an addendum to the NOTICE text from the Work, provided
120+ that such additional attribution notices cannot be construed
121+ as modifying the License.
122+ 
123+ You may add Your own copyright statement to Your modifications and
124+ may provide additional or different license terms and conditions
125+ for use, reproduction, or distribution of Your modifications, or
126+ for any such Derivative Works as a whole, provided Your use,
127+ reproduction, and distribution of the Work otherwise complies with
128+ the conditions stated in this License.
129+ 
130+ 5. Submission of Contributions. Unless You explicitly state otherwise,
131+ any Contribution intentionally submitted for inclusion in the Work
132+ by You to the Licensor shall be under the terms and conditions of
133+ this License, without any additional terms or conditions.
134+ Notwithstanding the above, nothing herein shall supersede or modify
135+ the terms of any separate license agreement you may have executed
136+ with Licensor regarding such Contributions.
137+ 
138+ 6. Trademarks. This License does not grant permission to use the trade
139+ names, trademarks, service marks, or product names of the Licensor,
140+ except as required for reasonable and customary use in describing the
141+ origin of the Work and reproducing the content of the NOTICE file.
142+ 
143+ 7. Disclaimer of Warranty. Unless required by applicable law or
144+ agreed to in writing, Licensor provides the Work (and each
145+ Contributor provides its Contributions) on an "AS IS" BASIS,
146+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
147+ implied, including, without limitation, any warranties or conditions
148+ of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
149+ PARTICULAR PURPOSE. You are solely responsible for determining the
150+ appropriateness of using or redistributing the Work and assume any
151+ risks associated with Your exercise of permissions under this License.
152+ 
153+ 8. Limitation of Liability. In no event and under no legal theory,
154+ whether in tort (including negligence), contract, or otherwise,
155+ unless required by applicable law (such as deliberate and grossly
156+ negligent acts) or agreed to in writing, shall any Contributor be
157+ liable to You for damages, including any direct, indirect, special,
158+ incidental, or consequential damages of any character arising as a
159+ result of this License or out of the use or inability to use the
160+ Work (including but not limited to damages for loss of goodwill,
161+ work stoppage, computer failure or malfunction, or any and all
162+ other commercial damages or losses), even if such Contributor
163+ has been advised of the possibility of such damages.
164+ 
165+ 9. Accepting Warranty or Additional Liability. While redistributing
166+ the Work or Derivative Works thereof, You may choose to offer,
167+ and charge a fee for, acceptance of support, warranty, indemnity,
168+ or other liability obligations and/or rights consistent with this
169+ License. However, in accepting such obligations, You may act only
170+ on Your own behalf and on Your sole responsibility, not on behalf
171+ of any other Contributor, and only if You agree to indemnify,
172+ defend, and hold each Contributor harmless for any liability
173+ incurred by, or claims asserted against, such Contributor by reason
174+ of your accepting any such warranty or additional liability.
175+ 
176+ END OF TERMS AND CONDITIONS
177+ 
178+ APPENDIX: How to apply the Apache License to your work.
179+ 
180+ To apply the Apache License to your work, attach the following
181+ boilerplate notice, with the fields enclosed by brackets "[]"
182+ replaced with your own identifying information. (Don't include
183+ the brackets!) The text should be enclosed in the appropriate
184+ comment syntax for the file format. We also recommend that a
185+ file or class name and description of purpose be included on the
186+ same "printed page" as the copyright notice for easier
187+ identification within third-party archives.
188+ 
189+ Copyright [yyyy] [name of copyright owner]
190+ 
191+ Licensed under the Apache License, Version 2.0 (the "License");
192+ you may not use this file except in compliance with the License.
193+ You may obtain a copy of the License at
194+ 
195+ http://www.apache.org/licenses/LICENSE-2.0
196+ 
197+ Unless required by applicable law or agreed to in writing, software
198+ distributed under the License is distributed on an "AS IS" BASIS,
199+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
200+ See the License for the specific language governing permissions and
201+ limitations under the License.
202+ 
203+## Some of TensorFlow's code is derived from Caffe, which is subject to the following copyright notice:
204+ 
205+COPYRIGHT
206+ 
207+All contributions by the University of California:
208+ 
209+Copyright (c) 2014, The Regents of the University of California (Regents)
210+All rights reserved.
211+ 
212+All other contributions:
213+ 
214+Copyright (c) 2014, the respective contributors
215+All rights reserved.
216+ 
217+Caffe uses a shared copyright model: each contributor holds copyright over
218+their contributions to Caffe. The project versioning records all such
219+contribution and copyright details. If a contributor wants to further mark
220+their specific copyright on a particular contribution, they should indicate
221+their copyright solely in the commit message of the change when it is
222+committed.
223+ 
224+LICENSE
225+ 
226+Redistribution and use in source and binary forms, with or without
227+modification, are permitted provided that the following conditions are met:
228+ 
229+1. Redistributions of source code must retain the above copyright notice, this
230+ list of conditions and the following disclaimer.
231+ 
232+2. Redistributions in binary form must reproduce the above copyright notice,
233+ this list of conditions and the following disclaimer in the documentation
234+ and/or other materials provided with the distribution.
235+ 
236+ THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
237+ ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
238+ WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
239+ DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
240+ ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
241+ (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
242+ LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
243+ ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
244+ (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
245+ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
246+ 
247+CONTRIBUTION AGREEMENT
248+ 
249+By contributing to the BVLC/caffe repository through pull-request, comment,
250+or otherwise, the contributor releases their content to the
251+license and copyright terms herein.
Acontrib/Tensorflow/MUNIT_ID0953_for_TensorFlow/MUNIT.py+697-0
@@ -0,0 +1,697 @@
1+# Copyright 2017 The TensorFlow Authors. All Rights Reserved.
2+#
3+# Licensed under the Apache License, Version 2.0 (the "License");
4+# you may not use this file except in compliance with the License.
5+# You may obtain a copy of the License at
6+#
7+# http://www.apache.org/licenses/LICENSE-2.0
8+#
9+# Unless required by applicable law or agreed to in writing, software
10+# distributed under the License is distributed on an "AS IS" BASIS,
11+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12+# See the License for the specific language governing permissions and
13+# limitations under the License.
14+# ============================================================================
15+# Copyright 2021 Huawei Technologies Co., Ltd
16+#
17+# Licensed under the Apache License, Version 2.0 (the "License");
18+# you may not use this file except in compliance with the License.
19+# You may obtain a copy of the License at
20+#
21+# http://www.apache.org/licenses/LICENSE-2.0
22+#
23+# Unless required by applicable law or agreed to in writing, software
24+# distributed under the License is distributed on an "AS IS" BASIS,
25+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
26+# See the License for the specific language governing permissions and
27+# limitations under the License.
28+ 
29+from npu_bridge.npu_init import *
30+from ops import *
31+from utils import *
32+from glob import glob
33+import time
34+from tensorflow.contrib.data import batch_and_drop_remainder
35+ 
36+class MUNIT(object) :
37+ def __init__(self, sess, args):
38+ self.model_name = 'MUNIT'
39+ self.sess = sess
40+ self.checkpoint_dir = args.checkpoint_dir
41+ self.result_dir = args.result_dir
42+ self.log_dir = args.log_dir
43+ self.sample_dir = args.sample_dir
44+ #self.dataset_name = args.dataset
45+ self.dataset_name = args.data_path
46+ self.augment_flag = args.augment_flag
47+ 
48+ self.epoch = args.epoch
49+ self.iteration = args.iteration
50+ 
51+ self.gan_type = args.gan_type
52+ 
53+ self.batch_size = args.batch_size
54+ self.print_freq = args.print_freq
55+ self.save_freq = args.save_freq
56+ self.num_style = args.num_style # for test
57+ self.guide_img = args.guide_img
58+ self.direction = args.direction
59+ 
60+ self.img_h = args.img_h
61+ self.img_w = args.img_w
62+ self.img_ch = args.img_ch
63+ 
64+ self.init_lr = args.lr
65+ self.ch = args.ch
66+ 
67+ """ Weight """
68+ self.gan_w = args.gan_w
69+ self.recon_x_w = args.recon_x_w
70+ self.recon_s_w = args.recon_s_w
71+ self.recon_c_w = args.recon_c_w
72+ self.recon_x_cyc_w = args.recon_x_cyc_w
73+ 
74+ """ Generator """
75+ self.n_res = args.n_res
76+ self.mlp_dim = pow(2, args.n_sample) * args.ch # default : 256
77+ 
78+ self.n_downsample = args.n_sample
79+ self.n_upsample = args.n_sample
80+ self.style_dim = args.style_dim
81+ 
82+ """ Discriminator """
83+ self.n_dis = args.n_dis
84+ self.n_scale = args.n_scale
85+ 
86+ self.sample_dir = os.path.join(args.sample_dir, self.model_dir)
87+ check_folder(self.sample_dir)
88+ 
89+ #self.trainA_dataset = glob('./dataset/{}/*.*'.format(self.dataset_name + '/trainA'))
90+ #self.trainB_dataset = glob('./dataset/{}/*.*'.format(self.dataset_name + '/trainB'))
91+ self.trainA_dataset = glob('{}/*.*'.format(self.dataset_name + '/trainA'))
92+ self.trainB_dataset = glob('{}/*.*'.format(self.dataset_name + '/trainB'))
93+ self.dataset_num = max(len(self.trainA_dataset), len(self.trainB_dataset))
94+ 
95+ # lossScale相关
96+ self.bert_loss_scale = args.bert_loss_scale
97+ self.mmgr = {}
98+ 
99+ print("##### Information #####")
100+ print("# gan type : ", self.gan_type)
101+ print("# dataset : ", self.dataset_name)
102+ print("# max dataset number : ", self.dataset_num)
103+ print("# batch_size : ", self.batch_size)
104+ print("# epoch : ", self.epoch)
105+ print("# iteration per epoch : ", self.iteration)
106+ print("# style in test phase : ", self.num_style)
107+ 
108+ print()
109+ 
110+ print("##### Generator #####")
111+ print("# residual blocks : ", self.n_res)
112+ print("# Style dimension : ", self.style_dim)
113+ print("# MLP dimension : ", self.mlp_dim)
114+ print("# Down sample : ", self.n_downsample)
115+ print("# Up sample : ", self.n_upsample)
116+ 
117+ print()
118+ 
119+ print("##### Discriminator #####")
120+ print("# Discriminator layer : ", self.n_dis)
121+ print("# Multi-scale Dis : ", self.n_scale)
122+ 
123+ ##################################################################################
124+ # Encoder and Decoders
125+ ##################################################################################
126+ 
127+ def Style_Encoder(self, x, reuse=False, scope='style_encoder'):
128+ # IN removes the original feature mean and variance that represent important style information
129+ channel = self.ch
130+ with tf.variable_scope(scope, reuse=reuse) :
131+ x = conv(x, channel, kernel=7, stride=1, pad=3, pad_type='reflect', scope='conv_0')
132+ x = relu(x)
133+ 
134+ for i in range(2) :
135+ x = conv(x, channel*2, kernel=4, stride=2, pad=1, pad_type='reflect', scope='conv_'+str(i+1))
136+ x = relu(x)
137+ 
138+ channel = channel * 2
139+ 
140+ for i in range(2) :
141+ x = conv(x, channel, kernel=4, stride=2, pad=1, pad_type='reflect', scope='down_conv_'+str(i))
142+ x = relu(x)
143+ 
144+ x = adaptive_avg_pooling(x) # global average pooling
145+ x = conv(x, self.style_dim, kernel=1, stride=1, scope='SE_logit')
146+ 
147+ return x
148+ 
149+ def Content_Encoder(self, x, reuse=False, scope='content_encoder'):
150+ channel = self.ch
151+ with tf.variable_scope(scope, reuse=reuse) :
152+ x = conv(x, channel, kernel=7, stride=1, pad=3, pad_type='reflect', scope='conv_0')
153+ x = instance_norm(x, scope='ins_0')
154+ x = relu(x)
155+ 
