NCTCMumbai commited on
Commit
0b8359d
1 Parent(s): 084d982

Upload 2571 files

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. .gitattributes +7 -0
  2. CTH_CODE_MAP.csv +3 -0
  3. CTH_Description.csv +0 -0
  4. CTH_WISE_DUTY_RATE.csv +0 -0
  5. Checkpoint/assets/vocab.txt +0 -0
  6. Checkpoint/keras_metadata.pb +3 -0
  7. Checkpoint/saved_model.pb +3 -0
  8. Checkpoint/variables/variables.data-00000-of-00001 +3 -0
  9. Checkpoint/variables/variables.index +0 -0
  10. app.py +220 -0
  11. models/.github/ISSUE_TEMPLATE/00-official-bug-report-issue.md +59 -0
  12. models/.github/ISSUE_TEMPLATE/10-official-documentation-issue.md +20 -0
  13. models/.github/ISSUE_TEMPLATE/20-official-feature-request-issue.md +26 -0
  14. models/.github/ISSUE_TEMPLATE/30-research-bug-report-issue.md +58 -0
  15. models/.github/ISSUE_TEMPLATE/40-research-documentation-issue.md +20 -0
  16. models/.github/ISSUE_TEMPLATE/50-research-feature-request-issue.md +26 -0
  17. models/.github/ISSUE_TEMPLATE/60-questions-help-issue.md +14 -0
  18. models/.github/ISSUE_TEMPLATE/config.yml +1 -0
  19. models/.github/PULL_REQUEST_TEMPLATE.md +41 -0
  20. models/.github/README_TEMPLATE.md +122 -0
  21. models/.gitignore +98 -0
  22. models/AUTHORS +10 -0
  23. models/CODEOWNERS +61 -0
  24. models/CONTRIBUTING.md +10 -0
  25. models/ISSUES.md +24 -0
  26. models/LICENSE +203 -0
  27. models/README.md +39 -0
  28. models/official/LICENSE +203 -0
  29. models/official/README-TPU.md +25 -0
  30. models/official/README.md +142 -0
  31. models/official/__init__.py +0 -0
  32. models/official/__pycache__/__init__.cpython-39.pyc +0 -0
  33. models/official/benchmark/__init__.py +0 -0
  34. models/official/benchmark/benchmark_wrappers.py +97 -0
  35. models/official/benchmark/bert_benchmark.py +365 -0
  36. models/official/benchmark/bert_benchmark_utils.py +127 -0
  37. models/official/benchmark/bert_pretrain_benchmark.py +179 -0
  38. models/official/benchmark/bert_squad_benchmark.py +608 -0
  39. models/official/benchmark/datastore/schema/benchmark_metric.json +56 -0
  40. models/official/benchmark/datastore/schema/benchmark_run.json +368 -0
  41. models/official/benchmark/datastore/schema/benchmark_run_status.json +14 -0
  42. models/official/benchmark/keras_benchmark.py +98 -0
  43. models/official/benchmark/keras_cifar_benchmark.py +402 -0
  44. models/official/benchmark/keras_imagenet_benchmark.py +1724 -0
  45. models/official/benchmark/models/__init__.py +0 -0
  46. models/official/benchmark/models/cifar_preprocessing.py +159 -0
  47. models/official/benchmark/models/resnet_cifar_main.py +284 -0
  48. models/official/benchmark/models/resnet_cifar_model.py +262 -0
  49. models/official/benchmark/models/resnet_cifar_test.py +180 -0
  50. models/official/benchmark/models/resnet_imagenet_main.py +301 -0
.gitattributes CHANGED
@@ -33,3 +33,10 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
 
 
 
 
 
 
 
 
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
+ Checkpoint/variables/variables.data-00000-of-00001 filter=lfs diff=lfs merge=lfs -text
37
+ CTH_CODE_MAP.csv filter=lfs diff=lfs merge=lfs -text
38
+ models/research/compression/image_encoder/example.png filter=lfs diff=lfs merge=lfs -text
39
+ models/research/deeplab/testing/pascal_voc_seg/val-00000-of-00001.tfrecord filter=lfs diff=lfs merge=lfs -text
40
+ models/research/lfads/synth_data/trained_itb/model-65000.meta filter=lfs diff=lfs merge=lfs -text
41
+ models/research/object_detection/g3doc/img/kites_with_segment_overlay.png filter=lfs diff=lfs merge=lfs -text
42
+ models/research/object_detection/test_images/image2.jpg filter=lfs diff=lfs merge=lfs -text
CTH_CODE_MAP.csv ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:64a308da335ad4906fc42757542b4283d4025a29fe6afec47148b53187c74a42
3
+ size 218205006
CTH_Description.csv ADDED
The diff for this file is too large to render. See raw diff
 
CTH_WISE_DUTY_RATE.csv ADDED
The diff for this file is too large to render. See raw diff
 
Checkpoint/assets/vocab.txt ADDED
The diff for this file is too large to render. See raw diff
 
Checkpoint/keras_metadata.pb ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:38bb0f1231a198848366566e176c8948dceab7b085b658d00550a83f784731f5
3
+ size 11535
Checkpoint/saved_model.pb ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b0054848283b4fb79fefcebe71830bedb75e023ad04c5655adbc6a2ddd1e2c60
3
+ size 11477628
Checkpoint/variables/variables.data-00000-of-00001 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b4ec6776ca3161577663eaa115fb9f965304670a1af8db7a37e9499a23082e67
3
+ size 1389095096
Checkpoint/variables/variables.index ADDED
Binary file (46.6 kB). View file
 
app.py ADDED
@@ -0,0 +1,220 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import pandas as pd
2
+ import numpy as np
3
+ import tensorflow as tf
4
+ import tensorflow_hub as hub
5
+ import sys
6
+ import random
7
+ sys.path.append('models')
8
+ from official.nlp.data import classifier_data_lib
9
+ from official.nlp.bert import tokenization
10
+ from official.nlp import optimization
11
+ tf.get_logger().setLevel('ERROR')
12
+
13
+ import math
14
+
15
+ import gradio as gr
16
+
17
+ config = tf.compat.v1.ConfigProto(
18
+ device_count = {'cpu': 0}
19
+ )
20
+ sess = tf.compat.v1.Session(config=config)
21
+ num_warmup_steps=1
22
+ num_train_steps=1
23
+ init_lr = 3e-5
24
+ optimizer = optimization.create_optimizer(init_lr=init_lr,
25
+ num_train_steps=num_train_steps,
26
+ num_warmup_steps=num_warmup_steps,
27
+ optimizer_type='adamw')
28
+
29
+ ### Load Model
30
+ checkpoint_filepath=r'./Checkpoint'
31
+ model = tf.keras.models.load_model(checkpoint_filepath, custom_objects={'KerasLayer':hub.KerasLayer , 'AdamWeightDecay': optimizer})
32
+
33
+
34
+
35
+ df_report = pd.read_csv('./CTH_Description.csv')
36
+ df_report['CTH Code'] = df_report['CTH Code'].astype(str).str.zfill(8)
37
+
38
+ df_report_DUTY = pd.read_csv('./CTH_WISE_DUTY_RATE.csv')
39
+ df_report_DUTY['CTH'] = df_report_DUTY['CTH'].astype(str).str.zfill(8)
40
+
41
+ #print(df_report_DUTY)
42
+
43
+ df = pd.read_csv("./CTH_CODE_MAP.csv")
44
+ df['CTH'] = df['CTH'].astype(str).str.zfill(8)
45
+ df = df[['CTH', 'code']]
46
+
47
+ class_names=df[['CTH','code']].drop_duplicates(subset='CTH').sort_values(by='code',ignore_index=True)['CTH'].values.tolist()
48
+ label_list=list(range(0,len(class_names)))
49
+ max_seq_length = 200 # maximum length of (token) input sequences . it can be any number
50
+ train_batch_size = 32 # batch size ( 16 choosen to avoid Out-Of-Memory errors)
51
+
52
+ # Get BERT layer and tokenizer:
53
+ # More details here: https://tfhub.dev/tensorflow/bert_en_uncased_L-12_H-768_A-12/4
54
+ bert_layer = hub.KerasLayer("https://tfhub.dev/tensorflow/bert_en_uncased_L-12_H-768_A-12/4" , trainable = True)
55
+ vocab_file = bert_layer.resolved_object.vocab_file.asset_path.numpy()
56
+ do_lower_case = bert_layer.resolved_object.do_lower_case.numpy()
57
+ tokenizer = tokenization.FullTokenizer(vocab_file , do_lower_case)
58
+
59
+ # This provides a function to convert each row to input features and label ( as required by BERT)
60
+
61
+ max_seq_length = 200 # maximum length of (token) input sequences . it can be any number
62
+ def to_feature(text, label, label_list=label_list, max_seq_length=max_seq_length, tokenizer=tokenizer):
63
+ example = classifier_data_lib.InputExample(guid = None,
64
+ text_a = text.numpy(),
65
+ text_b = None,
66
+ label = label.numpy())
67
+ feature = classifier_data_lib.convert_single_example(0 , example , label_list , max_seq_length , tokenizer)
68
+
69
+ return (feature.input_ids , feature.input_mask , feature.segment_ids , feature.label_id)
70
+
71
+
72
+ def to_feature_map(text, label):
73
+ input_ids , input_mask , segment_ids , label_id = tf.py_function(to_feature , inp = [text , label],
74
+ Tout = [tf.int32 , tf.int32 , tf.int32 , tf.int32])
75
+
76
+ input_ids.set_shape([max_seq_length])
77
+ input_mask.set_shape([max_seq_length])
78
+ segment_ids.set_shape([max_seq_length])
79
+ label_id.set_shape([])
80
+
81
+ x = {
82
+ "input_word_ids": input_ids,
83
+ "input_mask": input_mask,
84
+ "input_type_ids": segment_ids
85
+ }
86
+
87
+ return(x,label_id)
88
+
89
+
90
+
91
+ def print3largest(arr, arr_size):
92
+ third = first = second = -sys.maxsize
93
+ for i in range(0, arr_size):
94
+
95
+ if (arr[i] > first):
96
+ third = second
97
+ second = first
98
+ first = arr[i]
99
+ elif (arr[i] > second):
100
+ third = second
101
+ second = arr[i]
102
+ elif (arr[i] > third):
103
+ third = arr[i]
104
+ pred_value_max_three=[first, second, third]
105
+ return pred_value_max_three
106
+
107
+ def count_special_character(string):
108
+ special_char= 0
109
+ for i in range(len(string)):
110
+ ch = string[i]
111
+ if (string[i].isalpha()):
112
+ continue
113
+ else:
114
+ special_char += 1
115
+
116
+ if len(string)==special_char:
117
+ return False
118
+ else:
119
+ return True
120
+
121
+ def predict_CTH(txt):
122
+ print('Desc: ',txt)
123
+ if (txt!='') and len(txt)>=3 and (count_special_character(txt)):
124
+ valid_data = tf.data.Dataset.from_tensor_slices(([txt] , [1])) # 1 refers to 'entertainment' and 2 refers to 'sport'
125
+ valid_data = (valid_data.map(to_feature_map).batch(1))
126
+ preds = model.predict(valid_data)
127
+ predicted_values = tf.nn.softmax(preds)
128
+ arr = predicted_values.numpy().tolist()[0]
129
+ n = len(arr)
130
+ pred_value_max_three=print3largest(arr, n)
131
+
132
+
133
+
134
+ sum_all = pred_value_max_three[0] + pred_value_max_three[1] + pred_value_max_three[2]
135
+
136
+ val_1 = pred_value_max_three[0]/sum_all
137
+ val_2 = pred_value_max_three[1]/sum_all
138
+ val_3 = pred_value_max_three[2]/sum_all
139
+
140
+ #val_1= 97 #random.randrange(95, 99, 1)
141
+ #val_2=(pred_value_max_three[1]/pred_value_max_three[0])*val_1
142
+ #val_3=(pred_value_max_three[2]/pred_value_max_three[0])*val_1
143
+
144
+ if pred_value_max_three[0]<=0.000131:
145
+ Var_CTH=[]
146
+ Var_desc=[]
147
+ Var_duty=[]
148
+ pred_duty=''
149
+ pred_desc=''
150
+ pred_CTH=''
151
+
152
+ return{'Not a adequate description':float(1.0)}
153
+ else:
154
+ Var_CTH=[]
155
+ Var_desc=[]
156
+ Var_duty=[]
157
+ pred_duty=''
158
+ pred_desc=''
159
+ pred_CTH=''
160
+
161
+
162
+ for i in pred_value_max_three:
163
+ #i=pred_value_max_three[0]
164
+ predicted_code=np.where(predicted_values.numpy()==i)[1][0]
165
+ pred_CTH=df[df['code'] == predicted_code]['CTH'].iloc[0]
166
+
167
+ try:
168
+ pred_duty=df_report_DUTY[df_report_DUTY['CTH']==str(pred_CTH)]['DUTY_RATE'].iloc[0]
169
+ pred_desc=df_report[df_report['CTH Code']==str(pred_CTH)]['Concat Description'].iloc[0]
170
+ except:
171
+ pass
172
+
173
+ Var_CTH.append(pred_CTH)
174
+ Var_desc.append(pred_desc)
175
+ Var_duty.append(pred_duty)
176
+
177
+ P1 ='CTH: '+str(Var_CTH[0])+' Duty Rate(%): '+ str(Var_duty[0])
178
+ P2 ='CTH: '+str(Var_CTH[1])+' Duty Rate(%): '+ str(Var_duty[1])
179
+ P3 ='CTH: '+str(Var_CTH[2])+' Duty Rate(%): '+ str(Var_duty[2])
180
+
181
+
182
+ Q1='Desc: '+str(Var_desc[0])
183
+ Q2='Desc: '+str(Var_desc[1])
184
+ Q3='Desc: '+str(Var_desc[2])
185
+
186
+
187
+ return {str(P1):float(val_1),str(Q1):float(val_1),
188
+ str(P2):float(val_2),str(Q2):float(val_2),
189
+ str(P3):float(val_3),str(Q3):float(val_3),}
190
+ else:
191
+ return{'Enter Correct Description':float(1.0)}
192
+
193
+
194
+ input_txt=gr.Textbox(
195
+ label='Enter Your Product Descrption',
196
+ lines=3,
197
+ )
198
+ description="<p style='color:blue;text-align:justify;font-size:1vw;'>AdvaitBERT is modified version of BERT (Bidirectional Encoder Representation for Transformers), \
199
+ finetuned on the Text corpus of Indian Customs Declarations. It is trained for performing \
200
+ downstream tasks like automating the tariff classification and validation process of Customs \
201
+ declarations in realtime. This model may help Customs administration to efficiently use AI assisted \
202
+ NLP in realtime Customs process like Assessment, Post Clearance Audit, thereby highlighting classification \
203
+ inconsistencies and help in revenue augmentation.</a></p>"
204
+
205
+ title="<h1 style='color:green;text-align:center;font-size:2vw;'>AdvaitBERT </a></h1>"
206
+ article="<p style='color:black;text-align:right;font-size:1vw;'>Powered by NCTC </a></p>"
207
+
208
+ #css=".gradio-container {background-color: papayawhip}",
209
+
210
+ gr.Interface(
211
+ predict_CTH,
212
+ inputs=input_txt,
213
+ outputs="label",
214
+ interpretation="default",
215
+ description=description,
216
+ #live=True,
217
+ examples = ['200 SI/SI/SI LPO ALUMINIUM LIDS (QTY: 8820000 PCS/PRICE: 21.'],
218
+ title=title,
219
+ article=article,
220
+ ).launch()
models/.github/ISSUE_TEMPLATE/00-official-bug-report-issue.md ADDED
@@ -0,0 +1,59 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ name: "[Official Model] Bug Report"
3
+ about: Use this template for reporting a bug for the “official” directory
4
+ labels: type:bug,models:official
5
+
6
+ ---
7
+
8
+ # Prerequisites
9
+
10
+ Please answer the following questions for yourself before submitting an issue.
11
+
12
+ - [ ] I am using the latest TensorFlow Model Garden release and TensorFlow 2.
13
+ - [ ] I am reporting the issue to the correct repository. (Model Garden official or research directory)
14
+ - [ ] I checked to make sure that this issue has not been filed already.
15
+
16
+ ## 1. The entire URL of the file you are using
17
+
18
+ https://github.com/tensorflow/models/tree/master/official/...
19
+
20
+ ## 2. Describe the bug
21
+
22
+ A clear and concise description of what the bug is.
23
+
24
+ ## 3. Steps to reproduce
25
+
26
+ Steps to reproduce the behavior.
27
+
28
+ ## 4. Expected behavior
29
+
30
+ A clear and concise description of what you expected to happen.
31
+
32
+ ## 5. Additional context
33
+
34
+ Include any logs that would be helpful to diagnose the problem.
35
+
36
+ ## 6. System information
37
+
38
+ - OS Platform and Distribution (e.g., Linux Ubuntu 16.04):
39
+ - Mobile device name if the issue happens on a mobile device:
40
+ - TensorFlow installed from (source or binary):
41
+ - TensorFlow version (use command below):
42
+ - Python version:
43
+ - Bazel version (if compiling from source):
44
+ - GCC/Compiler version (if compiling from source):
45
+ - CUDA/cuDNN version:
46
+ - GPU model and memory:
47
+
48
+ <!--
49
+ Collect system information using our environment capture script.
50
+ https://github.com/tensorflow/tensorflow/tree/master/tools/tf_env_collect.sh
51
+
52
+ You can also obtain the TensorFlow version with:
53
+
54
+ 1. TensorFlow 1.0
55
+ `python -c "import tensorflow as tf; print(tf.GIT_VERSION, tf.VERSION)"`
56
+
57
+ 2. TensorFlow 2.0
58
+ `python -c "import tensorflow as tf; print(tf.version.GIT_VERSION, tf.version.VERSION)"`
59
+ -->
models/.github/ISSUE_TEMPLATE/10-official-documentation-issue.md ADDED
@@ -0,0 +1,20 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ name: "[Official Model] Documentation Issue"
3
+ about: Use this template for reporting a documentation issue for the “official” directory
4
+ labels: type:docs,models:official
5
+
6
+ ---
7
+
8
+ # Prerequisites
9
+
10
+ Please answer the following question for yourself before submitting an issue.
11
+
12
+ - [ ] I checked to make sure that this issue has not been filed already.
13
+
14
+ ## 1. The entire URL of the documentation with the issue
15
+
16
+ https://github.com/tensorflow/models/tree/master/official/...
17
+
18
+ ## 2. Describe the issue
19
+
20
+ A clear and concise description of what needs to be changed.
models/.github/ISSUE_TEMPLATE/20-official-feature-request-issue.md ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ name: "[Official Model] Feature request"
3
+ about: Use this template for raising a feature request for the “official” directory
4
+ labels: type:feature,models:official
5
+
6
+ ---
7
+
8
+ # Prerequisites
9
+
10
+ Please answer the following question for yourself before submitting an issue.
11
+
12
+ - [ ] I checked to make sure that this feature has not been requested already.
13
+
14
+ ## 1. The entire URL of the file you are using
15
+
16
+ https://github.com/tensorflow/models/tree/master/official/...
17
+
18
+ ## 2. Describe the feature you request
19
+
20
+ A clear and concise description of what you want to happen.
21
+
22
+ ## 3. Additional context
23
+
24
+ Add any other context about the feature request here.
25
+
26
+ ## 4. Are you willing to contribute it? (Yes or No)
models/.github/ISSUE_TEMPLATE/30-research-bug-report-issue.md ADDED
@@ -0,0 +1,58 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ name: "[Research Model] Bug Report"
3
+ about: Use this template for reporting a bug for the “research” directory
4
+ labels: type:bug,models:research
5
+
6
+ ---
7
+ # Prerequisites
8
+
9
+ Please answer the following questions for yourself before submitting an issue.
10
+
11
+ - [ ] I am using the latest TensorFlow Model Garden release and TensorFlow 2.
12
+ - [ ] I am reporting the issue to the correct repository. (Model Garden official or research directory)
13
+ - [ ] I checked to make sure that this issue has not already been filed.
14
+
15
+ ## 1. The entire URL of the file you are using
16
+
17
+ https://github.com/tensorflow/models/tree/master/research/...
18
+
19
+ ## 2. Describe the bug
20
+
21
+ A clear and concise description of what the bug is.
22
+
23
+ ## 3. Steps to reproduce
24
+
25
+ Steps to reproduce the behavior.
26
+
27
+ ## 4. Expected behavior
28
+
29
+ A clear and concise description of what you expected to happen.
30
+
31
+ ## 5. Additional context
32
+
33
+ Include any logs that would be helpful to diagnose the problem.
34
+
35
+ ## 6. System information
36
+
37
+ - OS Platform and Distribution (e.g., Linux Ubuntu 16.04):
38
+ - Mobile device name if the issue happens on a mobile device:
39
+ - TensorFlow installed from (source or binary):
40
+ - TensorFlow version (use command below):
41
+ - Python version:
42
+ - Bazel version (if compiling from source):
43
+ - GCC/Compiler version (if compiling from source):
44
+ - CUDA/cuDNN version:
45
+ - GPU model and memory:
46
+
47
+ <!--
48
+ Collect system information using our environment capture script.
49
+ https://github.com/tensorflow/tensorflow/tree/master/tools/tf_env_collect.sh
50
+
51
+ You can also obtain the TensorFlow version with:
52
+
53
+ 1. TensorFlow 1.0
54
+ `python -c "import tensorflow as tf; print(tf.GIT_VERSION, tf.VERSION)"`
55
+
56
+ 2. TensorFlow 2.0
57
+ `python -c "import tensorflow as tf; print(tf.version.GIT_VERSION, tf.version.VERSION)"`
58
+ -->
models/.github/ISSUE_TEMPLATE/40-research-documentation-issue.md ADDED
@@ -0,0 +1,20 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ name: "[Research Model] Documentation Issue"
3
+ about: Use this template for reporting a documentation issue for the “research” directory
4
+ labels: type:docs,models:research
5
+
6
+ ---
7
+
8
+ # Prerequisites
9
+
10
+ Please answer the following question for yourself before submitting an issue.
11
+
12
+ - [ ] I checked to make sure that this issue has not been filed already.
13
+
14
+ ## 1. The entire URL of the documentation with the issue
15
+
16
+ https://github.com/tensorflow/models/tree/master/research/...
17
+
18
+ ## 2. Describe the issue
19
+
20
+ A clear and concise description of what needs to be changed.
models/.github/ISSUE_TEMPLATE/50-research-feature-request-issue.md ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ name: "[Research Model] Feature Request"
3
+ about: Use this template for raising a feature request for the “research” directory
4
+ labels: type:feature,models:research
5
+
6
+ ---
7
+
8
+ # Prerequisites
9
+
10
+ Please answer the following question for yourself before submitting an issue.
11
+
12
+ - [ ] I checked to make sure that this feature has not been requested already.
13
+
14
+ ## 1. The entire URL of the file you are using
15
+
16
+ https://github.com/tensorflow/models/tree/master/research/...
17
+
18
+ ## 2. Describe the feature you request
19
+
20
+ A clear and concise description of what you want to happen.
21
+
22
+ ## 3. Additional context
23
+
24
+ Add any other context about the feature request here.
25
+
26
+ ## 4. Are you willing to contribute it? (Yes or No)
models/.github/ISSUE_TEMPLATE/60-questions-help-issue.md ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ name: Questions and Help
3
+ about: Use this template for Questions and Help.
4
+ labels: type:support
5
+
6
+ ---
7
+ <!--
8
+ As per our GitHub Policy (https://github.com/tensorflow/models/blob/master/ISSUES.md), we only address code bugs, documentation issues, and feature requests on GitHub.
9
+
10
+ We will automatically close questions and help related issues.
11
+
12
+ Please go to Stack Overflow (http://stackoverflow.com/questions/tagged/tensorflow-model-garden) for questions and help.
13
+
14
+ -->
models/.github/ISSUE_TEMPLATE/config.yml ADDED
@@ -0,0 +1 @@
 
 
1
+ blank_issues_enabled: false
models/.github/PULL_REQUEST_TEMPLATE.md ADDED
@@ -0,0 +1,41 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Description
2
+
3
+ > :memo: Please include a summary of the change.
4
+ >
5
+ > * Please also include relevant motivation and context.
6
+ > * List any dependencies that are required for this change.
7
+
8
+ ## Type of change
9
+
10
+ For a new feature or function, please create an issue first to discuss it
11
+ with us before submitting a pull request.
12
+
13
+ Note: Please delete options that are not relevant.
14
+
15
+ - [ ] Bug fix (non-breaking change which fixes an issue)
16
+ - [ ] Documentation update
17
+ - [ ] TensorFlow 2 migration
18
+ - [ ] New feature (non-breaking change which adds functionality)
19
+ - [ ] Breaking change (fix or feature that would cause existing functionality to not work as expected)
20
+ - [ ] A new research paper code implementation
21
+ - [ ] Other (Specify)
22
+
23
+ ## Tests
24
+
25
+ > :memo: Please describe the tests that you ran to verify your changes.
26
+ >
27
+ > * Provide instructions so we can reproduce.
28
+ > * Please also list any relevant details for your test configuration.
29
+
30
+ **Test Configuration**:
31
+
32
+ ## Checklist
33
+
34
+ - [ ] I have signed the [Contributor License Agreement](https://github.com/tensorflow/models/wiki/Contributor-License-Agreements).
35
+ - [ ] I have read [guidelines for pull request](https://github.com/tensorflow/models/wiki/Submitting-a-pull-request).
36
+ - [ ] My code follows the [coding guidelines](https://github.com/tensorflow/models/wiki/Coding-guidelines).
37
+ - [ ] I have performed a self [code review](https://github.com/tensorflow/models/wiki/Code-review) of my own code.
38
+ - [ ] I have commented my code, particularly in hard-to-understand areas.
39
+ - [ ] I have made corresponding changes to the documentation.
40
+ - [ ] My changes generate no new warnings.
41
+ - [ ] I have added tests that prove my fix is effective or that my feature works.
models/.github/README_TEMPLATE.md ADDED
@@ -0,0 +1,122 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ > :memo: A README.md template for releasing a paper code implementation to a GitHub repository.
2
+ >
3
+ > * Template version: 1.0.2020.170
4
+ > * Please modify sections depending on needs.
5
+
6
+ # Model name, Paper title, or Project Name
7
+
8
+ > :memo: Add a badge for the ArXiv identifier of your paper (arXiv:YYMM.NNNNN)
9
+
10
+ [![Paper](http://img.shields.io/badge/Paper-arXiv.YYMM.NNNNN-B3181B?logo=arXiv)](https://arxiv.org/abs/...)
11
+
12
+ This repository is the official or unofficial implementation of the following paper.
13
+
14
+ * Paper title: [Paper Title](https://arxiv.org/abs/YYMM.NNNNN)
15
+
16
+ ## Description
17
+
18
+ > :memo: Provide description of the model.
19
+ >
20
+ > * Provide brief information of the algorithms used.
21
+ > * Provide links for demos, blog posts, etc.
22
+
23
+ ## History
24
+
25
+ > :memo: Provide a changelog.
26
+
27
+ ## Authors or Maintainers
28
+
29
+ > :memo: Provide maintainer information.
30
+
31
+ * Full name ([@GitHub username](https://github.com/username))
32
+ * Full name ([@GitHub username](https://github.com/username))
33
+
34
+ ## Table of Contents
35
+
36
+ > :memo: Provide a table of contents to help readers navigate a lengthy README document.
37
+
38
+ ## Requirements
39
+
40
+ [![TensorFlow 2.1](https://img.shields.io/badge/TensorFlow-2.1-FF6F00?logo=tensorflow)](https://github.com/tensorflow/tensorflow/releases/tag/v2.1.0)
41
+ [![Python 3.6](https://img.shields.io/badge/Python-3.6-3776AB)](https://www.python.org/downloads/release/python-360/)
42
+
43
+ > :memo: Provide details of the software required.
44
+ >
45
+ > * Add a `requirements.txt` file to the root directory for installing the necessary dependencies.
46
+ > * Describe how to install requirements using pip.
47
+ > * Alternatively, create INSTALL.md.
48
+
49
+ To install requirements:
50
+
51
+ ```setup
52
+ pip install -r requirements.txt
53
+ ```
54
+
55
+ ## Results
56
+
57
+ > :memo: Provide a table with results. (e.g., accuracy, latency)
58
+ >
59
+ > * Provide links to the pre-trained models (checkpoint, SavedModel files).
60
+ > * Publish TensorFlow SavedModel files on TensorFlow Hub (tfhub.dev) if possible.
61
+ > * Add links to [TensorBoard.dev](https://tensorboard.dev/) for visualizing metrics.
62
+ >
63
+ > An example table for image classification results
64
+ >
65
+ > ### Image Classification
66
+ >
67
+ > | Model name | Download | Top 1 Accuracy | Top 5 Accuracy |
68
+ > |------------|----------|----------------|----------------|
69
+ > | Model name | [Checkpoint](https://drive.google.com/...), [SavedModel](https://tfhub.dev/...) | xx% | xx% |
70
+
71
+ ## Dataset
72
+
73
+ > :memo: Provide information of the dataset used.
74
+
75
+ ## Training
76
+
77
+ > :memo: Provide training information.
78
+ >
79
+ > * Provide details for preprocessing, hyperparameters, random seeds, and environment.
80
+ > * Provide a command line example for training.
81
+
82
+ Please run this command line for training.
83
+
84
+ ```shell
85
+ python3 ...
86
+ ```
87
+
88
+ ## Evaluation
89
+
90
+ > :memo: Provide an evaluation script with details of how to reproduce results.
91
+ >
92
+ > * Describe data preprocessing / postprocessing steps.
93
+ > * Provide a command line example for evaluation.
94
+
95
+ Please run this command line for evaluation.
96
+
97
+ ```shell
98
+ python3 ...
99
+ ```
100
+
101
+ ## References
102
+
103
+ > :memo: Provide links to references.
104
+
105
+ ## License
106
+
107
+ [![License](https://img.shields.io/badge/License-Apache%202.0-blue.svg)](https://opensource.org/licenses/Apache-2.0)
108
+
109
+ > :memo: Place your license text in a file named LICENSE in the root of the repository.
110
+ >
111
+ > * Include information about your license.
112
+ > * Reference: [Adding a license to a repository](https://help.github.com/en/github/building-a-strong-community/adding-a-license-to-a-repository)
113
+
114
+ This project is licensed under the terms of the **Apache License 2.0**.
115
+
116
+ ## Citation
117
+
118
+ > :memo: Make your repository citable.
119
+ >
120
+ > * Reference: [Making Your Code Citable](https://guides.github.com/activities/citable-code/)
121
+
122
+ If you want to cite this repository in your research paper, please use the following information.
models/.gitignore ADDED
@@ -0,0 +1,98 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Byte-compiled / optimized / DLL files
2
+ __pycache__/
3
+ *.py[cod]
4
+ *$py.class
5
+
6
+ # C extensions
7
+ *.so
8
+
9
+ # Distribution / packaging
10
+ .Python
11
+ env/
12
+ build/
13
+ develop-eggs/
14
+ dist/
15
+ downloads/
16
+ eggs/
17
+ .eggs/
18
+ lib/
19
+ lib64/
20
+ parts/
21
+ sdist/
22
+ var/
23
+ *.egg-info/
24
+ .installed.cfg
25
+ *.egg
26
+
27
+ # PyInstaller
28
+ # Usually these files are written by a python script from a template
29
+ # before PyInstaller builds the exe, so as to inject date/other infos into it.
30
+ *.manifest
31
+ *.spec
32
+
33
+ # Installer logs
34
+ pip-log.txt
35
+ pip-delete-this-directory.txt
36
+
37
+ # Unit test / coverage reports
38
+ htmlcov/
39
+ .tox/
40
+ .coverage
41
+ .coverage.*
42
+ .cache
43
+ nosetests.xml
44
+ coverage.xml
45
+ *,cover
46
+ .hypothesis/
47
+
48
+ # Translations
49
+ *.mo
50
+ *.pot
51
+
52
+ # Django stuff:
53
+ *.log
54
+ local_settings.py
55
+
56
+ # Flask stuff:
57
+ instance/
58
+ .webassets-cache
59
+
60
+ # Scrapy stuff:
61
+ .scrapy
62
+
63
+ # Sphinx documentation
64
+ docs/_build/
65
+
66
+ # PyBuilder
67
+ target/
68
+
69
+ # IPython Notebook
70
+ .ipynb_checkpoints
71
+
72
+ # pyenv
73
+ .python-version
74
+
75
+ # mypy
76
+ .mypy_cache
77
+
78
+ # celery beat schedule file
79
+ celerybeat-schedule
80
+
81
+ # dotenv
82
+ .env
83
+
84
+ # virtualenv
85
+ venv/
86
+ ENV/
87
+
88
+ # Spyder project settings
89
+ .spyderproject
90
+
91
+ # Rope project settings
92
+ .ropeproject
93
+
94
+ # PyCharm
95
+ .idea/
96
+
97
+ # For mac
98
+ .DS_Store
models/AUTHORS ADDED
@@ -0,0 +1,10 @@
 
 
 
 
 
 
 
 
 
 
 