156+ for i in range(self.n_downsample) :
157+ x = conv(x, channel*2, kernel=4, stride=2, pad=1, pad_type='reflect', scope='conv_'+str(i+1))
158+ x = instance_norm(x, scope='ins_'+str(i+1))
159+ x = relu(x)
160+ 
161+ channel = channel * 2
162+ 
163+ for i in range(self.n_res) :
164+ x = resblock(x, channel, scope='resblock_'+str(i))
165+ 
166+ return x
167+ 
168+ def generator(self, contents, style, reuse=False, scope="decoder"):
169+ channel = self.mlp_dim
170+ with tf.variable_scope(scope, reuse=reuse) :
171+ mu, var = self.MLP(style)
172+ x = contents
173+ 
174+ for i in range(self.n_res) :
175+ idx = 2 * i
176+ x = adaptive_resblock(x, channel, mu[idx], var[idx], mu[idx + 1], var[idx + 1], scope='adaptive_resblock_'+str(i))
177+ 
178+ for i in range(self.n_upsample) :
179+ # # IN removes the original feature mean and variance that represent important style information
180+ x = up_sample(x, scale_factor=2)
181+ x = conv(x, channel//2, kernel=5, stride=1, pad=2, pad_type='reflect', scope='conv_'+str(i))
182+ x = layer_norm(x, scope='layer_norm_'+str(i))
183+ x = relu(x)
184+ 
185+ channel = channel // 2
186+ 
187+ x = conv(x, channels=self.img_ch, kernel=7, stride=1, pad=3, pad_type='reflect', scope='G_logit')
188+ x = tanh(x)
189+ 
190+ return x
191+ 
192+ def MLP(self, style, scope='MLP'):
193+ channel = self.mlp_dim
194+ with tf.variable_scope(scope) :
195+ x = style
196+ 
197+ for i in range(2):
198+ x = fully_connected(x, channel, scope='FC_' + str(i))
199+ x = relu(x)
200+ 
201+ mu_list = []
202+ var_list = []
203+ 
204+ for i in range(self.n_res * 2):
205+ mu = fully_connected(x, channel, scope='FC_mu_' + str(i))
206+ var = fully_connected(x, channel, scope='FC_var_' + str(i))
207+ 
208+ mu = tf.reshape(mu, shape=[-1, 1, 1, channel])
209+ var = tf.reshape(var, shape=[-1, 1, 1, channel])
210+ 
211+ mu_list.append(mu)
212+ var_list.append(var)
213+ 
214+ return mu_list, var_list
215+ 
216+ ##################################################################################
217+ # Discriminator
218+ ##################################################################################
219+ 
220+ def discriminator(self, x_init, reuse=False, scope="discriminator"):
221+ D_logit = []
222+ with tf.variable_scope(scope, reuse=reuse) :
223+ for scale in range(self.n_scale) :
224+ channel = self.ch
225+ x = conv(x_init, channel, kernel=4, stride=2, pad=1, pad_type='reflect', scope='ms_' + str(scale) + 'conv_0')
226+ x = lrelu(x, 0.2)
227+ 
228+ for i in range(1, self.n_dis):
229+ x = conv(x, channel * 2, kernel=4, stride=2, pad=1, pad_type='reflect', scope='ms_' + str(scale) +'conv_' + str(i))
230+ x = lrelu(x, 0.2)
231+ 
232+ channel = channel * 2
233+ 
234+ x = conv(x, channels=1, kernel=1, stride=1, scope='ms_' + str(scale) + 'D_logit')
235+ D_logit.append(x)
236+ 
237+ x_init = down_sample(x_init)
238+ 
239+ return D_logit
240+ 
241+ ##################################################################################
242+ # Model
243+ ##################################################################################
244+ 
245+ def Encoder_A(self, x_A, reuse=False):
246+ style_A = self.Style_Encoder(x_A, reuse=reuse, scope='style_encoder_A')
247+ content_A = self.Content_Encoder(x_A, reuse=reuse, scope='content_encoder_A')
248+ 
249+ return content_A, style_A
250+ 
251+ def Encoder_B(self, x_B, reuse=False):
252+ style_B = self.Style_Encoder(x_B, reuse=reuse, scope='style_encoder_B')
253+ content_B = self.Content_Encoder(x_B, reuse=reuse, scope='content_encoder_B')
254+ 
255+ return content_B, style_B
256+ 
257+ def Decoder_A(self, content_B, style_A, reuse=False):
258+ x_ba = self.generator(contents=content_B, style=style_A, reuse=reuse, scope='decoder_A')
259+ 
260+ return x_ba
261+ 
262+ def Decoder_B(self, content_A, style_B, reuse=False):
263+ x_ab = self.generator(contents=content_A, style=style_B, reuse=reuse, scope='decoder_B')
264+ 
265+ return x_ab
266+ 
267+ def discriminate_real(self, x_A, x_B):
268+ real_A_logit = self.discriminator(x_A, scope="discriminator_A")
269+ real_B_logit = self.discriminator(x_B, scope="discriminator_B")
270+ 
271+ return real_A_logit, real_B_logit
272+ 
273+ def discriminate_fake(self, x_ba, x_ab):
274+ fake_A_logit = self.discriminator(x_ba, reuse=True, scope="discriminator_A")
275+ fake_B_logit = self.discriminator(x_ab, reuse=True, scope="discriminator_B")
276+ 
277+ return fake_A_logit, fake_B_logit
278+ 
279+ def build_model(self):
280+ self.lr = tf.placeholder(tf.float32, name='learning_rate')
281+ 
282+ """ Input Image"""
283+ Image_Data_Class = ImageData(self.img_h, self.img_w, self.img_ch, self.augment_flag)
284+ 
285+ trainA = tf.data.Dataset.from_tensor_slices(self.trainA_dataset)
286+ trainB = tf.data.Dataset.from_tensor_slices(self.trainB_dataset)
287+ 
288+ trainA = trainA.prefetch(self.batch_size).shuffle(self.dataset_num).map(Image_Data_Class.image_processing, num_parallel_calls=8).apply(batch_and_drop_remainder(self.batch_size)).repeat()
289+ trainB = trainB.prefetch(self.batch_size).shuffle(self.dataset_num).map(Image_Data_Class.image_processing, num_parallel_calls=8).apply(batch_and_drop_remainder(self.batch_size)).repeat()
290+ 
291+ trainA_iterator = trainA.make_one_shot_iterator()
292+ trainB_iterator = trainB.make_one_shot_iterator()
293+ 
294+ 
295+ self.domain_A = trainA_iterator.get_next()
296+ self.domain_B = trainB_iterator.get_next()
297+ 
298+ 
299+ """ Define Encoder, Generator, Discriminator """
300+ self.style_a = tf.placeholder(tf.float32, shape=[self.batch_size, 1, 1, self.style_dim], name='style_a')
301+ self.style_b = tf.placeholder(tf.float32, shape=[self.batch_size, 1, 1, self.style_dim], name='style_b')
302+ 
303+ # encode
304+ content_a, style_a_prime = self.Encoder_A(self.domain_A)
305+ content_b, style_b_prime = self.Encoder_B(self.domain_B)
306+ 
307+ # decode (within domain)
308+ x_aa = self.Decoder_A(content_B=content_a, style_A=style_a_prime)
309+ x_bb = self.Decoder_B(content_A=content_b, style_B=style_b_prime)
310+ 
311+ # decode (cross domain)
312+ x_ba = self.Decoder_A(content_B=content_b, style_A=self.style_a, reuse=True)
313+ x_ab = self.Decoder_B(content_A=content_a, style_B=self.style_b, reuse=True)
314+ 
315+ # encode again
316+ content_b_, style_a_ = self.Encoder_A(x_ba, reuse=True)
317+ content_a_, style_b_ = self.Encoder_B(x_ab, reuse=True)
318+ 
319+ # decode again (if needed)
320+ if self.recon_x_cyc_w > 0 :
321+ x_aba = self.Decoder_A(content_B=content_a_, style_A=style_a_prime, reuse=True)
322+ x_bab = self.Decoder_B(content_A=content_b_, style_B=style_b_prime, reuse=True)
323+ 
324+ cyc_recon_A = L1_loss(x_aba, self.domain_A)
325+ cyc_recon_B = L1_loss(x_bab, self.domain_B)
326+ 
327+ else :
328+ cyc_recon_A = 0.0
329+ cyc_recon_B = 0.0
330+ 
331+ real_A_logit, real_B_logit = self.discriminate_real(self.domain_A, self.domain_B)
332+ fake_A_logit, fake_B_logit = self.discriminate_fake(x_ba, x_ab)
333+ 
334+ """ Define Loss """
335+ G_ad_loss_a = generator_loss(self.gan_type, fake_A_logit)
336+ G_ad_loss_b = generator_loss(self.gan_type, fake_B_logit)
337+ 
338+ D_ad_loss_a = discriminator_loss(self.gan_type, real_A_logit, fake_A_logit)
339+ D_ad_loss_b = discriminator_loss(self.gan_type, real_B_logit, fake_B_logit)
340+ 
341+ recon_A = L1_loss(x_aa, self.domain_A) # reconstruction
342+ recon_B = L1_loss(x_bb, self.domain_B) # reconstruction
343+ 
344+ # The style reconstruction loss encourages
345+ # diverse outputs given different style codes
346+ recon_style_A = L1_loss(style_a_, self.style_a)
347+ recon_style_B = L1_loss(style_b_, self.style_b)
348+ 
349+ # The content reconstruction loss encourages
350+ # the translated image to preserve semantic content of the input image
351+ recon_content_A = L1_loss(content_a_, content_a)
352+ recon_content_B = L1_loss(content_b_, content_b)
353+ 
354+ 
355+ Generator_A_loss = self.gan_w * G_ad_loss_a + \
356+ self.recon_x_w * recon_A + \
357+ self.recon_s_w * recon_style_A + \
358+ self.recon_c_w * recon_content_A + \
359+ self.recon_x_cyc_w * cyc_recon_A
360+ 
361+ Generator_B_loss = self.gan_w * G_ad_loss_b + \
362+ self.recon_x_w * recon_B + \
363+ self.recon_s_w * recon_style_B + \
364+ self.recon_c_w * recon_content_B + \
365+ self.recon_x_cyc_w * cyc_recon_B
366+ 
367+ Discriminator_A_loss = self.gan_w * D_ad_loss_a
368+ Discriminator_B_loss = self.gan_w * D_ad_loss_b
369+ 
370+ self.Generator_loss = Generator_A_loss + Generator_B_loss + regularization_loss('encoder') + regularization_loss('decoder')
371+ self.Discriminator_loss = Discriminator_A_loss + Discriminator_B_loss + regularization_loss('discriminator')
372+ 
373+ """ Training """
374+ t_vars = tf.trainable_variables()
375+ G_vars = [var for var in t_vars if 'decoder' in var.name or 'encoder' in var.name]
376+ D_vars = [var for var in t_vars if 'discriminator' in var.name]
377+ 
378+ 
379+ # self.G_optim = tf.train.AdamOptimizer(self.lr, beta1=0.5, beta2=0.999).minimize(self.Generator_loss, var_list=G_vars)
H
Hhtwang2024年12月18日

无用的注释代码可以删除

likedislike
380+ # self.D_optim = tf.train.AdamOptimizer(self.lr, beta1=0.5, beta2=0.999).minimize(self.Discriminator_loss, var_list=D_vars)
381+ 
382+ # 开启loss_scale
383+ self.G_optim = tf.train.AdamOptimizer(self.lr, beta1=0.5, beta2=0.999)
384+ self.D_optim = tf.train.AdamOptimizer(self.lr, beta1=0.5, beta2=0.999)
385+ # 开启loss_scale
386+ self.G_optim = self.open_loss_scale(self.G_optim, 'G')
387+ self.D_optim = self.open_loss_scale(self.D_optim, 'D')
388+ 
389+ self.G_optim = self.G_optim.minimize(self.Generator_loss, var_list=G_vars)
390+ self.D_optim = self.D_optim.minimize(self.Discriminator_loss, var_list=D_vars)
391+ 
392+ 
393+ 
394+ """" Summary """
395+ self.all_G_loss = tf.summary.scalar("Generator_loss", self.Generator_loss)
396+ self.all_D_loss = tf.summary.scalar("Discriminator_loss", self.Discriminator_loss)
397+ self.G_A_loss = tf.summary.scalar("G_A_loss", Generator_A_loss)
398+ self.G_B_loss = tf.summary.scalar("G_B_loss", Generator_B_loss)
399+ self.D_A_loss = tf.summary.scalar("D_A_loss", Discriminator_A_loss)
400+ self.D_B_loss = tf.summary.scalar("D_B_loss", Discriminator_B_loss)
401+ 
402+ self.G_loss = tf.summary.merge([self.G_A_loss, self.G_B_loss, self.all_G_loss])
403+ self.D_loss = tf.summary.merge([self.D_A_loss, self.D_B_loss, self.all_D_loss])
404+ 
405+ """ Image """
406+ self.fake_A = x_ba
407+ self.fake_B = x_ab
408+ 
409+ self.real_A = self.domain_A
410+ self.real_B = self.domain_B
411+ 
412+ """ Test """
413+ self.test_image = tf.placeholder(tf.float32, [1, self.img_h, self.img_w, self.img_ch], name='test_image')
414+ self.test_style = tf.placeholder(tf.float32, [1, 1, 1, self.style_dim], name='test_style')
415+ 
416+ test_content_a, _ = self.Encoder_A(self.test_image, reuse=True)
417+ test_content_b, _ = self.Encoder_B(self.test_image, reuse=True)
418+ 
419+ self.test_fake_A = self.Decoder_A(content_B=test_content_b, style_A=self.test_style, reuse=True)
420+ self.test_fake_B = self.Decoder_B(content_A=test_content_a, style_B=self.test_style, reuse=True)
421+ 
422+ """ Guided Image Translation """
423+ self.content_image = tf.placeholder(tf.float32, [1, self.img_h, self.img_w, self.img_ch], name='content_image')
424+ self.style_image = tf.placeholder(tf.float32, [1, self.img_h, self.img_w, self.img_ch], name='guide_style_image')
425+ 
426+ if self.direction == 'a2b' :
427+ guide_content_A, guide_style_A = self.Encoder_A(self.content_image, reuse=True)
428+ guide_content_B, guide_style_B = self.Encoder_B(self.style_image, reuse=True)
429+ 
430+ else :
431+ guide_content_B, guide_style_B = self.Encoder_B(self.content_image, reuse=True)
432+ guide_content_A, guide_style_A = self.Encoder_A(self.style_image, reuse=True)
433+ 
434+ self.guide_fake_A = self.Decoder_A(content_B=guide_content_B, style_A=guide_style_A, reuse=True)
435+ self.guide_fake_B = self.Decoder_B(content_A=guide_content_A, style_B=guide_style_B, reuse=True)
436+ 
437+ def train(self):
438+ # initialize all variables
439+ tf.global_variables_initializer().run()
440+ 
441+ # saver to save model
442+ self.saver = tf.train.Saver()
443+ 
444+ # summary writer
445+ self.writer = tf.summary.FileWriter(self.log_dir + '/' + self.model_dir, self.sess.graph)
446+ 
447+ # restore check-point if it exits
448+ could_load, checkpoint_counter = self.load(self.checkpoint_dir)
449+ if could_load:
450+ start_epoch = (int)(checkpoint_counter / self.iteration)
451+ start_batch_id = checkpoint_counter - start_epoch * self.iteration
452+ counter = checkpoint_counter
453+ print(" [*] Load SUCCESS")
454+ else:
455+ start_epoch = 0
456+ start_batch_id = 0
457+ counter = 1
458+ print(" [!] Load failed...")