1
+ # This is the official list of authors for copyright purposes.
2
+ # This file is distinct from the CONTRIBUTORS files.
3
+ # See the latter for an explanation.
4
+
5
+ # Names should be added to this file as:
6
+ # Name or Organization <email address>
7
+ # The email address is not required for organizations.
8
+
9
+ Google Inc.
10
+ David Dao <daviddao@broad.mit.edu>
models/CODEOWNERS ADDED
@@ -0,0 +1,61 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ * @tensorflow/tf-garden-team @tensorflow/tf-model-garden-team
2
+ /official/ @rachellj218 @saberkun @jaeyounkim
3
+ /official/nlp/ @saberkun @chenGitHuber @lehougoogle @rachellj218
4
+ /official/vision/ @pengchongjin @xianzhidu @yeqingli @arashwan @saberkun @rachellj218
5
+ /research/adv_imagenet_models/ @alexeykurakin
6
+ /research/adversarial_crypto/ @dave-andersen
7
+ /research/adversarial_logit_pairing/ @alexeykurakin
8
+ /research/adversarial_text/ @rsepassi @a-dai
9
+ /research/attention_ocr/ @xavigibert
10
+ /research/audioset/ @plakal @dpwe
11
+ /research/autoaugment/* @barretzoph
12
+ /research/autoencoders/ @snurkabill
13
+ /research/brain_coder/ @danabo
14
+ /research/cognitive_mapping_and_planning/ @s-gupta
15
+ /research/compression/ @nmjohn
16
+ /research/cvt_text/ @clarkkev @lmthang
17
+ /research/deep_contextual_bandits/ @rikel
18
+ /research/deep_speech/ @yhliang2018
19
+ /research/deeplab/ @aquariusjay @yknzhu @gpapan
20
+ /research/delf/ @andrefaraujo
21
+ /research/domain_adaptation/ @bousmalis @dmrd
22
+ /research/efficient-hrl/ @ofirnachum
23
+ /research/feelvos/ @pvoigtlaender @yuningchai @aquariusjay
24
+ /research/fivo/ @dieterichlawson
25
+ /research/global_objectives/ @mackeya-google
26
+ /research/im2txt/ @cshallue
27
+ /research/inception/ @shlens @vincentvanhoucke
28
+ /research/keypointnet/ @mnorouzi
29
+ /research/learned_optimizer/ @olganw @nirum
30
+ /research/learning_to_remember_rare_events/ @lukaszkaiser @ofirnachum
31
+ /research/learning_unsupervised_learning/ @lukemetz @nirum
32
+ /research/lexnet_nc/ @vered1986 @waterson
33
+ /research/lfads/ @jazcollins @sussillo
34
+ /research/lm_1b/ @oriolvinyals @panyx0718
35
+ /research/lm_commonsense/ @thtrieu
36
+ /research/lstm_object_detection/ @yinxiaoli @yongzhe2160
37
+ /research/marco/ @vincentvanhoucke
38
+ /research/maskgan/ @liamb315 @a-dai
39
+ /research/namignizer/ @knathanieltucker
40
+ /research/neural_gpu/ @lukaszkaiser
41
+ /research/neural_programmer/ @arvind2505
42
+ /research/next_frame_prediction/ @panyx0718
43
+ /research/object_detection/ @jch1 @tombstone @pkulzc
44
+ /research/pcl_rl/ @ofirnachum
45
+ /research/ptn/ @xcyan @arkanath @hellojas @honglaklee
46
+ /research/qa_kg/ @yuyuz
47
+ /research/real_nvp/ @laurent-dinh
48
+ /research/rebar/ @gjtucker
49
+ /research/sentiment_analysis/ @sculd
50
+ /research/seq2species/ @apbusia @depristo
51
+ /research/skip_thoughts/ @cshallue
52
+ /research/slim/ @sguada @marksandler2
53
+ /research/steve/ @buckman-google
54
+ /research/street/ @theraysmith
55
+ /research/struct2depth/ @aneliaangelova
56
+ /research/swivel/ @waterson
57
+ /research/tcn/ @coreylynch @sermanet
58
+ /research/textsum/ @panyx0718 @peterjliu
59
+ /research/transformer/ @daviddao
60
+ /research/vid2depth/ @rezama
61
+ /research/video_prediction/ @cbfinn
models/CONTRIBUTING.md ADDED
@@ -0,0 +1,10 @@
 
 
 
 
 
 
 
 
 
 
 
1
+ # How to contribute
2
+
3
+ ![Contributors](https://img.shields.io/github/contributors/tensorflow/models)
4
+
5
+ We encourage you to contribute to the TensorFlow Model Garden.
6
+
7
+ Please read our [guidelines](../../wiki/How-to-contribute) for details.
8
+
9
+ **NOTE**: Only [code owners](./CODEOWNERS) are allowed to merge a pull request.
10
+ Please contact the code owners of each model to merge your pull request.
models/ISSUES.md ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # If you open a GitHub issue, here is our policy.
2
+
3
+ * It must be a **bug**, a **feature request**, or a significant problem
4
+ with **documentation**.
5
+ * Please send a pull request instead for small documentation fixes.
6
+ * The required form must be filled out.
7
+ * The issue should be related to the repository it is created in.
8
+
9
+ General help and support should be sought on [Stack Overflow](https://stackoverflow.com/questions/tagged/tensorflow-model-garden) or other non-GitHub channels.
10
+
11
+ [![](https://img.shields.io/stackexchange/stackoverflow/t/tensorflow-model-garden)](https://stackoverflow.com/questions/tagged/tensorflow-model-garden)
12
+
13
+ TensorFlow developers respond to issues.
14
+ We want to focus on work that benefits the whole community such as fixing bugs
15
+ and adding new features.
16
+ It helps us to address bugs and feature requests in a timely manner.
17
+
18
+ ---
19
+
20
+ Please understand that research models in the [research directory](https://github.com/tensorflow/models/tree/master/research)
21
+ included in this repository are experimental and research-style code.
22
+ They are not officially supported by the TensorFlow team.
23
+
24
+
models/LICENSE ADDED
@@ -0,0 +1,203 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Copyright 2016 The TensorFlow Authors. All rights reserved.
2
+
3
+ Apache License
4
+ Version 2.0, January 2004
5
+ http://www.apache.org/licenses/
6
+
7
+ TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
8
+
9
+ 1. Definitions.
10
+
11
+ "License" shall mean the terms and conditions for use, reproduction,
12
+ and distribution as defined by Sections 1 through 9 of this document.
13
+
14
+ "Licensor" shall mean the copyright owner or entity authorized by
15
+ the copyright owner that is granting the License.
16
+
17
+ "Legal Entity" shall mean the union of the acting entity and all
18
+ other entities that control, are controlled by, or are under common
19
+ control with that entity. For the purposes of this definition,
20
+ "control" means (i) the power, direct or indirect, to cause the
21
+ direction or management of such entity, whether by contract or
22
+ otherwise, or (ii) ownership of fifty percent (50%) or more of the
23
+ outstanding shares, or (iii) beneficial ownership of such entity.
24
+
25
+ "You" (or "Your") shall mean an individual or Legal Entity
26
+ exercising permissions granted by this License.
27
+
28
+ "Source" form shall mean the preferred form for making modifications,
29
+ including but not limited to software source code, documentation
30
+ source, and configuration files.
31
+
32
+ "Object" form shall mean any form resulting from mechanical
33
+ transformation or translation of a Source form, including but
34
+ not limited to compiled object code, generated documentation,
35
+ and conversions to other media types.
36
+
37
+ "Work" shall mean the work of authorship, whether in Source or
38
+ Object form, made available under the License, as indicated by a
39
+ copyright notice that is included in or attached to the work
40
+ (an example is provided in the Appendix below).
41
+
42
+ "Derivative Works" shall mean any work, whether in Source or Object
43
+ form, that is based on (or derived from) the Work and for which the
44
+ editorial revisions, annotations, elaborations, or other modifications
45
+ represent, as a whole, an original work of authorship. For the purposes
46
+ of this License, Derivative Works shall not include works that remain
47
+ separable from, or merely link (or bind by name) to the interfaces of,
48
+ the Work and Derivative Works thereof.
49
+
50
+ "Contribution" shall mean any work of authorship, including
51
+ the original version of the Work and any modifications or additions
52
+ to that Work or Derivative Works thereof, that is intentionally
53
+ submitted to Licensor for inclusion in the Work by the copyright owner
54
+ or by an individual or Legal Entity authorized to submit on behalf of
55
+ the copyright owner. For the purposes of this definition, "submitted"
56
+ means any form of electronic, verbal, or written communication sent
57
+ to the Licensor or its representatives, including but not limited to
58
+ communication on electronic mailing lists, source code control systems,
59
+ and issue tracking systems that are managed by, or on behalf of, the
60
+ Licensor for the purpose of discussing and improving the Work, but
61
+ excluding communication that is conspicuously marked or otherwise
62
+ designated in writing by the copyright owner as "Not a Contribution."
63
+
64
+ "Contributor" shall mean Licensor and any individual or Legal Entity
65
+ on behalf of whom a Contribution has been received by Licensor and
66
+ subsequently incorporated within the Work.
67
+
68
+ 2. Grant of Copyright License. Subject to the terms and conditions of
69
+ this License, each Contributor hereby grants to You a perpetual,
70
+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
71
+ copyright license to reproduce, prepare Derivative Works of,
72
+ publicly display, publicly perform, sublicense, and distribute the
73
+ Work and such Derivative Works in Source or Object form.
74
+
75
+ 3. Grant of Patent License. Subject to the terms and conditions of
76
+ this License, each Contributor hereby grants to You a perpetual,
77
+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
78
+ (except as stated in this section) patent license to make, have made,
79
+ use, offer to sell, sell, import, and otherwise transfer the Work,
80
+ where such license applies only to those patent claims licensable
81
+ by such Contributor that are necessarily infringed by their
82
+ Contribution(s) alone or by combination of their Contribution(s)
83
+ with the Work to which such Contribution(s) was submitted. If You
84
+ institute patent litigation against any entity (including a
85
+ cross-claim or counterclaim in a lawsuit) alleging that the Work
86
+ or a Contribution incorporated within the Work constitutes direct
87
+ or contributory patent infringement, then any patent licenses
88
+ granted to You under this License for that Work shall terminate
89
+ as of the date such litigation is filed.
90
+
91
+ 4. Redistribution. You may reproduce and distribute copies of the
92
+ Work or Derivative Works thereof in any medium, with or without
93
+ modifications, and in Source or Object form, provided that You
94
+ meet the following conditions:
95
+
96
+ (a) You must give any other recipients of the Work or
97
+ Derivative Works a copy of this License; and
98
+
99
+ (b) You must cause any modified files to carry prominent notices
100
+ stating that You changed the files; and
101
+
102
+ (c) You must retain, in the Source form of any Derivative Works
103
+ that You distribute, all copyright, patent, trademark, and
104
+ attribution notices from the Source form of the Work,
105
+ excluding those notices that do not pertain to any part of
106
+ the Derivative Works; and
107
+
108
+ (d) If the Work includes a "NOTICE" text file as part of its
109
+ distribution, then any Derivative Works that You distribute must
110
+ include a readable copy of the attribution notices contained
111
+ within such NOTICE file, excluding those notices that do not
112
+ pertain to any part of the Derivative Works, in at least one
113
+ of the following places: within a NOTICE text file distributed
114
+ as part of the Derivative Works; within the Source form or
115
+ documentation, if provided along with the Derivative Works; or,
116
+ within a display generated by the Derivative Works, if and
117
+ wherever such third-party notices normally appear. The contents
118
+ of the NOTICE file are for informational purposes only and
119
+ do not modify the License. You may add Your own attribution
120
+ notices within Derivative Works that You distribute, alongside
121
+ or as an addendum to the NOTICE text from the Work, provided
122
+ that such additional attribution notices cannot be construed
123
+ as modifying the License.
124
+
125
+ You may add Your own copyright statement to Your modifications and
126
+ may provide additional or different license terms and conditions
127
+ for use, reproduction, or distribution of Your modifications, or
128
+ for any such Derivative Works as a whole, provided Your use,
129
+ reproduction, and distribution of the Work otherwise complies with
130
+ the conditions stated in this License.
131
+
132
+ 5. Submission of Contributions. Unless You explicitly state otherwise,
133
+ any Contribution intentionally submitted for inclusion in the Work
134
+ by You to the Licensor shall be under the terms and conditions of
135
+ this License, without any additional terms or conditions.
136
+ Notwithstanding the above, nothing herein shall supersede or modify
137
+ the terms of any separate license agreement you may have executed
138
+ with Licensor regarding such Contributions.
139
+
140
+ 6. Trademarks. This License does not grant permission to use the trade
141
+ names, trademarks, service marks, or product names of the Licensor,
142
+ except as required for reasonable and customary use in describing the
143
+ origin of the Work and reproducing the content of the NOTICE file.
144
+
145
+ 7. Disclaimer of Warranty. Unless required by applicable law or
146
+ agreed to in writing, Licensor provides the Work (and each
147
+ Contributor provides its Contributions) on an "AS IS" BASIS,
148
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
149
+ implied, including, without limitation, any warranties or conditions
150
+ of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
151
+ PARTICULAR PURPOSE. You are solely responsible for determining the
152
+ appropriateness of using or redistributing the Work and assume any
153
+ risks associated with Your exercise of permissions under this License.
154
+
155
+ 8. Limitation of Liability. In no event and under no legal theory,
156
+ whether in tort (including negligence), contract, or otherwise,
157
+ unless required by applicable law (such as deliberate and grossly
158
+ negligent acts) or agreed to in writing, shall any Contributor be
159
+ liable to You for damages, including any direct, indirect, special,
160
+ incidental, or consequential damages of any character arising as a
161
+ result of this License or out of the use or inability to use the
162
+ Work (including but not limited to damages for loss of goodwill,
163
+ work stoppage, computer failure or malfunction, or any and all
164
+ other commercial damages or losses), even if such Contributor
165
+ has been advised of the possibility of such damages.
166
+
167
+ 9. Accepting Warranty or Additional Liability. While redistributing
168
+ the Work or Derivative Works thereof, You may choose to offer,
169
+ and charge a fee for, acceptance of support, warranty, indemnity,
170
+ or other liability obligations and/or rights consistent with this
171
+ License. However, in accepting such obligations, You may act only
172
+ on Your own behalf and on Your sole responsibility, not on behalf
173
+ of any other Contributor, and only if You agree to indemnify,
174
+ defend, and hold each Contributor harmless for any liability
175
+ incurred by, or claims asserted against, such Contributor by reason
176
+ of your accepting any such warranty or additional liability.
177
+
178
+ END OF TERMS AND CONDITIONS
179
+
180
+ APPENDIX: How to apply the Apache License to your work.
181
+
182
+ To apply the Apache License to your work, attach the following
183
+ boilerplate notice, with the fields enclosed by brackets "[]"
184
+ replaced with your own identifying information. (Don't include
185
+ the brackets!) The text should be enclosed in the appropriate
186
+ comment syntax for the file format. We also recommend that a
187
+ file or class name and description of purpose be included on the
188
+ same "printed page" as the copyright notice for easier
189
+ identification within third-party archives.
190
+
191
+ Copyright 2016, The Authors.
192
+
193
+ Licensed under the Apache License, Version 2.0 (the "License");
194
+ you may not use this file except in compliance with the License.
195
+ You may obtain a copy of the License at
196
+
197
+ http://www.apache.org/licenses/LICENSE-2.0
198
+
199
+ Unless required by applicable law or agreed to in writing, software
200
+ distributed under the License is distributed on an "AS IS" BASIS,
201
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
202
+ See the License for the specific language governing permissions and
203
+ limitations under the License.
models/README.md ADDED
@@ -0,0 +1,39 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ![Logo](https://storage.googleapis.com/model_garden_artifacts/TF_Model_Garden.png)
2
+
3
+ # Welcome to the Model Garden for TensorFlow
4
+
5
+ The TensorFlow Model Garden is a repository with a number of different implementations of state-of-the-art (SOTA) models and modeling solutions for TensorFlow users. We aim to demonstrate the best practices for modeling so that TensorFlow users
6
+ can take full advantage of TensorFlow for their research and product development.
7
+
8
+ | Directory | Description |
9
+ |-----------|-------------|
10
+ | [official](official) | • A collection of example implementations for SOTA models using the latest TensorFlow 2's high-level APIs<br />• Officially maintained, supported, and kept up to date with the latest TensorFlow 2 APIs by TensorFlow<br />• Reasonably optimized for fast performance while still being easy to read |
11
+ | [research](research) | • A collection of research model implementations in TensorFlow 1 or 2 by researchers<br />• Maintained and supported by researchers |
12
+ | [community](community) | • A curated list of the GitHub repositories with machine learning models and implementations powered by TensorFlow 2 |
13
+
14
+ ## [Announcements](https://github.com/tensorflow/models/wiki/Announcements)
15
+
16
+ | Date | News |
17
+ |------|------|
18
+ | June 17, 2020 | [Context R-CNN: Long Term Temporal Context for Per-Camera Object Detection](https://github.com/tensorflow/models/tree/master/research/object_detection#june-17th-2020) released
19
+ | May 21, 2020 | [Unifying Deep Local and Global Features for Image Search (DELG)](https://github.com/tensorflow/models/tree/master/research/delf#delg) code released
20
+ | May 19, 2020 | [MobileDets: Searching for Object Detection Architectures for Mobile Accelerators](https://github.com/tensorflow/models/tree/master/research/object_detection#may-19th-2020) released
21
+ | May 7, 2020 | [MnasFPN with MobileNet-V2 backbone](https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/detection_model_zoo.md#mobile-models) released for object detection
22
+ | May 1, 2020 | [DELF: DEep Local Features](https://github.com/tensorflow/models/tree/master/research/delf) updated to support TensorFlow 2.1
23
+ | March 31, 2020 | [Introducing the Model Garden for TensorFlow 2](https://blog.tensorflow.org/2020/03/introducing-model-garden-for-tensorflow-2.html) ([Tweet](https://twitter.com/TensorFlow/status/1245029834633297921)) |
24
+
25
+ ## [Milestones](https://github.com/tensorflow/models/milestones)
26
+
27
+ | Date | Milestone |
28
+ |------|-----------|
29
+ | July 7, 2020 | [![GitHub milestone](https://img.shields.io/github/milestones/progress/tensorflow/models/1)](https://github.com/tensorflow/models/milestone/1) |
30
+
31
+ ## Contributions
32
+
33
+ [![help wanted:paper implementation](https://img.shields.io/github/issues/tensorflow/models/help%20wanted%3Apaper%20implementation)](https://github.com/tensorflow/models/labels/help%20wanted%3Apaper%20implementation)
34
+
35
+ If you want to contribute, please review the [contribution guidelines](https://github.com/tensorflow/models/wiki/How-to-contribute).
36
+
37
+ ## License
38
+
39
+ [Apache License 2.0](LICENSE)
models/official/LICENSE ADDED
@@ -0,0 +1,203 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Copyright 2015 The TensorFlow Authors. All rights reserved.
2
+
3
+ Apache License
4
+ Version 2.0, January 2004
5
+ http://www.apache.org/licenses/
6
+
7
+ TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
8
+
9
+ 1. Definitions.
10
+
11
+ "License" shall mean the terms and conditions for use, reproduction,
12
+ and distribution as defined by Sections 1 through 9 of this document.
13
+
14
+ "Licensor" shall mean the copyright owner or entity authorized by
15
+ the copyright owner that is granting the License.
16
+
17
+ "Legal Entity" shall mean the union of the acting entity and all
18
+ other entities that control, are controlled by, or are under common
19
+ control with that entity. For the purposes of this definition,
20
+ "control" means (i) the power, direct or indirect, to cause the
21
+ direction or management of such entity, whether by contract or
22
+ otherwise, or (ii) ownership of fifty percent (50%) or more of the
23
+ outstanding shares, or (iii) beneficial ownership of such entity.
24
+
25
+ "You" (or "Your") shall mean an individual or Legal Entity
26
+ exercising permissions granted by this License.
27
+
28
+ "Source" form shall mean the preferred form for making modifications,
29
+ including but not limited to software source code, documentation
30
+ source, and configuration files.
31
+
32
+ "Object" form shall mean any form resulting from mechanical
33
+ transformation or translation of a Source form, including but
34
+ not limited to compiled object code, generated documentation,
35
+ and conversions to other media types.
36
+
37
+ "Work" shall mean the work of authorship, whether in Source or
38
+ Object form, made available under the License, as indicated by a
39
+ copyright notice that is included in or attached to the work
40
+ (an example is provided in the Appendix below).
41
+
42
+ "Derivative Works" shall mean any work, whether in Source or Object
43
+ form, that is based on (or derived from) the Work and for which the
44
+ editorial revisions, annotations, elaborations, or other modifications
45
+ represent, as a whole, an original work of authorship. For the purposes
46
+ of this License, Derivative Works shall not include works that remain
47
+ separable from, or merely link (or bind by name) to the interfaces of,
48
+ the Work and Derivative Works thereof.
49
+
50
+ "Contribution" shall mean any work of authorship, including
51
+ the original version of the Work and any modifications or additions
52
+ to that Work or Derivative Works thereof, that is intentionally
53
+ submitted to Licensor for inclusion in the Work by the copyright owner
54
+ or by an individual or Legal Entity authorized to submit on behalf of
55
+ the copyright owner. For the purposes of this definition, "submitted"
56
+ means any form of electronic, verbal, or written communication sent
57
+ to the Licensor or its representatives, including but not limited to
58
+ communication on electronic mailing lists, source code control systems,
59
+ and issue tracking systems that are managed by, or on behalf of, the
60
+ Licensor for the purpose of discussing and improving the Work, but
61
+ excluding communication that is conspicuously marked or otherwise
62
+ designated in writing by the copyright owner as "Not a Contribution."
63
+
64
+ "Contributor" shall mean Licensor and any individual or Legal Entity
65
+ on behalf of whom a Contribution has been received by Licensor and
66
+ subsequently incorporated within the Work.
67
+
68
+ 2. Grant of Copyright License. Subject to the terms and conditions of
69
+ this License, each Contributor hereby grants to You a perpetual,
70
+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
71
+ copyright license to reproduce, prepare Derivative Works of,
72
+ publicly display, publicly perform, sublicense, and distribute the
73
+ Work and such Derivative Works in Source or Object form.
74
+
75
+ 3. Grant of Patent License. Subject to the terms and conditions of
76
+ this License, each Contributor hereby grants to You a perpetual,
77
+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
78
+ (except as stated in this section) patent license to make, have made,
79
+ use, offer to sell, sell, import, and otherwise transfer the Work,
80
+ where such license applies only to those patent claims licensable
81
+ by such Contributor that are necessarily infringed by their
82
+ Contribution(s) alone or by combination of their Contribution(s)
83
+ with the Work to which such Contribution(s) was submitted. If You
84
+ institute patent litigation against any entity (including a
85
+ cross-claim or counterclaim in a lawsuit) alleging that the Work
86
+ or a Contribution incorporated within the Work constitutes direct
87
+ or contributory patent infringement, then any patent licenses
88
+ granted to You under this License for that Work shall terminate
89
+ as of the date such litigation is filed.
90
+
91
+ 4. Redistribution. You may reproduce and distribute copies of the
92
+ Work or Derivative Works thereof in any medium, with or without
93
+ modifications, and in Source or Object form, provided that You
94
+ meet the following conditions:
95
+
96
+ (a) You must give any other recipients of the Work or
97
+ Derivative Works a copy of this License; and
98
+
99
+ (b) You must cause any modified files to carry prominent notices
100
+ stating that You changed the files; and
101
+
102
+ (c) You must retain, in the Source form of any Derivative Works
103
+ that You distribute, all copyright, patent, trademark, and
104
+ attribution notices from the Source form of the Work,
105
+ excluding those notices that do not pertain to any part of
106
+ the Derivative Works; and
107
+
108
+ (d) If the Work includes a "NOTICE" text file as part of its
109
+ distribution, then any Derivative Works that You distribute must
110
+ include a readable copy of the attribution notices contained
111
+ within such NOTICE file, excluding those notices that do not
112
+ pertain to any part of the Derivative Works, in at least one
113
+ of the following places: within a NOTICE text file distributed
114
+ as part of the Derivative Works; within the Source form or
115
+ documentation, if provided along with the Derivative Works; or,
116
+ within a display generated by the Derivative Works, if and
117
+ wherever such third-party notices normally appear. The contents
118
+ of the NOTICE file are for informational purposes only and
119
+ do not modify the License. You may add Your own attribution
120
+ notices within Derivative Works that You distribute, alongside
121
+ or as an addendum to the NOTICE text from the Work, provided
122
+ that such additional attribution notices cannot be construed
123
+ as modifying the License.
124
+
125
+ You may add Your own copyright statement to Your modifications and
126
+ may provide additional or different license terms and conditions
127
+ for use, reproduction, or distribution of Your modifications, or
128
+ for any such Derivative Works as a whole, provided Your use,
129
+ reproduction, and distribution of the Work otherwise complies with
130
+ the conditions stated in this License.
131
+
132
+ 5. Submission of Contributions. Unless You explicitly state otherwise,
133
+ any Contribution intentionally submitted for inclusion in the Work
134
+ by You to the Licensor shall be under the terms and conditions of
135
+ this License, without any additional terms or conditions.
136
+ Notwithstanding the above, nothing herein shall supersede or modify
137
+ the terms of any separate license agreement you may have executed
138
+ with Licensor regarding such Contributions.
139
+
140
+ 6. Trademarks. This License does not grant permission to use the trade
141
+ names, trademarks, service marks, or product names of the Licensor,
142
+ except as required for reasonable and customary use in describing the
143
+ origin of the Work and reproducing the content of the NOTICE file.
144
+
145
+ 7. Disclaimer of Warranty. Unless required by applicable law or
146
+ agreed to in writing, Licensor provides the Work (and each
147
+ Contributor provides its Contributions) on an "AS IS" BASIS,
148
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
149
+ implied, including, without limitation, any warranties or conditions
150
+ of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
151
+ PARTICULAR PURPOSE. You are solely responsible for determining the
152
+ appropriateness of using or redistributing the Work and assume any
153
+ risks associated with Your exercise of permissions under this License.
154
+
155
+ 8. Limitation of Liability. In no event and under no legal theory,
156
+ whether in tort (including negligence), contract, or otherwise,
157
+ unless required by applicable law (such as deliberate and grossly
158
+ negligent acts) or agreed to in writing, shall any Contributor be
159
+ liable to You for damages, including any direct, indirect, special,
160
+ incidental, or consequential damages of any character arising as a
161
+ result of this License or out of the use or inability to use the
162
+ Work (including but not limited to damages for loss of goodwill,
163
+ work stoppage, computer failure or malfunction, or any and all
164
+ other commercial damages or losses), even if such Contributor
165
+ has been advised of the possibility of such damages.
166
+
167
+ 9. Accepting Warranty or Additional Liability. While redistributing
168
+ the Work or Derivative Works thereof, You may choose to offer,
169
+ and charge a fee for, acceptance of support, warranty, indemnity,
170
+ or other liability obligations and/or rights consistent with this
171
+ License. However, in accepting such obligations, You may act only
172
+ on Your own behalf and on Your sole responsibility, not on behalf
173
+ of any other Contributor, and only if You agree to indemnify,
174
+ defend, and hold each Contributor harmless for any liability
175
+ incurred by, or claims asserted against, such Contributor by reason
176
+ of your accepting any such warranty or additional liability.
177
+
178
+ END OF TERMS AND CONDITIONS
179
+
180
+ APPENDIX: How to apply the Apache License to your work.
181
+
182
+ To apply the Apache License to your work, attach the following
183
+ boilerplate notice, with the fields enclosed by brackets "[]"
184
+ replaced with your own identifying information. (Don't include
185
+ the brackets!) The text should be enclosed in the appropriate
186
+ comment syntax for the file format. We also recommend that a
187
+ file or class name and description of purpose be included on the
188
+ same "printed page" as the copyright notice for easier
189
+ identification within third-party archives.
190
+
191
+ Copyright 2015, The TensorFlow Authors.
192
+
193
+ Licensed under the Apache License, Version 2.0 (the "License");
194
+ you may not use this file except in compliance with the License.
195
+ You may obtain a copy of the License at
196
+
197
+ http://www.apache.org/licenses/LICENSE-2.0
198
+
199
+ Unless required by applicable law or agreed to in writing, software
200
+ distributed under the License is distributed on an "AS IS" BASIS,
201
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
202
+ See the License for the specific language governing permissions and
203
+ limitations under the License.
models/official/README-TPU.md ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Offically Supported TensorFlow 2.1+ Models on Cloud TPU
2
+
3
+ ## Natural Language Processing
4
+
5
+ * [bert](nlp/bert): A powerful pre-trained language representation model:
6
+ BERT, which stands for Bidirectional Encoder Representations from
7
+ Transformers.
8
+ [BERT FineTuning with Cloud TPU](https://cloud.google.com/tpu/docs/tutorials/bert-2.x) provides step by step instructions on Cloud TPU training. You can look [Bert MNLI Tensorboard.dev metrics](https://tensorboard.dev/experiment/LijZ1IrERxKALQfr76gndA) for MNLI fine tuning task.
9
+ * [transformer](nlp/transformer): A transformer model to translate the WMT
10
+ English to German dataset.
11
+ [Training transformer on Cloud TPU](https://cloud.google.com/tpu/docs/tutorials/transformer-2.x) for step by step instructions on Cloud TPU training.
12
+
13
+ ## Computer Vision
14
+
15
+ * [efficientnet](vision/image_classification): A family of convolutional
16
+ neural networks that scale by balancing network depth, width, and
17
+ resolution and can be used to classify ImageNet's dataset of 1000 classes.
18
+ See [Tensorboard.dev training metrics](https://tensorboard.dev/experiment/KnaWjrq5TXGfv0NW5m7rpg/#scalars).
19
+ * [mnist](vision/image_classification): A basic model to classify digits
20
+ from the MNIST dataset. See [Running MNIST on Cloud TPU](https://cloud.google.com/tpu/docs/tutorials/mnist-2.x) tutorial and [Tensorboard.dev metrics](https://tensorboard.dev/experiment/mIah5lppTASvrHqWrdr6NA).
21
+ * [mask-rcnn](vision/detection): An object detection and instance segmentation model. See [Tensorboard.dev training metrics](https://tensorboard.dev/experiment/LH7k0fMsRwqUAcE09o9kPA).
22
+ * [resnet](vision/image_classification): A deep residual network that can
23
+ be used to classify ImageNet's dataset of 1000 classes.
24
+ See [Training ResNet on Cloud TPU](https://cloud.google.com/tpu/docs/tutorials/resnet-2.x) tutorial and [Tensorboard.dev metrics](https://tensorboard.dev/experiment/CxlDK8YMRrSpYEGtBRpOhg).
25
+ * [retinanet](vision/detection): A fast and powerful object detector. See [Tensorboard.dev training metrics](https://tensorboard.dev/experiment/b8NRnWU3TqG6Rw0UxueU6Q).
models/official/README.md ADDED
@@ -0,0 +1,142 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ![Logo](https://storage.googleapis.com/model_garden_artifacts/TF_Model_Garden.png)
2
+
3
+ # TensorFlow Official Models
4
+
5
+ The TensorFlow official models are a collection of models
6
+ that use TensorFlow’s high-level APIs.
7
+ They are intended to be well-maintained, tested, and kept up to date
8
+ with the latest TensorFlow API.
9
+
10
+ They should also be reasonably optimized for fast performance while still
11
+ being easy to read.
12
+ These models are used as end-to-end tests, ensuring that the models run
13
+ with the same or improved speed and performance with each new TensorFlow build.
14
+
15
+ ## More models to come!
16
+
17
+ The team is actively developing new models.
18
+ In the near future, we will add:
19
+
20
+ * State-of-the-art language understanding models:
21
+ More members in Transformer family
22
+ * Start-of-the-art image classification models:
23
+ EfficientNet, MnasNet, and variants
24
+ * A set of excellent objection detection models.
25
+
26
+ ## Table of Contents
27
+
28
+ - [Models and Implementations](#models-and-implementations)
29
+ * [Computer Vision](#computer-vision)
30
+ + [Image Classification](#image-classification)
31
+ + [Object Detection and Segmentation](#object-detection-and-segmentation)
32
+ * [Natural Language Processing](#natural-language-processing)
33
+ * [Recommendation](#recommendation)
34
+ - [How to get started with the official models](#how-to-get-started-with-the-official-models)
35
+
36
+ ## Models and Implementations
37
+
38
+ ### Computer Vision
39
+
40
+ #### Image Classification
41
+
42
+ | Model | Reference (Paper) |
43
+ |-------|-------------------|
44
+ | [MNIST](vision/image_classification) | A basic model to classify digits from the [MNIST dataset](http://yann.lecun.com/exdb/mnist/) |
45
+ | [ResNet](vision/image_classification) | [Deep Residual Learning for Image Recognition](https://arxiv.org/abs/1512.03385) |
46
+ | [EfficientNet](vision/image_classification) | [EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks](https://arxiv.org/abs/1905.11946) |
47
+
48
+ #### Object Detection and Segmentation
49
+
50
+ | Model | Reference (Paper) |
51
+ |-------|-------------------|
52
+ | [RetinaNet](vision/detection) | [Focal Loss for Dense Object Detection](https://arxiv.org/abs/1708.02002) |
53
+ | [Mask R-CNN](vision/detection) | [Mask R-CNN](https://arxiv.org/abs/1703.06870) |
54
+ | [ShapeMask](vision/detection) | [ShapeMask: Learning to Segment Novel Objects by Refining Shape Priors](https://arxiv.org/abs/1904.03239) |
55
+
56
+ ### Natural Language Processing
57
+
58
+ | Model | Reference (Paper) |
59
+ |-------|-------------------|
60
+ | [ALBERT (A Lite BERT)](nlp/albert) | [ALBERT: A Lite BERT for Self-supervised Learning of Language Representations](https://arxiv.org/abs/1909.11942) |
61
+ | [BERT (Bidirectional Encoder Representations from Transformers)](nlp/bert) | [BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding](https://arxiv.org/abs/1810.04805) |
62
+ | [NHNet (News Headline generation model)](nlp/nhnet) | [Generating Representative Headlines for News Stories](https://arxiv.org/abs/2001.09386) |
63
+ | [Transformer](nlp/transformer) | [Attention Is All You Need](https://arxiv.org/abs/1706.03762) |
64
+ | [XLNet](nlp/xlnet) | [XLNet: Generalized Autoregressive Pretraining for Language Understanding](https://arxiv.org/abs/1906.08237) |
65
+
66
+ ### Recommendation
67
+
68
+ | Model | Reference (Paper) |
69
+ |-------|-------------------|
70
+ | [NCF](recommendation) | [Neural Collaborative Filtering](https://arxiv.org/abs/1708.05031) |
71
+
72
+ ## How to get started with the official models
73
+
74
+ * The models in the master branch are developed using TensorFlow 2,
75
+ and they target the TensorFlow [nightly binaries](https://github.com/tensorflow/tensorflow#installation)
76
+ built from the
77
+ [master branch of TensorFlow](https://github.com/tensorflow/tensorflow/tree/master).
78
+ * The stable versions targeting releases of TensorFlow are available
79
+ as tagged branches or [downloadable releases](https://github.com/tensorflow/models/releases).
80
+ * Model repository version numbers match the target TensorFlow release,
81
+ such that
82
+ [release v2.2.0](https://github.com/tensorflow/models/releases/tag/v2.2.0)
83
+ are compatible with
84
+ [TensorFlow v2.2.0](https://github.com/tensorflow/tensorflow/releases/tag/v2.2.0).
85
+
86
+ Please follow the below steps before running models in this repository.
87
+
88
+ ### Requirements
89
+
90
+ * The latest TensorFlow Model Garden release and TensorFlow 2
91
+ * If you are on a version of TensorFlow earlier than 2.2, please
92
+ upgrade your TensorFlow to [the latest TensorFlow 2](https://www.tensorflow.org/install/).
93
+
94
+ ```shell
95
+ pip3 install tf-nightly
96
+ ```
97
+
98
+ ### Installation
99
+
100
+ #### Method 1: Install the TensorFlow Model Garden pip package
101
+
102
+ **tf-models-nightly** is the nightly Model Garden package
103
+ created daily automatically. pip will install all models
104
+ and dependencies automatically.
105
+
106
+ ```shell
107
+ pip install tf-models-nightly
108
+ ```
109
+
110
+ Please check out our [example](colab/fine_tuning_bert.ipynb)
111
+ to learn how to use a PIP package.
112
+
113
+ #### Method 2: Clone the source
114
+
115
+ 1. Clone the GitHub repository:
116
+
117
+ ```shell
118
+ git clone https://github.com/tensorflow/models.git
119
+ ```
120
+
121
+ 2. Add the top-level ***/models*** folder to the Python path.
122
+
123
+ ```shell
124
+ export PYTHONPATH=$PYTHONPATH:/path/to/models
125
+ ```
126
+
127
+ If you are using a Colab notebook, please set the Python path with os.environ.
128
+
129
+ ```python
130
+ import os
131
+ os.environ['PYTHONPATH'] += ":/path/to/models"
132
+ ```
133
+
134
+ 3. Install other dependencies
135
+
136
+ ```shell
137
+ pip3 install --user -r official/requirements.txt
138
+ ```
139
+
140
+ ## Contributions
141
+
142
+ If you want to contribute, please review the [contribution guidelines](https://github.com/tensorflow/models/wiki/How-to-contribute).
models/official/__init__.py ADDED
File without changes
models/official/__pycache__/__init__.cpython-39.pyc ADDED
Binary file (139 Bytes). View file
 