459+ 
460+ # loop for epoch
461+ start_time = time.time()
462+ for epoch in range(start_epoch, self.epoch):
463+ 
464+ lr = self.init_lr * pow(0.5, epoch)
465+ 
466+ for idx in range(start_batch_id, self.iteration):
467+ style_a = np.random.normal(loc=0.0, scale=1.0, size=[self.batch_size, 1, 1, self.style_dim])
468+ style_b = np.random.normal(loc=0.0, scale=1.0, size=[self.batch_size, 1, 1, self.style_dim])
469+ 
470+ train_feed_dict = {
471+ self.style_a : style_a,
472+ self.style_b : style_b,
473+ self.lr : lr
474+ }
475+ 
476+ # Update D
477+ # _, d_loss, summary_str = \
478+ # self.sess.run([self.D_optim, self.Discriminator_loss, self.D_loss], feed_dict = train_feed_dict)
479+
480+ # 保存scale value，并打印到日志里，来观察整网的溢出情况
481+ _, d_loss, summary_str, d_scale = \
482+ self.sess.run([self.D_optim, self.Discriminator_loss, self.D_loss, self.mmgr['D'].get_loss_scale()],
483+ feed_dict=train_feed_dict)
484+ 
485+ self.writer.add_summary(summary_str, counter)
486+ 
487+ # Update G
488+ #batch_A_images, batch_B_images, fake_A, fake_B, _, g_loss, summary_str = self.sess.run([self.real_A, self.real_B, self.fake_A, self.fake_B, self.G_optim, self.Generator_loss, self.G_loss], feed_dict = train_feed_dict)
489+ batch_A_images, batch_B_images, fake_A, fake_B, _, g_loss, summary_str, g_scale = self.sess.run(
490+ [self.real_A, self.real_B, self.fake_A, self.fake_B, self.G_optim, self.Generator_loss,
491+ self.G_loss, self.mmgr['G'].get_loss_scale()], feed_dict=train_feed_dict)
492+ self.writer.add_summary(summary_str, counter)
493+ 
494+ # display training status
495+ counter += 1
496+ print("Epoch: [%2d] [%6d/%6d] time: %4.4f d_loss: %.8f, g_loss: %.8f, d_scale: %d, g_scale: %d\n" \
497+ % (epoch, idx, self.iteration, time.time() - start_time, d_loss, g_loss, d_scale, g_scale))
498+ # print("Epoch: [%2d] [%6d/%6d] time: %4.4f d_loss: %.8f, g_loss: %.8f\n" \
499+ # % (epoch, idx, self.iteration, time.time() - start_time, d_loss, g_loss))
500+ 
501+ if np.mod(idx+1, self.print_freq) == 0 :
502+ save_images(batch_A_images, [self.batch_size, 1],
503+ '{}/real_A_{:02d}_{:06d}.jpg'.format(self.sample_dir, epoch, idx+1))
504+ # save_images(batch_B_images, [self.batch_size, 1],
505+ # './{}/real_B_{}_{:02d}_{:06d}.jpg'.format(self.sample_dir, gpu_id, epoch, idx+1))
506+ 
507+ # save_images(fake_A, [self.batch_size, 1],
508+ # './{}/fake_A_{}_{:02d}_{:06d}.jpg'.format(self.sample_dir, gpu_id, epoch, idx+1))
509+ save_images(fake_B, [self.batch_size, 1],
510+ '{}/fake_B_{:02d}_{:06d}.jpg'.format(self.sample_dir, epoch, idx+1))
511+ 
512+ if np.mod(idx+1, self.save_freq) == 0 :
513+ self.save(self.checkpoint_dir, counter)
514+ 
515+ # After an epoch, start_batch_id is set to zero
516+ # non-zero value is only for the first epoch after loading pre-trained model
517+ start_batch_id = 0
518+ 
519+ # save model for final step
520+ self.save(self.checkpoint_dir, counter)
521+ 
522+ 
523+ @property
524+ def model_dir(self):
525+ return "{}_{}".format(self.model_name, self.gan_type)
526+ #return "{}_{}_{}".format(self.model_name, self.dataset_name, self.gan_type)
527+ 
528+ def save(self, checkpoint_dir, step):
529+ checkpoint_dir = os.path.join(checkpoint_dir, self.model_dir)
530+ 
531+ if not os.path.exists(checkpoint_dir):
532+ os.makedirs(checkpoint_dir)
533+ 
534+ self.saver.save(self.sess, os.path.join(checkpoint_dir, self.model_name + '.model'), global_step=step)
535+ 
536+ def load(self, checkpoint_dir):
537+ import re
538+ print(" [*] Reading checkpoints...")
539+ checkpoint_dir = os.path.join(checkpoint_dir, self.model_dir)
540+ 
541+ ckpt = tf.train.get_checkpoint_state(checkpoint_dir)
542+ if ckpt and ckpt.model_checkpoint_path:
543+ ckpt_name = os.path.basename(ckpt.model_checkpoint_path)
544+ self.saver.restore(self.sess, os.path.join(checkpoint_dir, ckpt_name))
545+ counter = int(next(re.finditer("(\d+)(?!.*\d)", ckpt_name)).group(0))
546+ print(" [*] Success to read {}".format(ckpt_name))
547+ return True, counter
548+ else:
549+ print(" [*] Failed to find a checkpoint")
550+ return False, 0
551+ 
552+ def test(self):
553+ tf.global_variables_initializer().run()
554+ 
555+ #test_A_files = glob('{}/*.*'.format(self.dataset_name + '/testA'))
556+ #test_B_files = glob('{}/*.*'.format(self.dataset_name + '/testB'))
557+ test_A_files = self.trainA_dataset
558+ test_B_files = self.trainB_dataset
559+ 
560+ self.saver = tf.train.Saver()
561+ could_load, checkpoint_counter = self.load(self.checkpoint_dir)
562+ self.result_dir = os.path.join(self.result_dir, self.model_dir)
563+ check_folder(self.result_dir)
564+ 
565+ if could_load :
566+ print(" [*] Load SUCCESS")
567+ else :
568+ print(" [!] Load failed...")
569+ 
570+ # write html for visual comparison
571+ index_path = os.path.join(self.result_dir, 'index.html')
572+ index = open(index_path, 'w')
573+ index.write("<html><body><table><tr>")
574+ index.write("<th>name</th><th>input</th><th>output</th></tr>")
575+ 
576+ cnt = 0
577+ for sample_file in test_A_files : # A -> B
578+ print('Processing A image: ' + sample_file)
579+ sample_image = np.asarray(load_test_data(sample_file, size_h=self.img_h, size_w=self.img_w))
580+ file_name = os.path.basename(sample_file).split(".")[0]
581+ file_extension = os.path.basename(sample_file).split(".")[1]
582+ 
583+ for i in range(self.num_style) :
584+ test_style = np.random.normal(loc=0.0, scale=1.0, size=[1, 1, 1, self.style_dim])
585+ image_path = os.path.join(self.result_dir, '{}_style{}.{}'.format(file_name, i, file_extension))
586+ 
587+ fake_img = self.sess.run(self.test_fake_B, feed_dict = {self.test_image : sample_image, self.test_style : test_style})
588+ save_images(fake_img, [1, 1], image_path)
589+ 
590+ index.write("<td>%s</td>" % os.path.basename(image_path))
591+ index.write("<td><img src='%s' width='%d' height='%d'></td>" % (sample_file if os.path.isabs(sample_file) else (
592+ '../..' + os.path.sep + sample_file), self.img_w, self.img_h))
593+ index.write("<td><img src='%s' width='%d' height='%d'></td>" % (image_path if os.path.isabs(image_path) else (
594+ '../..' + os.path.sep + image_path), self.img_w, self.img_h))
595+ index.write("</tr>")
596+ cnt += 1
597+ if cnt == 1: break
598+ 
599+ cnt = 0
600+ for sample_file in test_B_files : # B -> A
601+ print('Processing B image: ' + sample_file)
602+ sample_image = np.asarray(load_test_data(sample_file, size_h=self.img_h, size_w=self.img_w))
603+ file_name = os.path.basename(sample_file).split(".")[0]
604+ file_extension = os.path.basename(sample_file).split(".")[1]
605+ 
606+ for i in range(self.num_style):
607+ test_style = np.random.normal(loc=0.0, scale=1.0, size=[1, 1, 1, self.style_dim])
608+ image_path = os.path.join(self.result_dir, '{}_style{}.{}'.format(file_name, i, file_extension))
609+ 
610+ fake_img = self.sess.run(self.test_fake_A, feed_dict={self.test_image: sample_image, self.test_style: test_style})
611+ save_images(fake_img, [1, 1], image_path)
612+ 
613+ index.write("<td>%s</td>" % os.path.basename(image_path))
614+ index.write("<td><img src='%s' width='%d' height='%d'></td>" % (sample_file if os.path.isabs(sample_file) else (
615+ '../..' + os.path.sep + sample_file), self.img_w, self.img_h))
616+ index.write("<td><img src='%s' width='%d' height='%d'></td>" % (image_path if os.path.isabs(image_path) else (
617+ '../..' + os.path.sep + image_path), self.img_w, self.img_h))
618+ index.write("</tr>")
619+ cnt += 1
620+ if cnt == 1: break
621+ index.close()
622+ 
623+ def style_guide_test(self):
624+ tf.global_variables_initializer().run()
625+ test_A_files = glob('./dataset/{}/*.*'.format(self.dataset_name + '/testA'))
626+ test_B_files = glob('./dataset/{}/*.*'.format(self.dataset_name + '/testB'))
627+ 
628+ style_file = np.asarray(load_test_data(self.guide_img, size_h=self.img_h, size_w=self.img_w))
629+ 
630+ self.saver = tf.train.Saver()
631+ could_load, checkpoint_counter = self.load(self.checkpoint_dir)
632+ self.result_dir = os.path.join(self.result_dir, self.model_dir, 'guide')
633+ check_folder(self.result_dir)
634+ 
635+ if could_load:
636+ print(" [*] Load SUCCESS")
637+ else:
638+ print(" [!] Load failed...")
639+ 
640+ # write html for visual comparison
641+ index_path = os.path.join(self.result_dir, 'index.html')
642+ index = open(index_path, 'w')
643+ index.write("<html><body><table><tr>")
644+ index.write("<th>name</th><th>input</th><th>output</th></tr>")
645+ 
646+ if self.direction == 'a2b' :
647+ for sample_file in test_A_files: # A -> B
648+ print('Processing A image: ' + sample_file)
649+ sample_image = np.asarray(load_test_data(sample_file, size_h=self.img_h, size_w=self.img_w))
650+ image_path = os.path.join(self.result_dir, '{}'.format(os.path.basename(sample_file)))
651+ 
652+ fake_img = self.sess.run(self.guide_fake_B, feed_dict={self.content_image: sample_image, self.style_image : style_file})
653+ save_images(fake_img, [1, 1], image_path)
654+ 
655+ index.write("<td>%s</td>" % os.path.basename(image_path))
656+ index.write("<td><img src='%s' width='%d' height='%d'></td>" % (sample_file if os.path.isabs(sample_file) else (
657+ '../../..' + os.path.sep + sample_file), self.img_w, self.img_h))
658+ index.write("<td><img src='%s' width='%d' height='%d'></td>" % (image_path if os.path.isabs(image_path) else (
659+ '../../..' + os.path.sep + image_path), self.img_w, self.img_h))
660+ index.write("</tr>")
661+ 
662+ else :
663+ for sample_file in test_B_files: # B -> A
664+ print('Processing B image: ' + sample_file)
665+ sample_image = np.asarray(load_test_data(sample_file, size_h=self.img_h, size_w=self.img_w))
666+ image_path = os.path.join(self.result_dir, '{}'.format(os.path.basename(sample_file)))
667+ 
668+ fake_img = self.sess.run(self.guide_fake_A, feed_dict={self.content_image: sample_image, self.style_image : style_file})
669+ save_images(fake_img, [1, 1], image_path)
670+ 
671+ index.write("<td>%s</td>" % os.path.basename(image_path))
672+ index.write("<td><img src='%s' width='%d' height='%d'></td>" % (sample_file if os.path.isabs(sample_file) else (
673+ '../../..' + os.path.sep + sample_file), self.img_w, self.img_h))
674+ index.write("<td><img src='%s' width='%d' height='%d'></td>" % (image_path if os.path.isabs(image_path) else (
675+ '../../..' + os.path.sep + image_path), self.img_w, self.img_h))
676+ index.write("</tr>")
677+ index.close()
678+ 
679+ def open_loss_scale(self, opt, key):
680+ opt_tmp = opt
681+ if self.bert_loss_scale == 0:
682+ # loss_scale_manager = ExponentialUpdateLossScaleManager(init_loss_scale=2 ** 32, incr_every_n_steps=1000,
683+ # decr_every_n_nan_or_inf=2, decr_ratio=0.5)
684+ loss_scale_manager = ExponentialUpdateLossScaleManager(init_loss_scale = 2 ** 10, incr_every_n_steps = 100, decr_every_n_nan_or_inf = 2, decr_ratio = 0.8)
685+ print("lossScale type: exponential")
686+ elif self.bert_loss_scale >= 1:
687+ loss_scale_manager = FixedLossScaleManager(loss_scale=self.bert_loss_scale)
688+ else:
689+ raise ValueError("Invalid loss scale: %d" % self.bert_loss_scale)
690+ self.mmgr[key] = loss_scale_manager
691+ # device数是否大于1，如果大于1，进行分布式训练
692+ # if ops_adapter.size() > 1:
693+ # opt_tmp = NPUDistributedOptimizer(opt_tmp)
694+ # opt = NPULossScaleOptimizer(opt_tmp, loss_scale_manager, is_distributed=True)
695+ # else:
696+ opt = NPULossScaleOptimizer(opt_tmp, loss_scale_manager)
697+ return opt
Acontrib/Tensorflow/MUNIT_ID0953_for_TensorFlow/README.md+188-0
@@ -0,0 +1,188 @@
1+- [基本信息](#基本信息.md)
2+- [概述](#概述.md)
3+- [训练环境准备](#训练环境准备.md)
4+- [快速上手](#快速上手.md)
5+- [训练结果](#训练结果.md)
6+- [高级参考](#高级参考.md)
7+<h2 id="基本信息.md">基本信息</h2>
8+ 
9+**发布者（Publisher）：Huawei**
10+ 
11+**应用领域（Application Domain）：Computer Vision**
12+ 
13+**版本（Version）：1.2**
14+ 
15+**修改时间（Modified） ：2022.06.12**
16+ 
17+**大小（Size）：104KB**
18+ 
19+**框架（Framework）：TensorFlow 1.15.0**
20+ 
21+**模型格式（Model Format）：ckpt**
22+ 
23+**精度（Precision）：Mixed**
24+ 
25+**处理器（Processor）：昇腾910**
26+ 
27+**应用级别（Categories）：Official**
28+ 
29+**描述（Description）：基于TensorFlow框架的图像迁移算法**
30+ 
31+<h2 id="概述.md">概述</h2>
32+Munit是2018年提出的多模态无监督图像转换框架，可以从给定的源域图像生成不同风格的目标域图像输出。
33+ 
34+ 
35+- 参考论文：
36+ 
37+ https://arxiv.org/abs/1804.04732
38+ 
39+- 参考实现：
40+ 
41+ https://github.com/taki0112/MUNIT-Tensorflow
42+ 
43+- 适配昇腾 AI 处理器的实现：
44+
45+ https://gitee.com/harry-zzh/modelzoo/edit/master/contrib/Tensorflow/MUNIT_ID0953_for_TensorFlow/
46+
47+ 
48+ 
49+- 通过Git获取对应commit\_id的代码方法如下：
50+
51+ ```
52+ git clone {repository_url} # 克隆仓库的代码
53+ cd {repository_name} # 切换到模型的代码仓目录
54+ git checkout {branch} # 切换到对应分支
55+ git reset --hard ｛commit_id｝ # 代码设置到对应的commit_id
56+ cd ｛code_path｝ # 切换到模型代码所在路径，若仓库下只有该模型，则无需切换
57+ ```
58+ 
59+## 默认配置<a name="section91661242121611"></a>
60+ 
61+- 训练超参
62+ 
63+ - Batch size： 1
64+ - Train epoch: 1
65+ - Train step: 100000
66+ 
67+ 
68+## 支持特性<a name="section1899153513554"></a>
69+ 
70+| 特性列表 | 是否支持 |
71+|-------|------|
72+| 分布式训练 | 否 |
73+| 混合精度 | 否 |
74+| 并行数据 | 否 |
75+ 
76+ 
77+<h2 id="训练环境准备.md">训练环境准备</h2>
78+ 
79+1. 硬件环境准备请参见各硬件产品文档"[驱动和固件安装升级指南]( https://support.huawei.com/enterprise/zh/category/ai-computing-platform-pid-1557196528909)"。需要在硬件设备上安装与CANN版本配套的固件与驱动。
80+2. 宿主机上需要安装Docker并登录[Ascend Hub中心](https://ascendhub.huawei.com/#/detail?name=ascend-tensorflow-arm)获取镜像。
81+ 
82+ 当前模型支持的镜像列表如[表1](#zh-cn_topic_0000001074498056_table1519011227314)所示。
83+ 
84+ **表 1** 镜像列表
85+ 
86+ <a name="zh-cn_topic_0000001074498056_table1519011227314"></a>
87+ <table><thead align="left"><tr id="zh-cn_topic_0000001074498056_row0190152218319"><th class="cellrowborder" valign="top" width="47.32%" id="mcps1.2.4.1.1"><p id="zh-cn_topic_0000001074498056_p1419132211315"><a name="zh-cn_topic_0000001074498056_p1419132211315"></a><a name="zh-cn_topic_0000001074498056_p1419132211315"></a><em id="i1522884921219"><a name="i1522884921219"></a><a name="i1522884921219"></a>镜像名称</em></p>
88+ </th>
89+ <th class="cellrowborder" valign="top" width="25.52%" id="mcps1.2.4.1.2"><p id="zh-cn_topic_0000001074498056_p75071327115313"><a name="zh-cn_topic_0000001074498056_p75071327115313"></a><a name="zh-cn_topic_0000001074498056_p75071327115313"></a><em id="i1522994919122"><a name="i1522994919122"></a><a name="i1522994919122"></a>镜像版本</em></p>
90+ </th>
91+ <th class="cellrowborder" valign="top" width="27.16%" id="mcps1.2.4.1.3"><p id="zh-cn_topic_0000001074498056_p1024411406234"><a name="zh-cn_topic_0000001074498056_p1024411406234"></a><a name="zh-cn_topic_0000001074498056_p1024411406234"></a><em id="i723012493123"><a name="i723012493123"></a><a name="i723012493123"></a>配套CANN版本</em></p>
92+ </th>
93+ </tr>
94+ </thead>
95+ <tbody><tr id="zh-cn_topic_0000001074498056_row71915221134"><td class="cellrowborder" valign="top" width="47.32%" headers="mcps1.2.4.1.1 "><a name="zh-cn_topic_0000001074498056_ul81691515131910"></a><a name="zh-cn_topic_0000001074498056_ul81691515131910"></a><ul id="zh-cn_topic_0000001074498056_ul81691515131910"><li><em id="i82326495129"><a name="i82326495129"></a><a name="i82326495129"></a>ARM架构：<a href="https://ascend.huawei.com/ascendhub/#/detail?name=ascend-tensorflow-arm" target="_blank" rel="noopener noreferrer">ascend-tensorflow-arm</a></em></li><li><em id="i18233184918125"><a name="i18233184918125"></a><a name="i18233184918125"></a>x86架构：<a href="https://ascend.huawei.com/ascendhub/#/detail?name=ascend-tensorflow-x86" target="_blank" rel="noopener noreferrer">ascend-tensorflow-x86</a></em></li></ul>
96+ </td>
97+ <td class="cellrowborder" valign="top" width="25.52%" headers="mcps1.2.4.1.2 "><p id="zh-cn_topic_0000001074498056_p1450714271532"><a name="zh-cn_topic_0000001074498056_p1450714271532"></a><a name="zh-cn_topic_0000001074498056_p1450714271532"></a><em id="i72359495125"><a name="i72359495125"></a><a name="i72359495125"></a>20.2.0</em></p>
98+ </td>
99+ <td class="cellrowborder" valign="top" width="27.16%" headers="mcps1.2.4.1.3 "><p id="zh-cn_topic_0000001074498056_p18244640152312"><a name="zh-cn_topic_0000001074498056_p18244640152312"></a><a name="zh-cn_topic_0000001074498056_p18244640152312"></a><em id="i162363492129"><a name="i162363492129"></a><a name="i162363492129"></a><a href="https://support.huawei.com/enterprise/zh/ascend-computing/cann-pid-251168373/software" target="_blank" rel="noopener noreferrer">20.2</a></em></p>
100+ </td>
101+ </tr>
102+ </tbody>
103+ </table>
104+ 
105+ 
106+<h2 id="快速上手.md">快速上手</h2>
107+ 
108+- 数据集准备
109+1. 模型训练使用edges2shoes数据集，数据集请用户自行获取。
110+ 
111+## 模型训练<a name="section715881518135"></a>
112+ 
113+- 单击“立即下载”，并选择合适的下载方式下载源码包。
114+ 
115+- 启动训练之前，首先要配置程序运行相关环境变量。
116+ 
117+ 环境变量配置信息参见：
118+ 
119+ [Ascend 910训练平台环境变量设置](https://gitee.com/ascend/modelzoo/wikis/Ascend%20910%E8%AE%AD%E7%BB%83%E5%B9%B3%E5%8F%B0%E7%8E%AF%E5%A2%83%E5%8F%98%E9%87%8F%E8%AE%BE%E7%BD%AE?sort_id=3148819)
120+ 
121+- 单卡训练
122+ 
123+ 1. 配置训练参数。
124+ 
125+ 首先在脚本test/train_full_1p.sh中，配置batch_size、steps、epochs、data_path等参数，请用户根据实际路径配置data_path，或者在启动训练的命令行中以参数形式下发。
126+ 
127+ ```
128+ batch_size=1
129+ train_steps=100000
130+ epochs=1
131+ data_path="./dataset/edges2shoes/train"
132+ ```
133+ 
134+ 2. 启动训练。
135+ 
136+ 启动单卡训练 （脚本为MUNIT_ID0953_for_TensorFlow/test/train_full_1p.sh）
137+ 
138+ ```
139+ bash train_full_1p.sh
140+ ```
141+ 
142+<h2 id="训练结果.md">训练结果</h2>
143+ 
144+- 精度结果比对
145+ 
146+取训练最后1000个steps的loss，计算平均值，进行结果比对。
147+ 
148+|精度指标项|GPU实测|NPU实测|
149+|---|---|---|
150+|d_loss|2.619421507950002|2.7996314894200007|
151+|g_loss|4.192780654629998|4.389258856830003|
152+ 
153+ 
154+<h2 id="高级参考.md">高级参考</h2>
155+ 
156+## 脚本和示例代码<a name="section08421615141513"></a>
157+ 
158+```
159+├── MUNIT.py //网络训练与测试代码
160+├── main.py //主函数设置代码
161+├── ops.py //基础模块代码
162+├── utils.py //工具函数代码
163+├── README.md //代码说明文档
164+├── test
165+│ ├──train_performance_1p.sh //单卡训练验证性能启动脚本
166+│ ├──train_full_1p.sh //单卡全量训练启动脚本
167+ 
168+```
169+ 
170+## 脚本参数<a name="section6669162441511"></a>
171+ 
172+```
173+--data_path 数据集路径，默认：./dataset/edges2shoes/train
174+--phase 运行模式，默认：train
175+--batch_size 每个NPU的batch size，默认：1
176+--learing_rate 初始学习率，默认：0.001
177+--iteration 每个epcoh训练步数，默认：100000
178+--epoch 训练epcoh数量，默认：1
179+--result 结果输出路径，默认：./test/output/${ASCEND_DEVICE_ID}
180+```
181+ 
182+## 训练过程<a name="section1589455252218"></a>
183+ 
184+1. 通过“模型训练”中的训练指令启动单卡卡训练。
185+ 
186+2. 参考脚本的模型存储路径为./test/output/${ASCEND_DEVICE_ID}/checkpoint/MUNIT_lsgan。
187+ 
188+ 
Acontrib/Tensorflow/MUNIT_ID0953_for_TensorFlow/main.py+212-0
@@ -0,0 +1,212 @@
1+# Copyright 2017 The TensorFlow Authors. All Rights Reserved.