models/official/benchmark/__init__.py ADDED
File without changes
models/official/benchmark/benchmark_wrappers.py ADDED
@@ -0,0 +1,97 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Lint as: python3
2
+ # Copyright 2019 The TensorFlow Authors. All Rights Reserved.
3
+ #
4
+ # Licensed under the Apache License, Version 2.0 (the "License");
5
+ # you may not use this file except in compliance with the License.
6
+ # You may obtain a copy of the License at
7
+ #
8
+ # http://www.apache.org/licenses/LICENSE-2.0
9
+ #
10
+ # Unless required by applicable law or agreed to in writing, software
11
+ # distributed under the License is distributed on an "AS IS" BASIS,
12
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
+ # See the License for the specific language governing permissions and
14
+ # limitations under the License.
15
+ # ==============================================================================
16
+ """Utils to annotate and trace benchmarks."""
17
+
18
+ from __future__ import absolute_import
19
+ from __future__ import division
20
+ from __future__ import print_function
21
+
22
+ from absl import flags
23
+ from absl import logging
24
+ from absl.testing import flagsaver
25
+
26
+ FLAGS = flags.FLAGS
27
+
28
+ flags.DEFINE_multi_string(
29
+ 'benchmark_method_flags', None,
30
+ 'Optional list of runtime flags of the form key=value. Specify '
31
+ 'multiple times to specify different flags. These will override the FLAGS '
32
+ 'object directly after hardcoded settings in individual benchmark methods '
33
+ 'before they call _run_and_report benchmark. Example if we set '
34
+ '--benchmark_method_flags=train_steps=10 and a benchmark method hardcodes '
35
+ 'FLAGS.train_steps=10000 and later calls _run_and_report_benchmark, '
36
+ 'it\'ll only run for 10 steps. This is useful for '
37
+ 'debugging/profiling workflows.')
38
+
39
+
40
+ def enable_runtime_flags(decorated_func):
41
+ """Sets attributes from --benchmark_method_flags for method execution.
42
+
43
+ @enable_runtime_flags decorator temporarily adds flags passed in via
44
+ --benchmark_method_flags and runs the decorated function in that context.
45
+
46
+ A user can set --benchmark_method_flags=train_steps=5 to run the benchmark
47
+ method in the snippet below with FLAGS.train_steps=5 for debugging (without
48
+ modifying the benchmark code).
49
+
50
+ class ModelBenchmark():
51
+
52
+ @benchmark_wrappers.enable_runtime_flags
53
+ def _run_and_report_benchmark(self):
54
+ # run benchmark ...
55
+ # report benchmark results ...
56
+
57
+ def benchmark_method(self):
58
+ FLAGS.train_steps = 1000
59
+ ...
60
+ self._run_and_report_benchmark()
61
+
62
+ Args:
63
+ decorated_func: The method that runs the benchmark after previous setup
64
+ execution that set some flags.
65
+
66
+ Returns:
67
+ new_func: The same method which executes in a temporary context where flag
68
+ overrides from --benchmark_method_flags are active.
69
+ """
70
+
71
+ def runner(*args, **kwargs):
72
+ """Creates a temporary context to activate --benchmark_method_flags."""
73
+ if FLAGS.benchmark_method_flags:
74
+ saved_flag_values = flagsaver.save_flag_values()
75
+ for key_value in FLAGS.benchmark_method_flags:
76
+ key, value = key_value.split('=', 1)
77
+ try:
78
+ numeric_float = float(value)
79
+ numeric_int = int(numeric_float)
80
+ if abs(numeric_int) == abs(numeric_float):
81
+ flag_value = numeric_int
82
+ else:
83
+ flag_value = numeric_float
84
+ except ValueError:
85
+ flag_value = value
86
+ logging.info('Setting --%s=%s', key, flag_value)
87
+ setattr(FLAGS, key, flag_value)
88
+ else:
89
+ saved_flag_values = None
90
+ try:
91
+ result = decorated_func(*args, **kwargs)
92
+ return result
93
+ finally:
94
+ if saved_flag_values:
95
+ flagsaver.restore_flag_values(saved_flag_values)
96
+
97
+ return runner
models/official/benchmark/bert_benchmark.py ADDED
@@ -0,0 +1,365 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2019 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
+ """Executes BERT benchmarks and accuracy tests."""
16
+
17
+ from __future__ import absolute_import
18
+ from __future__ import division
19
+ from __future__ import print_function
20
+
21
+ import functools
22
+ import json
23
+ import math
24
+ import os
25
+ import time
26
+
27
+ # pylint: disable=g-bad-import-order
28
+ from absl import flags
29
+ from absl.testing import flagsaver
30
+ import tensorflow as tf
31
+ # pylint: enable=g-bad-import-order
32
+
33
+ from official.benchmark import bert_benchmark_utils as benchmark_utils
34
+ from official.benchmark import owner_utils
35
+ from official.nlp.bert import configs
36
+ from official.nlp.bert import run_classifier
37
+ from official.utils.misc import distribution_utils
38
+ from official.benchmark import benchmark_wrappers
39
+
40
+ # pylint: disable=line-too-long
41
+ PRETRAINED_CHECKPOINT_PATH = 'gs://cloud-tpu-checkpoints/bert/keras_bert/uncased_L-24_H-1024_A-16/bert_model.ckpt'
42
+ CLASSIFIER_TRAIN_DATA_PATH = 'gs://tf-perfzero-data/bert/classification/mrpc_train.tf_record'
43
+ CLASSIFIER_EVAL_DATA_PATH = 'gs://tf-perfzero-data/bert/classification/mrpc_eval.tf_record'
44
+ CLASSIFIER_INPUT_META_DATA_PATH = 'gs://tf-perfzero-data/bert/classification/mrpc_meta_data'
45
+ MODEL_CONFIG_FILE_PATH = 'gs://cloud-tpu-checkpoints/bert/keras_bert/uncased_L-24_H-1024_A-16/bert_config.json'
46
+ # pylint: enable=line-too-long
47
+
48
+ TMP_DIR = os.getenv('TMPDIR')
49
+ FLAGS = flags.FLAGS
50
+
51
+
52
+ class BertClassifyBenchmarkBase(benchmark_utils.BertBenchmarkBase):
53
+ """Base class to hold methods common to test classes in the module."""
54
+
55
+ def __init__(self, output_dir=None, tpu=None):
56
+ super(BertClassifyBenchmarkBase, self).__init__(output_dir, tpu=tpu)
57
+ self.num_epochs = None
58
+ self.num_steps_per_epoch = None
59
+ FLAGS.steps_per_loop = 1
60
+
61
+ @flagsaver.flagsaver
62
+ def _run_bert_classifier(self, callbacks=None, use_ds=True):
63
+ """Starts BERT classification task."""
64
+ with tf.io.gfile.GFile(FLAGS.input_meta_data_path, 'rb') as reader:
65
+ input_meta_data = json.loads(reader.read().decode('utf-8'))
66
+
67
+ bert_config = configs.BertConfig.from_json_file(FLAGS.bert_config_file)
68
+ epochs = self.num_epochs if self.num_epochs else FLAGS.num_train_epochs
69
+ if self.num_steps_per_epoch:
70
+ steps_per_epoch = self.num_steps_per_epoch
71
+ else:
72
+ train_data_size = input_meta_data['train_data_size']
73
+ steps_per_epoch = int(train_data_size / FLAGS.train_batch_size)
74
+ warmup_steps = int(epochs * steps_per_epoch * 0.1)
75
+ eval_steps = int(
76
+ math.ceil(input_meta_data['eval_data_size'] / FLAGS.eval_batch_size))
77
+ if self.tpu:
78
+ strategy = distribution_utils.get_distribution_strategy(
79
+ distribution_strategy='tpu', tpu_address=self.tpu)
80
+ else:
81
+ strategy = distribution_utils.get_distribution_strategy(
82
+ distribution_strategy='mirrored' if use_ds else 'off',
83
+ num_gpus=self.num_gpus)
84
+
85
+ max_seq_length = input_meta_data['max_seq_length']
86
+ train_input_fn = run_classifier.get_dataset_fn(
87
+ FLAGS.train_data_path,
88
+ max_seq_length,
89
+ FLAGS.train_batch_size,
90
+ is_training=True)
91
+ eval_input_fn = run_classifier.get_dataset_fn(
92
+ FLAGS.eval_data_path,
93
+ max_seq_length,
94
+ FLAGS.eval_batch_size,
95
+ is_training=False)
96
+ _, summary = run_classifier.run_bert_classifier(
97
+ strategy,
98
+ bert_config,
99
+ input_meta_data,
100
+ FLAGS.model_dir,
101
+ epochs,
102
+ steps_per_epoch,
103
+ FLAGS.steps_per_loop,
104
+ eval_steps,
105
+ warmup_steps,
106
+ FLAGS.learning_rate,
107
+ FLAGS.init_checkpoint,
108
+ train_input_fn,
109
+ eval_input_fn,
110
+ training_callbacks=False,
111
+ custom_callbacks=callbacks)
112
+ return summary
113
+
114
+
115
+ class BertClassifyBenchmarkReal(BertClassifyBenchmarkBase):
116
+ """Short benchmark performance tests for BERT model.
117
+
118
+ Tests BERT classification performance in different GPU, TPU configurations.
119
+ The naming convention of below test cases follow
120
+ `benchmark_(number of gpus)_gpu_(dataset type)` for GPUs and
121
+ `benchmark_(topology)_tpu_(dataset type)` for TPUs.
122
+ """
123
+
124
+ def __init__(self, output_dir=TMP_DIR, tpu=None, **kwargs):
125
+ super(BertClassifyBenchmarkReal, self).__init__(
126
+ output_dir=output_dir, tpu=tpu)
127
+
128
+ self.train_data_path = CLASSIFIER_TRAIN_DATA_PATH
129
+ self.eval_data_path = CLASSIFIER_EVAL_DATA_PATH
130
+ self.bert_config_file = MODEL_CONFIG_FILE_PATH
131
+ self.input_meta_data_path = CLASSIFIER_INPUT_META_DATA_PATH
132
+
133
+ # Since we only care about performance metrics, we limit
134
+ # the number of training steps and epochs to prevent unnecessarily
135
+ # long tests.
136
+ self.num_steps_per_epoch = 100
137
+ self.num_epochs = 1
138
+
139
+ @benchmark_wrappers.enable_runtime_flags
140
+ def _run_and_report_benchmark(self,
141
+ training_summary_path,
142
+ min_accuracy=0,
143
+ max_accuracy=1,
144
+ use_ds=True):
145
+ """Starts BERT performance benchmark test."""
146
+ start_time_sec = time.time()
147
+ summary = self._run_bert_classifier(
148
+ callbacks=[self.timer_callback], use_ds=use_ds)
149
+ wall_time_sec = time.time() - start_time_sec
150
+
151
+ # Since we do not load from any pretrained checkpoints, we ignore all
152
+ # accuracy metrics.
153
+ summary.pop('eval_metrics', None)
154
+ summary['start_time_sec'] = start_time_sec
155
+
156
+ super(BertClassifyBenchmarkReal, self)._report_benchmark(
157
+ stats=summary,
158
+ wall_time_sec=wall_time_sec,
159
+ min_accuracy=min_accuracy,
160
+ max_accuracy=max_accuracy)
161
+
162
+ def benchmark_1_gpu_mrpc(self):
163
+ """Test BERT model performance with 1 GPU."""
164
+
165
+ self._setup()
166
+ self.num_gpus = 1
167
+ FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu_mrpc')
168
+ FLAGS.train_data_path = self.train_data_path
169
+ FLAGS.eval_data_path = self.eval_data_path
170
+ FLAGS.input_meta_data_path = self.input_meta_data_path
171
+ FLAGS.bert_config_file = self.bert_config_file
172
+ FLAGS.train_batch_size = 4
173
+ FLAGS.eval_batch_size = 4
174
+
175
+ summary_path = os.path.join(FLAGS.model_dir,
176
+ 'summaries/training_summary.txt')
177
+ self._run_and_report_benchmark(summary_path)
178
+
179
+ def benchmark_1_gpu_mrpc_xla(self):
180
+ """Test BERT model performance with 1 GPU."""
181
+
182
+ self._setup()
183
+ self.num_gpus = 1
184
+ FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu_mrpc_xla')
185
+ FLAGS.train_data_path = self.train_data_path
186
+ FLAGS.eval_data_path = self.eval_data_path
187
+ FLAGS.input_meta_data_path = self.input_meta_data_path
188
+ FLAGS.bert_config_file = self.bert_config_file
189
+ FLAGS.train_batch_size = 4
190
+ FLAGS.eval_batch_size = 4
191
+ FLAGS.enable_xla = True
192
+
193
+ summary_path = os.path.join(FLAGS.model_dir,
194
+ 'summaries/training_summary.txt')
195
+ self._run_and_report_benchmark(summary_path)
196
+
197
+ def benchmark_1_gpu_mrpc_no_dist_strat(self):
198
+ """Test BERT model performance with 1 GPU, no distribution strategy."""
199
+
200
+ self._setup()
201
+ self.num_gpus = 1
202
+ FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu_mrpc_no_dist_strat')
203
+ FLAGS.train_data_path = self.train_data_path
204
+ FLAGS.eval_data_path = self.eval_data_path
205
+ FLAGS.input_meta_data_path = self.input_meta_data_path
206
+ FLAGS.bert_config_file = self.bert_config_file
207
+ FLAGS.train_batch_size = 4
208
+ FLAGS.eval_batch_size = 4
209
+
210
+ summary_path = os.path.join(FLAGS.model_dir,
211
+ 'summaries/training_summary.txt')
212
+ self._run_and_report_benchmark(summary_path, use_ds=False)
213
+
214
+ @owner_utils.Owner('tf-model-garden')
215
+ def benchmark_8_gpu_mrpc(self):
216
+ """Test BERT model performance with 8 GPUs."""
217
+
218
+ self._setup()
219
+ FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_mrpc')
220
+ FLAGS.train_data_path = self.train_data_path
221
+ FLAGS.eval_data_path = self.eval_data_path
222
+ FLAGS.input_meta_data_path = self.input_meta_data_path
223
+ FLAGS.bert_config_file = self.bert_config_file
224
+
225
+ summary_path = os.path.join(FLAGS.model_dir,
226
+ 'summaries/training_summary.txt')
227
+ self._run_and_report_benchmark(summary_path)
228
+
229
+ def benchmark_1_gpu_amp_mrpc_no_dist_strat(self):
230
+ """Performance for 1 GPU no DS with automatic mixed precision."""
231
+ self._setup()
232
+ self.num_gpus = 1
233
+ FLAGS.model_dir = self._get_model_dir(
234
+ 'benchmark_1_gpu_amp_mrpc_no_dist_strat')
235
+ FLAGS.train_data_path = self.train_data_path
236
+ FLAGS.eval_data_path = self.eval_data_path
237
+ FLAGS.input_meta_data_path = self.input_meta_data_path
238
+ FLAGS.bert_config_file = self.bert_config_file
239
+ FLAGS.train_batch_size = 4
240
+ FLAGS.eval_batch_size = 4
241
+ FLAGS.dtype = 'fp16'
242
+ FLAGS.fp16_implementation = 'graph_rewrite'
243
+
244
+ summary_path = os.path.join(FLAGS.model_dir,
245
+ 'summaries/training_summary.txt')
246
+ self._run_and_report_benchmark(summary_path, use_ds=False)
247
+
248
+ def benchmark_8_gpu_amp_mrpc(self):
249
+ """Test BERT model performance with 8 GPUs with automatic mixed precision."""
250
+
251
+ self._setup()
252
+ self.num_gpus = 8
253
+ FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_amp_mrpc')
254
+ FLAGS.train_data_path = self.train_data_path
255
+ FLAGS.eval_data_path = self.eval_data_path
256
+ FLAGS.input_meta_data_path = self.input_meta_data_path
257
+ FLAGS.bert_config_file = self.bert_config_file
258
+ FLAGS.train_batch_size = 32
259
+ FLAGS.eval_batch_size = 32
260
+ FLAGS.dtype = 'fp16'
261
+ FLAGS.fp16_implementation = 'graph_rewrite'
262
+
263
+ summary_path = os.path.join(FLAGS.model_dir,
264
+ 'summaries/training_summary.txt')
265
+ self._run_and_report_benchmark(summary_path, use_ds=False)
266
+
267
+ @owner_utils.Owner('tf-model-garden')
268
+ def benchmark_2x2_tpu_mrpc(self):
269
+ """Test BERT model performance with 2x2 TPU."""
270
+
271
+ self._setup()
272
+ FLAGS.steps_per_loop = 50
273
+ FLAGS.model_dir = self._get_model_dir('benchmark_2x2_tpu_mrpc')
274
+ FLAGS.train_data_path = self.train_data_path
275
+ FLAGS.eval_data_path = self.eval_data_path
276
+ FLAGS.input_meta_data_path = self.input_meta_data_path
277
+ FLAGS.bert_config_file = self.bert_config_file
278
+ FLAGS.train_batch_size = 32
279
+ FLAGS.eval_batch_size = 32
280
+
281
+ summary_path = os.path.join(FLAGS.model_dir,
282
+ 'summaries/training_summary.txt')
283
+ self._run_and_report_benchmark(summary_path, use_ds=False)
284
+
285
+
286
+ class BertClassifyAccuracy(BertClassifyBenchmarkBase):
287
+ """Short accuracy test for BERT model.
288
+
289
+ Tests BERT classification task model accuracy. The naming
290
+ convention of below test cases follow
291
+ `benchmark_(number of gpus)_gpu_(dataset type)` format.
292
+ """
293
+
294
+ def __init__(self, output_dir=TMP_DIR, tpu=None, **kwargs):
295
+ self.train_data_path = CLASSIFIER_TRAIN_DATA_PATH
296
+ self.eval_data_path = CLASSIFIER_EVAL_DATA_PATH
297
+ self.bert_config_file = MODEL_CONFIG_FILE_PATH
298
+ self.input_meta_data_path = CLASSIFIER_INPUT_META_DATA_PATH
299
+ self.pretrained_checkpoint_path = PRETRAINED_CHECKPOINT_PATH
300
+
301
+ super(BertClassifyAccuracy, self).__init__(output_dir=output_dir, tpu=tpu)
302
+
303
+ @benchmark_wrappers.enable_runtime_flags
304
+ def _run_and_report_benchmark(self,
305
+ training_summary_path,
306
+ min_accuracy=0.84,
307
+ max_accuracy=0.88):
308
+ """Starts BERT accuracy benchmark test."""
309
+
310
+ start_time_sec = time.time()
311
+ summary = self._run_bert_classifier(callbacks=[self.timer_callback])
312
+ wall_time_sec = time.time() - start_time_sec
313
+
314
+ super(BertClassifyAccuracy, self)._report_benchmark(
315
+ stats=summary,
316
+ wall_time_sec=wall_time_sec,
317
+ min_accuracy=min_accuracy,
318
+ max_accuracy=max_accuracy)
319
+
320
+ def _setup(self):
321
+ super(BertClassifyAccuracy, self)._setup()
322
+ FLAGS.train_data_path = self.train_data_path
323
+ FLAGS.eval_data_path = self.eval_data_path
324
+ FLAGS.input_meta_data_path = self.input_meta_data_path
325
+ FLAGS.bert_config_file = self.bert_config_file
326
+ FLAGS.init_checkpoint = self.pretrained_checkpoint_path
327
+
328
+ @owner_utils.Owner('tf-model-garden')
329
+ def benchmark_8_gpu_mrpc(self):
330
+ """Run BERT model accuracy test with 8 GPUs.
331
+
332
+ Due to comparatively small cardinality of MRPC dataset, training
333
+ accuracy metric has high variance between trainings. As so, we
334
+ set the wide range of allowed accuracy (84% to 88%).
335
+ """
336
+ self._setup()
337
+ FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_mrpc')
338
+
339
+ summary_path = os.path.join(FLAGS.model_dir,
340
+ 'summaries/training_summary.txt')
341
+ self._run_and_report_benchmark(summary_path)
342
+
343
+ def benchmark_8_gpu_mrpc_xla(self):
344
+ """Run BERT model accuracy test with 8 GPUs with XLA."""
345
+ self._setup()
346
+ FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_mrpc_xla')
347
+ FLAGS.enable_xla = True
348
+ summary_path = os.path.join(FLAGS.model_dir,
349
+ 'summaries/training_summary.txt')
350
+ self._run_and_report_benchmark(summary_path)
351
+
352
+ @owner_utils.Owner('tf-model-garden')
353
+ def benchmark_2x2_tpu_mrpc(self):
354
+ """Run BERT model accuracy test on 2x2 TPU."""
355
+ self._setup()
356
+ FLAGS.steps_per_loop = 50
357
+ FLAGS.model_dir = self._get_model_dir('benchmark_2x2_tpu_mrpc')
358
+
359
+ summary_path = os.path.join(FLAGS.model_dir,
360
+ 'summaries/training_summary.txt')
361
+ self._run_and_report_benchmark(summary_path)
362
+
363
+
364
+ if __name__ == '__main__':
365
+ tf.test.main()
models/official/benchmark/bert_benchmark_utils.py ADDED
@@ -0,0 +1,127 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2019 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
+ """Utility functions or classes shared between BERT benchmarks."""
16
+
17
+ from __future__ import absolute_import
18
+ from __future__ import division
19
+ from __future__ import print_function
20
+
21
+ import time
22
+
23
+ # pylint: disable=g-bad-import-order
24
+ import numpy as np
25
+ from absl import flags
26
+ import tensorflow as tf
27
+ # pylint: enable=g-bad-import-order
28
+
29
+ from official.utils.flags import core as flags_core
30
+ from official.benchmark.perfzero_benchmark import PerfZeroBenchmark
31
+
32
+ FLAGS = flags.FLAGS
33
+
34
+
35
+ class BenchmarkTimerCallback(tf.keras.callbacks.Callback):
36
+ """Callback that records time it takes to run each batch."""
37
+
38
+ def __init__(self, num_batches_to_skip=10):
39
+ super(BenchmarkTimerCallback, self).__init__()
40
+ self.batch_start_times = {}
41
+ self.batch_stop_times = {}
42
+
43
+ def on_batch_begin(self, batch, logs=None):
44
+ self.batch_start_times[batch] = time.time()
45
+
46
+ def on_batch_end(self, batch, logs=None):
47
+ # If there are multiple steps_per_loop, the end batch index will not be the
48
+ # same as the starting index. Use the last starting index instead.
49
+ if batch not in self.batch_start_times:
50
+ batch = max(self.batch_start_times.keys())
51
+
52
+ self.batch_stop_times[batch] = time.time()
53
+
54
+ def get_examples_per_sec(self, batch_size, num_batches_to_skip=1):
55
+ batch_durations = []
56
+ for batch in self.batch_start_times:
57
+ if batch in self.batch_stop_times and batch >= num_batches_to_skip:
58
+ batch_durations.append(self.batch_stop_times[batch] -
59
+ self.batch_start_times[batch])
60
+ return batch_size / np.mean(batch_durations)
61
+
62
+ def get_startup_time(self, program_start_time):
63
+ return self.batch_start_times[0] - program_start_time
64
+
65
+
66
+ class BertBenchmarkBase(PerfZeroBenchmark):
67
+ """Base class to hold methods common to test classes."""
68
+ local_flags = None
69
+
70
+ def __init__(self, output_dir=None, tpu=None, **kwargs):
71
+ super(BertBenchmarkBase, self).__init__(
72
+ output_dir=output_dir, tpu=tpu, **kwargs)
73
+ self.num_gpus = 8
74
+ self.timer_callback = None
75
+
76
+ def _setup(self):
77
+ """Sets up and resets flags before each test."""
78
+ super(BertBenchmarkBase, self)._setup()
79
+ self.timer_callback = BenchmarkTimerCallback()
80
+
81
+ def _report_benchmark(self, stats, wall_time_sec, min_accuracy, max_accuracy):
82
+ """Report benchmark results by writing to local protobuf file.
83
+
84
+ Args:
85
+ stats: dict returned from BERT models with known entries.
86
+ wall_time_sec: the during of the benchmark execution in seconds
87
+ min_accuracy: Minimum classification accuracy constraint to verify
88
+ correctness of the model.
89
+ max_accuracy: Maximum classification accuracy constraint to verify
90
+ correctness of the model.
91
+ """
92
+ metrics = [{
93
+ 'name': 'training_loss',
94
+ 'value': stats['train_loss'],
95
+ }]
96
+ if self.timer_callback:
97
+ metrics.append({
98
+ 'name':
99
+ 'exp_per_second',
100
+ 'value':
101
+ self.timer_callback.get_examples_per_sec(FLAGS.train_batch_size *
102
+ FLAGS.steps_per_loop)
103
+ })
104
+ else:
105
+ metrics.append({
106
+ 'name': 'exp_per_second',
107
+ 'value': 0.0,
108
+ })
109
+ if self.timer_callback and 'start_time_sec' in stats:
110
+ metrics.append({
111
+ 'name': 'startup_time',
112
+ 'value': self.timer_callback.get_startup_time(stats['start_time_sec'])
113
+ })
114
+
115
+ if 'eval_metrics' in stats:
116
+ metrics.append({
117
+ 'name': 'eval_accuracy',
118
+ 'value': stats['eval_metrics'],
119
+ 'min_value': min_accuracy,
120
+ 'max_value': max_accuracy,
121
+ })
122
+ flags_str = flags_core.get_nondefault_flags_as_str()
123
+ self.report_benchmark(
124
+ iters=stats['total_training_steps'],
125
+ wall_time=wall_time_sec,
126
+ metrics=metrics,
127
+ extras={'flags': flags_str})
models/official/benchmark/bert_pretrain_benchmark.py ADDED
@@ -0,0 +1,179 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Lint as: python3
2
+ # Copyright 2020 The TensorFlow Authors. All Rights Reserved.
3
+ #
4
+ # Licensed under the Apache License, Version 2.0 (the "License");
5
+ # you may not use this file except in compliance with the License.
6
+ # You may obtain a copy of the License at
7
+ #
8
+ # http://www.apache.org/licenses/LICENSE-2.0
9
+ #
10
+ # Unless required by applicable law or agreed to in writing, software
11
+ # distributed under the License is distributed on an "AS IS" BASIS,
12
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
+ # See the License for the specific language governing permissions and
14
+ # limitations under the License.
15
+ # ==============================================================================
16
+ """Executes benchmark testing for bert pretraining."""
17
+ # pylint: disable=line-too-long
18
+ from __future__ import print_function
19
+
20
+ import json
21
+ import os
22
+ import time
23
+ from typing import Optional
24
+
25
+ from absl import flags
26
+ from absl import logging
27
+ import tensorflow as tf # pylint: disable=g-bad-import-order
28
+
29
+ from official.benchmark import benchmark_wrappers
30
+ from official.benchmark import bert_benchmark_utils
31
+ from official.benchmark import owner_utils
32
+ from official.nlp.bert import run_pretraining
33
+ from official.utils.flags import core as flags_core
34
+ from official.utils.misc import distribution_utils
35
+
36
+ # Pretrain masked lanauge modeling accuracy range:
37
+ MIN_MLM_ACCURACY = 0.635
38
+ MAX_MLM_ACCURACY = 0.645
39
+
40
+ # Pretrain next sentence prediction accuracy range:
41
+ MIN_NSP_ACCURACY = 0.94
42
+ MAX_NSP_ACCURACY = 0.96
43
+
44
+ BERT_PRETRAIN_FILES_SEQ128 = 'gs://mlcompass-data/bert/pretraining_data/seq_128/wikipedia.tfrecord*,gs://mlcompass-data/bert/pretraining_data/seq_128/books.tfrecord*'
45
+ BERT_BASE_CONFIG_FILE = 'gs://cloud-tpu-checkpoints/bert/keras_bert/uncased_L-12_H-768_A-12/bert_config.json'
46
+
47
+ FLAGS = flags.FLAGS
48
+
49
+
50
+ class BertPretrainAccuracyBenchmark(bert_benchmark_utils.BertBenchmarkBase):
51
+ """Benchmark accuracy tests for BERT Pretraining."""
52
+
53
+ def __init__(self,
54
+ output_dir: Optional[str] = None,
55
+ tpu: Optional[str] = None,
56
+ **kwargs):
57
+ """Inits BertPretrainAccuracyBenchmark class.
58
+
59
+ Args:
60
+ output_dir: Directory where to output e.g. log files
61
+ tpu: TPU name to use in a TPU benchmark.
62
+ **kwargs: Additional keyword arguments.
63
+ """
64
+ super(BertPretrainAccuracyBenchmark, self).__init__(
65
+ output_dir=output_dir, tpu=tpu, **kwargs)
66
+
67
+ @benchmark_wrappers.enable_runtime_flags
68
+ def _run_and_report_benchmark(self, summary_path: str, report_accuracy: bool):
69
+ """Runs and reports the benchmark given the provided configuration."""
70
+ distribution = distribution_utils.get_distribution_strategy(
71
+ distribution_strategy='tpu', tpu_address=self.tpu)
72
+ logging.info('Flags: %s', flags_core.get_nondefault_flags_as_str())
73
+ start_time_sec = time.time()
74
+ run_pretraining.run_bert_pretrain(
75
+ strategy=distribution, custom_callbacks=self.timer_callback)
76
+ wall_time_sec = time.time() - start_time_sec
77
+
78
+ with tf.io.gfile.GFile(summary_path, 'rb') as reader:
79
+ summary = json.loads(reader.read().decode('utf-8'))
80
+ self._report_benchmark(summary, start_time_sec, wall_time_sec,
81
+ report_accuracy)
82
+
83
+ def _report_benchmark(self, summary, start_time_sec, wall_time_sec,
84
+ report_accuracy):
85
+ metrics = [{
86
+ 'name': 'train_loss',
87
+ 'value': summary['train_loss'],
88
+ }, {
89
+ 'name':
90
+ 'exp_per_second',
91
+ 'value':
92
+ self.timer_callback.get_examples_per_sec(FLAGS.train_batch_size *
93
+ FLAGS.steps_per_loop)
94
+ }, {
95
+ 'name': 'startup_time',
96
+ 'value': self.timer_callback.get_startup_time(start_time_sec)
97
+ }]
98
+ if report_accuracy:
99
+ metrics.extend([{
100
+ 'name': 'masked_lm_accuracy',
101
+ 'value': summary['masked_lm_accuracy'],
102
+ 'min_value': MIN_MLM_ACCURACY,
103
+ 'max_value': MAX_MLM_ACCURACY,
104
+ }, {
105
+ 'name': 'next_sentence_accuracy',
106
+ 'value': summary['next_sentence_accuracy'],
107
+ 'min_value': MIN_NSP_ACCURACY,
108
+ 'max_value': MAX_NSP_ACCURACY,
109
+ }])
110
+ self.report_benchmark(
111
+ iters=summary['total_training_steps'],
112
+ wall_time=wall_time_sec,
113
+ metrics=metrics,
114
+ extras={'flags': flags_core.get_nondefault_flags_as_str()})
115
+
116
+ def _specify_common_flags(self):
117
+ FLAGS.bert_config_file = BERT_BASE_CONFIG_FILE
118
+ FLAGS.train_batch_size = 512
119
+ FLAGS.learning_rate = 1e-4
120
+ FLAGS.warmup_steps = 10000
121
+ FLAGS.steps_per_loop = 10000
122
+ FLAGS.distribution_strategy = 'tpu'
123
+ FLAGS.input_files = BERT_PRETRAIN_FILES_SEQ128
124
+ FLAGS.max_seq_length = 128
125
+ FLAGS.max_predictions_per_seq = 20
126
+ FLAGS.dtype = 'bf16'
127
+
128
+ @owner_utils.Owner('tf-model-garden')
129
+ def benchmark_accuracy_8x8_tpu_bf16_seq128_500k_steps(self):
130
+ """Test bert pretraining with 8x8 TPU for 500k steps."""
131
+ # This is used for accuracy test.
132
+ self._setup()
133
+ self._specify_common_flags()
134
+ FLAGS.num_steps_per_epoch = 500000
135
+ FLAGS.num_train_epochs = 1
136
+ FLAGS.model_dir = self._get_model_dir(
137
+ 'benchmark_accuracy_8x8_tpu_bf16_seq128_500k_steps')
138
+ summary_path = os.path.join(FLAGS.model_dir,
139
+ 'summaries/training_summary.txt')