2+#
3+# Licensed under the Apache License, Version 2.0 (the "License");
4+# you may not use this file except in compliance with the License.
5+# You may obtain a copy of the License at
6+#
7+# http://www.apache.org/licenses/LICENSE-2.0
8+#
9+# Unless required by applicable law or agreed to in writing, software
10+# distributed under the License is distributed on an "AS IS" BASIS,
11+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12+# See the License for the specific language governing permissions and
13+# limitations under the License.
14+# ============================================================================
15+# Copyright 2021 Huawei Technologies Co., Ltd
16+#
17+# Licensed under the Apache License, Version 2.0 (the "License");
18+# you may not use this file except in compliance with the License.
19+# You may obtain a copy of the License at
20+#
21+# http://www.apache.org/licenses/LICENSE-2.0
22+#
23+# Unless required by applicable law or agreed to in writing, software
24+# distributed under the License is distributed on an "AS IS" BASIS,
25+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
26+# See the License for the specific language governing permissions and
27+# limitations under the License.
28+ 
29+from npu_bridge.npu_init import *
30+import tensorflow as tf
31+from MUNIT import MUNIT
32+import argparse
33+from utils import *
34+# from help_modelarts import modelarts_result2obs
35+import precision_tool.tf_config as npu_tf_config
36+ 
37+"""parsing and configuration"""
38+def parse_args():
39+ desc = "Tensorflow implementation of MUNIT"
40+ parser = argparse.ArgumentParser(description=desc)
41+ parser.add_argument('--code_dir', type=str, default='code', help='code_dir')
42+ parser.add_argument('--phase', type=str, default='train', help='train or test or guide')
43+ parser.add_argument('--dataset', type=str, default='summer2winter', help='dataset_name')
44+ parser.add_argument('--data_path', type=str, default='summer2winter', help='dataset_name')
45+ parser.add_argument('--augment_flag', type=bool, default=False, help='Image augmentation use or not')
46+ parser.add_argument('--obs_dir', type=str, default='./output/', help='obs_dir')
47+ 
48+ parser.add_argument('--epoch', type=int, default=10, help='The number of epochs to run')
49+ parser.add_argument('--iteration', type=int, default=100000, help='The number of training iterations')
50+ parser.add_argument('--batch_size', type=int, default=1, help='The batch size')
51+ parser.add_argument('--print_freq', type=int, default=1000, help='The number of image_print_freq')
52+ parser.add_argument('--save_freq', type=int, default=1000, help='The number of ckpt_save_freq')
53+ parser.add_argument('--num_style', type=int, default=3, help='number of styles to sample')
54+ parser.add_argument('--direction', type=str, default='a2b', help='direction of style guided image translation')
55+ parser.add_argument('--guide_img', type=str, default='guide.jpg', help='Style guided image translation')
56+ 
57+ parser.add_argument('--gan_type', type=str, default='lsgan', help='GAN loss type [gan / lsgan]')
58+ 
59+ parser.add_argument('--lr', type=float, default=0.0001, help='The learning rate')
60+ parser.add_argument('--gan_w', type=float, default=1.0, help='weight of adversarial loss')
61+ parser.add_argument('--recon_x_w', type=float, default=10.0, help='weight of image reconstruction loss')
62+ parser.add_argument('--recon_s_w', type=float, default=1.0, help='weight of style reconstruction loss')
63+ parser.add_argument('--recon_c_w', type=float, default=1.0, help='weight of content reconstruction loss')
64+ parser.add_argument('--recon_x_cyc_w', type=float, default=0.0, help='weight of explicit style augmented cycle consistency loss')
65+ 
66+ parser.add_argument('--ch', type=int, default=64, help='base channel number per layer')
67+ parser.add_argument('--style_dim', type=int, default=8, help='length of style code')
68+ parser.add_argument('--n_sample', type=int, default=2, help='number of sampling layers in content encoder')
69+ parser.add_argument('--n_res', type=int, default=4, help='number of residual blocks in content encoder/decoder')
70+ 
71+ parser.add_argument('--n_dis', type=int, default=4, help='number of discriminator layer')
72+ parser.add_argument('--n_scale', type=int, default=3, help='number of scales')
73+ 
74+ parser.add_argument('--img_h', type=int, default=256, help='The size of image hegiht')
75+ parser.add_argument('--img_w', type=int, default=256, help='The size of image width')
76+ parser.add_argument('--img_ch', type=int, default=3, help='The size of image channel')
77+ 
78+ parser.add_argument('--result', type=str, default='results',
79+ help='Directory name to save the results')
80+ parser.add_argument('--checkpoint_dir', type=str, default='checkpoint',
81+ help='Directory name to save the checkpoints')
82+ parser.add_argument('--result_dir', type=str, default='results',
83+ help='Directory name to save the generated images')
84+ parser.add_argument('--log_dir', type=str, default='logs',
85+ help='Directory name to save training logs')
86+ parser.add_argument('--sample_dir', type=str, default='samples',
87+ help='Directory name to save the samples on training')
88+ 
89+ # parser.add_argument('--use_fp16', type=bool, default=True)
90+ parser.add_argument('--bert_loss_scale', type=int, default=0)
91+ 
92+ 
93+ return check_args(parser.parse_args())
94+ 
95+"""checking arguments"""
96+def check_args(args):
97+ # --checkpoint_dir
98+ args.checkpoint_dir = os.path.join(args.result, args.checkpoint_dir)
99+ check_folder(args.checkpoint_dir)
100+ 
101+ # --result_dir
102+ args.result_dir = os.path.join(args.result, args.result_dir)
103+ check_folder(args.result_dir)
104+ 
105+ # --result_dir
106+ args.log_dir = os.path.join(args.result, args.log_dir)
107+ check_folder(args.log_dir)
108+ 
109+ # --dump_dir
110+ args.dump_dir = os.path.join(args.result, "dump")
111+ check_folder(args.dump_dir)
112+ 
113+ # --sample_dir
114+ args.sample_dir = os.path.join(args.result, args.sample_dir)
115+ check_folder(args.sample_dir)
116+ 
117+ # --epoch
118+ try:
119+ assert args.epoch >= 1
120+ except:
121+ print('number of epochs must be larger than or equal to one')
122+ 
123+ # --batch_size
124+ try:
125+ assert args.batch_size >= 1
126+ except:
127+ print('batch size must be larger than or equal to one')
128+ return args
129+ 
130+"""main"""
131+def main():
132+ # parse arguments
133+ args = parse_args()
134+ if args is None:
135+ exit()
136+ 
137+ ############################## npu modify #########################
138+ config = tf.ConfigProto(allow_soft_placement=True)
139+ custom_op = config.graph_options.rewrite_options.custom_optimizers.add()
140+ custom_op.name = "NpuOptimizer"
141+ custom_op.parameter_map["use_off_line"].b = True
142+ # # 混合精度
143+ custom_op.parameter_map["precision_mode"].s = tf.compat.as_bytes("allow_fp32_to_fp16")
144+ #custom_op.parameter_map["precision_mode"].s = tf.compat.as_bytes("force_fp32")
145+ # 算子黑名单
146+ #custom_op.parameter_map["modify_mixlist"].s = tf.compat.as_bytes("./ops_info.json")
147+ # custom_op.parameter_map["modify_mixlist"].s = tf.compat.as_bytes(os.path.join(args.code_dir, "ops_info.json"))
148+ # print(os.path.isfile(os.path.join(args.code_dir, "ops_info.json")))
149+ # print(os.path.join(args.code_dir, "ops_info.json"))
150+ 
151+ # 判断是否溢出
152+ # # dump_path：dump数据存放路径，该参数指定的目录需要在启动训练的环境上（容器或Host侧）提前创建且确保安装时配置的运行用户具有读写权限
153+ # custom_op.parameter_map["dump_path"].s = tf.compat.as_bytes(args.dump_dir)
154+ # # enable_dump_debug：是否开启溢出检测功能
155+ # custom_op.parameter_map["enable_dump_debug"].b = True
156+ # # dump_debug_mode：溢出检测模式，取值：all/aicore_overflow/atomic_overflow
157+ # custom_op.parameter_map["dump_debug_mode"].s = tf.compat.as_bytes("all")
158+ # custom_op = npu_tf_config.update_custom_op(custom_op, action='overflow')
159+ 
160+ # # 关闭全部融合规则
161+ # config = npu_tf_config.session_dump_config(config, action='fusion_off')
162+ 
163+ config.graph_options.rewrite_options.remapping = RewriterConfig.OFF #off remap
164+ config = npu_config_proto(config_proto=config)
165+ 
166+ 
167+ 
168+ # if args.use_fp16 and (args.bert_loss_scale not in [None, -1]):
169+ # opt_tmp = custom_op
170+ # if args.bert_loss_scale == 0:
171+ # loss_scale_manager = ExponentialUpdateLossScaleManager(init_loss_scale=2 ** 32, incr_every_n_steps=1000,
172+ # decr_every_n_nan_or_inf=2, decr_ratio=0.5)
173+ # elif args.bert_loss_scale >= 1:
174+ # loss_scale_manager = FixedLossScaleManager(loss_scale=args.bert_loss_scale)
175+ # else:
176+ # raise ValueError("Invalid loss scale: %d" % args.bert_loss_scale)
177+ # # device数是否大于1，如果大于1，进行分布式训练
178+ # # if ops_adapter.size() > 1:
179+ # # opt_tmp = NPUDistributedOptimizer(opt_tmp)
180+ # # custom_op = NPULossScaleOptimizer(opt_tmp, loss_scale_manager, is_distributed=True)
181+ # # else:
182+ # custom_op = NPULossScaleOptimizer(opt_tmp, loss_scale_manager)
183+ 
184+ # open session
185+ with tf.Session(config=config) as sess:
186+ gan = MUNIT(sess, args)
187+ ############################## npu modify #########################
188+ 
189+ # build graph
190+ gan.build_model()
191+ 
192+ # show network architecture
193+ show_all_variables()
194+ 
195+ if args.phase == 'train' :
196+ # launch the graph in a session
197+ gan.train()
198+ print(" [*] Training finished!")
199+ 
200+ if args.phase == 'test' :
201+ gan.test()
202+ print(" [*] Test finished!")
203+ 
204+ if args.phase == 'guide' :
205+ gan.style_guide_test()
206+ print(" [*] Guide finished!")
207+ 
208+ #modelarts_result2obs(args)
209+ 
210+if __name__ == '__main__':
211+ main()
212+ 
Acontrib/Tensorflow/MUNIT_ID0953_for_TensorFlow/modelarts_entry_acc.py+63-0
@@ -0,0 +1,63 @@
1+# Copyright 2017 The TensorFlow Authors. All Rights Reserved.
2+#
3+# Licensed under the Apache License, Version 2.0 (the "License");
4+# you may not use this file except in compliance with the License.
5+# You may obtain a copy of the License at
6+#
7+# http://www.apache.org/licenses/LICENSE-2.0
8+#
9+# Unless required by applicable law or agreed to in writing, software
10+# distributed under the License is distributed on an "AS IS" BASIS,
11+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12+# See the License for the specific language governing permissions and
13+# limitations under the License.
14+# ============================================================================
15+# Copyright 2021 Huawei Technologies Co., Ltd
16+#
17+# Licensed under the Apache License, Version 2.0 (the "License");
18+# you may not use this file except in compliance with the License.
19+# You may obtain a copy of the License at
20+#
21+# http://www.apache.org/licenses/LICENSE-2.0
22+#
23+# Unless required by applicable law or agreed to in writing, software
24+# distributed under the License is distributed on an "AS IS" BASIS,
25+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
26+# See the License for the specific language governing permissions and
27+# limitations under the License.
28+ 
29+import os
30+import argparse
31+import sys
32+ 
33+# 解析输入参数data_url
34+parser = argparse.ArgumentParser()
35+parser.add_argument("--data_url", type=str, default="/home/ma-user/modelarts/inputs/data_url_0")
36+parser.add_argument("--train_url", type=str, default="/home/ma-user/modelarts/outputs/train_url_0/")
37+config = parser.parse_args()
38+ 
39+print("[CANN-Modelzoo] code_dir path is [%s]" % (sys.path[0]))
40+code_dir = sys.path[0]
41+os.chdir(code_dir)
42+print("[CANN-Modelzoo] work_dir path is [%s]" % (os.getcwd()))
43+ 
44+print("[CANN-Modelzoo] before train - list my run files:")
45+os.system("ls -al /usr/local/Ascend/ascend-toolkit/")
46+ 
47+print("[CANN-Modelzoo] before train - list my dataset files:")
48+os.system("ls -al %s" % config.data_url)
49+ 
50+print("[CANN-Modelzoo] start run train shell")
51+# 设置sh文件格式为linux可执行
52+os.system("dos2unix ./test/*")
53+ 
54+# 执行train_full_1p.sh或者train_performance_1p.sh，需要用户自己指定
55+# full和performance的差异，performance只需要执行很少的step，控制在15分钟以内，主要关注性能FPS
56+os.system("bash ./test/train_full_1p.sh --data_path=%s --output_path=%s " % (config.data_url, config.train_url))
57+ 
58+print("[CANN-Modelzoo] finish run train shell")
59+ 
60+# 将当前执行目录所有文件拷贝到obs的output进行备份
61+print("[CANN-Modelzoo] after train - list my output files:")
62+os.system("cp -r %s %s " % (code_dir, config.train_url))
63+os.system("ls -al %s" % config.train_url)
Acontrib/Tensorflow/MUNIT_ID0953_for_TensorFlow/modelarts_entry_perf.py+63-0
@@ -0,0 +1,63 @@
1+# Copyright 2017 The TensorFlow Authors. All Rights Reserved.
2+#
3+# Licensed under the Apache License, Version 2.0 (the "License");
4+# you may not use this file except in compliance with the License.
5+# You may obtain a copy of the License at
6+#
7+# http://www.apache.org/licenses/LICENSE-2.0
8+#
9+# Unless required by applicable law or agreed to in writing, software
10+# distributed under the License is distributed on an "AS IS" BASIS,
11+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12+# See the License for the specific language governing permissions and
13+# limitations under the License.
14+# ============================================================================
15+# Copyright 2021 Huawei Technologies Co., Ltd
16+#
17+# Licensed under the Apache License, Version 2.0 (the "License");
18+# you may not use this file except in compliance with the License.
19+# You may obtain a copy of the License at
20+#
21+# http://www.apache.org/licenses/LICENSE-2.0
22+#
23+# Unless required by applicable law or agreed to in writing, software
24+# distributed under the License is distributed on an "AS IS" BASIS,
25+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
26+# See the License for the specific language governing permissions and
27+# limitations under the License.
28+ 
29+import os
30+import argparse
31+import sys
32+ 
33+# 解析输入参数data_url
34+parser = argparse.ArgumentParser()
35+parser.add_argument("--data_url", type=str, default="/home/ma-user/modelarts/inputs/data_url_0")
36+parser.add_argument("--train_url", type=str, default="/home/ma-user/modelarts/outputs/train_url_0/")
37+config = parser.parse_args()
38+ 
39+print("[CANN-Modelzoo] code_dir path is [%s]" % (sys.path[0]))
40+code_dir = sys.path[0]
41+os.chdir(code_dir)
42+print("[CANN-Modelzoo] work_dir path is [%s]" % (os.getcwd()))
43+ 
44+print("[CANN-Modelzoo] before train - list my run files:")
45+os.system("ls -al /usr/local/Ascend/ascend-toolkit/")
46+ 
47+print("[CANN-Modelzoo] before train - list my dataset files:")
48+os.system("ls -al %s" % config.data_url)
49+ 
50+print("[CANN-Modelzoo] start run train shell")
51+# 设置sh文件格式为linux可执行
52+os.system("dos2unix ./test/*")
53+ 
54+# 执行train_full_1p.sh或者train_performance_1p.sh，需要用户自己指定
55+# full和performance的差异，performance只需要执行很少的step，控制在15分钟以内，主要关注性能FPS
56+os.system("bash ./test/train_performance_1p.sh --data_path=%s --output_path=%s " % (config.data_url, config.train_url))
57+ 
58+print("[CANN-Modelzoo] finish run train shell")
59+ 
60+# 将当前执行目录所有文件拷贝到obs的output进行备份
61+print("[CANN-Modelzoo] after train - list my output files:")
62+os.system("cp -r %s %s " % (code_dir, config.train_url))
63+os.system("ls -al %s" % config.train_url)
Acontrib/Tensorflow/MUNIT_ID0953_for_TensorFlow/modelzoo_level.txt+3-0
@@ -0,0 +1,3 @@
1+FuncStatus:OK
2+PerfStatus:NOK
3+PrecisionStatus:OK
Acontrib/Tensorflow/MUNIT_ID0953_for_TensorFlow/ops.py+243-0
@@ -0,0 +1,243 @@
1+# Copyright 2017 The TensorFlow Authors. All Rights Reserved.
2+#
3+# Licensed under the Apache License, Version 2.0 (the "License");
4+# you may not use this file except in compliance with the License.
5+# You may obtain a copy of the License at
6+#
7+# http://www.apache.org/licenses/LICENSE-2.0
8+#
9+# Unless required by applicable law or agreed to in writing, software
10+# distributed under the License is distributed on an "AS IS" BASIS,
11+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12+# See the License for the specific language governing permissions and
13+# limitations under the License.
14+# ============================================================================
15+# Copyright 2021 Huawei Technologies Co., Ltd
16+#
17+# Licensed under the Apache License, Version 2.0 (the "License");
18+# you may not use this file except in compliance with the License.
19+# You may obtain a copy of the License at
20+#
21+# http://www.apache.org/licenses/LICENSE-2.0
22+#
23+# Unless required by applicable law or agreed to in writing, software
24+# distributed under the License is distributed on an "AS IS" BASIS,
25+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
26+# See the License for the specific language governing permissions and
27+# limitations under the License.
28+ 
29+from npu_bridge.npu_init import *
30+import tensorflow as tf
31+import tensorflow.contrib as tf_contrib
32+from utils import pytorch_kaiming_weight_factor
33+ 
34+factor, mode, uniform = pytorch_kaiming_weight_factor(a=0.0, uniform=False)
35+weight_init = tf_contrib.layers.variance_scaling_initializer(factor=factor, mode=mode, uniform=uniform)
36+weight_regularizer = tf_contrib.layers.l2_regularizer(scale=0.0001)
37+ 
38+##################################################################################
39+# Layer
40+##################################################################################
41+ 
42+def conv(x, channels, kernel=4, stride=2, pad=0, pad_type='zero', use_bias=True, scope='conv'):
43+ with tf.variable_scope(scope):
44+ if scope.__contains__("discriminator") :
45+ weight_init = tf.random_normal_initializer(mean=0.0, stddev=0.02)
46+ else :
47+ weight_init = tf_contrib.layers.variance_scaling_initializer()
48+ 
49+ if pad > 0:
50+ h = x.get_shape().as_list()[1]
51+ if h % stride == 0:
52+ pad = pad * 2
53+ else:
54+ pad = max(kernel - (h % stride), 0)
55+ 
56+ pad_top = pad // 2
57+ pad_bottom = pad - pad_top
58+ pad_left = pad // 2
59+ pad_right = pad - pad_left
60+ 
61+ if pad_type == 'zero':
62+ x = tf.pad(x, [[0, 0], [pad_top, pad_bottom], [pad_left, pad_right], [0, 0]])
63+ if pad_type == 'reflect':
64+ x = tf.pad(x, [[0, 0], [pad_top, pad_bottom], [pad_left, pad_right], [0, 0]], mode='REFLECT')
65+ 
66+ x = tf.layers.conv2d(inputs=x, filters=channels,
67+ kernel_size=kernel, kernel_initializer=weight_init,
68+ kernel_regularizer=weight_regularizer,
69+ strides=stride, use_bias=use_bias)
70+ 
71+ return x
72+ 
73+def fully_connected(x, units, use_bias=True, scope='fully_connected'):
74+ with tf.variable_scope(scope):
75+ x = flatten(x)
76+ x = tf.layers.dense(x, units=units, kernel_initializer=weight_init,
77+ kernel_regularizer=weight_regularizer,
78+ use_bias=use_bias)
79+ 
80+ return x
81+ 
82+def flatten(x) :
83+ return tf.layers.flatten(x)
84+ 
85+##################################################################################
86+# Residual-block
87+##################################################################################
88+ 
89+def resblock(x_init, channels, use_bias=True, scope='resblock'):
90+ with tf.variable_scope(scope):
91+ with tf.variable_scope('res1'):
92+ x = conv(x_init, channels, kernel=3, stride=1, pad=1, pad_type='reflect', use_bias=use_bias)
93+ x = instance_norm(x)
94+ x = relu(x)
95+ 
96+ with tf.variable_scope('res2'):
97+ x = conv(x, channels, kernel=3, stride=1, pad=1, pad_type='reflect', use_bias=use_bias)
98+ x = instance_norm(x)
99+ 
100+ return x + x_init
101+ 
102+def adaptive_resblock(x_init, channels, gamma1, beta1, gamma2, beta2, use_bias=True, scope='adaptive_resblock') :
103+ with tf.variable_scope(scope):
104+ with tf.variable_scope('res1'):
105+ x = conv(x_init, channels, kernel=3, stride=1, pad=1, pad_type='reflect', use_bias=use_bias)
106+ x = adaptive_instance_norm(x, gamma1, beta1)
107+ x = relu(x)
108+ 
109+ with tf.variable_scope('res2'):
110+ x = conv(x, channels, kernel=3, stride=1, pad=1, pad_type='reflect', use_bias=use_bias)
111+ x = adaptive_instance_norm(x, gamma2, beta2)
112+ 
113+ return x + x_init
114+ 
115+##################################################################################
116+# Sampling
117+##################################################################################
118+ 
119+def down_sample(x) :
120+ return tf.layers.average_pooling2d(x, pool_size=3, strides=2, padding='SAME')
121+ 
122+def up_sample(x, scale_factor=2):
123+ _, h, w, _ = x.get_shape().as_list()
124+ new_size = [h * scale_factor, w * scale_factor]
125+ return tf.image.resize_nearest_neighbor(x, size=new_size)
126+ 
127+def adaptive_avg_pooling(x):
128+ # global average pooling
129+ gap = tf.reduce_mean(x, axis=[1, 2], keep_dims=True)
130+ 
131+ return gap
132+ 
133+##################################################################################
134+# Activation function
135+##################################################################################
136+ 
137+def lrelu(x, alpha=0.01):
138+ # pytorch alpha is 0.01
139+ return tf.nn.leaky_relu(x, alpha)
140+ 
141+ 
142+def relu(x):
143+ return tf.nn.relu(x)
144+ 
145+ 
146+def tanh(x):
147+ return tf.tanh(x)
148+ 
149+##################################################################################
150+# Normalization function
151+##################################################################################
152+ 
153+def adaptive_instance_norm(content, gamma, beta, epsilon=1e-5):
154+ # gamma, beta = style_mean, style_std from MLP
155+ 
156+ c_mean, c_var = tf.nn.moments(content, axes=[1, 2], keep_dims=True)
157+ c_std = tf.sqrt(c_var + epsilon)
158+ 
159+ return gamma * ((content - c_mean) / c_std) + beta
160+ 
161+ 
162+def instance_norm(x, scope='instance_norm'):
163+ return tf_contrib.layers.instance_norm(x,
164+ epsilon=1e-05,
165+ center=True, scale=True,
166+ scope=scope)
167+ 
168+def layer_norm(x, scope='layer_norm') :
169+ return tf_contrib.layers.layer_norm(x,
170+ center=True, scale=True,
171+ scope=scope)
172+ 
173+##################################################################################
174+# Loss function
175+##################################################################################
176+ 
177+"""
178+ 
179+Author use LSGAN
180+For LSGAN, multiply each of G and D by 0.5.
181+However, MUNIT authors did not do this.