140
+ # Set train_summary_interval to -1 to disable training summary, because
141
+ # writing summary to gcs may fail and summaries are not needed for this
142
+ # accuracy benchmark test.
143
+ FLAGS.train_summary_interval = -1
144
+ self._run_and_report_benchmark(summary_path=summary_path,
145
+ report_accuracy=True)
146
+
147
+ @owner_utils.Owner('tf-model-garden')
148
+ def benchmark_perf_4x4_tpu_bf16_seq128_10k_steps(self):
149
+ """Test bert pretraining with 4x4 TPU for 10000 steps."""
150
+ self._setup()
151
+ self._specify_common_flags()
152
+ FLAGS.num_steps_per_epoch = 5000
153
+ FLAGS.num_train_epochs = 2
154
+ FLAGS.model_dir = self._get_model_dir(
155
+ 'benchmark_perf_4x4_tpu_bf16_seq128_10k_steps')
156
+ summary_path = os.path.join(FLAGS.model_dir,
157
+ 'summaries/training_summary.txt')
158
+ # Disable accuracy check.
159
+ self._run_and_report_benchmark(
160
+ summary_path=summary_path, report_accuracy=False)
161
+
162
+ @owner_utils.Owner('tf-model-garden')
163
+ def benchmark_perf_8x8_tpu_bf16_seq128_10k_steps(self):
164
+ """Test bert pretraining with 8x8 TPU for 10000 steps."""
165
+ self._setup()
166
+ self._specify_common_flags()
167
+ FLAGS.num_steps_per_epoch = 5000
168
+ FLAGS.num_train_epochs = 2
169
+ FLAGS.model_dir = self._get_model_dir(
170
+ 'benchmark_perf_8x8_tpu_bf16_seq128_10k_steps')
171
+ summary_path = os.path.join(FLAGS.model_dir,
172
+ 'summaries/training_summary.txt')
173
+ # Disable accuracy check.
174
+ self._run_and_report_benchmark(summary_path=summary_path,
175
+ report_accuracy=False)
176
+
177
+
178
+ if __name__ == '__main__':
179
+ tf.test.main()
models/official/benchmark/bert_squad_benchmark.py ADDED
@@ -0,0 +1,608 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2019 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
+ """Executes BERT SQuAD benchmarks and accuracy tests."""
16
+
17
+ from __future__ import absolute_import
18
+ from __future__ import division
19
+ from __future__ import print_function
20
+
21
+ import json
22
+ import os
23
+ import time
24
+
25
+ # pylint: disable=g-bad-import-order
26
+ from absl import flags
27
+ from absl import logging
28
+ from absl.testing import flagsaver
29
+ import tensorflow as tf
30
+ # pylint: enable=g-bad-import-order
31
+
32
+ from official.benchmark import bert_benchmark_utils as benchmark_utils
33
+ from official.benchmark import owner_utils
34
+ from official.nlp.bert import run_squad
35
+ from official.utils.misc import distribution_utils
36
+ from official.utils.misc import keras_utils
37
+ from official.benchmark import benchmark_wrappers
38
+
39
+
40
+ # pylint: disable=line-too-long
41
+ PRETRAINED_CHECKPOINT_PATH = 'gs://cloud-tpu-checkpoints/bert/keras_bert/uncased_L-24_H-1024_A-16/bert_model.ckpt'
42
+ SQUAD_TRAIN_DATA_PATH = 'gs://tf-perfzero-data/bert/squad/squad_train.tf_record'
43
+ SQUAD_PREDICT_FILE = 'gs://tf-perfzero-data/bert/squad/dev-v1.1.json'
44
+ SQUAD_VOCAB_FILE = 'gs://tf-perfzero-data/bert/squad/vocab.txt'
45
+ SQUAD_MEDIUM_INPUT_META_DATA_PATH = 'gs://tf-perfzero-data/bert/squad/squad_medium_meta_data'
46
+ SQUAD_LONG_INPUT_META_DATA_PATH = 'gs://tf-perfzero-data/bert/squad/squad_long_meta_data'
47
+ SQUAD_FULL_INPUT_META_DATA_PATH = 'gs://tf-perfzero-data/bert/squad/squad_full_meta_data'
48
+ MODEL_CONFIG_FILE_PATH = 'gs://cloud-tpu-checkpoints/bert/keras_bert/uncased_L-24_H-1024_A-16/bert_config.json'
49
+ # pylint: enable=line-too-long
50
+
51
+ TMP_DIR = os.getenv('TMPDIR')
52
+ FLAGS = flags.FLAGS
53
+
54
+
55
+ class BertSquadBenchmarkBase(benchmark_utils.BertBenchmarkBase):
56
+ """Base class to hold methods common to test classes in the module."""
57
+
58
+ def __init__(self, output_dir=None, tpu=None):
59
+ super(BertSquadBenchmarkBase, self).__init__(output_dir=output_dir, tpu=tpu)
60
+
61
+ def _read_training_summary_from_file(self):
62
+ """Reads the training summary from a file."""
63
+ summary_path = os.path.join(FLAGS.model_dir,
64
+ 'summaries/training_summary.txt')
65
+ with tf.io.gfile.GFile(summary_path, 'rb') as reader:
66
+ return json.loads(reader.read().decode('utf-8'))
67
+
68
+ def _read_input_meta_data_from_file(self):
69
+ """Reads the input metadata from a file."""
70
+ with tf.io.gfile.GFile(FLAGS.input_meta_data_path, 'rb') as reader:
71
+ return json.loads(reader.read().decode('utf-8'))
72
+
73
+ def _get_distribution_strategy(self, ds_type='mirrored'):
74
+ """Gets the distribution strategy.
75
+
76
+ Args:
77
+ ds_type: String, the distribution strategy type to be used. Can be
78
+ 'mirrored', 'multi_worker_mirrored', 'tpu' and 'off'.
79
+
80
+ Returns:
81
+ A `tf.distribute.DistibutionStrategy` object.
82
+ """
83
+ if self.tpu or ds_type == 'tpu':
84
+ return distribution_utils.get_distribution_strategy(
85
+ distribution_strategy='tpu', tpu_address=self.tpu)
86
+ elif ds_type == 'multi_worker_mirrored':
87
+ # Configures cluster spec for multi-worker distribution strategy.
88
+ _ = distribution_utils.configure_cluster(FLAGS.worker_hosts,
89
+ FLAGS.task_index)
90
+ return distribution_utils.get_distribution_strategy(
91
+ distribution_strategy=ds_type,
92
+ num_gpus=self.num_gpus,
93
+ all_reduce_alg=FLAGS.all_reduce_alg)
94
+
95
+ def _init_gpu_and_data_threads(self):
96
+ """Set env variables before any TF calls."""
97
+ if FLAGS.tf_gpu_thread_mode:
98
+ keras_utils.set_gpu_thread_mode_and_count(
99
+ per_gpu_thread_count=FLAGS.per_gpu_thread_count,
100
+ gpu_thread_mode=FLAGS.tf_gpu_thread_mode,
101
+ num_gpus=self.num_gpus,
102
+ datasets_num_private_threads=FLAGS.datasets_num_private_threads)
103
+
104
+ @flagsaver.flagsaver
105
+ def _train_squad(self, run_eagerly=False, ds_type='mirrored'):
106
+ """Runs BERT SQuAD training. Uses mirrored strategy by default."""
107
+ self._init_gpu_and_data_threads()
108
+ input_meta_data = self._read_input_meta_data_from_file()
109
+ strategy = self._get_distribution_strategy(ds_type)
110
+
111
+ run_squad.train_squad(
112
+ strategy=strategy,
113
+ input_meta_data=input_meta_data,
114
+ run_eagerly=run_eagerly,
115
+ custom_callbacks=[self.timer_callback])
116
+
117
+ @flagsaver.flagsaver
118
+ def _evaluate_squad(self, ds_type='mirrored'):
119
+ """Runs BERT SQuAD evaluation. Uses mirrored strategy by default."""
120
+ self._init_gpu_and_data_threads()
121
+ input_meta_data = self._read_input_meta_data_from_file()
122
+ strategy = self._get_distribution_strategy(ds_type)
123
+
124
+ if input_meta_data.get('version_2_with_negative', False):
125
+ logging.error('In memory evaluation result for SQuAD v2 is not accurate')
126
+ eval_metrics = run_squad.eval_squad(strategy=strategy,
127
+ input_meta_data=input_meta_data)
128
+ # Use F1 score as reported evaluation metric.
129
+ self.eval_metrics = eval_metrics['final_f1']
130
+
131
+
132
+ class BertSquadBenchmarkReal(BertSquadBenchmarkBase):
133
+ """Short benchmark performance tests for BERT SQuAD model.
134
+
135
+ Tests BERT SQuAD performance in different GPU configurations.
136
+ The naming convention of below test cases follow
137
+ `benchmark_(number of gpus)_gpu` format for GPUs and
138
+ `benchmark_(topology)_tpu` format for TPUs.
139
+ """
140
+
141
+ def __init__(self, output_dir=TMP_DIR, tpu=None, **kwargs):
142
+ super(BertSquadBenchmarkReal, self).__init__(output_dir=output_dir, tpu=tpu)
143
+
144
+ def _setup(self):
145
+ """Sets up the benchmark and SQuAD flags."""
146
+ super(BertSquadBenchmarkReal, self)._setup()
147
+ FLAGS.train_data_path = SQUAD_TRAIN_DATA_PATH
148
+ FLAGS.predict_file = SQUAD_PREDICT_FILE
149
+ FLAGS.vocab_file = SQUAD_VOCAB_FILE
150
+ FLAGS.bert_config_file = MODEL_CONFIG_FILE_PATH
151
+ FLAGS.num_train_epochs = 1
152
+ FLAGS.steps_per_loop = 100
153
+
154
+ @benchmark_wrappers.enable_runtime_flags
155
+ def _run_and_report_benchmark(self,
156
+ run_eagerly=False,
157
+ ds_type='mirrored'):
158
+ """Runs the benchmark and reports various metrics."""
159
+ if FLAGS.train_batch_size <= 4 or run_eagerly:
160
+ FLAGS.input_meta_data_path = SQUAD_MEDIUM_INPUT_META_DATA_PATH
161
+ else:
162
+ FLAGS.input_meta_data_path = SQUAD_LONG_INPUT_META_DATA_PATH
163
+ start_time_sec = time.time()
164
+ self._train_squad(run_eagerly=run_eagerly, ds_type=ds_type)
165
+ wall_time_sec = time.time() - start_time_sec
166
+
167
+ summary = self._read_training_summary_from_file()
168
+ summary['start_time_sec'] = start_time_sec
169
+
170
+ super(BertSquadBenchmarkReal, self)._report_benchmark(
171
+ stats=summary,
172
+ wall_time_sec=wall_time_sec,
173
+ min_accuracy=0,
174
+ max_accuracy=1)
175
+
176
+ def benchmark_1_gpu(self):
177
+ """Tests BERT SQuAD model performance with 1 GPU."""
178
+
179
+ self._setup()
180
+ self.num_gpus = 1
181
+ FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu_squad')
182
+ FLAGS.train_batch_size = 4
183
+
184
+ self._run_and_report_benchmark()
185
+
186
+ def benchmark_1_gpu_eager(self):
187
+ """Tests BERT SQuAD model performance with 1 GPU."""
188
+
189
+ self._setup()
190
+ self.num_gpus = 1
191
+ FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu_squad_eager')
192
+ FLAGS.train_batch_size = 2
193
+
194
+ self._run_and_report_benchmark(run_eagerly=True)
195
+
196
+ def benchmark_1_gpu_xla(self):
197
+ """Tests BERT SQuAD model performance with 1 GPU with XLA."""
198
+
199
+ self._setup()
200
+ self.num_gpus = 1
201
+ FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu_xla_squad')
202
+ # XLA runs out of memory when running with batch size 4.
203
+ FLAGS.train_batch_size = 3
204
+ FLAGS.enable_xla = True
205
+
206
+ self._run_and_report_benchmark()
207
+
208
+ def benchmark_1_gpu_no_dist_strat(self):
209
+ """Tests BERT SQuAD model performance with 1 GPU without DS."""
210
+
211
+ self._setup()
212
+ self.num_gpus = 1
213
+ FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu_no_dist_strat_squad')
214
+ FLAGS.train_batch_size = 4
215
+
216
+ self._run_and_report_benchmark(ds_type='off')
217
+
218
+ def benchmark_1_gpu_eager_no_dist_strat(self):
219
+ """Tests BERT SQuAD model performance with 1 GPU with eager execution."""
220
+
221
+ self._setup()
222
+ self.num_gpus = 1
223
+ FLAGS.model_dir = self._get_model_dir(
224
+ 'benchmark_1_gpu_eager_no_dist_strat_squad')
225
+ FLAGS.train_batch_size = 4
226
+
227
+ self._run_and_report_benchmark(ds_type='off', run_eagerly=True)
228
+
229
+ @owner_utils.Owner('tf-model-garden')
230
+ def benchmark_8_gpu(self):
231
+ """Tests BERT SQuAD model performance with 8 GPUs."""
232
+
233
+ self._setup()
234
+ self.num_gpus = 8
235
+ FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_squad')
236
+ FLAGS.train_batch_size = 24
237
+ FLAGS.tf_gpu_thread_mode = 'gpu_private'
238
+
239
+ self._run_and_report_benchmark()
240
+
241
+ def benchmark_1_gpu_fp16_eager(self):
242
+ """Tests BERT SQuAD model performance with 1 GPU and FP16."""
243
+
244
+ self._setup()
245
+ self.num_gpus = 1
246
+ FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu_squad_fp16_eager')
247
+ FLAGS.train_batch_size = 4
248
+ FLAGS.dtype = 'fp16'
249
+ FLAGS.loss_scale = 'dynamic'
250
+
251
+ self._run_and_report_benchmark(run_eagerly=True)
252
+
253
+ def benchmark_1_gpu_fp16(self):
254
+ """Tests BERT SQuAD model performance with 1 GPU and FP16."""
255
+
256
+ self._setup()
257
+ self.num_gpus = 1
258
+ FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu_squad_fp16')
259
+ FLAGS.train_batch_size = 4
260
+ FLAGS.dtype = 'fp16'
261
+ FLAGS.loss_scale = 'dynamic'
262
+
263
+ self._run_and_report_benchmark()
264
+
265
+ def benchmark_1_gpu_xla_fp16(self):
266
+ """Tests BERT SQuAD model performance with 1 GPU with XLA and FP16."""
267
+
268
+ self._setup()
269
+ self.num_gpus = 1
270
+ FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu_xla_squad_fp16')
271
+ FLAGS.train_batch_size = 4
272
+ FLAGS.enable_xla = True
273
+ FLAGS.dtype = 'fp16'
274
+ FLAGS.loss_scale = 'dynamic'
275
+
276
+ self._run_and_report_benchmark()
277
+
278
+ def benchmark_8_gpu_fp16(self):
279
+ """Tests BERT SQuAD model performance with 8 GPUs."""
280
+
281
+ self._setup()
282
+ self.num_gpus = 8
283
+ FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_squad_fp16')
284
+ FLAGS.train_batch_size = 32
285
+ FLAGS.dtype = 'fp16'
286
+ FLAGS.loss_scale = 'dynamic'
287
+ FLAGS.tf_gpu_thread_mode = 'gpu_private'
288
+
289
+ self._run_and_report_benchmark()
290
+
291
+ def benchmark_8_gpu_xla_fp16(self):
292
+ """Tests BERT SQuAD model performance with 8 GPUs with XLA."""
293
+
294
+ self._setup()
295
+ self.num_gpus = 8
296
+ FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_squad_fp16')
297
+ FLAGS.train_batch_size = 32
298
+ FLAGS.enable_xla = True
299
+ FLAGS.dtype = 'fp16'
300
+ FLAGS.loss_scale = 'dynamic'
301
+
302
+ self._run_and_report_benchmark()
303
+
304
+ def benchmark_1_gpu_amp(self):
305
+ """Tests BERT SQuAD model performance with 1 GPU with automatic mixed precision."""
306
+
307
+ self._setup()
308
+ self.num_gpus = 1
309
+ FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu_amp_squad')
310
+ FLAGS.train_batch_size = 4
311
+ FLAGS.dtype = 'fp16'
312
+ FLAGS.fp16_implementation = 'graph_rewrite'
313
+
314
+ self._run_and_report_benchmark()
315
+
316
+ def benchmark_8_gpu_amp(self):
317
+ """Tests BERT SQuAD model performance with 1 GPU with automatic mixed precision."""
318
+
319
+ self._setup()
320
+ self.num_gpus = 8
321
+ FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_amp_squad')
322
+ FLAGS.train_batch_size = 32
323
+ FLAGS.dtype = 'fp16'
324
+ FLAGS.fp16_implementation = 'graph_rewrite'
325
+ FLAGS.tf_gpu_thread_mode = 'gpu_private'
326
+
327
+ self._run_and_report_benchmark()
328
+
329
+ @owner_utils.Owner('tf-model-garden')
330
+ def benchmark_2x2_tpu(self):
331
+ """Tests BERT SQuAD model performance with 2x2 TPU."""
332
+
333
+ self._setup()
334
+ FLAGS.model_dir = self._get_model_dir('benchmark_2x2_tpu')
335
+ FLAGS.train_batch_size = 48
336
+ FLAGS.predict_batch_size = 48
337
+ FLAGS.mode = 'train'
338
+ FLAGS.learning_rate = 8e-5
339
+ FLAGS.num_train_epochs = 1
340
+ FLAGS.steps_per_loop = 100
341
+ FLAGS.do_lower_case = True
342
+ FLAGS.init_checkpoint = PRETRAINED_CHECKPOINT_PATH
343
+ self._run_and_report_benchmark()
344
+
345
+
346
+ class BertSquadAccuracy(BertSquadBenchmarkBase):
347
+ """Short accuracy test for BERT SQuAD model.
348
+
349
+ Tests BERT SQuAD accuracy. The naming convention of below test cases follow
350
+ `benchmark_(number of gpus)_gpu` format for GPUs and
351
+ `benchmark_(topology)_tpu` format for TPUs.
352
+ """
353
+
354
+ def __init__(self, output_dir=None, tpu=None, **kwargs):
355
+ super(BertSquadAccuracy, self).__init__(output_dir=output_dir, tpu=tpu)
356
+
357
+ def _setup(self):
358
+ """Sets up the benchmark and SQuAD flags."""
359
+ super(BertSquadAccuracy, self)._setup()
360
+ FLAGS.train_data_path = SQUAD_TRAIN_DATA_PATH
361
+ FLAGS.predict_file = SQUAD_PREDICT_FILE
362
+ FLAGS.vocab_file = SQUAD_VOCAB_FILE
363
+ FLAGS.input_meta_data_path = SQUAD_FULL_INPUT_META_DATA_PATH
364
+ FLAGS.bert_config_file = MODEL_CONFIG_FILE_PATH
365
+ FLAGS.init_checkpoint = PRETRAINED_CHECKPOINT_PATH
366
+ FLAGS.num_train_epochs = 2
367
+ FLAGS.steps_per_loop = 100
368
+
369
+ @benchmark_wrappers.enable_runtime_flags
370
+ def _run_and_report_benchmark(self,
371
+ run_eagerly=False,
372
+ ds_type='mirrored'):
373
+ """Runs the benchmark and reports various metrics."""
374
+ start_time_sec = time.time()
375
+ self._train_squad(run_eagerly=run_eagerly, ds_type=ds_type)
376
+ self._evaluate_squad(ds_type=ds_type)
377
+ wall_time_sec = time.time() - start_time_sec
378
+
379
+ summary = self._read_training_summary_from_file()
380
+ summary['eval_metrics'] = self.eval_metrics
381
+ summary['start_time_sec'] = start_time_sec
382
+
383
+ super(BertSquadAccuracy, self)._report_benchmark(
384
+ stats=summary,
385
+ wall_time_sec=wall_time_sec,
386
+ min_accuracy=0.900,
387
+ max_accuracy=0.920)
388
+
389
+ def benchmark_1_gpu_eager(self):
390
+ """Tests BERT SQuAD model accuracy with 1 GPU with eager execution."""
391
+
392
+ self._setup()
393
+ self.num_gpus = 1
394
+ FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu_squad_eager')
395
+ FLAGS.train_batch_size = 4
396
+
397
+ self._run_and_report_benchmark(ds_type='off', run_eagerly=True)
398
+
399
+ @owner_utils.Owner('tf-model-garden')
400
+ def benchmark_8_gpu(self):
401
+ """Tests BERT SQuAD model accuracy with 8 GPUs."""
402
+
403
+ self._setup()
404
+ self.num_gpus = 8
405
+ FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_squad')
406
+ FLAGS.train_batch_size = 24
407
+ FLAGS.tf_gpu_thread_mode = 'gpu_private'
408
+
409
+ self._run_and_report_benchmark()
410
+
411
+ def benchmark_8_gpu_fp16(self):
412
+ """Tests BERT SQuAD model accuracy with 8 GPUs and FP16."""
413
+
414
+ self._setup()
415
+ self.num_gpus = 8
416
+ FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_squad_fp16')
417
+ FLAGS.train_batch_size = 32
418
+ FLAGS.dtype = 'fp16'
419
+ FLAGS.loss_scale = 'dynamic'
420
+ FLAGS.tf_gpu_thread_mode = 'gpu_private'
421
+
422
+ self._run_and_report_benchmark()
423
+
424
+ def benchmark_8_gpu_xla(self):
425
+ """Tests BERT SQuAD model accuracy with 8 GPUs."""
426
+
427
+ self._setup()
428
+ self.num_gpus = 8
429
+ FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_squad_xla')
430
+ FLAGS.train_batch_size = 32
431
+ FLAGS.enable_xla = True
432
+ FLAGS.tf_gpu_thread_mode = 'gpu_private'
433
+
434
+ self._run_and_report_benchmark()
435
+
436
+ @owner_utils.Owner('tf-model-garden')
437
+ def benchmark_2x2_tpu(self):
438
+ """Tests BERT SQuAD model accuracy with 2x2 TPU."""
439
+
440
+ self._setup()
441
+ FLAGS.model_dir = self._get_model_dir('benchmark_2x2_tpu')
442
+ FLAGS.train_batch_size = 48
443
+
444
+ self._run_and_report_benchmark()
445
+
446
+
447
+ class BertSquadMultiWorkerAccuracy(BertSquadBenchmarkBase):
448
+ """BERT SQuAD distributed accuracy tests with multiple workers."""
449
+
450
+ def __init__(self, output_dir=None, tpu=None, **kwargs):
451
+ super(BertSquadMultiWorkerAccuracy, self).__init__(
452
+ output_dir=output_dir, tpu=tpu)
453
+
454
+ def _setup(self):
455
+ """Sets up the benchmark and SQuAD flags."""
456
+ super(BertSquadMultiWorkerAccuracy, self)._setup()
457
+ FLAGS.train_data_path = SQUAD_TRAIN_DATA_PATH
458
+ FLAGS.predict_file = SQUAD_PREDICT_FILE
459
+ FLAGS.vocab_file = SQUAD_VOCAB_FILE
460
+ FLAGS.input_meta_data_path = SQUAD_FULL_INPUT_META_DATA_PATH
461
+ FLAGS.bert_config_file = MODEL_CONFIG_FILE_PATH
462
+ FLAGS.init_checkpoint = PRETRAINED_CHECKPOINT_PATH
463
+ FLAGS.num_train_epochs = 2
464
+ FLAGS.steps_per_loop = 100
465
+
466
+ @benchmark_wrappers.enable_runtime_flags
467
+ def _run_and_report_benchmark(self,
468
+ use_ds=True,
469
+ run_eagerly=False):
470
+ """Runs the benchmark and reports various metrics."""
471
+ start_time_sec = time.time()
472
+ self._train_squad(run_eagerly=run_eagerly,
473
+ ds_type='multi_worker_mirrored')
474
+ self._evaluate_squad(ds_type='multi_worker_mirrored')
475
+ wall_time_sec = time.time() - start_time_sec
476
+
477
+ summary = self._read_training_summary_from_file()
478
+ summary['eval_metrics'] = self.eval_metrics
479
+
480
+ super(BertSquadMultiWorkerAccuracy, self)._report_benchmark(
481
+ stats=summary,
482
+ wall_time_sec=wall_time_sec,
483
+ min_accuracy=0.900,
484
+ max_accuracy=0.920)
485
+
486
+ def _benchmark_common(self, num_workers, all_reduce_alg):
487
+ """Common to all benchmarks in this class."""
488
+ self._setup()
489
+
490
+ num_gpus = 8
491
+ FLAGS.num_gpus = num_gpus
492
+ FLAGS.dtype = 'fp16'
493
+ FLAGS.enable_xla = False
494
+ FLAGS.distribution_strategy = 'multi_worker_mirrored'
495
+ FLAGS.tf_gpu_thread_mode = 'gpu_private'
496
+ FLAGS.datasets_num_private_threads = 32
497
+ FLAGS.model_dir = self._get_model_dir(
498
+ 'benchmark_8_gpu_{}_worker_fp16_{}_tweaked'.format(
499
+ num_workers, all_reduce_alg))
500
+ FLAGS.train_batch_size = 4 * num_gpus * num_workers
501
+ FLAGS.all_reduce_alg = all_reduce_alg
502
+
503
+ self._run_and_report_benchmark()
504
+
505
+ def benchmark_eager_8_gpu_2_workers_fp16_ring_tweaked(self):
506
+ """8 GPUs per worker, 2 workers, fp16, ring all-reduce."""
507
+ self._benchmark_common(num_workers=2, all_reduce_alg='ring')
508
+
509
+ def benchmark_eager_8_gpu_2_workers_fp16_nccl_tweaked(self):
510
+ """8 GPUs per worker, 2 workers, fp16, nccl all-reduce."""
511
+ self._benchmark_common(num_workers=2, all_reduce_alg='nccl')
512
+
513
+ def benchmark_8_gpu_8_workers_fp16_ring_tweaked(self):
514
+ """8 GPUs per worker, 8 workers, fp16, ring all-reduce."""
515
+ self._benchmark_common(num_workers=8, all_reduce_alg='ring')
516
+
517
+ def benchmark_8_gpu_8_workers_fp16_nccl_tweaked(self):
518
+ """8 GPUs per worker, 8 workers, fp16, nccl all-reduce."""
519
+ self._benchmark_common(num_workers=8, all_reduce_alg='nccl')
520
+
521
+
522
+ class BertSquadMultiWorkerBenchmark(BertSquadBenchmarkBase):
523
+ """BERT SQuAD distributed benchmark tests with multiple workers."""
524
+
525
+ def __init__(self, output_dir=TMP_DIR, tpu=None, **kwargs):
526
+ super(BertSquadMultiWorkerBenchmark, self).__init__(
527
+ output_dir=output_dir, tpu=tpu)
528
+
529
+ def _setup(self):
530
+ """Sets up the benchmark and SQuAD flags."""
531
+ super(BertSquadMultiWorkerBenchmark, self)._setup()
532
+ FLAGS.train_data_path = SQUAD_TRAIN_DATA_PATH
533
+ FLAGS.predict_file = SQUAD_PREDICT_FILE
534
+ FLAGS.vocab_file = SQUAD_VOCAB_FILE
535
+ FLAGS.input_meta_data_path = SQUAD_FULL_INPUT_META_DATA_PATH
536
+ FLAGS.bert_config_file = MODEL_CONFIG_FILE_PATH
537
+ FLAGS.num_train_epochs = 1
538
+ FLAGS.steps_per_loop = 100
539
+
540
+ @benchmark_wrappers.enable_runtime_flags
541
+ def _run_and_report_benchmark(self,
542
+ use_ds=True,
543
+ run_eagerly=False):
544
+ """Runs the benchmark and reports various metrics."""
545
+ if FLAGS.train_batch_size <= 4 * 8:
546
+ FLAGS.input_meta_data_path = SQUAD_LONG_INPUT_META_DATA_PATH
547
+ else:
548
+ FLAGS.input_meta_data_path = SQUAD_FULL_INPUT_META_DATA_PATH
549
+ start_time_sec = time.time()
550
+ self._train_squad(run_eagerly=run_eagerly,
551
+ ds_type='multi_worker_mirrored')
552
+ wall_time_sec = time.time() - start_time_sec
553
+
554
+ summary = self._read_training_summary_from_file()
555
+ summary['start_time_sec'] = start_time_sec
556
+
557
+ super(BertSquadMultiWorkerBenchmark, self)._report_benchmark(
558
+ stats=summary,
559
+ wall_time_sec=wall_time_sec,
560
+ min_accuracy=0,
561
+ max_accuracy=1)
562
+
563
+ def _benchmark_common(self, num_workers, all_reduce_alg):
564
+ """Common to all benchmarks in this class."""
565
+ self._setup()
566
+
567
+ num_gpus = 8
568
+ FLAGS.num_gpus = num_gpus
569
+ FLAGS.dtype = 'fp16'
570
+ FLAGS.enable_xla = False
571
+ FLAGS.distribution_strategy = 'multi_worker_mirrored'
572
+ FLAGS.tf_gpu_thread_mode = 'gpu_private'
573
+ FLAGS.datasets_num_private_threads = 32
574
+ FLAGS.model_dir = self._get_model_dir(
575
+ 'benchmark_8_gpu_{}_worker_fp16_{}_tweaked'.format(
576
+ num_workers, all_reduce_alg))
577
+ FLAGS.train_batch_size = 4 * num_gpus * num_workers
578
+ FLAGS.all_reduce_alg = all_reduce_alg
579
+
580
+ self._run_and_report_benchmark()
581
+
582
+ def benchmark_8_gpu_1_worker_fp16_ring_tweaked(self):
583
+ """8 GPUs per worker, 1 worker, fp16, ring all-reduce."""
584
+ self._benchmark_common(num_workers=1, all_reduce_alg='ring')
585
+
586
+ def benchmark_8_gpu_1_worker_fp16_nccl_tweaked(self):
587
+ """8 GPUs per worker, 1 worker, fp16, nccl all-reduce."""
588
+ self._benchmark_common(num_workers=1, all_reduce_alg='nccl')
589
+
590
+ def benchmark_8_gpu_2_workers_fp16_ring_tweaked(self):
591
+ """8 GPUs per worker, 2 workers, fp16, ring all-reduce."""
592
+ self._benchmark_common(num_workers=2, all_reduce_alg='ring')
593
+
594
+ def benchmark_8_gpu_2_workers_fp16_nccl_tweaked(self):
595
+ """8 GPUs per worker, 2 workers, fp16, nccl all-reduce."""
596
+ self._benchmark_common(num_workers=2, all_reduce_alg='nccl')
597
+
598
+ def benchmark_8_gpu_8_workers_fp16_ring_tweaked(self):
599
+ """8 GPUs per worker, 8 workers, fp16, ring all-reduce."""
600
+ self._benchmark_common(num_workers=8, all_reduce_alg='ring')
601
+
602
+ def benchmark_8_gpu_8_workers_fp16_nccl_tweaked(self):
603
+ """8 GPUs per worker, 8 workers, fp16, nccl all-reduce."""
604
+ self._benchmark_common(num_workers=8, all_reduce_alg='nccl')
605
+
606
+
607
+ if __name__ == '__main__':
608
+ tf.test.main()
models/official/benchmark/datastore/schema/benchmark_metric.json ADDED
@@ -0,0 +1,56 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "description": "The ID of the benchmark run, where this metric should tie to.",
4
+ "mode": "REQUIRED",
5
+ "name": "run_id",
6
+ "type": "STRING"
7
+ },
8
+ {
9
+ "description": "The name of the metric, which should be descriptive. E.g. training_loss, accuracy.",
10
+ "mode": "REQUIRED",
11
+ "name": "name",
12
+ "type": "STRING"
13
+ },
14
+ {
15
+ "description": "The unit of the metric. E.g. MB per sec.",
16
+ "mode": "NULLABLE",
17
+ "name": "unit",
18
+ "type": "STRING"
19
+ },
20
+ {
21
+ "description": "The value of the metric.",
22
+ "mode": "NULLABLE",
23
+ "name": "value",
24
+ "type": "FLOAT"
25
+ },
26
+ {
27
+ "description": "The timestamp when the metric is recorded.",
28
+ "mode": "REQUIRED",
29
+ "name": "timestamp",
30
+ "type": "TIMESTAMP"
31
+ },
32
+ {
33
+ "description": "The global step when this metric is recorded.",
34
+ "mode": "NULLABLE",
35
+ "name": "global_step",
36
+ "type": "INTEGER"
37
+ },
38
+ {
39
+ "description": "Free format metadata for the extra information about the metric.",
40
+ "mode": "REPEATED",
41
+ "name": "extras",
42
+ "type": "RECORD",
43
+ "fields": [
44
+ {
45
+ "mode": "NULLABLE",
46
+ "name": "name",
47
+ "type": "STRING"
48
+ },
49
+ {
50
+ "mode": "NULLABLE",
51
+ "name": "value",
52
+ "type": "STRING"
53
+ }
54
+ ]
55
+ }
56
+ ]
models/official/benchmark/datastore/schema/benchmark_run.json ADDED
@@ -0,0 +1,368 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "description": "The UUID of the run for the benchmark.",
4
+ "mode": "REQUIRED",
5
+ "name": "model_id",
6
+ "type": "STRING"
7
+ },
8
+ {
9
+ "description": "The name of the model, E.g ResNet50, LeNet-5 etc.",
10
+ "mode": "REQUIRED",
11
+ "name": "model_name",
12
+ "type": "STRING"
13
+ },
14
+ {
15
+ "description": "The date when the test of the model is started",
16
+ "mode": "REQUIRED",
17
+ "name": "run_date",
18
+ "type": "TIMESTAMP"
19
+ },
20
+ {
21
+ "description": "The unique name for a test by the combination of key parameters, eg batch size, num of GPU, etc. It is hardware independent.",
22
+ "mode": "NULLABLE",
23
+ "name": "test_id",
24
+ "type": "STRING"
25
+ },
26
+ {
27
+ "description": "The tensorflow version information.",
28
+ "fields": [
29
+ {
30
+ "description": "Version of the tensorflow. E.g. 1.7.0-rc0",
31
+ "mode": "REQUIRED",
32
+ "name": "version",
33
+ "type": "STRING"
34
+ },
35
+ {
36
+ "description": "Git Hash of the tensorflow",
37
+ "mode": "NULLABLE",
38
+ "name": "git_hash",
39
+ "type": "STRING"
40
+ },
41
+ {
42
+ "description": "The channel of the tensorflow binary, eg, nightly, RC, final, custom.",
43
+ "mode": "NULLABLE",
44
+ "name": "channel",
45
+ "type": "STRING"
46
+ },
47
+ {
48
+ "description": "Identify anything special about the build, eg CUDA 10, NCCL, MKL, etc.",
49
+ "mode": "NULLABLE",
50
+ "name": "build_type",
51
+ "type": "STRING"
52
+ }
53
+ ],
54
+ "mode": "REQUIRED",
55
+ "name": "tensorflow_version",
56
+ "type": "RECORD"
57
+ },
58
+ {
59
+ "description": "The arbitrary attribute of the model.",
60
+ "fields": [
61
+ {
62
+ "description": "The name of the attribute.",
63
+ "mode": "REQUIRED",
64
+ "name": "name",
65
+ "type": "STRING"
66
+ },
67
+ {
68
+ "description": "The value of the attribute.",
69
+ "mode": "NULLABLE",
70
+ "name": "value",
71
+ "type": "STRING"
72
+ }
73
+ ],
74
+ "mode": "REPEATED",
75
+ "name": "attribute",
76
+ "type": "RECORD"
77
+ },
78
+ {
79
+ "description": "Environment variables when the benchmark run is executed.",
80
+ "fields": [
81
+ {
82
+ "description": "The name of the variable.",
83
+ "mode": "REQUIRED",
84
+ "name": "name",
85
+ "type": "STRING"
86
+ },
87
+ {
88
+ "description": "The value of the variable.",
89
+ "mode": "NULLABLE",
90
+ "name": "value",
91
+ "type": "STRING"
92
+ }
93
+ ],
94
+ "mode": "REPEATED",
95
+ "name": "environment_variable",
96
+ "type": "RECORD"
97
+ },
98
+ {
99
+ "description": "TF Environment variables when the benchmark run is executed.",
100
+ "fields": [
101
+ {
102
+ "description": "The name of the variable.",
103
+ "mode": "REQUIRED",
104
+ "name": "name",
105
+ "type": "STRING"
106
+ },
107
+ {
108
+ "description": "The value of the variable.",
109
+ "mode": "NULLABLE",
110
+ "name": "value",
111
+ "type": "STRING"
112
+ }
113
+ ],
114
+ "mode": "REPEATED",
115
+ "name": "tensorflow_environment_variables",
116
+ "type": "RECORD"
117
+ },
118
+ {
119
+ "description": "The list of parameters run with the model. It could contain hyperparameters or others.",
120
+ "fields": [
121
+ {
122
+ "description": "The name of the parameter.",
123
+ "mode": "REQUIRED",
124
+ "name": "name",
125
+ "type": "STRING"
126
+ },
127
+ {
128
+ "description": "The string value of the parameter.",
129
+ "mode": "NULLABLE",
130
+ "name": "string_value",
131
+ "type": "STRING"
132
+ },
133
+ {
134
+ "description": "The bool value of the parameter.",
135
+ "mode": "NULLABLE",
136
+ "name": "bool_value",
137
+ "type": "STRING"
138
+ },
139
+ {
140
+ "description": "The int/long value of the parameter.",
141
+ "mode": "NULLABLE",
142
+ "name": "long_value",
143
+ "type": "INTEGER"
144
+ },
145
+ {
146
+ "description": "The double/float value of parameter.",
147
+ "mode": "NULLABLE",
148
+ "name": "float_value",
149
+ "type": "FLOAT"
150
+ }
151
+ ],
152
+ "mode": "REPEATED",
153
+ "name": "run_parameters",
154
+ "type": "RECORD"
155
+ },
156
+ {
157
+ "description": "The dataset that run with the benchmark.",
158
+ "mode": "NULLABLE",
159
+ "name": "dataset",
160
+ "type": "RECORD",
161
+ "fields": [
162
+ {
163
+ "description": "The name of the dataset that the model is trained/validated with. E.g ImageNet, mnist.",
164
+ "mode": "REQUIRED",
165
+ "name": "name",
166
+ "type": "STRING"
167
+ },
168
+ {
169
+ "description": "The arbitrary attribute of the dataset.",
170
+ "fields": [
171
+ {
172
+ "description": "The name of the attribute.",
173
+ "mode": "REQUIRED",
174
+ "name": "name",
175
+ "type": "STRING"
176
+ },
177
+ {
178
+ "description": "The value of the attribute.",
179
+ "mode": "NULLABLE",
180
+ "name": "value",
181
+ "type": "STRING"
182
+ }
183
+ ],
184
+ "mode": "REPEATED",
185
+ "name": "attribute",
186
+ "type": "RECORD"
187
+ }
188
+ ]
189
+ },
190
+ {
191
+ "description": "Used to differentiate from AWS, GCE or DGX-1 at a high level",
192
+ "mode": "NULLABLE",
193
+ "name": "test_environment",
194
+ "type": "STRING"
195
+ },
196
+ {
197
+ "description": "The machine configuration of the benchmark run.",
198
+ "mode": "NULLABLE",
199
+ "name": "machine_config",
200
+ "type": "RECORD",
201
+ "fields": [
202
+ {
203
+ "description": "The platform information of the benchmark run.",
204
+ "mode": "NULLABLE",
205
+ "name": "platform_info",
206
+ "type": "RECORD",
207
+ "fields": [
208
+ {
209
+ "description": "Eg: 64bit.",
210
+ "mode": "NULLABLE",
211
+ "name": "bits",
212
+ "type": "STRING"
213
+ },
214
+ {
215
+ "description": "Eg: ELF.",
216
+ "mode": "NULLABLE",
217
+ "name": "linkage",
218
+ "type": "STRING"
219
+ },
220
+ {
221
+ "description": "Eg: i386.",
222
+ "mode": "NULLABLE",
223
+ "name": "machine",
224
+ "type": "STRING"
225
+ },
226
+ {
227
+ "description": "Eg: 3.13.0-76-generic.",
228
+ "mode": "NULLABLE",
229
+ "name": "release",
230
+ "type": "STRING"
231
+ },
232
+ {
233
+ "description": "Eg: Linux.",
234
+ "mode": "NULLABLE",
235
+ "name": "system",
236
+ "type": "STRING"
237
+ },
238
+ {
239
+ "description": "Eg: #120-Ubuntu SMP Mon Jan 18 15:59:10 UTC 2016.",
240
+ "mode": "NULLABLE",
241
+ "name": "version",
242
+ "type": "STRING"
243
+ }
244
+ ]
245
+ },
246
+ {
247
+ "description": "The CPU information of the benchmark run.",
248
+ "mode": "NULLABLE",
249
+ "name": "cpu_info",
250
+ "type": "RECORD",
251
+ "fields": [
252
+ {
253
+ "mode": "NULLABLE",
254
+ "name": "num_cores",
255
+ "type": "INTEGER"
256
+ },
257
+ {
258
+ "mode": "NULLABLE",
259
+ "name": "num_cores_allowed",
260
+ "type": "INTEGER"
261
+ },
262
+ {
263
+ "description" : "How fast are those CPUs.",
264
+ "mode": "NULLABLE",
265
+ "name": "mhz_per_cpu",
266
+ "type": "FLOAT"
267
+ },
268
+ {
269
+ "description" : "Additional CPU info, Eg: Intel Ivybridge with HyperThreading (24 cores).",
270
+ "mode": "NULLABLE",
271
+ "name": "cpu_info",
272
+ "type": "STRING"
273
+ },
274
+ {
275
+ "description" : "What kind of cpu scaling is enabled on the host. Eg performance, ondemand, conservative, mixed.",
276
+ "mode": "NULLABLE",
277
+ "name": "cpu_governor",
278
+ "type": "STRING"
279
+ },
280
+ {
281
+ "description": "Cache size of the CPUs.",
282
+ "mode": "NULLABLE",
283
+ "name": "cache_size",
284
+ "type": "RECORD",
285
+ "fields": [
286
+ {
287
+ "mode": "NULLABLE",
288
+ "name": "level",
289
+ "type": "STRING"
290
+ },
291
+ {
292
+ "mode": "NULLABLE",
293
+ "name": "size",
294
+ "type": "INTEGER"
295
+ }
296
+ ]
297
+ }
298
+ ]
299
+ },
300
+ {
301
+ "mode": "NULLABLE",
302
+ "name": "gpu_info",
303
+ "type": "RECORD",
304
+ "fields": [
305
+ {
306
+ "mode": "NULLABLE",
307
+ "name": "count",
308
+ "type": "INTEGER"
309
+ },
310
+ {
311
+ "mode": "NULLABLE",
312
+ "name": "model",
313
+ "type": "STRING"
314
+ },
315
+ {
316
+ "mode": "NULLABLE",
317
+ "name": "cuda_version",
318
+ "type": "STRING"
319
+ }
320
+ ]
321
+ },
322
+ {
323
+ "description": "The cloud instance inforation if the benchmark run is executed on cloud",
324
+ "mode": "NULLABLE",
325
+ "name": "cloud_info",
326
+ "type": "RECORD",
327
+ "fields": [
328
+ {
329
+ "description": "The instance type, E.g. n1-standard-4.",
330
+ "mode": "NULLABLE",
331
+ "name": "instance_type",
332
+ "type": "STRING"
333
+ },
334
+ {
335
+ "description": "The arbitrary attribute of the cloud info.",
336
+ "fields": [
337
+ {
338
+ "description": "The name of the attribute.",
339
+ "mode": "REQUIRED",
340
+ "name": "name",
341
+ "type": "STRING"
342
+ },
343
+ {
344
+ "description": "The value of the attribute.",
345
+ "mode": "NULLABLE",
346
+ "name": "value",
347
+ "type": "STRING"
348
+ }
349
+ ],
350
+ "mode": "REPEATED",
351
+ "name": "attribute",
352
+ "type": "RECORD"
353
+ }
354
+ ]
355
+ },
356
+ {
357
+ "mode": "NULLABLE",
358
+ "name": "memory_total",
359
+ "type": "INTEGER"
360
+ },
361
+ {
362
+ "mode": "NULLABLE",
363
+ "name": "memory_available",
364
+ "type": "STRING"
365
+ }
366
+ ]
367
+ }
368
+ ]
models/official/benchmark/datastore/schema/benchmark_run_status.json ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "description": "The UUID of the run for the benchmark.",
4
+ "mode": "REQUIRED",
5
+ "name": "run_id",
6
+ "type": "STRING"
7
+ },
8
+ {
9
+ "description": "The status of the run for the benchmark. Eg, running, failed, success",
10
+ "mode": "REQUIRED",
11
+ "name": "status",
12
+ "type": "STRING"
13
+ }
14
+ ]
models/official/benchmark/keras_benchmark.py ADDED
@@ -0,0 +1,98 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2018 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
+ """Executes Keras benchmarks and accuracy tests."""
16
+
17
+ from __future__ import absolute_import
18
+ from __future__ import division
19
+ from __future__ import print_function
20
+
21
+ import tensorflow as tf
22
+ from official.benchmark.perfzero_benchmark import PerfZeroBenchmark
23
+ from official.utils.flags import core as flags_core
24
+
25
+
26
+ class KerasBenchmark(PerfZeroBenchmark):
27
+ """Base benchmark class with methods to simplify testing."""
28
+
29
+ def __init__(self,
30
+ output_dir=None,
31
+ default_flags=None,
32
+ flag_methods=None,
33
+ tpu=None):
34
+ super(KerasBenchmark, self).__init__(
35
+ output_dir=output_dir,
36
+ default_flags=default_flags,
37
+ flag_methods=flag_methods,
38
+ tpu=tpu)
39
+
40
+ def _report_benchmark(self,
41
+ stats,
42
+ wall_time_sec,
43
+ top_1_max=None,
44
+ top_1_min=None,
45
+ log_steps=None,
46
+ total_batch_size=None,
47
+ warmup=1,
48
+ start_time_sec=None):
49
+ """Report benchmark results by writing to local protobuf file.
50
+
51
+ Args:
52
+ stats: dict returned from keras models with known entries.
53
+ wall_time_sec: the during of the benchmark execution in seconds
54
+ top_1_max: highest passing level for top_1 accuracy.
55
+ top_1_min: lowest passing level for top_1 accuracy.
56
+ log_steps: How often the log was created for stats['step_timestamp_log'].
57
+ total_batch_size: Global batch-size.
58
+ warmup: number of entries in stats['step_timestamp_log'] to ignore.
59
+ start_time_sec: the start time of the program in seconds since epoch
60
+ """
61
+
62
+ metrics = []
63
+ if 'accuracy_top_1' in stats:
64
+ metrics.append({'name': 'accuracy_top_1',
65
+ 'value': stats['accuracy_top_1'],
66
+ 'min_value': top_1_min,
67
+ 'max_value': top_1_max})
68
+ metrics.append({'name': 'top_1_train_accuracy',
69
+ 'value': stats['training_accuracy_top_1']})
70
+
71
+ if (warmup and 'step_timestamp_log' in stats and
72
+ len(stats['step_timestamp_log']) > warmup):
73
+ # first entry in the time_log is start of step 1. The rest of the
74
+ # entries are the end of each step recorded
75
+ time_log = stats['step_timestamp_log']
76
+ elapsed = time_log[-1].timestamp - time_log[warmup].timestamp
77
+ num_examples = (
78
+ total_batch_size * log_steps * (len(time_log) - warmup - 1))
79
+ examples_per_sec = num_examples / elapsed
80
+ metrics.append({'name': 'exp_per_second',
81
+ 'value': examples_per_sec})
82
+
83
+ if 'avg_exp_per_second' in stats:
84
+ metrics.append({'name': 'avg_exp_per_second',
85
+ 'value': stats['avg_exp_per_second']})
86
+
87
+ if start_time_sec and 'step_timestamp_log' in stats:
88
+ time_log = stats['step_timestamp_log']
89
+ # time_log[0] is recorded at the beginning of the first step.
90
+ startup_time = time_log[0].timestamp - start_time_sec
91
+ metrics.append({'name': 'startup_time', 'value': startup_time})
92
+
93
+ flags_str = flags_core.get_nondefault_flags_as_str()
94
+ self.report_benchmark(
95
+ iters=-1,
96
+ wall_time=wall_time_sec,
97
+ metrics=metrics,
98
+ extras={'flags': flags_str})
models/official/benchmark/keras_cifar_benchmark.py ADDED
@@ -0,0 +1,402 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2018 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
+ """Executes Keras benchmarks and accuracy tests."""
16
+ from __future__ import absolute_import
17
+ from __future__ import division
18
+ from __future__ import print_function
19
+
20
+ import os
21
+ import time
22
+ from absl import flags
23
+ import tensorflow as tf # pylint: disable=g-bad-import-order
24
+
25
+ from official.benchmark import keras_benchmark
26
+ from official.benchmark import benchmark_wrappers
27
+ from official.benchmark.models import resnet_cifar_main
28
+
29
+ MIN_TOP_1_ACCURACY = 0.929
30
+ MAX_TOP_1_ACCURACY = 0.938
31
+
32
+ FLAGS = flags.FLAGS
33
+ CIFAR_DATA_DIR_NAME = 'cifar-10-batches-bin'
34
+
35
+
36
+ class Resnet56KerasAccuracy(keras_benchmark.KerasBenchmark):
37
+ """Accuracy tests for ResNet56 Keras CIFAR-10."""
38
+
39
+ def __init__(self, output_dir=None, root_data_dir=None, **kwargs):
40
+ """A benchmark class.
41
+
42
+ Args:
43
+ output_dir: directory where to output e.g. log files
44
+ root_data_dir: directory under which to look for dataset
45
+ **kwargs: arbitrary named arguments. This is needed to make the
46
+ constructor forward compatible in case PerfZero provides more
47
+ named arguments before updating the constructor.
48
+ """
49
+
50
+ self.data_dir = os.path.join(root_data_dir, CIFAR_DATA_DIR_NAME)
51
+ flag_methods = [resnet_cifar_main.define_cifar_flags]
52
+
53
+ super(Resnet56KerasAccuracy, self).__init__(
54
+ output_dir=output_dir, flag_methods=flag_methods)
55
+
56
+ def _setup(self):
57
+ super(Resnet56KerasAccuracy, self)._setup()
58
+ FLAGS.use_tensor_lr = False
59
+
60
+ def benchmark_graph_1_gpu(self):
61
+ """Test keras based model with Keras fit and distribution strategies."""
62
+ self._setup()
63
+ FLAGS.num_gpus = 1
64
+ FLAGS.data_dir = self.data_dir
65
+ FLAGS.batch_size = 128
66
+ FLAGS.train_epochs = 182
67
+ FLAGS.model_dir = self._get_model_dir('benchmark_graph_1_gpu')
68
+ FLAGS.dtype = 'fp32'
69
+ self._run_and_report_benchmark()
70
+
71
+ def benchmark_1_gpu(self):
72
+ """Test keras based model with eager and distribution strategies."""
73
+ self._setup()
74
+ FLAGS.num_gpus = 1
75
+ FLAGS.data_dir = self.data_dir
76
+ FLAGS.batch_size = 128
77
+ FLAGS.train_epochs = 182
78
+ FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu')
79
+ FLAGS.dtype = 'fp32'
80
+ FLAGS.enable_eager = True
81
+ self._run_and_report_benchmark()
82
+
83
+ def benchmark_cpu(self):
84
+ """Test keras based model on CPU."""
85
+ self._setup()
86
+ FLAGS.num_gpus = 0
87
+ FLAGS.data_dir = self.data_dir
88
+ FLAGS.batch_size = 128
89
+ FLAGS.train_epochs = 182
90
+ FLAGS.model_dir = self._get_model_dir('benchmark_cpu')
91
+ FLAGS.dtype = 'fp32'
92
+ FLAGS.enable_eager = True
93
+ FLAGS.data_format = 'channels_last'
94
+ self._run_and_report_benchmark()
95
+
96
+ def benchmark_cpu_no_dist_strat(self):
97
+ """Test keras based model on CPU without distribution strategies."""
98
+ self._setup()
99
+ FLAGS.num_gpus = 0
100
+ FLAGS.data_dir = self.data_dir
101
+ FLAGS.batch_size = 128
102
+ FLAGS.train_epochs = 182
103
+ FLAGS.model_dir = self._get_model_dir('benchmark_cpu_no_dist_strat')
104
+ FLAGS.dtype = 'fp32'
105
+ FLAGS.enable_eager = True
106
+ FLAGS.distribution_strategy = 'off'
107
+ FLAGS.data_format = 'channels_last'
108
+ self._run_and_report_benchmark()
109
+
110
+ def benchmark_cpu_no_dist_strat_run_eagerly(self):
111
+ """Test keras based model on CPU w/forced eager and no dist_strat."""
112
+ self._setup()
113
+ FLAGS.num_gpus = 0
114
+ FLAGS.data_dir = self.data_dir
115
+ FLAGS.batch_size = 128
116
+ FLAGS.train_epochs = 182
117
+ FLAGS.model_dir = self._get_model_dir(
118
+ 'benchmark_cpu_no_dist_strat_run_eagerly')
119
+ FLAGS.dtype = 'fp32'
120
+ FLAGS.enable_eager = True
121
+ FLAGS.run_eagerly = True
122
+ FLAGS.distribution_strategy = 'off'
123
+ FLAGS.data_format = 'channels_last'
124
+ self._run_and_report_benchmark()
125
+
126
+ def benchmark_1_gpu_no_dist_strat(self):
127
+ """Test keras based model with eager and no dist strat."""
128
+ self._setup()
129
+ FLAGS.num_gpus = 1
130
+ FLAGS.data_dir = self.data_dir
131
+ FLAGS.batch_size = 128
132
+ FLAGS.train_epochs = 182
133
+ FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu_no_dist_strat')
134
+ FLAGS.dtype = 'fp32'
135
+ FLAGS.enable_eager = True
136
+ FLAGS.distribution_strategy = 'off'
137
+ self._run_and_report_benchmark()
138
+
139
+ def benchmark_1_gpu_no_dist_strat_run_eagerly(self):
140
+ """Test keras based model w/forced eager and no dist_strat."""
141
+ self._setup()
142
+ FLAGS.num_gpus = 1
143
+ FLAGS.data_dir = self.data_dir
144
+ FLAGS.batch_size = 128
145
+ FLAGS.train_epochs = 182
146
+ FLAGS.model_dir = self._get_model_dir(
147
+ 'benchmark_1_gpu_no_dist_strat_run_eagerly')
148
+ FLAGS.dtype = 'fp32'
149
+ FLAGS.enable_eager = True
150
+ FLAGS.run_eagerly = True
151
+ FLAGS.distribution_strategy = 'off'
152
+ self._run_and_report_benchmark()
153
+
154
+ def benchmark_graph_1_gpu_no_dist_strat(self):
155
+ """Test keras based model with Keras fit but not distribution strategies."""
156
+ self._setup()
157
+ FLAGS.distribution_strategy = 'off'
158
+ FLAGS.num_gpus = 1
159
+ FLAGS.data_dir = self.data_dir
160
+ FLAGS.batch_size = 128
161
+ FLAGS.train_epochs = 182
162
+ FLAGS.model_dir = self._get_model_dir('benchmark_graph_1_gpu_no_dist_strat')
163
+ FLAGS.dtype = 'fp32'
164
+ self._run_and_report_benchmark()
165
+
166
+ def benchmark_2_gpu(self):
167
+ """Test keras based model with eager and distribution strategies."""
168
+ self._setup()
169
+ FLAGS.num_gpus = 2
170
+ FLAGS.data_dir = self.data_dir
171
+ FLAGS.batch_size = 128
172
+ FLAGS.train_epochs = 182
173
+ FLAGS.model_dir = self._get_model_dir('benchmark_2_gpu')
174
+ FLAGS.dtype = 'fp32'
175
+ FLAGS.enable_eager = True
176
+ self._run_and_report_benchmark()
177
+
178
+ def benchmark_graph_2_gpu(self):
179
+ """Test keras based model with Keras fit and distribution strategies."""
180
+ self._setup()
181
+ FLAGS.num_gpus = 2
182
+ FLAGS.data_dir = self.data_dir
183
+ FLAGS.batch_size = 128
184
+ FLAGS.train_epochs = 182
185
+ FLAGS.model_dir = self._get_model_dir('benchmark_graph_2_gpu')
186
+ FLAGS.dtype = 'fp32'
187
+ self._run_and_report_benchmark()
188
+
189
+ @benchmark_wrappers.enable_runtime_flags
190
+ def _run_and_report_benchmark(self):
191
+ start_time_sec = time.time()
192
+ stats = resnet_cifar_main.run(FLAGS)
193
+ wall_time_sec = time.time() - start_time_sec
194
+
195
+ super(Resnet56KerasAccuracy, self)._report_benchmark(
196
+ stats,
197
+ wall_time_sec,
198
+ top_1_min=MIN_TOP_1_ACCURACY,
199
+ top_1_max=MAX_TOP_1_ACCURACY,
200
+ total_batch_size=FLAGS.batch_size,
201
+ log_steps=100)
202
+
203
+
204
+ class Resnet56KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
205
+ """Short performance tests for ResNet56 via Keras and CIFAR-10."""
206
+
207
+ def __init__(self, output_dir=None, default_flags=None):
208
+ flag_methods = [resnet_cifar_main.define_cifar_flags]
209
+
210
+ super(Resnet56KerasBenchmarkBase, self).__init__(
211
+ output_dir=output_dir,
212
+ flag_methods=flag_methods,
213
+ default_flags=default_flags)
214
+
215
+ @benchmark_wrappers.enable_runtime_flags
216
+ def _run_and_report_benchmark(self):
217
+ start_time_sec = time.time()
218
+ stats = resnet_cifar_main.run(FLAGS)
219
+ wall_time_sec = time.time() - start_time_sec
220
+
221
+ super(Resnet56KerasBenchmarkBase, self)._report_benchmark(
222
+ stats,
223
+ wall_time_sec,
224
+ total_batch_size=FLAGS.batch_size,
225
+ log_steps=FLAGS.log_steps)
226
+
227
+ def benchmark_1_gpu(self):
228
+ """Test 1 gpu."""
229
+ self._setup()
230
+ FLAGS.num_gpus = 1
231
+ FLAGS.enable_eager = True
232
+ FLAGS.distribution_strategy = 'one_device'
233
+ FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu')
234
+ FLAGS.batch_size = 128
235
+ self._run_and_report_benchmark()
236
+
237
+ def benchmark_1_gpu_xla(self):
238
+ """Test 1 gpu with xla enabled."""
239
+ self._setup()
240
+ FLAGS.num_gpus = 1
241
+ FLAGS.enable_eager = True
242
+ FLAGS.run_eagerly = False
243
+ FLAGS.enable_xla = True
244
+ FLAGS.distribution_strategy = 'one_device'
245
+ FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu_xla')
246
+ FLAGS.batch_size = 128
247
+ self._run_and_report_benchmark()
248
+
249
+ def benchmark_graph_1_gpu(self):
250
+ """Test 1 gpu graph."""
251
+ self._setup()
252
+ FLAGS.num_gpus = 1
253
+ FLAGS.enable_eager = False
254
+ FLAGS.run_eagerly = False
255
+ FLAGS.distribution_strategy = 'one_device'
256
+ FLAGS.model_dir = self._get_model_dir('benchmark_graph_1_gpu')
257
+ FLAGS.batch_size = 128
258
+ self._run_and_report_benchmark()
259
+
260
+ def benchmark_1_gpu_no_dist_strat(self):
261
+ """Test 1 gpu without distribution strategies."""
262
+ self._setup()
263
+ FLAGS.num_gpus = 1
264
+ FLAGS.enable_eager = True
265
+ FLAGS.distribution_strategy = 'off'
266
+ FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu_no_dist_strat')
267
+ FLAGS.batch_size = 128
268
+ self._run_and_report_benchmark()
269
+
270
+ def benchmark_graph_1_gpu_no_dist_strat(self):
271
+ """Test 1 gpu graph mode without distribution strategies."""
272
+ self._setup()
273
+ FLAGS.num_gpus = 1
274
+ FLAGS.enable_eager = False
275
+ FLAGS.distribution_strategy = 'off'
276
+ FLAGS.model_dir = self._get_model_dir('benchmark_graph_1_gpu_no_dist_strat')
277
+ FLAGS.batch_size = 128
278
+ self._run_and_report_benchmark()
279
+
280
+ def benchmark_1_gpu_no_dist_strat_run_eagerly(self):
281
+ """Test 1 gpu without distribution strategy and forced eager."""
282
+ self._setup()
283
+ FLAGS.num_gpus = 1
284
+ FLAGS.batch_size = 128
285
+ FLAGS.model_dir = self._get_model_dir(
286
+ 'benchmark_1_gpu_no_dist_strat_run_eagerly')
287
+ FLAGS.dtype = 'fp32'
288
+ FLAGS.enable_eager = True
289
+ FLAGS.run_eagerly = True
290
+ FLAGS.distribution_strategy = 'off'
291
+ self._run_and_report_benchmark()
292
+
293
+ def benchmark_2_gpu(self):
294
+ """Test 2 gpu."""
295
+ self._setup()
296
+ FLAGS.num_gpus = 2
297
+ FLAGS.enable_eager = True
298
+ FLAGS.run_eagerly = False
299
+ FLAGS.distribution_strategy = 'mirrored'
300
+ FLAGS.model_dir = self._get_model_dir('benchmark_2_gpu')
301
+ FLAGS.batch_size = 128 * 2 # 2 GPUs
302
+ self._run_and_report_benchmark()
303
+
304
+ def benchmark_graph_2_gpu(self):
305
+ """Test 2 gpu graph mode."""
306
+ self._setup()
307
+ FLAGS.num_gpus = 2
308
+ FLAGS.enable_eager = False
309
+ FLAGS.run_eagerly = False
310
+ FLAGS.distribution_strategy = 'mirrored'
311
+ FLAGS.model_dir = self._get_model_dir('benchmark_graph_2_gpu')
312
+ FLAGS.batch_size = 128 * 2 # 2 GPUs
313
+ self._run_and_report_benchmark()
314
+
315
+ def benchmark_cpu(self):
316
+ """Test cpu."""
317
+ self._setup()
318
+ FLAGS.num_gpus = 0
319
+ FLAGS.enable_eager = True
320
+ FLAGS.model_dir = self._get_model_dir('benchmark_cpu')
321
+ FLAGS.batch_size = 128
322
+ FLAGS.data_format = 'channels_last'
323
+ self._run_and_report_benchmark()
324
+
325
+ def benchmark_graph_cpu(self):
326
+ """Test cpu graph mode."""
327
+ self._setup()
328
+ FLAGS.num_gpus = 0
329
+ FLAGS.enable_eager = False
330
+ FLAGS.model_dir = self._get_model_dir('benchmark_graph_cpu')
331
+ FLAGS.batch_size = 128
332
+ FLAGS.data_format = 'channels_last'
333
+ self._run_and_report_benchmark()
334
+
335
+ def benchmark_cpu_no_dist_strat_run_eagerly(self):
336
+ """Test cpu without distribution strategy and forced eager."""
337
+ self._setup()
338
+ FLAGS.num_gpus = 0
339
+ FLAGS.distribution_strategy = 'off'
340
+ FLAGS.enable_eager = True
341
+ FLAGS.run_eagerly = True
342
+ FLAGS.model_dir = self._get_model_dir(
343
+ 'benchmark_cpu_no_dist_strat_run_eagerly')
344
+ FLAGS.batch_size = 128
345
+ FLAGS.data_format = 'channels_last'
346
+ self._run_and_report_benchmark()
347
+
348
+ def benchmark_cpu_no_dist_strat(self):
349
+ """Test cpu without distribution strategies."""
350
+ self._setup()
351
+ FLAGS.num_gpus = 0
352
+ FLAGS.enable_eager = True
353
+ FLAGS.distribution_strategy = 'off'
354
+ FLAGS.model_dir = self._get_model_dir('benchmark_cpu_no_dist_strat')
355
+ FLAGS.batch_size = 128
356
+ FLAGS.data_format = 'channels_last'
357
+ self._run_and_report_benchmark()
358
+
359
+ def benchmark_graph_cpu_no_dist_strat(self):
360
+ """Test cpu graph mode without distribution strategies."""
361
+ self._setup()
362
+ FLAGS.num_gpus = 0
363
+ FLAGS.enable_eager = False
364
+ FLAGS.distribution_strategy = 'off'
365
+ FLAGS.model_dir = self._get_model_dir('benchmark_graph_cpu_no_dist_strat')
366
+ FLAGS.batch_size = 128
367
+ FLAGS.data_format = 'channels_last'
368
+ self._run_and_report_benchmark()
369
+
370
+
371
+ class Resnet56KerasBenchmarkSynth(Resnet56KerasBenchmarkBase):
372
+ """Synthetic benchmarks for ResNet56 and Keras."""
373
+
374
+ def __init__(self, output_dir=None, root_data_dir=None, **kwargs):
375
+ default_flags = {}
376
+ default_flags['skip_eval'] = True
377
+ default_flags['use_synthetic_data'] = True
378
+ default_flags['train_steps'] = 110
379
+ default_flags['log_steps'] = 10
380
+ default_flags['use_tensor_lr'] = False
381
+
382
+ super(Resnet56KerasBenchmarkSynth, self).__init__(
383
+ output_dir=output_dir, default_flags=default_flags)
384
+
385
+
386
+ class Resnet56KerasBenchmarkReal(Resnet56KerasBenchmarkBase):
387
+ """Real data benchmarks for ResNet56 and Keras."""
388
+
389
+ def __init__(self, output_dir=None, root_data_dir=None, **kwargs):
390
+ default_flags = {}
391
+ default_flags['skip_eval'] = True
392
+ default_flags['data_dir'] = os.path.join(root_data_dir, CIFAR_DATA_DIR_NAME)
393
+ default_flags['train_steps'] = 110
394
+ default_flags['log_steps'] = 10
395
+ default_flags['use_tensor_lr'] = False
396
+
397
+ super(Resnet56KerasBenchmarkReal, self).__init__(
398
+ output_dir=output_dir, default_flags=default_flags)
399
+
400
+
401
+ if __name__ == '__main__':
402
+ tf.test.main()
models/official/benchmark/keras_imagenet_benchmark.py ADDED
@@ -0,0 +1,1724 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Lint as: python3
2
+ # Copyright 2018 The TensorFlow Authors. All Rights Reserved.
3
+ #
4
+ # Licensed under the Apache License, Version 2.0 (the "License");
5
+ # you may not use this file except in compliance with the License.
6
+ # You may obtain a copy of the License at
7
+ #
8
+ # http://www.apache.org/licenses/LICENSE-2.0
9
+ #
10
+ # Unless required by applicable law or agreed to in writing, software
11
+ # distributed under the License is distributed on an "AS IS" BASIS,
12
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
+ # See the License for the specific language governing permissions and
14
+ # limitations under the License.
15
+ # ==============================================================================
16
+ """Executes Keras benchmarks and accuracy tests."""
17
+ # pylint: disable=line-too-long
18
+ from __future__ import print_function
19
+
20
+ import json
21
+ import os
22
+ import time
23
+
24
+ from typing import Any, MutableMapping, Optional
25
+
26
+ from absl import flags
27
+ import tensorflow as tf # pylint: disable=g-bad-import-order
28
+
29
+ from official.benchmark import benchmark_wrappers
30
+ from official.benchmark import keras_benchmark
31
+ from official.benchmark.models import resnet_imagenet_main
32
+ from official.vision.image_classification import classifier_trainer
33
+
34
+ MIN_TOP_1_ACCURACY = 0.76
35
+ MAX_TOP_1_ACCURACY = 0.77
36
+
37
+ MOBILENET_V1_MIN_TOP_1_ACCURACY = 0.65
38
+ MOBILENET_V1_MAX_TOP_1_ACCURACY = 0.68
39
+
40
+ # Range of top-1 accracies for model optimization techniques.
41
+ # Each item indicates (MIN_TOP_1_ACCURACY, MAX_TOP_1_ACCURACY).
42
+ MODEL_OPTIMIZATION_TOP_1_ACCURACY = {
43
+ 'RESNET50_FINETUNE_PRUNING': (0.76, 0.77),
44
+ 'MOBILENET_V1_FINETUNE_PRUNING': (0.67, 0.68),
45
+ }
46
+
47
+ FLAGS = flags.FLAGS
48
+
49
+
50
+ def _get_classifier_parameters(
51
+ num_gpus: int = 0,
52