182+ 
183+"""
184+ 
185+def discriminator_loss(type, real, fake):
186+ n_scale = len(real)
187+ loss = []
188+ 
189+ real_loss = 0
190+ fake_loss = 0
191+ 
192+ for i in range(n_scale) :
193+ if type == 'lsgan' :
194+ real_loss = tf.reduce_mean(tf.squared_difference(real[i], 1.0))
195+ fake_loss = tf.reduce_mean(tf.square(fake[i]))
196+ 
197+ if type == 'gan' :
198+ real_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(labels=tf.ones_like(real[i]), logits=real[i]))
199+ fake_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(labels=tf.zeros_like(fake[i]), logits=fake[i]))
200+ 
201+ loss.append(real_loss + fake_loss)
202+ 
203+ return sum(loss)
204+ 
205+ 
206+def generator_loss(type, fake):
207+ n_scale = len(fake)
208+ loss = []
209+ 
210+ fake_loss = 0
211+ 
212+ for i in range(n_scale) :
213+ if type == 'lsgan' :
214+ fake_loss = tf.reduce_mean(tf.squared_difference(fake[i], 1.0))
215+ 
216+ if type == 'gan' :
217+ fake_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(labels=tf.ones_like(fake[i]), logits=fake[i]))
218+ 
219+ loss.append(fake_loss)
220+ 
221+ 
222+ return sum(loss)
223+ 
224+ 
225+def L1_loss(x, y):
226+ loss = tf.reduce_mean(tf.abs(x - y))
227+ 
228+ return loss
229+ 
230+def regularization_loss(scope_name) :
231+ """
232+ If you want to use "Regularization"
233+ g_loss += regularization_loss('generator')
234+ d_loss += regularization_loss('discriminator')
235+ """
236+ collection_regularization = tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)
237+ 
238+ loss = []
239+ for item in collection_regularization :
240+ if scope_name in item.name :
241+ loss.append(item)
242+ 
243+ return tf.reduce_sum(loss)
Acontrib/Tensorflow/MUNIT_ID0953_for_TensorFlow/test/train_full_1p.sh+225-0
@@ -0,0 +1,225 @@
1+#!/bin/bash
2+ 
3+##########################################################
4+#########第3行 至 100行，请一定不要、不要、不要修改##########
5+#########第3行 至 100行，请一定不要、不要、不要修改##########
6+#########第3行 至 100行，请一定不要、不要、不要修改##########
7+##########################################################
8+# shell脚本所在路径
9+cur_path=`echo $(cd $(dirname $0);pwd)`
10+ 
11+# 判断当前shell是否是performance
12+perf_flag=`echo $0 | grep performance | wc -l`
13+ 
14+# 当前执行网络的名称
15+Network=`echo $(cd $(dirname $0);pwd) | awk -F"/" '{print $(NF-1)}'`
16+ 
17+export RANK_SIZE=1
18+export RANK_ID=0
19+export JOB_ID=10087
20+ 
21+# 路径参数初始化
22+data_path=""
23+output_path=""
24+ 
25+# 帮助信息，不需要修改
26+if [[ $1 == --help || $1 == -h ]];then
27+ echo"usage:./train_performance_1P.sh <args>"
28+ echo " "
29+ echo "parameter explain:
30+ --data_path # dataset of training
31+ --output_path # output of training
32+ --train_steps # max_step for training
33+ --train_epochs # max_epoch for training
34+ --batch_size # batch size
35+ -h/--help show help message
36+ "
37+ exit 1
38+fi
39+ 
40+# 参数校验，不需要修改
41+for para in $*
42+do
43+ if [[ $para == --data_path* ]];then
44+ data_path=`echo ${para#*=}`
45+ elif [[ $para == --output_path* ]];then
46+ output_path=`echo ${para#*=}`
47+ elif [[ $para == --train_steps* ]];then
48+ train_steps=`echo ${para#*=}`
49+ elif [[ $para == --train_epochs* ]];then
50+ train_epochs=`echo ${para#*=}`
51+ elif [[ $para == --batch_size* ]];then
52+ batch_size=`echo ${para#*=}`
53+ fi
54+done
55+ 
56+# 校验是否传入data_path,不需要修改
57+# data_path="./dataset/edges2shoes/train"
58+if [[ $data_path == "" ]];then
59+ echo "[Error] para \"data_path\" must be config"
60+ exit 1
61+fi
62+ 
63+# 校验是否传入output_path,不需要修改
64+if [[ $output_path == "" ]];then
65+ output_path="./test/output/${ASCEND_DEVICE_ID}"
66+fi
67+ 
68+# 设置打屏日志文件名，请保留，文件名为${print_log}
69+print_log="./test/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log"
70+modelarts_flag=`cat /etc/passwd |grep ma-user`
71+if [ x"${modelarts_flag}" != x ];
72+then
73+ echo "running with modelarts_flag..."
74+ print_log_name=`ls /home/ma-user/modelarts/log/ | grep proc-rank`
75+ print_log="/home/ma-user/modelarts/log/${print_log_name}"
76+fi
77+echo "### get your log here : ${print_log}"
78+ 
79+CaseName=""
80+function get_casename()
81+{
82+ if [ x"${perf_flag}" = x1 ];
83+ then
84+ CaseName=${Network}_bs${batch_size}_${RANK_SIZE}'p'_'perf'
85+ else
86+ CaseName=${Network}_bs${batch_size}_${RANK_SIZE}'p'_'acc'
87+ fi
88+}
89+ 
90+# 跳转到code目录
91+cd ${cur_path}/../
92+rm -rf ./test/output/${ASCEND_DEVICE_ID}
93+mkdir -p ./test/output/${ASCEND_DEVICE_ID}
94+ 
95+# 训练开始时间记录，不需要修改
96+start_time=$(date +%s)
97+##########################################################
98+#########第3行 至 100行，请一定不要、不要、不要修改##########
99+#########第3行 至 100行，请一定不要、不要、不要修改##########
100+#########第3行 至 100行，请一定不要、不要、不要修改##########
101+##########################################################
102+ 
103+#=========================================================
104+#=========================================================
105+#========训练执行命令，需要根据您的网络进行修改==============
106+#=========================================================
107+#=========================================================
108+# 基础参数，需要模型审视修改
109+# 您的训练数据集在${data_path}路径下，请直接使用这个变量获取
110+# 您的训练输出目录在${output_path}路径下，请直接使用这个变量获取
111+# 您的其他基础参数，可以自定义增加，但是batch_size请保留，并且设置正确的值
112+ 
113+# batch_size=64
114+ 
115+# if [ x"${modelarts_flag}" != x ];
116+# then
117+# python3.7 ./LeNet.py --data_path=${data_path} --output_path=${output_path}
118+# else
119+# python3.7 ./LeNet.py --data_path=${data_path} --output_path=${output_path} 1>${print_log} 2>&1
120+# fi
121+ 
122+# # 性能相关数据计算
123+# StepTime=`grep "sec/step :" ${print_log} | tail -n 10 | awk '{print $NF}' | awk '{sum+=$1} END {print sum/NR}'`
124+# FPS=`awk 'BEGIN{printf "%.2f\n", '${batch_size}'/'${StepTime}'}'`
125+ 
126+# # 精度相关数据计算
127+# train_accuracy=`grep "Final Accuracy accuracy" ${print_log} | awk '{print $NF}'`
128+# # 提取所有loss打印信息
129+# grep "loss :" ${print_log} | awk -F ":" '{print $4}' | awk -F "-" '{print $1}' > ./test/output/${ASCEND_DEVICE_ID}/my_output_loss.txt
130+train_epochs=1
131+train_steps=100000
132+batch_size=1
133+epoch=1
134+ 
135+if [ x"${modelarts_flag}" != x ];
136+then
137+ ASCEND_VISIBLE_DEVICES=0 python3.7 ./main.py --data_path=${data_path} --phase train \
138+ --epoch ${epoch} \
139+ --iteration ${train_steps} \
140+ --result=${output_path} \
141+ --batch_size ${batch_size}
142+else
143+ ASCEND_VISIBLE_DEVICES=0 python3.7 ./main.py --data_path=${data_path} --phase train \
144+ --epoch ${epoch} \
145+ --iteration ${train_steps} \
146+ --result=${output_path} \
147+ --batch_size ${batch_size} 1>${print_log} 2>&1
148+fi
149+ 
150+# 性能相关数据计算
151+#StepTime=`grep "sec/step :" ${print_log} | tail -n 10 | awk '{print $NF}' | awk '{sum+=$1} END {print sum/NR}'`
152+step0=`grep time ${print_log} | awk -F"time: " '{print $2}' | awk -F" " 'END{print $1}'`
153+step1=`grep time ${print_log} | awk -F"time: " '{print $2}' | awk -F" " '{print $1}' | tail -2 | head -1`
154+StepTime=`awk 'BEGIN{printf "%.4f",('${step0}'-'${step1}')}'`
155+FPS=`awk 'BEGIN{printf "%.2f\n", '${batch_size}'/'${StepTime}'}'`
156+ 
157+# # 精度相关数据计算
158+# train_accuracy=`grep "Final Accuracy accuracy" ${print_log} | awk '{print $NF}'`
159+# # 提取所有loss打印信息
160+# grep "loss :" ${print_log} | awk -F ":" '{print $4}' | awk -F "-" '{print $1}' > ./test/output/${ASCEND_DEVICE_ID}/my_output_loss.txt
161+# 精度相关数据计算
162+# train_accuracy=`grep "Final Accuracy accuracy" ${print_log} | awk '{print $NF}'`
163+train_accuracy="None"
164+# 提取所有loss打印信息
165+#grep "loss :" ${print_log} | awk -F ":" '{print $4}' | awk -F "-" '{print $1}' > ./test/output/${ASCEND_DEVICE_ID}/my_output_loss.txt
166+grep d_loss ${print_log} | awk -F"d_loss: " '{print $2}' > ./test/output/${ASCEND_DEVICE_ID}/my_output_loss.txt
167+grep g_loss ${print_log} | awk -F"g_loss: " '{print $2}' > ./test/output/${ASCEND_DEVICE_ID}/my_output_loss.txt
168+ 
169+ 
170+###########################################################
171+#########后面的所有内容请不要修改###########################
172+#########后面的所有内容请不要修改###########################
173+#########后面的所有内容请不要修改###########################
174+###########################################################
175+ 
176+# 判断本次执行是否正确使用Ascend NPU
177+tf_flag=`echo ${Network} | grep TensorFlow | wc -l`
178+use_npu_flag=`grep "The model has been compiled on the Ascend AI processor" ${print_log} | wc -l`
179+if [ x"${use_npu_flag}" == x0 -a x"${tf_flag}" == x1 ];
180+then
181+ echo "------------------ ERROR NOTICE START ------------------"
182+ echo "ERROR, your task haven't used Ascend NPU, please check your npu Migration."
183+ echo "------------------ ERROR NOTICE END------------------"
184+else
185+ echo "------------------ INFO NOTICE START------------------"
186+ echo "INFO, your task have used Ascend NPU, please check your result."
187+ echo "------------------ INFO NOTICE END------------------"
188+fi
189+ 
190+# 获取最终的casename，请保留，case文件名为${CaseName}
191+get_casename
192+ 
193+# 重命名loss文件
194+if [ -f ./test/output/${ASCEND_DEVICE_ID}/my_output_loss.txt ];
195+then
196+ mv ./test/output/${ASCEND_DEVICE_ID}/my_output_loss.txt ./test/output/${ASCEND_DEVICE_ID}/${CaseName}_loss.txt
197+fi
198+ 
199+# 训练端到端耗时
200+end_time=$(date +%s)
201+e2e_time=$(( $end_time - $start_time ))
202+ 
203+echo "------------------ Final result ------------------"
204+# 输出性能FPS/单step耗时/端到端耗时
205+echo "Final Performance images/sec : $FPS"
206+echo "Final Performance sec/step : $StepTime"
207+echo "E2E Training Duration sec : $e2e_time"
208+ 
209+# 输出训练精度
210+echo "Final Train Accuracy : ${train_accuracy}"
211+ 
212+# 最后一个迭代loss值，不需要修改
213+ActualLoss=(`awk 'END {print $NF}' $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}_loss.txt`)
214+ 
215+#关键信息打印到${CaseName}.log中，不需要修改
216+echo "Network = ${Network}" > $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
217+echo "RankSize = ${RANK_SIZE}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
218+echo "BatchSize = ${batch_size}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
219+echo "DeviceType = `uname -m`" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
220+echo "CaseName = ${CaseName}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
221+echo "ActualFPS = ${FPS}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
222+echo "TrainingTime = ${StepTime}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
223+echo "ActualLoss = ${ActualLoss}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
224+echo "E2ETrainingTime = ${e2e_time}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
225+echo "TrainAccuracy = ${train_accuracy}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
Acontrib/Tensorflow/MUNIT_ID0953_for_TensorFlow/test/train_performance_1p.sh+207-0
@@ -0,0 +1,207 @@
1+#!/bin/bash
2+ 
3+##########################################################
4+#########第3行 至 100行，请一定不要、不要、不要修改##########
5+#########第3行 至 100行，请一定不要、不要、不要修改##########
6+#########第3行 至 100行，请一定不要、不要、不要修改##########
7+##########################################################
8+# shell脚本所在路径
9+cur_path=`echo $(cd $(dirname $0);pwd)`
10+ 
11+output_path="./edges2shoes_npu_allow_fp32_to_fp16"
12+# 判断当前shell是否是performance
13+perf_flag=`echo $0 | grep performance | wc -l`
14+ 
15+# 当前执行网络的名称
16+Network=`echo $(cd $(dirname $0);pwd) | awk -F"/" '{print $(NF-1)}'`
17+ 
18+export RANK_SIZE=1
19+export RANK_ID=0
20+export JOB_ID=10087
21+ 
22+# 路径参数初始化
23+data_path=""
24+output_path=""
25+ 
26+# 帮助信息，不需要修改
27+if [[ $1 == --help || $1 == -h ]];then
28+ echo"usage:./train_performance_1P.sh <args>"
29+ echo " "
30+ echo "parameter explain:
31+ --data_path # dataset of training
32+ --output_path # output of training
33+ --train_steps # max_step for training
34+ --train_epochs # max_epoch for training
35+ --batch_size # batch size
36+ -h/--help show help message
37+ "
38+ exit 1
39+fi
40+ 
41+# 参数校验，不需要修改
42+for para in $*
43+do
44+ if [[ $para == --data_path* ]];then
45+ data_path=`echo ${para#*=}`
46+ elif [[ $para == --output_path* ]];then
47+ output_path=`echo ${para#*=}`
48+ elif [[ $para == --train_steps* ]];then
49+ train_steps=`echo ${para#*=}`
50+ elif [[ $para == --train_epochs* ]];then
51+ train_epochs=`echo ${para#*=}`
52+ elif [[ $para == --batch_size* ]];then
53+ batch_size=`echo ${para#*=}`
54+ fi
55+done
56+ 
57+# 校验是否传入data_path,不需要修改
58+# data_path="./dataset/edges2shoes/train"
59+if [[ $data_path == "" ]];then
60+ echo "[Error] para \"data_path\" must be config"
61+ exit 1
62+fi
63+ 
64+# 校验是否传入output_path,不需要修改
65+if [[ $output_path == "" ]];then
66+ output_path="./test/output/${ASCEND_DEVICE_ID}"
67+fi
68+ 
69+# 设置打屏日志文件名，请保留，文件名为${print_log}
70+print_log="./test/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log"
71+modelarts_flag=`cat /etc/passwd |grep ma-user`
72+if [ x"${modelarts_flag}" != x ];
73+then
74+ echo "running with modelarts..."