+ builder: str = 'records',
53
+ skip_eval: bool = False,
54
+ distribution_strategy: str = 'mirrored',
55
+ per_replica_batch_size: int = 128,
56
+ epochs: int = 90,
57
+ steps: int = 0,
58
+ epochs_between_evals: int = 1,
59
+ dtype: str = 'float32',
60
+ enable_xla: bool = False,
61
+ run_eagerly: bool = False,
62
+ gpu_thread_mode: Optional[str] = None,
63
+ dataset_num_private_threads: Optional[int] = None,
64
+ loss_scale: Optional[str] = None,
65
+ report_metrics: bool = True,
66
+ batchnorm_spatial_persistent: bool = False) -> MutableMapping[str, Any]:
67
+ """Gets classifier trainer's ResNet parameters."""
68
+ return {
69
+ 'runtime': {
70
+ 'num_gpus': num_gpus,
71
+ 'distribution_strategy': distribution_strategy,
72
+ 'run_eagerly': run_eagerly,
73
+ 'enable_xla': enable_xla,
74
+ 'dataset_num_private_threads': dataset_num_private_threads,
75
+ 'gpu_thread_mode': gpu_thread_mode,
76
+ 'loss_scale': loss_scale,
77
+ 'batchnorm_spatial_persistent': batchnorm_spatial_persistent,
78
+ },
79
+ 'train_dataset': {
80
+ 'builder': builder,
81
+ 'use_per_replica_batch_size': True,
82
+ 'batch_size': per_replica_batch_size,
83
+ 'image_size': 224,
84
+ 'dtype': dtype,
85
+ },
86
+ 'validation_dataset': {
87
+ 'builder': builder,
88
+ 'batch_size': per_replica_batch_size,
89
+ 'use_per_replica_batch_size': True,
90
+ 'image_size': 224,
91
+ 'dtype': dtype,
92
+ },
93
+ 'train': {
94
+ 'epochs': epochs,
95
+ 'steps': steps,
96
+ 'callbacks': {
97
+ 'enable_tensorboard': False,
98
+ 'enable_checkpoint_and_export': False,
99
+ 'enable_time_history': True,
100
+ },
101
+ 'metrics': ['accuracy'] if report_metrics else [],
102
+ },
103
+ 'model': {
104
+ 'loss': {
105
+ 'label_smoothing': 0.1,
106
+ },
107
+ },
108
+ 'evaluation': {
109
+ 'epochs_between_evals': epochs_between_evals,
110
+ 'skip_eval': skip_eval,
111
+ },
112
+ }
113
+
114
+
115
+ class Resnet50KerasAccuracy(keras_benchmark.KerasBenchmark):
116
+ """Benchmark accuracy tests for ResNet50 in Keras."""
117
+
118
+ def __init__(self,
119
+ output_dir: Optional[str] = None,
120
+ root_data_dir: Optional[str] = None,
121
+ **kwargs):
122
+ """A benchmark class.
123
+
124
+ Args:
125
+ output_dir: directory where to output e.g. log files
126
+ root_data_dir: directory under which to look for dataset
127
+ **kwargs: arbitrary named arguments. This is needed to make the
128
+ constructor forward compatible in case PerfZero provides more
129
+ named arguments before updating the constructor.
130
+ """
131
+
132
+ flag_methods = [classifier_trainer.define_classifier_flags]
133
+
134
+ self.data_dir = os.path.join(root_data_dir, 'imagenet')
135
+ super(Resnet50KerasAccuracy, self).__init__(
136
+ output_dir=output_dir, flag_methods=flag_methods)
137
+
138
+ @benchmark_wrappers.enable_runtime_flags
139
+ def _run_and_report_benchmark(
140
+ self,
141
+ experiment_name: str,
142
+ top_1_min: float = MIN_TOP_1_ACCURACY,
143
+ top_1_max: float = MAX_TOP_1_ACCURACY,
144
+ num_gpus: int = 0,
145
+ distribution_strategy: str = 'mirrored',
146
+ per_replica_batch_size: int = 128,
147
+ epochs: int = 90,
148
+ steps: int = 0,
149
+ epochs_between_evals: int = 1,
150
+ dtype: str = 'float32',
151
+ enable_xla: bool = False,
152
+ run_eagerly: bool = False,
153
+ gpu_thread_mode: Optional[str] = None,
154
+ dataset_num_private_threads: Optional[int] = None,
155
+ loss_scale: Optional[str] = None):
156
+ """Runs and reports the benchmark given the provided configuration."""
157
+ FLAGS.model_type = 'resnet'
158
+ FLAGS.dataset = 'imagenet'
159
+ FLAGS.mode = 'train_and_eval'
160
+ FLAGS.data_dir = self.data_dir
161
+ FLAGS.model_dir = self._get_model_dir(experiment_name)
162
+ parameters = _get_classifier_parameters(
163
+ num_gpus=num_gpus,
164
+ distribution_strategy=distribution_strategy,
165
+ per_replica_batch_size=per_replica_batch_size,
166
+ epochs=epochs,
167
+ steps=steps,
168
+ epochs_between_evals=epochs_between_evals,
169
+ dtype=dtype,
170
+ enable_xla=enable_xla,
171
+ run_eagerly=run_eagerly,
172
+ gpu_thread_mode=gpu_thread_mode,
173
+ dataset_num_private_threads=dataset_num_private_threads,
174
+ report_metrics=True,
175
+ loss_scale=loss_scale,
176
+ batchnorm_spatial_persistent=True)
177
+ FLAGS.params_override = json.dumps(parameters)
178
+ total_batch_size = num_gpus * per_replica_batch_size
179
+
180
+ start_time_sec = time.time()
181
+ stats = classifier_trainer.run(flags.FLAGS)
182
+ wall_time_sec = time.time() - start_time_sec
183
+
184
+ super(Resnet50KerasAccuracy, self)._report_benchmark(
185
+ stats,
186
+ wall_time_sec,
187
+ top_1_min=top_1_min,
188
+ top_1_max=top_1_max,
189
+ total_batch_size=total_batch_size,
190
+ log_steps=100)
191
+
192
+ def benchmark_8_gpu(self):
193
+ """Tests Keras model with eager, dist_strat and 8 GPUs."""
194
+ self._setup()
195
+ self._run_and_report_benchmark(
196
+ experiment_name='benchmark_8_gpu',
197
+ num_gpus=8,
198
+ per_replica_batch_size=128,
199
+ epochs=90,
200
+ epochs_between_evals=10,
201
+ dtype='float32')
202
+
203
+ def benchmark_8_gpu_fp16(self):
204
+ """Tests Keras model with eager, dist_strat, 8 GPUs, and fp16."""
205
+ self._setup()
206
+ self._run_and_report_benchmark(
207
+ experiment_name='benchmark_8_gpu_fp16',
208
+ num_gpus=8,
209
+ per_replica_batch_size=256,
210
+ epochs=90,
211
+ epochs_between_evals=10,
212
+ dtype='float16')
213
+
214
+ def benchmark_xla_8_gpu_fp16(self):
215
+ """Tests Keras model with XLA, eager, dist_strat, 8 GPUs and fp16."""
216
+ self._setup()
217
+ self._run_and_report_benchmark(
218
+ experiment_name='benchmark_xla_8_gpu_fp16',
219
+ num_gpus=8,
220
+ per_replica_batch_size=256,
221
+ epochs=90,
222
+ epochs_between_evals=10,
223
+ dtype='float16',
224
+ enable_xla=True)
225
+
226
+ def benchmark_xla_8_gpu_fp16_dynamic(self):
227
+ """Tests Keras model with XLA, eager, dist_strat, 8 GPUs, dynamic fp16."""
228
+ self._setup()
229
+ self._run_and_report_benchmark(
230
+ experiment_name='benchmark_xla_8_gpu_fp16_dynamic',
231
+ top_1_min=0.736,
232
+ num_gpus=8,
233
+ per_replica_batch_size=256,
234
+ epochs=90,
235
+ epochs_between_evals=10,
236
+ dtype='float16',
237
+ loss_scale='dynamic')
238
+
239
+ def _get_model_dir(self, folder_name):
240
+ return os.path.join(self.output_dir, folder_name)
241
+
242
+
243
+ class MobilenetV1KerasAccuracy(keras_benchmark.KerasBenchmark):
244
+ """Benchmark accuracy tests for MobilenetV1 in Keras."""
245
+
246
+ def __init__(self, output_dir=None, root_data_dir=None, **kwargs):
247
+ """A benchmark class.
248
+
249
+ Args:
250
+ output_dir: directory where to output e.g. log files
251
+ root_data_dir: directory under which to look for dataset
252
+ **kwargs: arbitrary named arguments. This is needed to make the
253
+ constructor forward compatible in case PerfZero provides more
254
+ named arguments before updating the constructor.
255
+ """
256
+
257
+ flag_methods = [resnet_imagenet_main.define_imagenet_keras_flags]
258
+
259
+ self.data_dir = os.path.join(root_data_dir, 'imagenet')
260
+ super(MobilenetV1KerasAccuracy, self).__init__(
261
+ output_dir=output_dir,
262
+ flag_methods=flag_methods,
263
+ default_flags={
264
+ 'model': 'mobilenet',
265
+ 'optimizer': 'mobilenet_default',
266
+ 'initial_learning_rate_per_sample': 0.00039,
267
+ })
268
+
269
+ def benchmark_8_gpu(self):
270
+ """Test Keras model with eager, dist_strat and 8 GPUs."""
271
+ self._setup()
272
+ FLAGS.num_gpus = 8
273
+ FLAGS.data_dir = self.data_dir
274
+ FLAGS.batch_size = 128 * 8
275
+ FLAGS.train_epochs = 90
276
+ FLAGS.epochs_between_evals = 10
277
+ FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu')
278
+ FLAGS.dtype = 'fp32'
279
+ FLAGS.enable_eager = True
280
+ self._run_and_report_benchmark()
281
+
282
+ @benchmark_wrappers.enable_runtime_flags
283
+ def _run_and_report_benchmark(self,
284
+ top_1_min=MOBILENET_V1_MIN_TOP_1_ACCURACY,
285
+ top_1_max=MOBILENET_V1_MAX_TOP_1_ACCURACY):
286
+ start_time_sec = time.time()
287
+ stats = resnet_imagenet_main.run(flags.FLAGS)
288
+ wall_time_sec = time.time() - start_time_sec
289
+
290
+ super(MobilenetV1KerasAccuracy, self)._report_benchmark(
291
+ stats,
292
+ wall_time_sec,
293
+ top_1_min=top_1_min,
294
+ top_1_max=top_1_max,
295
+ total_batch_size=FLAGS.batch_size,
296
+ log_steps=100)
297
+
298
+ def _get_model_dir(self, folder_name):
299
+ return os.path.join(self.output_dir, folder_name)
300
+
301
+
302
+ class Resnet50KerasClassifierBenchmarkBase(keras_benchmark.KerasBenchmark):
303
+ """Resnet50 (classifier_trainer) benchmarks."""
304
+
305
+ def __init__(self, output_dir=None, default_flags=None,
306
+ tpu=None, dataset_builder='records', train_epochs=1,
307
+ train_steps=110, data_dir=None):
308
+ flag_methods = [classifier_trainer.define_classifier_flags]
309
+
310
+ self.dataset_builder = dataset_builder
311
+ self.train_epochs = train_epochs
312
+ self.train_steps = train_steps
313
+ self.data_dir = data_dir
314
+
315
+ super(Resnet50KerasClassifierBenchmarkBase, self).__init__(
316
+ output_dir=output_dir,
317
+ flag_methods=flag_methods,
318
+ default_flags=default_flags,
319
+ tpu=tpu)
320
+
321
+ @benchmark_wrappers.enable_runtime_flags
322
+ def _run_and_report_benchmark(
323
+ self,
324
+ experiment_name: str,
325
+ skip_steps: Optional[int] = None,
326
+ top_1_min: float = MIN_TOP_1_ACCURACY,
327
+ top_1_max: float = MAX_TOP_1_ACCURACY,
328
+ num_gpus: int = 0,
329
+ num_tpus: int = 0,
330
+ distribution_strategy: str = 'mirrored',
331
+ per_replica_batch_size: int = 128,
332
+ epochs_between_evals: int = 1,
333
+ dtype: str = 'float32',
334
+ enable_xla: bool = False,
335
+ run_eagerly: bool = False,
336
+ gpu_thread_mode: Optional[str] = None,
337
+ dataset_num_private_threads: Optional[int] = None,
338
+ loss_scale: Optional[str] = None):
339
+ """Runs and reports the benchmark given the provided configuration."""
340
+ FLAGS.model_type = 'resnet'
341
+ FLAGS.dataset = 'imagenet'
342
+ FLAGS.mode = 'train_and_eval'
343
+ FLAGS.data_dir = self.data_dir
344
+ FLAGS.model_dir = self._get_model_dir(experiment_name)
345
+ parameters = _get_classifier_parameters(
346
+ builder=self.dataset_builder,
347
+ skip_eval=True,
348
+ num_gpus=num_gpus,
349
+ distribution_strategy=distribution_strategy,
350
+ per_replica_batch_size=per_replica_batch_size,
351
+ epochs=self.train_epochs,
352
+ steps=self.train_steps,
353
+ epochs_between_evals=epochs_between_evals,
354
+ dtype=dtype,
355
+ enable_xla=enable_xla,
356
+ gpu_thread_mode=gpu_thread_mode,
357
+ dataset_num_private_threads=dataset_num_private_threads,
358
+ loss_scale=loss_scale,
359
+ report_metrics=False,
360
+ batchnorm_spatial_persistent=True)
361
+ FLAGS.params_override = json.dumps(parameters)
362
+ if distribution_strategy == 'tpu':
363
+ total_batch_size = num_tpus * per_replica_batch_size
364
+ else:
365
+ total_batch_size = num_gpus * per_replica_batch_size
366
+
367
+ start_time_sec = time.time()
368
+ stats = classifier_trainer.run(flags.FLAGS)
369
+ wall_time_sec = time.time() - start_time_sec
370
+ # Number of logged step time entries that are excluded in performance
371
+ # report. We keep results from last 100 batches, or skip the steps based on
372
+ # input skip_steps.
373
+ warmup = (skip_steps or (self.train_steps - 100)) // FLAGS.log_steps
374
+
375
+ super(Resnet50KerasClassifierBenchmarkBase, self)._report_benchmark(
376
+ stats,
377
+ wall_time_sec,
378
+ total_batch_size=total_batch_size,
379
+ log_steps=FLAGS.log_steps,
380
+ warmup=warmup,
381
+ start_time_sec=start_time_sec)
382
+
383
+ def benchmark_1_gpu_no_dist_strat(self):
384
+ """Tests Keras model with 1 GPU, no distribution strategy."""
385
+ self._setup()
386
+ self._run_and_report_benchmark(
387
+ experiment_name='benchmark_1_gpu_no_dist_strat',
388
+ num_gpus=1,
389
+ distribution_strategy='off',
390
+ per_replica_batch_size=128)
391
+
392
+ def benchmark_1_gpu_no_dist_strat_run_eagerly(self):
393
+ """Tests Keras model with 1 GPU, no distribution strategy, run eagerly."""
394
+ self._setup()
395
+ self._run_and_report_benchmark(
396
+ experiment_name='benchmark_1_gpu_no_dist_strat_run_eagerly',
397
+ num_gpus=1,
398
+ run_eagerly=True,
399
+ distribution_strategy='off',
400
+ per_replica_batch_size=64)
401
+
402
+ def benchmark_1_gpu_no_dist_strat_run_eagerly_fp16(self):
403
+ """Tests with 1 GPU, no distribution strategy, fp16, run eagerly."""
404
+ self._setup()
405
+ self._run_and_report_benchmark(
406
+ experiment_name='benchmark_1_gpu_no_dist_strat_run_eagerly_fp16',
407
+ num_gpus=1,
408
+ run_eagerly=True,
409
+ distribution_strategy='off',
410
+ dtype='float16',
411
+ per_replica_batch_size=128)
412
+
413
+ def benchmark_1_gpu(self):
414
+ """Tests Keras model with 1 GPU."""
415
+ self._setup()
416
+ self._run_and_report_benchmark(
417
+ experiment_name='benchmark_1_gpu',
418
+ num_gpus=1,
419
+ distribution_strategy='one_device',
420
+ per_replica_batch_size=128)
421
+
422
+ def benchmark_xla_1_gpu(self):
423
+ """Tests Keras model with XLA and 1 GPU."""
424
+ self._setup()
425
+ self._run_and_report_benchmark(
426
+ experiment_name='benchmark_xla_1_gpu',
427
+ num_gpus=1,
428
+ enable_xla=True,
429
+ distribution_strategy='one_device',
430
+ per_replica_batch_size=128)
431
+
432
+ def benchmark_1_gpu_fp16(self):
433
+ """Tests Keras model with 1 GPU and fp16."""
434
+ self._setup()
435
+ self._run_and_report_benchmark(
436
+ experiment_name='benchmark_1_gpu_fp16',
437
+ num_gpus=1,
438
+ distribution_strategy='one_device',
439
+ dtype='float16',
440
+ per_replica_batch_size=256)
441
+
442
+ def benchmark_1_gpu_fp16_dynamic(self):
443
+ """Tests Keras model with 1 GPU, fp16, and dynamic loss scaling."""
444
+ self._setup()
445
+ self._run_and_report_benchmark(
446
+ experiment_name='benchmark_1_gpu_fp16_dynamic',
447
+ num_gpus=1,
448
+ distribution_strategy='one_device',
449
+ dtype='float16',
450
+ per_replica_batch_size=256,
451
+ loss_scale='dynamic')
452
+
453
+ def benchmark_xla_1_gpu_fp16(self):
454
+ """Tests Keras model with XLA, 1 GPU and fp16."""
455
+ self._setup()
456
+ self._run_and_report_benchmark(
457
+ experiment_name='benchmark_xla_1_gpu_fp16',
458
+ num_gpus=1,
459
+ enable_xla=True,
460
+ distribution_strategy='one_device',
461
+ dtype='float16',
462
+ per_replica_batch_size=256)
463
+
464
+ def benchmark_xla_1_gpu_fp16_tweaked(self):
465
+ """Tests Keras model with XLA, 1 GPU, fp16, and manual config tuning."""
466
+ self._setup()
467
+ self._run_and_report_benchmark(
468
+ experiment_name='benchmark_xla_1_gpu_fp16_tweaked',
469
+ num_gpus=1,
470
+ enable_xla=True,
471
+ distribution_strategy='one_device',
472
+ dtype='float16',
473
+ per_replica_batch_size=256,
474
+ gpu_thread_mode='gpu_private')
475
+
476
+ def benchmark_xla_1_gpu_fp16_dynamic(self):
477
+ """Tests Keras model with XLA, 1 GPU, fp16, and dynamic loss scaling."""
478
+ self._setup()
479
+ self._run_and_report_benchmark(
480
+ experiment_name='benchmark_xla_1_gpu_fp16_dynamic',
481
+ num_gpus=1,
482
+ enable_xla=True,
483
+ distribution_strategy='one_device',
484
+ dtype='float16',
485
+ per_replica_batch_size=256,
486
+ loss_scale='dynamic')
487
+
488
+ def benchmark_8_gpu(self):
489
+ """Tests Keras model with 8 GPUs."""
490
+ self._setup()
491
+ self._run_and_report_benchmark(
492
+ experiment_name='benchmark_8_gpu',
493
+ num_gpus=8,
494
+ distribution_strategy='mirrored',
495
+ per_replica_batch_size=128)
496
+
497
+ def benchmark_8_gpu_tweaked(self):
498
+ """Tests Keras model with manual config tuning and 8 GPUs."""
499
+ self._setup()
500
+ self._run_and_report_benchmark(
501
+ experiment_name='benchmark_8_gpu_tweaked',
502
+ num_gpus=8,
503
+ distribution_strategy='mirrored',
504
+ per_replica_batch_size=128,
505
+ dataset_num_private_threads=14)
506
+
507
+ def benchmark_xla_8_gpu(self):
508
+ """Tests Keras model with XLA and 8 GPUs."""
509
+ self._setup()
510
+ self._run_and_report_benchmark(
511
+ experiment_name='benchmark_xla_8_gpu',
512
+ num_gpus=8,
513
+ enable_xla=True,
514
+ distribution_strategy='mirrored',
515
+ per_replica_batch_size=128)
516
+
517
+ def benchmark_xla_8_gpu_tweaked(self):
518
+ """Tests Keras model with manual config tuning, 8 GPUs, and XLA."""
519
+ self._setup()
520
+ self._run_and_report_benchmark(
521
+ experiment_name='benchmark_xla_8_gpu_tweaked',
522
+ num_gpus=8,
523
+ enable_xla=True,
524
+ distribution_strategy='mirrored',
525
+ per_replica_batch_size=128,
526
+ gpu_thread_mode='gpu_private',
527
+ dataset_num_private_threads=24)
528
+
529
+ def benchmark_8_gpu_fp16(self):
530
+ """Tests Keras model with 8 GPUs and fp16."""
531
+ self._setup()
532
+ self._run_and_report_benchmark(
533
+ experiment_name='benchmark_8_gpu_fp16',
534
+ num_gpus=8,
535
+ dtype='float16',
536
+ distribution_strategy='mirrored',
537
+ per_replica_batch_size=256)
538
+
539
+ def benchmark_8_gpu_fp16_tweaked(self):
540
+ """Tests Keras model with 8 GPUs, fp16, and manual config tuning."""
541
+ self._setup()
542
+ self._run_and_report_benchmark(
543
+ experiment_name='benchmark_8_gpu_fp16_tweaked',
544
+ num_gpus=8,
545
+ dtype='float16',
546
+ distribution_strategy='mirrored',
547
+ per_replica_batch_size=256,
548
+ gpu_thread_mode='gpu_private',
549
+ dataset_num_private_threads=40)
550
+
551
+ def benchmark_8_gpu_fp16_dynamic_tweaked(self):
552
+ """Tests Keras model with 8 GPUs, fp16, dynamic loss scaling, and tuned."""
553
+ self._setup()
554
+ self._run_and_report_benchmark(
555
+ experiment_name='benchmark_8_gpu_fp16_dynamic_tweaked',
556
+ num_gpus=8,
557
+ dtype='float16',
558
+ distribution_strategy='mirrored',
559
+ per_replica_batch_size=256,
560
+ loss_scale='dynamic',
561
+ gpu_thread_mode='gpu_private',
562
+ dataset_num_private_threads=40)
563
+
564
+ def benchmark_xla_8_gpu_fp16(self):
565
+ """Tests Keras model with XLA, 8 GPUs and fp16."""
566
+ self._setup()
567
+ self._run_and_report_benchmark(
568
+ experiment_name='benchmark_xla_8_gpu_fp16',
569
+ dtype='float16',
570
+ num_gpus=8,
571
+ enable_xla=True,
572
+ distribution_strategy='mirrored',
573
+ per_replica_batch_size=256)
574
+
575
+ def benchmark_xla_8_gpu_fp16_tweaked(self):
576
+ """Test Keras model with manual config tuning, XLA, 8 GPUs and fp16."""
577
+ self._setup()
578
+ self._run_and_report_benchmark(
579
+ experiment_name='benchmark_xla_8_gpu_fp16_tweaked',
580
+ dtype='float16',
581
+ num_gpus=8,
582
+ enable_xla=True,
583
+ distribution_strategy='mirrored',
584
+ per_replica_batch_size=256,
585
+ gpu_thread_mode='gpu_private',
586
+ dataset_num_private_threads=48)
587
+
588
+ def benchmark_xla_8_gpu_fp16_tweaked_delay_measure(self):
589
+ """Tests with manual config tuning, XLA, 8 GPUs and fp16.
590
+
591
+ Delay performance measurement for stable performance on 96 vCPU platforms.
592
+ """
593
+ self._setup()
594
+ self._run_and_report_benchmark(
595
+ experiment_name='benchmark_xla_8_gpu_fp16_tweaked_delay_measure',
596
+ dtype='float16',
597
+ num_gpus=8,
598
+ enable_xla=True,
599
+ distribution_strategy='mirrored',
600
+ per_replica_batch_size=256,
601
+ gpu_thread_mode='gpu_private',
602
+ dataset_num_private_threads=48,
603
+ steps=310)
604
+
605
+ def benchmark_xla_8_gpu_fp16_dynamic_tweaked(self):
606
+ """Tests Keras model with config tuning, XLA, 8 GPUs and dynamic fp16."""
607
+ self._setup()
608
+ self._run_and_report_benchmark(
609
+ experiment_name='benchmark_xla_8_gpu_fp16_dynamic_tweaked',
610
+ dtype='float16',
611
+ num_gpus=8,
612
+ enable_xla=True,
613
+ distribution_strategy='mirrored',
614
+ per_replica_batch_size=256,
615
+ gpu_thread_mode='gpu_private',
616
+ loss_scale='dynamic',
617
+ dataset_num_private_threads=48)
618
+
619
+ def benchmark_2x2_tpu_bf16(self):
620
+ """Test Keras model with 2x2 TPU, bf16."""
621
+ self._setup()
622
+ self._run_and_report_benchmark(
623
+ experiment_name='benchmark_2x2_tpu_bf16',
624
+ dtype='bfloat16',
625
+ num_tpus=8,
626
+ distribution_strategy='tpu',
627
+ per_replica_batch_size=128)
628
+
629
+ def benchmark_4x4_tpu_bf16(self):
630
+ """Test Keras model with 4x4 TPU, bf16."""
631
+ self._setup()
632
+ self._run_and_report_benchmark(
633
+ experiment_name='benchmark_4x4_tpu_bf16',
634
+ dtype='bfloat16',
635
+ num_tpus=32,
636
+ distribution_strategy='tpu',
637
+ per_replica_batch_size=128)
638
+
639
+ def benchmark_8x8_tpu_bf16(self):
640
+ """Test Keras model with 8x8 TPU, bf16."""
641
+ self._setup()
642
+ self._run_and_report_benchmark(
643
+ experiment_name='benchmark_8x8_tpu_bf16',
644
+ dtype='bfloat16',
645
+ num_tpus=128,
646
+ distribution_strategy='tpu',
647
+ per_replica_batch_size=64)
648
+
649
+ def fill_report_object(self, stats):
650
+ super(Resnet50KerasClassifierBenchmarkBase, self).fill_report_object(
651
+ stats,
652
+ total_batch_size=FLAGS.batch_size,
653
+ log_steps=FLAGS.log_steps)
654
+
655
+
656
+ class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
657
+ """Resnet50 benchmarks."""
658
+
659
+ def __init__(self, output_dir=None, default_flags=None, tpu=None):
660
+ flag_methods = [resnet_imagenet_main.define_imagenet_keras_flags]
661
+
662
+ super(Resnet50KerasBenchmarkBase, self).__init__(
663
+ output_dir=output_dir,
664
+ flag_methods=flag_methods,
665
+ default_flags=default_flags,
666
+ tpu=tpu)
667
+
668
+ @benchmark_wrappers.enable_runtime_flags
669
+ def _run_and_report_benchmark(self, skip_steps=None):
670
+ start_time_sec = time.time()
671
+ stats = resnet_imagenet_main.run(FLAGS)
672
+ wall_time_sec = time.time() - start_time_sec
673
+ # Number of logged step time entries that are excluded in performance
674
+ # report. We keep results from last 100 batches, or skip the steps based on
675
+ # input skip_steps.
676
+ warmup = (skip_steps or (FLAGS.train_steps - 100)) // FLAGS.log_steps
677
+
678
+ super(Resnet50KerasBenchmarkBase, self)._report_benchmark(
679
+ stats,
680
+ wall_time_sec,
681
+ total_batch_size=FLAGS.batch_size,
682
+ log_steps=FLAGS.log_steps,
683
+ warmup=warmup,
684
+ start_time_sec=start_time_sec)
685
+
686
+ def benchmark_1_gpu_no_dist_strat(self):
687
+ """Test Keras model with 1 GPU, no distribution strategy."""
688
+ self._setup()
689
+
690
+ FLAGS.num_gpus = 1
691
+ FLAGS.enable_eager = True
692
+ FLAGS.distribution_strategy = 'off'
693
+ FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu_no_dist_strat')
694
+ FLAGS.batch_size = 128
695
+ self._run_and_report_benchmark()
696
+
697
+ def benchmark_1_gpu_no_dist_strat_run_eagerly(self):
698
+ """Test Keras model with 1 GPU, no distribution strategy, run eagerly."""
699
+ self._setup()
700
+
701
+ FLAGS.num_gpus = 1
702
+ FLAGS.enable_eager = True
703
+ FLAGS.run_eagerly = True
704
+ FLAGS.distribution_strategy = 'off'
705
+ FLAGS.model_dir = self._get_model_dir(
706
+ 'benchmark_1_gpu_no_dist_strat_run_eagerly')
707
+ FLAGS.batch_size = 64
708
+ self._run_and_report_benchmark()
709
+
710
+ def benchmark_1_gpu_no_dist_strat_run_eagerly_tweaked(self):
711
+ """Test Keras model with 1 GPU, no distribution strategy, run eagerly."""
712
+ self._setup()
713
+
714
+ FLAGS.num_gpus = 1
715
+ FLAGS.enable_eager = True
716
+ FLAGS.run_eagerly = True
717
+ FLAGS.explicit_gpu_placement = True
718
+ FLAGS.distribution_strategy = 'off'
719
+ FLAGS.model_dir = self._get_model_dir(
720
+ 'benchmark_1_gpu_no_dist_strat_run_eagerly_tweaked')
721
+ FLAGS.batch_size = 64
722
+ self._run_and_report_benchmark()
723
+
724
+ def benchmark_1_gpu_no_dist_strat_run_eagerly_fp16(self):
725
+ """Test with 1 GPU, no distribution strategy, fp16, run eagerly."""
726
+ self._setup()
727
+
728
+ FLAGS.num_gpus = 1
729
+ FLAGS.enable_eager = True
730
+ FLAGS.run_eagerly = True
731
+ FLAGS.distribution_strategy = 'off'
732
+ FLAGS.model_dir = self._get_model_dir(
733
+ 'benchmark_1_gpu_no_dist_strat_run_eagerly_fp16')
734
+ FLAGS.dtype = 'fp16'
735
+ FLAGS.batch_size = 128
736
+ self._run_and_report_benchmark()
737
+
738
+ def benchmark_1_gpu_no_dist_strat_run_eagerly_fp16_tweaked(self):
739
+ """Test with 1 GPU, no distribution strategy, fp16, run eagerly."""
740
+ self._setup()
741
+
742
+ FLAGS.num_gpus = 1
743
+ FLAGS.enable_eager = True
744
+ FLAGS.run_eagerly = True
745
+ FLAGS.explicit_gpu_placement = True
746
+ FLAGS.distribution_strategy = 'off'
747
+ FLAGS.model_dir = self._get_model_dir(
748
+ 'benchmark_1_gpu_no_dist_strat_run_eagerly_fp16_tweaked')
749
+ FLAGS.dtype = 'fp16'
750
+ FLAGS.batch_size = 128
751
+ self._run_and_report_benchmark()
752
+
753
+ def benchmark_1_gpu(self):
754
+ """Test Keras model with 1 GPU."""
755
+ self._setup()
756
+
757
+ FLAGS.num_gpus = 1
758
+ FLAGS.enable_eager = True
759
+ FLAGS.distribution_strategy = 'one_device'
760
+ FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu')
761
+ FLAGS.batch_size = 128
762
+ self._run_and_report_benchmark()
763
+
764
+ def benchmark_1_gpu_amp(self):
765
+ """Test Keras model with 1 GPU with automatic mixed precision."""
766
+ self._setup()
767
+
768
+ FLAGS.num_gpus = 1
769
+ FLAGS.enable_eager = True
770
+ FLAGS.dtype = 'fp16'
771
+ FLAGS.fp16_implementation = 'graph_rewrite'
772
+ FLAGS.distribution_strategy = 'one_device'
773
+ FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu_amp')
774
+ FLAGS.batch_size = 256
775
+ self._run_and_report_benchmark()
776
+
777
+ def benchmark_xla_1_gpu(self):
778
+ """Test Keras model with XLA and 1 GPU."""
779
+ self._setup()
780
+
781
+ FLAGS.num_gpus = 1
782
+ FLAGS.enable_eager = True
783
+ FLAGS.enable_xla = True
784
+ FLAGS.distribution_strategy = 'one_device'
785
+ FLAGS.model_dir = self._get_model_dir('benchmark_xla_1_gpu')
786
+ FLAGS.batch_size = 128
787
+ self._run_and_report_benchmark()
788
+
789
+ def benchmark_xla_1_gpu_amp(self):
790
+ """Test Keras model with XLA and 1 GPU with automatic mixed precision."""
791
+ self._setup()
792
+
793
+ FLAGS.num_gpus = 1
794
+ FLAGS.enable_eager = True
795
+ FLAGS.dtype = 'fp16'
796
+ FLAGS.fp16_implementation = 'graph_rewrite'
797
+ FLAGS.enable_xla = True
798
+ FLAGS.distribution_strategy = 'one_device'
799
+ FLAGS.model_dir = self._get_model_dir('benchmark_xla_1_gpu_amp')
800
+ FLAGS.batch_size = 256
801
+ self._run_and_report_benchmark()
802
+
803
+ def benchmark_1_gpu_fp16(self):
804
+ """Test Keras model with 1 GPU and fp16."""
805
+ self._setup()
806
+
807
+ FLAGS.num_gpus = 1
808
+ FLAGS.enable_eager = True
809
+ FLAGS.distribution_strategy = 'one_device'
810
+ FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu_fp16')
811
+ FLAGS.dtype = 'fp16'
812
+ FLAGS.batch_size = 256
813
+ self._run_and_report_benchmark()
814
+
815
+ def benchmark_1_gpu_fp16_dynamic(self):
816
+ """Test Keras model with 1 GPU, fp16, and dynamic loss scaling."""
817
+ self._setup()
818
+
819
+ FLAGS.num_gpus = 1
820
+ FLAGS.enable_eager = True
821
+ FLAGS.distribution_strategy = 'one_device'
822
+ FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu_fp16_dynamic')
823
+ FLAGS.dtype = 'fp16'
824
+ FLAGS.batch_size = 256
825
+ FLAGS.loss_scale = 'dynamic'
826
+ self._run_and_report_benchmark()
827
+
828
+ def benchmark_xla_1_gpu_fp16(self):
829
+ """Test Keras model with XLA, 1 GPU and fp16."""
830
+ self._setup()
831
+
832
+ FLAGS.num_gpus = 1
833
+ FLAGS.enable_eager = True
834
+ FLAGS.enable_xla = True
835
+ FLAGS.distribution_strategy = 'one_device'
836
+ FLAGS.model_dir = self._get_model_dir('benchmark_xla_1_gpu_fp16')
837
+ FLAGS.dtype = 'fp16'
838
+ FLAGS.batch_size = 256
839
+ self._run_and_report_benchmark()
840
+
841
+ def benchmark_xla_1_gpu_fp16_tweaked(self):
842
+ """Test Keras model with XLA, 1 GPU, fp16, and manual config tuning."""
843
+ self._setup()
844
+
845
+ FLAGS.num_gpus = 1
846
+ FLAGS.enable_eager = True
847
+ FLAGS.enable_xla = True
848
+ FLAGS.distribution_strategy = 'one_device'
849
+ FLAGS.model_dir = self._get_model_dir('benchmark_xla_1_gpu_fp16_tweaked')
850
+ FLAGS.dtype = 'fp16'
851
+ FLAGS.batch_size = 256
852
+ FLAGS.tf_gpu_thread_mode = 'gpu_private'
853
+ self._run_and_report_benchmark()
854
+
855
+ def benchmark_xla_1_gpu_fp16_dynamic(self):
856
+ """Test Keras model with XLA, 1 GPU, fp16, and dynamic loss scaling."""
857
+ self._setup()
858
+
859
+ FLAGS.num_gpus = 1
860
+ FLAGS.enable_eager = True
861
+ FLAGS.enable_xla = True
862
+ FLAGS.distribution_strategy = 'one_device'
863
+ FLAGS.model_dir = self._get_model_dir('benchmark_xla_1_gpu_fp16_dynamic')
864
+ FLAGS.dtype = 'fp16'
865
+ FLAGS.batch_size = 256
866
+ FLAGS.loss_scale = 'dynamic'
867
+ self._run_and_report_benchmark()
868
+
869
+ def benchmark_8_gpu(self):
870
+ """Test Keras model with 8 GPUs."""
871
+ self._setup()
872
+
873
+ FLAGS.num_gpus = 8
874
+ FLAGS.enable_eager = True
875
+ FLAGS.distribution_strategy = 'mirrored'
876
+ FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu')
877
+ FLAGS.batch_size = 128 * 8 # 8 GPUs
878
+ self._run_and_report_benchmark()
879
+
880
+ def benchmark_8_gpu_amp(self):
881
+ """Test Keras model with 8 GPUs with automatic mixed precision."""
882
+ self._setup()
883
+
884
+ FLAGS.num_gpus = 8
885
+ FLAGS.enable_eager = True
886
+ FLAGS.dtype = 'fp16'
887
+ FLAGS.fp16_implementation = 'graph_rewrite'
888
+ FLAGS.distribution_strategy = 'mirrored'
889
+ FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_amp')
890
+ FLAGS.batch_size = 256 * 8 # 8 GPUs
891
+ self._run_and_report_benchmark()
892
+
893
+ def benchmark_8_gpu_tweaked(self):
894
+ """Test Keras model with manual config tuning and 8 GPUs."""
895
+ self._setup()
896
+
897
+ FLAGS.num_gpus = 8
898
+ FLAGS.enable_eager = True
899
+ FLAGS.distribution_strategy = 'mirrored'
900
+ FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_tweaked')
901
+ FLAGS.batch_size = 128 * 8 # 8 GPUs
902
+ FLAGS.datasets_num_private_threads = 14
903
+ self._run_and_report_benchmark()
904
+
905
+ def benchmark_xla_8_gpu(self):
906
+ """Test Keras model with XLA and 8 GPUs."""
907
+ self._setup()
908
+
909
+ FLAGS.num_gpus = 8
910
+ FLAGS.enable_eager = True
911
+ FLAGS.enable_xla = True
912
+ FLAGS.distribution_strategy = 'mirrored'
913
+ FLAGS.model_dir = self._get_model_dir('benchmark_xla_8_gpu')
914
+ FLAGS.batch_size = 128 * 8 # 8 GPUs
915
+ self._run_and_report_benchmark()
916
+
917
+ def benchmark_xla_8_gpu_amp(self):
918
+ """Test Keras model with XLA and 8 GPUs with automatic mixed precision."""
919
+ self._setup()
920
+
921
+ FLAGS.num_gpus = 8
922
+ FLAGS.enable_eager = True
923
+ FLAGS.dtype = 'fp16'
924
+ FLAGS.fp16_implementation = 'graph_rewrite'
925
+ FLAGS.enable_xla = True
926
+ FLAGS.distribution_strategy = 'mirrored'
927
+ FLAGS.model_dir = self._get_model_dir('benchmark_xla_8_gpu_amp')
928
+ FLAGS.batch_size = 256 * 8 # 8 GPUs
929
+ self._run_and_report_benchmark()
930
+
931
+ def benchmark_xla_8_gpu_tweaked(self):
932
+ """Test Keras model with manual config tuning, 8 GPUs, and XLA."""
933
+ self._setup()
934
+
935
+ FLAGS.num_gpus = 8
936
+ FLAGS.enable_eager = True
937
+ FLAGS.enable_xla = True
938
+ FLAGS.distribution_strategy = 'mirrored'
939
+ FLAGS.model_dir = self._get_model_dir('benchmark_xla_8_gpu_tweaked')
940
+ FLAGS.batch_size = 128 * 8
941
+ FLAGS.tf_gpu_thread_mode = 'gpu_private'
942
+ FLAGS.datasets_num_private_threads = 24
943
+ self._run_and_report_benchmark()
944
+
945
+ def benchmark_8_gpu_fp16(self):
946
+ """Test Keras model with 8 GPUs and fp16."""
947
+ self._setup()
948
+
949
+ FLAGS.num_gpus = 8
950
+ FLAGS.dtype = 'fp16'
951
+ FLAGS.enable_eager = True
952
+ FLAGS.distribution_strategy = 'mirrored'
953
+ FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_fp16')
954
+ FLAGS.batch_size = 256 * 8 # 8 GPUs
955
+ self._run_and_report_benchmark()
956
+
957
+ def benchmark_8_gpu_fp16_tweaked(self):
958
+ """Test Keras model with 8 GPUs, fp16, and manual config tuning."""
959
+ self._setup()
960
+
961
+ FLAGS.num_gpus = 8
962
+ FLAGS.dtype = 'fp16'
963
+ FLAGS.enable_eager = True
964
+ FLAGS.distribution_strategy = 'mirrored'
965
+ FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_fp16_tweaked')
966
+ FLAGS.batch_size = 256 * 8 # 8 GPUs
967
+ FLAGS.tf_gpu_thread_mode = 'gpu_private'
968
+ FLAGS.dataset_num_private_threads = 40
969
+ self._run_and_report_benchmark()
970
+
971
+ def benchmark_8_gpu_fp16_dynamic_tweaked(self):
972
+ """Test Keras model with 8 GPUs, fp16, dynamic loss scaling, and tuned."""
973
+ self._setup()
974
+
975
+ FLAGS.num_gpus = 8
976
+ FLAGS.dtype = 'fp16'
977
+ FLAGS.enable_eager = True
978
+ FLAGS.distribution_strategy = 'mirrored'
979
+ FLAGS.model_dir = self._get_model_dir(
980
+ 'benchmark_8_gpu_fp16_dynamic_tweaked')
981
+ FLAGS.batch_size = 256 * 8 # 8 GPUs
982
+ FLAGS.loss_scale = 'dynamic'
983
+ FLAGS.tf_gpu_thread_mode = 'gpu_private'
984
+ FLAGS.dataset_num_private_threads = 40
985
+ self._run_and_report_benchmark()
986
+
987
+ def benchmark_xla_8_gpu_fp16(self):
988
+ """Test Keras model with XLA, 8 GPUs and fp16."""
989
+ self._setup()
990
+
991
+ FLAGS.num_gpus = 8
992
+ FLAGS.dtype = 'fp16'
993
+ FLAGS.enable_eager = True
994
+ FLAGS.enable_xla = True
995
+ FLAGS.distribution_strategy = 'mirrored'
996
+ FLAGS.model_dir = self._get_model_dir('benchmark_xla_8_gpu_fp16')
997
+ FLAGS.batch_size = 256 * 8 # 8 GPUs
998
+ self._run_and_report_benchmark()
999
+
1000
+ def benchmark_xla_8_gpu_fp16_tweaked(self):
1001
+ """Test Keras model with manual config tuning, XLA, 8 GPUs and fp16."""
1002
+ self._setup()
1003
+
1004
+ FLAGS.num_gpus = 8
1005
+ FLAGS.dtype = 'fp16'
1006
+ FLAGS.enable_eager = True
1007
+ FLAGS.enable_xla = True
1008
+ FLAGS.distribution_strategy = 'mirrored'
1009
+ FLAGS.model_dir = self._get_model_dir('benchmark_xla_8_gpu_fp16_tweaked')
1010
+ FLAGS.batch_size = 256 * 8 # 8 GPUs
1011
+ FLAGS.tf_gpu_thread_mode = 'gpu_private'
1012
+ FLAGS.datasets_num_private_threads = 48
1013
+ self._run_and_report_benchmark()
1014
+
1015
+ def benchmark_xla_8_gpu_fp16_tweaked_delay_measure(self):
1016
+ """Test with manual config tuning, XLA, 8 GPUs and fp16.
1017
+
1018
+ Delay performance measurement for stable performance on 96 vCPU platforms.
1019
+ """
1020
+ self._setup()
1021
+
1022
+ FLAGS.num_gpus = 8
1023
+ FLAGS.dtype = 'fp16'
1024
+ FLAGS.enable_eager = True
1025
+ FLAGS.enable_xla = True
1026
+ FLAGS.distribution_strategy = 'mirrored'
1027
+ FLAGS.model_dir = self._get_model_dir(
1028
+ 'benchmark_xla_8_gpu_fp16_tweaked_delay_measure')
1029
+ FLAGS.batch_size = 256 * 8
1030
+ FLAGS.tf_gpu_thread_mode = 'gpu_private'
1031
+ FLAGS.datasets_num_private_threads = 48
1032
+ FLAGS.train_steps = 310
1033
+ self._run_and_report_benchmark()
1034
+
1035
+ def benchmark_xla_8_gpu_fp16_dynamic_tweaked(self):
1036
+ """Test Keras model with config tuning, XLA, 8 GPUs and dynamic fp16."""
1037
+ self._setup()
1038
+
1039
+ FLAGS.num_gpus = 8
1040
+ FLAGS.dtype = 'fp16'
1041
+ FLAGS.enable_eager = True
1042
+ FLAGS.enable_xla = True
1043
+ FLAGS.distribution_strategy = 'mirrored'
1044
+ FLAGS.model_dir = self._get_model_dir(
1045
+ 'benchmark_xla_8_gpu_fp16_dynamic_tweaked')
1046
+ FLAGS.batch_size = 256 * 8 # 8 GPUs
1047
+ FLAGS.loss_scale = 'dynamic'
1048
+ FLAGS.tf_gpu_thread_mode = 'gpu_private'
1049
+ FLAGS.datasets_num_private_threads = 48
1050
+ self._run_and_report_benchmark()
1051
+
1052
+ def benchmark_2x2_tpu_bf16(self):
1053
+ """Test Keras model with 2x2 TPU, bf16."""
1054
+ self._setup()
1055
+
1056
+ FLAGS.dtype = 'bf16'
1057
+ FLAGS.distribution_strategy = 'tpu'
1058
+ FLAGS.model_dir = self._get_model_dir('benchmark_2x2_tpu_bf16')
1059
+ FLAGS.batch_size = 1024
1060
+ self._run_and_report_benchmark()
1061
+
1062
+ def benchmark_4x4_tpu_bf16(self):
1063
+ """Test Keras model with 4x4 TPU, bf16."""
1064
+ self._setup()
1065
+
1066
+ FLAGS.dtype = 'bf16'
1067
+ FLAGS.distribution_strategy = 'tpu'
1068
+ FLAGS.model_dir = self._get_model_dir('benchmark_4x4_tpu_bf16')
1069
+ FLAGS.batch_size = 4096
1070
+ self._run_and_report_benchmark()
1071
+
1072
+ def benchmark_8x8_tpu_bf16(self):
1073
+ """Test Keras model with 8x8 TPU, bf16."""
1074
+ self._setup()
1075
+
1076
+ FLAGS.dtype = 'bf16'
1077
+ FLAGS.distribution_strategy = 'tpu'
1078
+ FLAGS.model_dir = self._get_model_dir('benchmark_8x8_tpu_bf16')
1079
+ FLAGS.batch_size = 8192
1080
+ self._run_and_report_benchmark()
1081
+
1082
+ def fill_report_object(self, stats):
1083
+ super(Resnet50KerasBenchmarkBase, self).fill_report_object(
1084
+ stats,
1085
+ total_batch_size=FLAGS.batch_size,
1086
+ log_steps=FLAGS.log_steps)
1087
+
1088
+
1089
+ class Resnet50KerasBenchmarkSynth(Resnet50KerasClassifierBenchmarkBase):
1090
+ """Resnet50 synthetic benchmark tests."""
1091
+
1092
+ def __init__(self, output_dir=None, root_data_dir=None, tpu=None, **kwargs):
1093
+ def_flags = {}
1094
+ def_flags['log_steps'] = 10
1095
+
1096
+ super(Resnet50KerasBenchmarkSynth, self).__init__(
1097
+ output_dir=output_dir, default_flags=def_flags, tpu=tpu,
1098
+ dataset_builder='synthetic', train_epochs=1, train_steps=110)
1099
+
1100
+
1101
+ class Resnet50KerasBenchmarkReal(Resnet50KerasClassifierBenchmarkBase):
1102
+ """Resnet50 real data benchmark tests."""
1103
+
1104
+ def __init__(self, output_dir=None, root_data_dir=None, tpu=None, **kwargs):
1105
+ data_dir = os.path.join(root_data_dir, 'imagenet')
1106
+ def_flags = {}
1107
+ def_flags['log_steps'] = 10
1108
+
1109
+ super(Resnet50KerasBenchmarkReal, self).__init__(
1110
+ output_dir=output_dir, default_flags=def_flags, tpu=tpu,
1111
+ dataset_builder='records', train_epochs=1, train_steps=110,
1112
+ data_dir=data_dir)
1113
+
1114
+
1115
+ class Resnet50KerasBenchmarkRemoteData(Resnet50KerasBenchmarkBase):
1116
+ """Resnet50 real data (stored in remote storage) benchmark tests."""
1117
+
1118
+ def __init__(self, output_dir=None, root_data_dir=None, **kwargs):
1119
+ def_flags = {}
1120
+ def_flags['skip_eval'] = True
1121
+ def_flags['report_accuracy_metrics'] = False
1122
+ def_flags['data_dir'] = os.path.join(root_data_dir, 'imagenet')
1123
+ # Defining multiple epochs overrides the train_steps setting in benchmarks.
1124
+ def_flags['train_epochs'] = 2
1125
+ # Cache dataset so performance is stable after the first epoch.
1126
+ def_flags['training_dataset_cache'] = True
1127
+ def_flags['log_steps'] = 100
1128
+ # Note that for single GPU and pure eager tests which are less likely to be
1129
+ # input bound and more stable, these tests will run for shorter time by
1130
+ # overriding FLAGS.train_epochs, train_seteps, log_steps in benchmark
1131
+ # methods, and skip_steps in _run_and_report_benchmark().
1132
+
1133
+ super(Resnet50KerasBenchmarkRemoteData, self).__init__(
1134
+ output_dir=output_dir, default_flags=def_flags)
1135
+
1136
+ def _override_flags_to_run_test_shorter(self):
1137
+ FLAGS.train_epochs = 1
1138
+ FLAGS.train_steps = 300
1139
+ FLAGS.log_steps = 10
1140
+
1141
+ def benchmark_1_gpu_no_dist_strat(self):
1142
+ """Test Keras model with 1 GPU, no distribution strategy."""
1143
+ self._setup()
1144
+
1145
+ FLAGS.num_gpus = 1
1146
+ FLAGS.enable_eager = True
1147
+ FLAGS.distribution_strategy = 'off'
1148
+ FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu_no_dist_strat')
1149
+ FLAGS.batch_size = 128
1150
+ self._override_flags_to_run_test_shorter()
1151
+ self._run_and_report_benchmark()
1152
+
1153
+ def benchmark_1_gpu_no_dist_strat_run_eagerly(self):
1154
+ """Test Keras model with 1 GPU, no distribution strategy, run eagerly."""
1155
+ self._setup()
1156
+
1157
+ FLAGS.num_gpus = 1
1158
+ FLAGS.enable_eager = True
1159
+ FLAGS.run_eagerly = True
1160
+ FLAGS.distribution_strategy = 'off'
1161
+ FLAGS.model_dir = self._get_model_dir(
1162
+ 'benchmark_1_gpu_no_dist_strat_run_eagerly')
1163
+ FLAGS.batch_size = 64
1164
+ self._override_flags_to_run_test_shorter()
1165
+ self._run_and_report_benchmark()
1166
+
1167
+ def benchmark_1_gpu_no_dist_strat_run_eagerly_tweaked(self):
1168
+ """Test Keras model with 1 GPU, no distribution strategy, run eagerly."""
1169
+ self._setup()
1170
+
1171
+ FLAGS.num_gpus = 1
1172
+ FLAGS.enable_eager = True
1173
+ FLAGS.run_eagerly = True
1174
+ FLAGS.explicit_gpu_placement = True
1175
+ FLAGS.distribution_strategy = 'off'
1176
+ FLAGS.model_dir = self._get_model_dir(
1177
+ 'benchmark_1_gpu_no_dist_strat_run_eagerly_tweaked')
1178
+ FLAGS.batch_size = 64
1179
+ self._override_flags_to_run_test_shorter()
1180
+ self._run_and_report_benchmark()
1181
+
1182
+ def benchmark_1_gpu_no_dist_strat_run_eagerly_fp16(self):
1183
+ """Test with 1 GPU, no distribution strategy, fp16, run eagerly."""
1184
+ self._setup()
1185
+
1186
+ FLAGS.num_gpus = 1
1187
+ FLAGS.enable_eager = True
1188
+ FLAGS.run_eagerly = True
1189
+ FLAGS.distribution_strategy = 'off'
1190
+ FLAGS.model_dir = self._get_model_dir(
1191
+ 'benchmark_1_gpu_no_dist_strat_run_eagerly_fp16')
1192
+ FLAGS.dtype = 'fp16'
1193
+ FLAGS.batch_size = 128
1194
+ self._override_flags_to_run_test_shorter()
1195
+ self._run_and_report_benchmark()
1196
+
1197
+ def benchmark_1_gpu_no_dist_strat_run_eagerly_fp16_tweaked(self):
1198
+ """Test with 1 GPU, no distribution strategy, fp16, run eagerly."""
1199
+ self._setup()
1200
+
1201
+ FLAGS.num_gpus = 1
1202
+ FLAGS.enable_eager = True
1203
+ FLAGS.run_eagerly = True
1204
+ FLAGS.explicit_gpu_placement = True
1205
+ FLAGS.distribution_strategy = 'off'
1206
+ FLAGS.model_dir = self._get_model_dir(
1207
+ 'benchmark_1_gpu_no_dist_strat_run_eagerly_fp16_tweaked')
1208
+ FLAGS.dtype = 'fp16'
1209
+ FLAGS.batch_size = 128
1210
+ self._override_flags_to_run_test_shorter()
1211
+ self._run_and_report_benchmark()
1212
+
1213
+ def benchmark_1_gpu(self):
1214
+ """Test Keras model with 1 GPU."""
1215
+ self._setup()
1216
+
1217
+ FLAGS.num_gpus = 1
1218
+ FLAGS.enable_eager = True
1219
+ FLAGS.distribution_strategy = 'one_device'
1220
+ FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu')
1221
+ FLAGS.batch_size = 128
1222
+ self._override_flags_to_run_test_shorter()
1223
+ self._run_and_report_benchmark()
1224
+
1225
+ def benchmark_1_gpu_amp(self):
1226
+ """Test Keras model with 1 GPU with automatic mixed precision."""
1227
+ self._setup()
1228
+
1229
+ FLAGS.num_gpus = 1
1230
+ FLAGS.enable_eager = True
1231
+ FLAGS.dtype = 'fp16'
1232
+ FLAGS.fp16_implementation = 'graph_rewrite'
1233
+ FLAGS.distribution_strategy = 'one_device'
1234
+ FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu_amp')
1235
+ FLAGS.batch_size = 256
1236
+ self._override_flags_to_run_test_shorter()
1237
+ self._run_and_report_benchmark()
1238
+
1239
+ def benchmark_xla_1_gpu(self):
1240
+ """Test Keras model with XLA and 1 GPU."""
1241
+ self._setup()
1242
+
1243
+ FLAGS.num_gpus = 1
1244
+ FLAGS.enable_eager = True
1245
+ FLAGS.enable_xla = True
1246
+ FLAGS.distribution_strategy = 'one_device'
1247
+ FLAGS.model_dir = self._get_model_dir('benchmark_xla_1_gpu')
1248
+ FLAGS.batch_size = 128
1249
+ self._override_flags_to_run_test_shorter()
1250
+ self._run_and_report_benchmark()
1251
+
1252
+ def benchmark_xla_1_gpu_amp(self):
1253
+ """Test Keras model with XLA and 1 GPU with automatic mixed precision."""
1254
+ self._setup()
1255
+
1256
+ FLAGS.num_gpus = 1
1257
+ FLAGS.enable_eager = True
1258
+ FLAGS.dtype = 'fp16'
1259
+ FLAGS.fp16_implementation = 'graph_rewrite'
1260
+ FLAGS.enable_xla = True
1261
+ FLAGS.distribution_strategy = 'one_device'
1262
+ FLAGS.model_dir = self._get_model_dir('benchmark_xla_1_gpu_amp')
1263
+ FLAGS.batch_size = 256
1264
+ self._override_flags_to_run_test_shorter()
1265
+ self._run_and_report_benchmark()
1266
+
1267
+ def benchmark_1_gpu_fp16(self):
1268
+ """Test Keras model with 1 GPU and fp16."""
1269
+ self._setup()
1270
+
1271
+ FLAGS.num_gpus = 1
1272
+ FLAGS.enable_eager = True
1273
+ FLAGS.distribution_strategy = 'one_device'
1274
+ FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu_fp16')
1275
+ FLAGS.dtype = 'fp16'
1276
+ FLAGS.batch_size = 256
1277
+ self._override_flags_to_run_test_shorter()
1278
+ self._run_and_report_benchmark()
1279
+
1280
+ def benchmark_1_gpu_fp16_dynamic(self):
1281
+ """Test Keras model with 1 GPU, fp16, and dynamic loss scaling."""
1282
+ self._setup()
1283
+
1284
+ FLAGS.num_gpus = 1
1285
+ FLAGS.enable_eager = True
1286
+ FLAGS.distribution_strategy = 'one_device'
1287
+ FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu_fp16_dynamic')
1288
+ FLAGS.dtype = 'fp16'
1289
+ FLAGS.batch_size = 256
1290
+ FLAGS.loss_scale = 'dynamic'
1291
+ self._override_flags_to_run_test_shorter()
1292
+ self._run_and_report_benchmark()
1293
+
1294
+ def benchmark_xla_1_gpu_fp16(self):
1295
+ """Test Keras model with XLA, 1 GPU and fp16."""
1296
+ self._setup()
1297
+
1298
+ FLAGS.num_gpus = 1
1299
+ FLAGS.enable_eager = True
1300
+ FLAGS.enable_xla = True
1301
+ FLAGS.distribution_strategy = 'one_device'
1302
+ FLAGS.model_dir = self._get_model_dir('benchmark_xla_1_gpu_fp16')
1303
+ FLAGS.dtype = 'fp16'
1304
+ FLAGS.batch_size = 256
1305
+ self._override_flags_to_run_test_shorter()
1306
+ self._run_and_report_benchmark()
1307
+
1308
+ def benchmark_xla_1_gpu_fp16_tweaked(self):
1309
+ """Test Keras model with XLA, 1 GPU, fp16, and manual config tuning."""
1310
+ self._setup()
1311
+
1312
+ FLAGS.num_gpus = 1
1313
+ FLAGS.enable_eager = True
1314
+ FLAGS.enable_xla = True
1315
+ FLAGS.distribution_strategy = 'one_device'
1316
+ FLAGS.model_dir = self._get_model_dir('benchmark_xla_1_gpu_fp16_tweaked')
1317
+ FLAGS.dtype = 'fp16'
1318
+ FLAGS.batch_size = 256
1319
+ FLAGS.tf_gpu_thread_mode = 'gpu_private'
1320
+ self._override_flags_to_run_test_shorter()
1321
+ self._run_and_report_benchmark()
1322
+
1323
+ def benchmark_xla_1_gpu_fp16_dynamic(self):
1324
+ """Test Keras model with XLA, 1 GPU, fp16, and dynamic loss scaling."""
1325
+ self._setup()
1326
+
1327
+ FLAGS.num_gpus = 1
1328
+ FLAGS.enable_eager = True
1329
+ FLAGS.enable_xla = True
1330
+ FLAGS.distribution_strategy = 'one_device'
1331
+ FLAGS.model_dir = self._get_model_dir('benchmark_xla_1_gpu_fp16_dynamic')
1332
+ FLAGS.dtype = 'fp16'
1333
+ FLAGS.batch_size = 256
1334
+ FLAGS.loss_scale = 'dynamic'
1335
+ self._override_flags_to_run_test_shorter()
1336
+ self._run_and_report_benchmark()
1337
+
1338
+ @benchmark_wrappers.enable_runtime_flags
1339
+ def _run_and_report_benchmark(self):
1340
+ if FLAGS.num_gpus == 1 or FLAGS.run_eagerly:
1341
+ # For single GPU and pure eager tests which are less likely to be input
1342
+ # bound and more stable, run for shorter time and use the default
1343
+ # skip_steps.
1344
+ skip_steps = None
1345
+ else:
1346
+ # skip the first epoch for performance measurement.
1347
+ skip_steps = 600
1348
+ super(Resnet50KerasBenchmarkRemoteData,
1349
+ self)._run_and_report_benchmark(skip_steps=skip_steps)
1350
+
1351
+
1352
+ class TrivialKerasBenchmarkReal(keras_benchmark.KerasBenchmark):
1353
+ """Trivial model with real data benchmark tests."""
1354
+
1355
+ def __init__(self, output_dir=None, root_data_dir=None, **kwargs):
1356
+ flag_methods = [resnet_imagenet_main.define_imagenet_keras_flags]
1357
+
1358
+ def_flags = {}
1359
+ def_flags['use_trivial_model'] = True
1360
+ def_flags['skip_eval'] = True
1361
+ def_flags['report_accuracy_metrics'] = False
1362
+ def_flags['dtype'] = 'fp16'
1363
+ def_flags['data_dir'] = os.path.join(root_data_dir, 'imagenet')
1364
+ def_flags['train_steps'] = 600
1365
+ def_flags['log_steps'] = 100
1366
+ def_flags['distribution_strategy'] = 'mirrored'
1367
+
1368
+ super(TrivialKerasBenchmarkReal, self).__init__(
1369
+ output_dir=output_dir,
1370
+ flag_methods=flag_methods,
1371
+ default_flags=def_flags)
1372
+
1373
+ @benchmark_wrappers.enable_runtime_flags
1374
+ def _run_and_report_benchmark(self):
1375
+ start_time_sec = time.time()
1376
+ stats = resnet_imagenet_main.run(FLAGS)
1377
+ wall_time_sec = time.time() - start_time_sec
1378
+
1379
+ super(TrivialKerasBenchmarkReal, self)._report_benchmark(
1380
+ stats,
1381
+ wall_time_sec,
1382
+ total_batch_size=FLAGS.batch_size,
1383
+ log_steps=FLAGS.log_steps)
1384
+
1385
+ def benchmark_8_gpu_warmup(self):
1386
+ """Dummy test that runs over an epoch to warmup the machine."""
1387
+ self._setup()
1388
+
1389
+ FLAGS.num_gpus = 8
1390
+ FLAGS.enable_eager = True
1391
+ FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_warmup')
1392
+ FLAGS.batch_size = 256 * 8
1393
+ FLAGS.train_steps = 700
1394
+ self._run_and_report_benchmark()
1395
+
1396
+ def fill_report_object(self, stats):
1397
+ super(TrivialKerasBenchmarkReal, self).fill_report_object(
1398
+ stats,
1399
+ total_batch_size=FLAGS.batch_size,
1400
+ log_steps=FLAGS.log_steps)
1401
+
1402
+
1403
+ class Resnet50MultiWorkerKerasAccuracy(keras_benchmark.KerasBenchmark):
1404
+ """Resnet50 distributed accuracy tests with multiple workers."""
1405
+
1406
+ def __init__(self, output_dir=None, root_data_dir=None, **kwargs):
1407
+ flag_methods = [classifier_trainer.define_imagenet_keras_flags]
1408
+ self.data_dir = os.path.join(root_data_dir, 'imagenet')
1409
+ super(Resnet50MultiWorkerKerasAccuracy, self).__init__(
1410
+ output_dir=output_dir, flag_methods=flag_methods)
1411
+
1412
+ def _benchmark_common(self, eager, num_workers, all_reduce_alg):
1413
+ """Common to all benchmarks in this class."""
1414
+ self._setup()
1415
+
1416
+ num_gpus = 8
1417
+ FLAGS.num_gpus = num_gpus
1418
+ FLAGS.data_dir = self.data_dir
1419
+ FLAGS.train_epochs = 90
1420
+ FLAGS.epochs_between_evals = 10
1421
+ FLAGS.dtype = 'fp16'
1422
+ FLAGS.enable_eager = eager
1423
+ FLAGS.enable_xla = False
1424
+ FLAGS.distribution_strategy = 'multi_worker_mirrored'
1425
+ FLAGS.tf_gpu_thread_mode = 'gpu_private'
1426
+ FLAGS.datasets_num_private_threads = 32
1427
+ FLAGS.model_dir = self._get_model_dir(
1428
+ 'benchmark_{}_8_gpu_{}_worker_fp16_{}_tweaked'.format(
1429
+ 'eager' if eager else 'graph', num_workers, all_reduce_alg))
1430
+ FLAGS.batch_size = 256 * num_gpus * num_workers
1431
+ FLAGS.all_reduce_alg = all_reduce_alg
1432
+
1433
+ self._run_and_report_benchmark()
1434
+
1435
+ @benchmark_wrappers.enable_runtime_flags
1436
+ def _run_and_report_benchmark(self,
1437
+ top_1_min=MIN_TOP_1_ACCURACY,
1438
+ top_1_max=MAX_TOP_1_ACCURACY):
1439
+ start_time_sec = time.time()
1440
+ stats = classifier_trainer.run(flags.FLAGS)
1441
+ wall_time_sec = time.time() - start_time_sec
1442
+
1443
+ super(Resnet50MultiWorkerKerasAccuracy, self)._report_benchmark(
1444
+ stats,
1445
+ wall_time_sec,
1446
+ top_1_min=top_1_min,
1447
+ top_1_max=top_1_max,
1448
+ total_batch_size=FLAGS.batch_size,
1449
+ log_steps=100)
1450
+
1451
+ def _get_model_dir(self, folder_name):
1452
+ return os.path.join(self.output_dir, folder_name)
1453
+
1454
+ def benchmark_eager_8_gpu_2_workers_fp16_ring_tweaked(self):
1455
+ """Eager, 8 GPUs per worker, 2 workers, fp16, ring all-reduce."""
1456
+ self._benchmark_common(eager=True, num_workers=2, all_reduce_alg='ring')
1457
+
1458
+ def benchmark_eager_8_gpu_2_workers_fp16_nccl_tweaked(self):
1459
+ """Eager, 8 GPUs per worker, 2 workers, fp16, nccl all-reduce."""
1460
+ self._benchmark_common(eager=True, num_workers=2, all_reduce_alg='nccl')
1461
+
1462
+ def benchmark_eager_8_gpu_8_workers_fp16_ring_tweaked(self):
1463
+ """Eager, 8 GPUs per worker, 8 workers, fp16, ring all-reduce."""
1464
+ self._benchmark_common(eager=True, num_workers=8, all_reduce_alg='ring')
1465
+
1466
+ def benchmark_eager_8_gpu_8_workers_fp16_nccl_tweaked(self):
1467
+ """Eager, 8 GPUs per worker, 8 workers, fp16, nccl all-reduce."""
1468
+ self._benchmark_common(eager=True, num_workers=8, all_reduce_alg='nccl')
1469
+
1470
+
1471
+ class Resnet50MultiWorkerKerasBenchmark(Resnet50KerasBenchmarkBase):
1472
+ """Resnet50 distributed benchmark tests with multiple workers."""
1473
+
1474
+ def __init__(self, output_dir=None, default_flags=None):
1475
+ super(Resnet50MultiWorkerKerasBenchmark, self).__init__(
1476
+ output_dir=output_dir, default_flags=default_flags)
1477
+
1478
+ def _benchmark_common(self, eager, num_workers, all_reduce_alg):
1479
+ """Common to all benchmarks in this class."""
1480
+ self._setup()
1481
+
1482
+ num_gpus = 8
1483
+ FLAGS.num_gpus = num_gpus
1484
+ FLAGS.dtype = 'fp16'
1485
+ FLAGS.enable_eager = eager
1486
+ FLAGS.enable_xla = False
1487
+ FLAGS.distribution_strategy = 'multi_worker_mirrored'
1488
+ FLAGS.tf_gpu_thread_mode = 'gpu_private'
1489
+ FLAGS.datasets_num_private_threads = 32
1490
+ FLAGS.model_dir = self._get_model_dir(
1491
+ 'benchmark_{}_8_gpu_{}_worker_fp16_{}_tweaked'.format(
1492
+ 'eager' if eager else 'graph', num_workers, all_reduce_alg))
1493
+ FLAGS.batch_size = 256 * num_gpus * num_workers
1494
+ FLAGS.all_reduce_alg = all_reduce_alg
1495
+
1496
+ self._run_and_report_benchmark()
1497
+
1498
+ def benchmark_eager_8_gpu_1_worker_fp16_ring_tweaked(self):
1499
+ """Eager, 8 GPUs per worker, 1 worker, fp16, ring all-reduce."""
1500
+ self._benchmark_common(eager=True, num_workers=1, all_reduce_alg='ring')
1501
+
1502
+ def benchmark_eager_8_gpu_1_worker_fp16_nccl_tweaked(self):
1503
+ """Eager, 8 GPUs per worker, 1 worker, fp16, nccl all-reduce."""
1504
+ self._benchmark_common(eager=True, num_workers=1, all_reduce_alg='nccl')
1505
+
1506
+ def benchmark_eager_8_gpu_2_workers_fp16_ring_tweaked(self):
1507
+ """Eager, 8 GPUs per worker, 2 workers, fp16, ring all-reduce."""
1508
+ self._benchmark_common(eager=True, num_workers=2, all_reduce_alg='ring')
1509
+
1510
+ def benchmark_eager_8_gpu_2_workers_fp16_nccl_tweaked(self):
1511
+ """Eager, 8 GPUs per worker, 2 workers, fp16, nccl all-reduce."""
1512
+ self._benchmark_common(eager=True, num_workers=2, all_reduce_alg='nccl')
1513
+
1514
+ def benchmark_eager_8_gpu_8_workers_fp16_ring_tweaked(self):
1515
+ """Eager, 8 GPUs per worker, 8 workers, fp16, ring all-reduce."""
1516
+ self._benchmark_common(eager=True, num_workers=8, all_reduce_alg='ring')
1517
+
1518
+ def benchmark_eager_8_gpu_8_workers_fp16_nccl_tweaked(self):
1519
+ """Eager, 8 GPUs per worker, 8 workers, fp16, nccl all-reduce."""
1520
+ self._benchmark_common(eager=True, num_workers=8, all_reduce_alg='nccl')
1521
+
1522
+
1523
+ class Resnet50MultiWorkerKerasBenchmarkSynth(Resnet50MultiWorkerKerasBenchmark):
1524
+ """Resnet50 multi-worker synthetic data benchmark tests."""
1525
+
1526
+ def __init__(self, output_dir=None, root_data_dir=None, **kwargs):
1527
+ def_flags = {}
1528
+ def_flags['skip_eval'] = True
1529
+ def_flags['report_accuracy_metrics'] = False
1530
+ def_flags['use_synthetic_data'] = True
1531
+ def_flags['train_steps'] = 110
1532
+ def_flags['log_steps'] = 10
1533
+
1534
+ super(Resnet50MultiWorkerKerasBenchmarkSynth, self).__init__(
1535
+ output_dir=output_dir, default_flags=def_flags)
1536
+
1537
+
1538
+ class Resnet50MultiWorkerKerasBenchmarkReal(Resnet50MultiWorkerKerasBenchmark):