75+ print_log_name=`ls /home/ma-user/modelarts/log/ | grep proc-rank`
76+ print_log="/home/ma-user/modelarts/log/${print_log_name}"
77+fi
78+echo "### get your log here : ${print_log}"
79+ 
80+CaseName=""
81+function get_casename()
82+{
83+ if [ x"${perf_flag}" = x1 ];
84+ then
85+ CaseName=${Network}_bs${batch_size}_${RANK_SIZE}'p'_'perf'
86+ else
87+ CaseName=${Network}_bs${batch_size}_${RANK_SIZE}'p'_'acc'
88+ fi
89+}
90+ 
91+# 跳转到code目录
92+cd ${cur_path}/../
93+rm -rf ./test/output/${ASCEND_DEVICE_ID}
94+mkdir -p ./test/output/${ASCEND_DEVICE_ID}
95+ 
96+# 训练开始时间记录，不需要修改
97+start_time=$(date +%s)
98+##########################################################
99+#########第3行 至 100行，请一定不要、不要、不要修改##########
100+#########第3行 至 100行，请一定不要、不要、不要修改##########
101+#########第3行 至 100行，请一定不要、不要、不要修改##########
102+##########################################################
103+ 
104+#=========================================================
105+#=========================================================
106+#========训练执行命令，需要根据您的网络进行修改==============
107+#=========================================================
108+#=========================================================
109+# 基础参数，需要模型审视修改
110+# 您的训练数据集在${data_path}路径下，请直接使用这个变量获取
111+# 您的训练输出目录在${output_path}路径下，请直接使用这个变量获取
112+# 您的其他基础参数，可以自定义增加，但是batch_size请保留，并且设置正确的值
113+train_epochs=1
114+train_steps=100000
115+batch_size=1
116+epoch=1
117+ 
118+if [ x"${modelarts_flag}" != x ];
119+then
120+ python3.7 ./main.py --data_path=${data_path} --phase train \
121+ --epoch ${epoch} \
122+ --iteration ${train_steps} \
123+ --result=${output_path} \
124+ --batch_size ${batch_size}
125+else
126+ python3.7 ./main.py --data_path=${data_path} --phase train \
127+ --epoch ${epoch} \
128+ --iteration ${train_steps} \
129+ --result=${output_path} \
130+ --batch_size ${batch_size} 1>${print_log} 2>&1
131+fi
132+ 
133+# 性能相关数据计算
134+#StepTime=`grep "sec/step :" ${print_log} | tail -n 10 | awk '{print $NF}' | awk '{sum+=$1} END {print sum/NR}'`
135+step0=`grep time ${print_log} | awk -F"time: " '{print $2}' | awk -F" " 'END{print $1}'`
136+step1=`grep time ${print_log} | awk -F"time: " '{print $2}' | awk -F" " '{print $1}' | tail -2 | head -1`
137+StepTime=`awk 'BEGIN{printf "%.4f",('${step0}'-'${step1}')}'`
138+FPS=`awk 'BEGIN{printf "%.2f\n", '${batch_size}'/'${StepTime}'}'`
139+ 
140+# # 精度相关数据计算
141+# train_accuracy=`grep "Final Accuracy accuracy" ${print_log} | awk '{print $NF}'`
142+# # 提取所有loss打印信息
143+# grep "loss :" ${print_log} | awk -F ":" '{print $4}' | awk -F "-" '{print $1}' > ./test/output/${ASCEND_DEVICE_ID}/my_output_loss.txt
144+# 精度相关数据计算
145+# train_accuracy=`grep "Final Accuracy accuracy" ${print_log} | awk '{print $NF}'`
146+train_accuracy="None"
147+# 提取所有loss打印信息
148+#grep "loss :" ${print_log} | awk -F ":" '{print $4}' | awk -F "-" '{print $1}' > ./test/output/${ASCEND_DEVICE_ID}/my_output_loss.txt
149+grep d_loss ${print_log} | awk -F"d_loss: " '{print $2}' > ./test/output/${ASCEND_DEVICE_ID}/my_output_loss.txt
150+grep g_loss ${print_log} | awk -F"g_loss: " '{print $2}' > ./test/output/${ASCEND_DEVICE_ID}/my_output_loss.txt
151+ 
152+ 
153+###########################################################
154+#########后面的所有内容请不要修改###########################
155+#########后面的所有内容请不要修改###########################
156+#########后面的所有内容请不要修改###########################
157+###########################################################
158+ 
159+# 判断本次执行是否正确使用Ascend NPU
160+tf_flag=`echo ${Network} | grep TensorFlow | wc -l`
161+use_npu_flag=`grep "The model has been compiled on the Ascend AI processor" ${print_log} | wc -l`
162+if [ x"${use_npu_flag}" == x0 -a x"${tf_flag}" == x1 ];
163+then
164+ echo "------------------ ERROR NOTICE START ------------------"
165+ echo "ERROR, your task haven't used Ascend NPU, please check your npu Migration."
166+ echo "------------------ ERROR NOTICE END------------------"
167+else
168+ echo "------------------ INFO NOTICE START------------------"
169+ echo "INFO, your task have used Ascend NPU, please check your result."
170+ echo "------------------ INFO NOTICE END------------------"
171+fi
172+ 
173+# 获取最终的casename，请保留，case文件名为${CaseName}
174+get_casename
175+ 
176+# 重命名loss文件
177+if [ -f ./test/output/${ASCEND_DEVICE_ID}/my_output_loss.txt ];
178+then
179+ mv ./test/output/${ASCEND_DEVICE_ID}/my_output_loss.txt ./test/output/${ASCEND_DEVICE_ID}/${CaseName}_loss.txt
180+fi
181+ 
182+# 训练端到端耗时
183+end_time=$(date +%s)
184+e2e_time=$(( $end_time - $start_time ))
185+ 
186+echo "------------------ Final result ------------------"
187+# 输出性能FPS/单step耗时/端到端耗时
188+echo "Final Performance images/sec : $FPS"
189+echo "Final Performance sec/step : $StepTime"
190+echo "E2E Training Duration sec : $e2e_time"
191+ 
192+# 输出训练精度
193+echo "Final Train Accuracy : ${train_accuracy}"
194+ 
195+# 最后一个迭代loss值，不需要修改
196+ActualLoss=(`awk 'END {print $NF}' $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}_loss.txt`)
197+ 
198+#关键信息打印到${CaseName}.log中，不需要修改
199+echo "Network = ${Network}" > $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
200+echo "RankSize = ${RANK_SIZE}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
201+echo "BatchSize = ${batch_size}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
202+echo "DeviceType = `uname -m`" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
203+echo "CaseName = ${CaseName}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
204+echo "ActualFPS = ${FPS}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
205+echo "TrainingTime = ${StepTime}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
206+echo "ActualLoss = ${ActualLoss}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
207+echo "E2ETrainingTime = ${e2e_time}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
Acontrib/Tensorflow/MUNIT_ID0953_for_TensorFlow/utils.py+145-0
@@ -0,0 +1,145 @@
1+# Copyright 2017 The TensorFlow Authors. All Rights Reserved.
2+#
3+# Licensed under the Apache License, Version 2.0 (the "License");
4+# you may not use this file except in compliance with the License.
5+# You may obtain a copy of the License at
6+#
7+# http://www.apache.org/licenses/LICENSE-2.0
8+#
9+# Unless required by applicable law or agreed to in writing, software
10+# distributed under the License is distributed on an "AS IS" BASIS,
11+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12+# See the License for the specific language governing permissions and
13+# limitations under the License.
14+# ============================================================================
15+# Copyright 2021 Huawei Technologies Co., Ltd
16+#
17+# Licensed under the Apache License, Version 2.0 (the "License");
18+# you may not use this file except in compliance with the License.
19+# You may obtain a copy of the License at
20+#
21+# http://www.apache.org/licenses/LICENSE-2.0
22+#
23+# Unless required by applicable law or agreed to in writing, software
24+# distributed under the License is distributed on an "AS IS" BASIS,
25+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
26+# See the License for the specific language governing permissions and
27+# limitations under the License.
28+ 
29+from npu_bridge.npu_init import *
30+import tensorflow as tf
31+from tensorflow.contrib import slim
32+from scipy import misc
33+import os, random
34+import numpy as np
35+import imageio
36+from skimage.transform import resize
37+ 
38+# https://people.eecs.berkeley.edu/~taesung_park/CycleGAN/datasets/
39+# https://people.eecs.berkeley.edu/~tinghuiz/projects/pix2pix/datasets/
H
Hhtwang2024年12月18日

代码中的链接需要做公网地址申明，避免安全风险

likedislike
40+ 
41+class ImageData:
42+ 
43+ def __init__(self, img_h, img_w, channels, augment_flag=False):
44+ self.img_h = img_h
45+ self.img_w = img_w
46+ self.channels = channels
47+ self.augment_flag = augment_flag
48+ 
49+ def image_processing(self, filename):
50+ x = tf.read_file(filename)
51+ x_decode = tf.image.decode_jpeg(x, channels=self.channels)
52+ img = tf.image.resize_images(x_decode, [self.img_h, self.img_w])
53+ img = tf.cast(img, tf.float32) / 127.5 - 1
54+ 
55+ if self.augment_flag :
56+ augment_size_h = self.img_h + (30 if self.img_h == 256 else 15)
57+ augment_size_w = self.img_w + (30 if self.img_w == 256 else 15)
58+ p = random.random()
59+ if p > 0.5:
60+ img = augmentation(img, augment_size_h, augment_size_w)
61+ 
62+ return img
63+ 
64+ 
65+def load_test_data(image_path, size_h=256, size_w=256):
66+ #img = misc.imread(image_path, mode='RGB')
67+ img = imageio.imread(image_path, pilmode= 'RGB')
68+ #img = misc.imresize(img, [size_h, size_w])
69+ img = resize(img, output_shape=(size_h, size_w))
70+ img = np.expand_dims(img, axis=0)
71+ img = preprocessing(img)
72+ 
73+ return img
74+ 
75+def preprocessing(x):
76+ x = x/127.5 - 1 # -1 ~ 1
77+ return x
78+ 
79+def augmentation(image, aug_img_h, aug_img_w):
80+ seed = random.randint(0, 2 ** 31 - 1)
81+ ori_image_shape = tf.shape(image)
82+ image = tf.image.random_flip_left_right(image, seed=seed)
83+ image = tf.image.resize_images(image, [aug_img_h, aug_img_w])
84+ image = tf.random_crop(image, ori_image_shape, seed=seed)
85+ return image
86+ 
87+def save_images(images, size, image_path):
88+ return imsave(inverse_transform(images), size, image_path)
89+ 
90+def inverse_transform(images):
91+ return (images+1.) / 2
92+ 
93+def imsave(images, size, path):
94+ return imageio.imwrite(path, merge(images, size))
95+ 
96+def merge(images, size):
97+ h, w = images.shape[1], images.shape[2]
98+ img = np.zeros((h * size[0], w * size[1], 3))
99+ for idx, image in enumerate(images):
100+ i = idx % size[1]
101+ j = idx // size[1]
102+ img[h*j:h*(j+1), w*i:w*(i+1), :] = image
103+ 
104+ return img
105+ 
106+def show_all_variables():
107+ model_vars = tf.trainable_variables()
108+ slim.model_analyzer.analyze_vars(model_vars, print_info=True)
109+ 
110+def check_folder(log_dir):
111+ if not os.path.exists(log_dir):
112+ os.makedirs(log_dir)
113+ return log_dir
114+ 
115+def pytorch_xavier_weight_factor(gain=0.02, uniform=False) :
116+ 
117+ if uniform :
118+ factor = gain * gain
119+ mode = 'FAN_AVG'
120+ else :
121+ factor = (gain * gain) / 1.3
122+ mode = 'FAN_AVG'
123+ 
124+ return factor, mode, uniform
125+ 
126+def pytorch_kaiming_weight_factor(a=0.0, activation_function='relu', uniform=False) :
127+ 
128+ if activation_function == 'relu' :
129+ gain = np.sqrt(2.0)
130+ elif activation_function == 'leaky_relu' :
131+ gain = np.sqrt(2.0 / (1 + a ** 2))
132+ elif activation_function =='tanh' :
133+ gain = 5.0 / 3
134+ else :
135+ gain = 1.0
136+ 
137+ if uniform :
138+ factor = gain * gain
139+ mode = 'FAN_IN'
140+ else :
141+ factor = (gain * gain) / 1.3
142+ mode = 'FAN_IN'
143+ 
144+ return factor, mode, uniform
145+