1539
+ """Resnet50 multi-worker real data benchmark tests."""
1540
+
1541
+ def __init__(self, output_dir=None, root_data_dir=None, **kwargs):
1542
+ def_flags = {}
1543
+ def_flags['skip_eval'] = True
1544
+ def_flags['report_accuracy_metrics'] = False
1545
+ def_flags['data_dir'] = os.path.join(root_data_dir, 'imagenet')
1546
+ def_flags['train_steps'] = 110
1547
+ def_flags['log_steps'] = 10
1548
+
1549
+ super(Resnet50MultiWorkerKerasBenchmarkReal, self).__init__(
1550
+ output_dir=output_dir, default_flags=def_flags)
1551
+
1552
+
1553
+ # TODO(kimjaehong): It also should be also cover other metheods of model
1554
+ # optimization techniques. In that time, this class will change to something
1555
+ # like 'KerasModelOptimizationAccuracyBase'.
1556
+ class KerasPruningAccuracyBase(keras_benchmark.KerasBenchmark):
1557
+ """Benchmark accuracy tests for pruning method."""
1558
+
1559
+ def __init__(self,
1560
+ output_dir=None,
1561
+ root_data_dir=None,
1562
+ default_flags=None,
1563
+ **kwargs):
1564
+ """A accuracy benchmark class for pruning method.
1565
+
1566
+ Args:
1567
+ output_dir: directory where to output e.g. log files
1568
+ root_data_dir: directory under which to look for dataset
1569
+ default_flags: default flags
1570
+ **kwargs: arbitrary named arguments. This is needed to make the
1571
+ constructor forward compatible in case PerfZero provides more
1572
+ named arguments before updating the constructor.
1573
+ """
1574
+ if default_flags is None:
1575
+ default_flags = {}
1576
+ default_flags['pruning_method'] = 'polynomial_decay'
1577
+ default_flags['data_dir'] = os.path.join(root_data_dir, 'imagenet')
1578
+
1579
+ flag_methods = [resnet_imagenet_main.define_imagenet_keras_flags]
1580
+
1581
+ super(KerasPruningAccuracyBase, self).__init__(
1582
+ output_dir=output_dir,
1583
+ flag_methods=flag_methods,
1584
+ default_flags=default_flags,
1585
+ **kwargs)
1586
+
1587
+ def benchmark_8_gpu(self):
1588
+ """Test Keras model with eager, dist_strat and 8 GPUs."""
1589
+ self._setup()
1590
+ FLAGS.num_gpus = 8
1591
+ FLAGS.batch_size = 32 * 8
1592
+ FLAGS.train_epochs = 90
1593
+ FLAGS.epochs_between_evals = 10
1594
+ FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu')
1595
+ FLAGS.dtype = 'fp32'
1596
+ FLAGS.enable_eager = True
1597
+ self._run_and_report_benchmark()
1598
+
1599
+ @benchmark_wrappers.enable_runtime_flags
1600
+ def _run_and_report_benchmark(self,
1601
+ top_1_min=MODEL_OPTIMIZATION_TOP_1_ACCURACY[
1602
+ 'RESNET50_FINETUNE_PRUNING'][0],
1603
+ top_1_max=MODEL_OPTIMIZATION_TOP_1_ACCURACY[
1604
+ 'RESNET50_FINETUNE_PRUNING'][1]):
1605
+ start_time_sec = time.time()
1606
+ stats = resnet_imagenet_main.run(flags.FLAGS)
1607
+ wall_time_sec = time.time() - start_time_sec
1608
+
1609
+ super(KerasPruningAccuracyBase, self)._report_benchmark(
1610
+ stats,
1611
+ wall_time_sec,
1612
+ top_1_min=top_1_min,
1613
+ top_1_max=top_1_max,
1614
+ total_batch_size=FLAGS.batch_size,
1615
+ log_steps=100)
1616
+
1617
+
1618
+ class MobilenetV1KerasPruningAccuracy(KerasPruningAccuracyBase):
1619
+ """Benchmark accuracy tests for MobilenetV1 with pruning method."""
1620
+
1621
+ def __init__(self, root_data_dir=None, **kwargs):
1622
+ default_flags = {
1623
+ 'model': 'mobilenet',
1624
+ 'optimizer': 'mobilenet_default',
1625
+ 'initial_learning_rate_per_sample': 0.00007,
1626
+ 'pretrained_filepath': tf.train.latest_checkpoint(
1627
+ os.path.join(root_data_dir, 'mobilenet_v1')),
1628
+ 'pruning_begin_step': 0,
1629
+ 'pruning_end_step': 100000,
1630
+ 'pruning_initial_sparsity': 0.0,
1631
+ 'pruning_final_sparsity': 0.5,
1632
+ 'pruning_frequency': 100,
1633
+ }
1634
+ super(MobilenetV1KerasPruningAccuracy, self).__init__(
1635
+ root_data_dir=root_data_dir,
1636
+ default_flags=default_flags,
1637
+ **kwargs)
1638
+
1639
+ def _run_and_report_benchmark(self):
1640
+ super(MobilenetV1KerasPruningAccuracy, self)._run_and_report_benchmark(
1641
+ top_1_min=\
1642
+ MODEL_OPTIMIZATION_TOP_1_ACCURACY['MOBILENET_V1_FINETUNE_PRUNING'][0],
1643
+ top_1_max=\
1644
+ MODEL_OPTIMIZATION_TOP_1_ACCURACY['MOBILENET_V1_FINETUNE_PRUNING'][1])
1645
+
1646
+
1647
+ class Resnet50KerasPruningAccuracy(KerasPruningAccuracyBase):
1648
+ """Benchmark accuracy tests for resnet50 with pruning method."""
1649
+
1650
+ def __init__(self, root_data_dir=None, **kwargs):
1651
+ default_flags = {
1652
+ 'model': 'resnet50_v1.5',
1653
+ 'optimizer': 'mobilenet_default',
1654
+ 'initial_learning_rate_per_sample': 0.0000039,
1655
+ 'pretrained_filepath': tf.train.latest_checkpoint(
1656
+ os.path.join(root_data_dir, 'resnet50')),
1657
+ 'pruning_begin_step': 0,
1658
+ 'pruning_end_step': 50000,
1659
+ 'pruning_initial_sparsity': 0.0,
1660
+ 'pruning_final_sparsity': 0.5,
1661
+ 'pruning_frequency': 100,
1662
+ }
1663
+ super(Resnet50KerasPruningAccuracy, self).__init__(
1664
+ root_data_dir=root_data_dir,
1665
+ default_flags=default_flags,
1666
+ **kwargs)
1667
+
1668
+ def _run_and_report_benchmark(self):
1669
+ super(Resnet50KerasPruningAccuracy, self)._run_and_report_benchmark(
1670
+ top_1_min=\
1671
+ MODEL_OPTIMIZATION_TOP_1_ACCURACY['RESNET50_FINETUNE_PRUNING'][0],
1672
+ top_1_max=\
1673
+ MODEL_OPTIMIZATION_TOP_1_ACCURACY['RESNET50_FINETUNE_PRUNING'][1])
1674
+
1675
+
1676
+ class KerasPruningBenchmarkRealBase(Resnet50KerasBenchmarkBase):
1677
+ """Pruning method benchmarks."""
1678
+
1679
+ def __init__(self, root_data_dir=None, default_flags=None, **kwargs):
1680
+ if default_flags is None:
1681
+ default_flags = {}
1682
+ default_flags.update({
1683
+ 'skip_eval': True,
1684
+ 'report_accuracy_metrics': False,
1685
+ 'data_dir': os.path.join(root_data_dir, 'imagenet'),
1686
+ 'train_steps': 110,
1687
+ 'log_steps': 10,
1688
+ 'pruning_method': 'polynomial_decay',
1689
+ 'pruning_begin_step': 0,
1690
+ 'pruning_end_step': 50000,
1691
+ 'pruning_initial_sparsity': 0,
1692
+ 'pruning_final_sparsity': 0.5,
1693
+ 'pruning_frequency': 100,
1694
+ })
1695
+ super(KerasPruningBenchmarkRealBase, self).__init__(
1696
+ default_flags=default_flags, **kwargs)
1697
+
1698
+
1699
+ class MobilenetV1KerasPruningBenchmarkReal(KerasPruningBenchmarkRealBase):
1700
+ """Pruning method benchmarks for MobilenetV1."""
1701
+
1702
+ def __init__(self, **kwargs):
1703
+ default_flags = {
1704
+ 'model': 'mobilenet',
1705
+ 'optimizer': 'mobilenet_default',
1706
+ }
1707
+ super(MobilenetV1KerasPruningBenchmarkReal, self).__init__(
1708
+ default_flags=default_flags, **kwargs)
1709
+
1710
+
1711
+ class Resnet50KerasPruningBenchmarkReal(KerasPruningBenchmarkRealBase):
1712
+ """Pruning method benchmarks for resnet50."""
1713
+
1714
+ def __init__(self, **kwargs):
1715
+ default_flags = {
1716
+ 'model': 'resnet50_v1.5',
1717
+ 'optimizer': 'mobilenet_default',
1718
+ }
1719
+ super(Resnet50KerasPruningBenchmarkReal, self).__init__(
1720
+ default_flags=default_flags, **kwargs)
1721
+
1722
+
1723
+ if __name__ == '__main__':
1724
+ tf.test.main()
models/official/benchmark/models/__init__.py ADDED
File without changes
models/official/benchmark/models/cifar_preprocessing.py ADDED
@@ -0,0 +1,159 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2016 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
+ """Provides utilities to Cifar-10 dataset."""
16
+
17
+ from __future__ import absolute_import
18
+ from __future__ import division
19
+ from __future__ import print_function
20
+
21
+ import os
22
+ from absl import logging
23
+ import tensorflow as tf
24
+
25
+ from official.vision.image_classification.resnet import imagenet_preprocessing
26
+
27
+ HEIGHT = 32
28
+ WIDTH = 32
29
+ NUM_CHANNELS = 3
30
+ _DEFAULT_IMAGE_BYTES = HEIGHT * WIDTH * NUM_CHANNELS
31
+ # The record is the image plus a one-byte label
32
+ _RECORD_BYTES = _DEFAULT_IMAGE_BYTES + 1
33
+
34
+ # TODO(tobyboyd): Change to best practice 45K(train)/5K(val)/10K(test) splits.
35
+ NUM_IMAGES = {
36
+ 'train': 50000,
37
+ 'validation': 10000,
38
+ }
39
+ _NUM_DATA_FILES = 5
40
+ NUM_CLASSES = 10
41
+
42
+
43
+ def parse_record(raw_record, is_training, dtype):
44
+ """Parses a record containing a training example of an image.
45
+
46
+ The input record is parsed into a label and image, and the image is passed
47
+ through preprocessing steps (cropping, flipping, and so on).
48
+
49
+ This method converts the label to one hot to fit the loss function.
50
+
51
+ Args:
52
+ raw_record: scalar Tensor tf.string containing a serialized
53
+ Example protocol buffer.
54
+ is_training: A boolean denoting whether the input is for training.
55
+ dtype: Data type to use for input images.
56
+
57
+ Returns:
58
+ Tuple with processed image tensor and one-hot-encoded label tensor.
59
+ """
60
+ # Convert bytes to a vector of uint8 that is record_bytes long.
61
+ record_vector = tf.io.decode_raw(raw_record, tf.uint8)
62
+
63
+ # The first byte represents the label, which we convert from uint8 to int32
64
+ # and then to one-hot.
65
+ label = tf.cast(record_vector[0], tf.int32)
66
+
67
+ # The remaining bytes after the label represent the image, which we reshape
68
+ # from [depth * height * width] to [depth, height, width].
69
+ depth_major = tf.reshape(record_vector[1:_RECORD_BYTES],
70
+ [NUM_CHANNELS, HEIGHT, WIDTH])
71
+
72
+ # Convert from [depth, height, width] to [height, width, depth], and cast as
73
+ # float32.
74
+ image = tf.cast(tf.transpose(a=depth_major, perm=[1, 2, 0]), tf.float32)
75
+
76
+ image = preprocess_image(image, is_training)
77
+ image = tf.cast(image, dtype)
78
+
79
+ return image, label
80
+
81
+
82
+ def preprocess_image(image, is_training):
83
+ """Preprocess a single image of layout [height, width, depth]."""
84
+ if is_training:
85
+ # Resize the image to add four extra pixels on each side.
86
+ image = tf.image.resize_with_crop_or_pad(
87
+ image, HEIGHT + 8, WIDTH + 8)
88
+
89
+ # Randomly crop a [HEIGHT, WIDTH] section of the image.
90
+ image = tf.image.random_crop(image, [HEIGHT, WIDTH, NUM_CHANNELS])
91
+
92
+ # Randomly flip the image horizontally.
93
+ image = tf.image.random_flip_left_right(image)
94
+
95
+ # Subtract off the mean and divide by the variance of the pixels.
96
+ image = tf.image.per_image_standardization(image)
97
+ return image
98
+
99
+
100
+ def get_filenames(is_training, data_dir):
101
+ """Returns a list of filenames."""
102
+ assert tf.io.gfile.exists(data_dir), (
103
+ 'Run cifar10_download_and_extract.py first to download and extract the '
104
+ 'CIFAR-10 data.')
105
+
106
+ if is_training:
107
+ return [
108
+ os.path.join(data_dir, 'data_batch_%d.bin' % i)
109
+ for i in range(1, _NUM_DATA_FILES + 1)
110
+ ]
111
+ else:
112
+ return [os.path.join(data_dir, 'test_batch.bin')]
113
+
114
+
115
+ def input_fn(is_training,
116
+ data_dir,
117
+ batch_size,
118
+ dtype=tf.float32,
119
+ datasets_num_private_threads=None,
120
+ parse_record_fn=parse_record,
121
+ input_context=None,
122
+ drop_remainder=False):
123
+ """Input function which provides batches for train or eval.
124
+
125
+ Args:
126
+ is_training: A boolean denoting whether the input is for training.
127
+ data_dir: The directory containing the input data.
128
+ batch_size: The number of samples per batch.
129
+ dtype: Data type to use for images/features
130
+ datasets_num_private_threads: Number of private threads for tf.data.
131
+ parse_record_fn: Function to use for parsing the records.
132
+ input_context: A `tf.distribute.InputContext` object passed in by
133
+ `tf.distribute.Strategy`.
134
+ drop_remainder: A boolean indicates whether to drop the remainder of the
135
+ batches. If True, the batch dimension will be static.
136
+
137
+ Returns:
138
+ A dataset that can be used for iteration.
139
+ """
140
+ filenames = get_filenames(is_training, data_dir)
141
+ dataset = tf.data.FixedLengthRecordDataset(filenames, _RECORD_BYTES)
142
+
143
+ if input_context:
144
+ logging.info(
145
+ 'Sharding the dataset: input_pipeline_id=%d num_input_pipelines=%d',
146
+ input_context.input_pipeline_id, input_context.num_input_pipelines)
147
+ dataset = dataset.shard(input_context.num_input_pipelines,
148
+ input_context.input_pipeline_id)
149
+
150
+ return imagenet_preprocessing.process_record_dataset(
151
+ dataset=dataset,
152
+ is_training=is_training,
153
+ batch_size=batch_size,
154
+ shuffle_buffer=NUM_IMAGES['train'],
155
+ parse_record_fn=parse_record_fn,
156
+ dtype=dtype,
157
+ datasets_num_private_threads=datasets_num_private_threads,
158
+ drop_remainder=drop_remainder
159
+ )
models/official/benchmark/models/resnet_cifar_main.py ADDED
@@ -0,0 +1,284 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2018 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
+ """Runs a ResNet model on the Cifar-10 dataset."""
16
+
17
+ from __future__ import absolute_import
18
+ from __future__ import division
19
+ from __future__ import print_function
20
+
21
+ from absl import app
22
+ from absl import flags
23
+ from absl import logging
24
+ import numpy as np
25
+ import tensorflow as tf
26
+ from official.benchmark.models import cifar_preprocessing
27
+ from official.benchmark.models import resnet_cifar_model
28
+ from official.benchmark.models import synthetic_util
29
+ from official.utils.flags import core as flags_core
30
+ from official.utils.misc import distribution_utils
31
+ from official.utils.misc import keras_utils
32
+ from official.vision.image_classification.resnet import common
33
+
34
+
35
+ LR_SCHEDULE = [ # (multiplier, epoch to start) tuples
36
+ (0.1, 91), (0.01, 136), (0.001, 182)
37
+ ]
38
+
39
+
40
+ def learning_rate_schedule(current_epoch,
41
+ current_batch,
42
+ batches_per_epoch,
43
+ batch_size):
44
+ """Handles linear scaling rule and LR decay.
45
+
46
+ Scale learning rate at epoch boundaries provided in LR_SCHEDULE by the
47
+ provided scaling factor.
48
+
49
+ Args:
50
+ current_epoch: integer, current epoch indexed from 0.
51
+ current_batch: integer, current batch in the current epoch, indexed from 0.
52
+ batches_per_epoch: integer, number of steps in an epoch.
53
+ batch_size: integer, total batch sized.
54
+
55
+ Returns:
56
+ Adjusted learning rate.
57
+ """
58
+ del current_batch, batches_per_epoch # not used
59
+ initial_learning_rate = common.BASE_LEARNING_RATE * batch_size / 128
60
+ learning_rate = initial_learning_rate
61
+ for mult, start_epoch in LR_SCHEDULE:
62
+ if current_epoch >= start_epoch:
63
+ learning_rate = initial_learning_rate * mult
64
+ else:
65
+ break
66
+ return learning_rate
67
+
68
+
69
+ class LearningRateBatchScheduler(tf.keras.callbacks.Callback):
70
+ """Callback to update learning rate on every batch (not epoch boundaries).
71
+
72
+ N.B. Only support Keras optimizers, not TF optimizers.
73
+
74
+ Attributes:
75
+ schedule: a function that takes an epoch index and a batch index as input
76
+ (both integer, indexed from 0) and returns a new learning rate as
77
+ output (float).
78
+ """
79
+
80
+ def __init__(self, schedule, batch_size, steps_per_epoch):
81
+ super(LearningRateBatchScheduler, self).__init__()
82
+ self.schedule = schedule
83
+ self.steps_per_epoch = steps_per_epoch
84
+ self.batch_size = batch_size
85
+ self.epochs = -1
86
+ self.prev_lr = -1
87
+
88
+ def on_epoch_begin(self, epoch, logs=None):
89
+ if not hasattr(self.model.optimizer, 'learning_rate'):
90
+ raise ValueError('Optimizer must have a "learning_rate" attribute.')
91
+ self.epochs += 1
92
+
93
+ def on_batch_begin(self, batch, logs=None):
94
+ """Executes before step begins."""
95
+ lr = self.schedule(self.epochs,
96
+ batch,
97
+ self.steps_per_epoch,
98
+ self.batch_size)
99
+ if not isinstance(lr, (float, np.float32, np.float64)):
100
+ raise ValueError('The output of the "schedule" function should be float.')
101
+ if lr != self.prev_lr:
102
+ self.model.optimizer.learning_rate = lr # lr should be a float here
103
+ self.prev_lr = lr
104
+ logging.debug(
105
+ 'Epoch %05d Batch %05d: LearningRateBatchScheduler '
106
+ 'change learning rate to %s.', self.epochs, batch, lr)
107
+
108
+
109
+ def run(flags_obj):
110
+ """Run ResNet Cifar-10 training and eval loop using native Keras APIs.
111
+
112
+ Args:
113
+ flags_obj: An object containing parsed flag values.
114
+
115
+ Raises:
116
+ ValueError: If fp16 is passed as it is not currently supported.
117
+
118
+ Returns:
119
+ Dictionary of training and eval stats.
120
+ """
121
+ keras_utils.set_session_config(
122
+ enable_xla=flags_obj.enable_xla)
123
+
124
+ # Execute flag override logic for better model performance
125
+ if flags_obj.tf_gpu_thread_mode:
126
+ keras_utils.set_gpu_thread_mode_and_count(
127
+ per_gpu_thread_count=flags_obj.per_gpu_thread_count,
128
+ gpu_thread_mode=flags_obj.tf_gpu_thread_mode,
129
+ num_gpus=flags_obj.num_gpus,
130
+ datasets_num_private_threads=flags_obj.datasets_num_private_threads)
131
+ common.set_cudnn_batchnorm_mode()
132
+
133
+ dtype = flags_core.get_tf_dtype(flags_obj)
134
+ if dtype == 'fp16':
135
+ raise ValueError('dtype fp16 is not supported in Keras. Use the default '
136
+ 'value(fp32).')
137
+
138
+ data_format = flags_obj.data_format
139
+ if data_format is None:
140
+ data_format = ('channels_first' if tf.config.list_physical_devices('GPU')
141
+ else 'channels_last')
142
+ tf.keras.backend.set_image_data_format(data_format)
143
+
144
+ strategy = distribution_utils.get_distribution_strategy(
145
+ distribution_strategy=flags_obj.distribution_strategy,
146
+ num_gpus=flags_obj.num_gpus,
147
+ all_reduce_alg=flags_obj.all_reduce_alg,
148
+ num_packs=flags_obj.num_packs)
149
+
150
+ if strategy:
151
+ # flags_obj.enable_get_next_as_optional controls whether enabling
152
+ # get_next_as_optional behavior in DistributedIterator. If true, last
153
+ # partial batch can be supported.
154
+ strategy.extended.experimental_enable_get_next_as_optional = (
155
+ flags_obj.enable_get_next_as_optional
156
+ )
157
+
158
+ strategy_scope = distribution_utils.get_strategy_scope(strategy)
159
+
160
+ if flags_obj.use_synthetic_data:
161
+ synthetic_util.set_up_synthetic_data()
162
+ input_fn = common.get_synth_input_fn(
163
+ height=cifar_preprocessing.HEIGHT,
164
+ width=cifar_preprocessing.WIDTH,
165
+ num_channels=cifar_preprocessing.NUM_CHANNELS,
166
+ num_classes=cifar_preprocessing.NUM_CLASSES,
167
+ dtype=flags_core.get_tf_dtype(flags_obj),
168
+ drop_remainder=True)
169
+ else:
170
+ synthetic_util.undo_set_up_synthetic_data()
171
+ input_fn = cifar_preprocessing.input_fn
172
+
173
+ train_input_dataset = input_fn(
174
+ is_training=True,
175
+ data_dir=flags_obj.data_dir,
176
+ batch_size=flags_obj.batch_size,
177
+ parse_record_fn=cifar_preprocessing.parse_record,
178
+ datasets_num_private_threads=flags_obj.datasets_num_private_threads,
179
+ dtype=dtype,
180
+ # Setting drop_remainder to avoid the partial batch logic in normalization
181
+ # layer, which triggers tf.where and leads to extra memory copy of input
182
+ # sizes between host and GPU.
183
+ drop_remainder=(not flags_obj.enable_get_next_as_optional))
184
+
185
+ eval_input_dataset = None
186
+ if not flags_obj.skip_eval:
187
+ eval_input_dataset = input_fn(
188
+ is_training=False,
189
+ data_dir=flags_obj.data_dir,
190
+ batch_size=flags_obj.batch_size,
191
+ parse_record_fn=cifar_preprocessing.parse_record)
192
+
193
+ steps_per_epoch = (
194
+ cifar_preprocessing.NUM_IMAGES['train'] // flags_obj.batch_size)
195
+ lr_schedule = 0.1
196
+ if flags_obj.use_tensor_lr:
197
+ initial_learning_rate = common.BASE_LEARNING_RATE * flags_obj.batch_size / 128
198
+ lr_schedule = tf.keras.optimizers.schedules.PiecewiseConstantDecay(
199
+ boundaries=list(p[1] * steps_per_epoch for p in LR_SCHEDULE),
200
+ values=[initial_learning_rate] +
201
+ list(p[0] * initial_learning_rate for p in LR_SCHEDULE))
202
+
203
+ with strategy_scope:
204
+ optimizer = common.get_optimizer(lr_schedule)
205
+ model = resnet_cifar_model.resnet56(classes=cifar_preprocessing.NUM_CLASSES)
206
+ model.compile(
207
+ loss='sparse_categorical_crossentropy',
208
+ optimizer=optimizer,
209
+ metrics=(['sparse_categorical_accuracy']
210
+ if flags_obj.report_accuracy_metrics else None),
211
+ run_eagerly=flags_obj.run_eagerly)
212
+
213
+ train_epochs = flags_obj.train_epochs
214
+
215
+ callbacks = common.get_callbacks()
216
+
217
+ if not flags_obj.use_tensor_lr:
218
+ lr_callback = LearningRateBatchScheduler(
219
+ schedule=learning_rate_schedule,
220
+ batch_size=flags_obj.batch_size,
221
+ steps_per_epoch=steps_per_epoch)
222
+ callbacks.append(lr_callback)
223
+
224
+ # if mutliple epochs, ignore the train_steps flag.
225
+ if train_epochs <= 1 and flags_obj.train_steps:
226
+ steps_per_epoch = min(flags_obj.train_steps, steps_per_epoch)
227
+ train_epochs = 1
228
+
229
+ num_eval_steps = (cifar_preprocessing.NUM_IMAGES['validation'] //
230
+ flags_obj.batch_size)
231
+
232
+ validation_data = eval_input_dataset
233
+ if flags_obj.skip_eval:
234
+ if flags_obj.set_learning_phase_to_train:
235
+ # TODO(haoyuzhang): Understand slowdown of setting learning phase when
236
+ # not using distribution strategy.
237
+ tf.keras.backend.set_learning_phase(1)
238
+ num_eval_steps = None
239
+ validation_data = None
240
+
241
+ if not strategy and flags_obj.explicit_gpu_placement:
242
+ # TODO(b/135607227): Add device scope automatically in Keras training loop
243
+ # when not using distribition strategy.
244
+ no_dist_strat_device = tf.device('/device:GPU:0')
245
+ no_dist_strat_device.__enter__()
246
+
247
+ history = model.fit(train_input_dataset,
248
+ epochs=train_epochs,
249
+ steps_per_epoch=steps_per_epoch,
250
+ callbacks=callbacks,
251
+ validation_steps=num_eval_steps,
252
+ validation_data=validation_data,
253
+ validation_freq=flags_obj.epochs_between_evals,
254
+ verbose=2)
255
+ eval_output = None
256
+ if not flags_obj.skip_eval:
257
+ eval_output = model.evaluate(eval_input_dataset,
258
+ steps=num_eval_steps,
259
+ verbose=2)
260
+
261
+ if not strategy and flags_obj.explicit_gpu_placement:
262
+ no_dist_strat_device.__exit__()
263
+
264
+ stats = common.build_stats(history, eval_output, callbacks)
265
+ return stats
266
+
267
+
268
+ def define_cifar_flags():
269
+ common.define_keras_flags(dynamic_loss_scale=False)
270
+
271
+ flags_core.set_defaults(data_dir='/tmp/cifar10_data/cifar-10-batches-bin',
272
+ model_dir='/tmp/cifar10_model',
273
+ epochs_between_evals=10,
274
+ batch_size=128)
275
+
276
+
277
+ def main(_):
278
+ return run(flags.FLAGS)
279
+
280
+
281
+ if __name__ == '__main__':
282
+ logging.set_verbosity(logging.INFO)
283
+ define_cifar_flags()
284
+ app.run(main)
models/official/benchmark/models/resnet_cifar_model.py ADDED
@@ -0,0 +1,262 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2018 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
+ """ResNet56 model for Keras adapted from tf.keras.applications.ResNet50.
16
+
17
+ # Reference:
18
+ - [Deep Residual Learning for Image Recognition](
19
+ https://arxiv.org/abs/1512.03385)
20
+ Adapted from code contributed by BigMoyan.
21
+ """
22
+ from __future__ import absolute_import
23
+ from __future__ import division
24
+ from __future__ import print_function
25
+
26
+ import functools
27
+ import tensorflow as tf
28
+ from tensorflow.python.keras import backend
29
+ from tensorflow.python.keras import initializers
30
+ from tensorflow.python.keras import layers
31
+ from tensorflow.python.keras import regularizers
32
+
33
+
34
+ BATCH_NORM_DECAY = 0.997
35
+ BATCH_NORM_EPSILON = 1e-5
36
+ L2_WEIGHT_DECAY = 2e-4
37
+
38
+
39
+ def identity_building_block(input_tensor,
40
+ kernel_size,
41
+ filters,
42
+ stage,
43
+ block,
44
+ training=None):
45
+ """The identity block is the block that has no conv layer at shortcut.
46
+
47
+ Arguments:
48
+ input_tensor: input tensor
49
+ kernel_size: default 3, the kernel size of
50
+ middle conv layer at main path
51
+ filters: list of integers, the filters of 3 conv layer at main path
52
+ stage: integer, current stage label, used for generating layer names
53
+ block: current block label, used for generating layer names
54
+ training: Only used if training keras model with Estimator. In other
55
+ scenarios it is handled automatically.
56
+
57
+ Returns:
58
+ Output tensor for the block.
59
+ """
60
+ filters1, filters2 = filters
61
+ if backend.image_data_format() == 'channels_last':
62
+ bn_axis = 3
63
+ else:
64
+ bn_axis = 1
65
+ conv_name_base = 'res' + str(stage) + block + '_branch'
66
+ bn_name_base = 'bn' + str(stage) + block + '_branch'
67
+
68
+ x = layers.Conv2D(filters1, kernel_size,
69
+ padding='same', use_bias=False,
70
+ kernel_initializer='he_normal',
71
+ kernel_regularizer=regularizers.l2(L2_WEIGHT_DECAY),
72
+ name=conv_name_base + '2a')(input_tensor)
73
+ x = layers.BatchNormalization(
74
+ axis=bn_axis, momentum=BATCH_NORM_DECAY, epsilon=BATCH_NORM_EPSILON,
75
+ name=bn_name_base + '2a')(x, training=training)
76
+ x = layers.Activation('relu')(x)
77
+
78
+ x = layers.Conv2D(filters2, kernel_size,
79
+ padding='same', use_bias=False,
80
+ kernel_initializer='he_normal',
81
+ kernel_regularizer=regularizers.l2(L2_WEIGHT_DECAY),
82
+ name=conv_name_base + '2b')(x)
83
+ x = layers.BatchNormalization(
84
+ axis=bn_axis, momentum=BATCH_NORM_DECAY, epsilon=BATCH_NORM_EPSILON,
85
+ name=bn_name_base + '2b')(x, training=training)
86
+
87
+ x = layers.add([x, input_tensor])
88
+ x = layers.Activation('relu')(x)
89
+ return x
90
+
91
+
92
+ def conv_building_block(input_tensor,
93
+ kernel_size,
94
+ filters,
95
+ stage,
96
+ block,
97
+ strides=(2, 2),
98
+ training=None):
99
+ """A block that has a conv layer at shortcut.
100
+
101
+ Arguments:
102
+ input_tensor: input tensor
103
+ kernel_size: default 3, the kernel size of
104
+ middle conv layer at main path
105
+ filters: list of integers, the filters of 3 conv layer at main path
106
+ stage: integer, current stage label, used for generating layer names
107
+ block: current block label, used for generating layer names
108
+ strides: Strides for the first conv layer in the block.
109
+ training: Only used if training keras model with Estimator. In other
110
+ scenarios it is handled automatically.
111
+
112
+ Returns:
113
+ Output tensor for the block.
114
+
115
+ Note that from stage 3,
116
+ the first conv layer at main path is with strides=(2, 2)
117
+ And the shortcut should have strides=(2, 2) as well
118
+ """
119
+ filters1, filters2 = filters
120
+ if tf.keras.backend.image_data_format() == 'channels_last':
121
+ bn_axis = 3
122
+ else:
123
+ bn_axis = 1
124
+ conv_name_base = 'res' + str(stage) + block + '_branch'
125
+ bn_name_base = 'bn' + str(stage) + block + '_branch'
126
+
127
+ x = layers.Conv2D(filters1, kernel_size, strides=strides,
128
+ padding='same', use_bias=False,
129
+ kernel_initializer='he_normal',
130
+ kernel_regularizer=regularizers.l2(L2_WEIGHT_DECAY),
131
+ name=conv_name_base + '2a')(input_tensor)
132
+ x = layers.BatchNormalization(
133
+ axis=bn_axis, momentum=BATCH_NORM_DECAY, epsilon=BATCH_NORM_EPSILON,
134
+ name=bn_name_base + '2a')(x, training=training)
135
+ x = layers.Activation('relu')(x)
136
+
137
+ x = layers.Conv2D(filters2, kernel_size, padding='same', use_bias=False,
138
+ kernel_initializer='he_normal',
139
+ kernel_regularizer=regularizers.l2(L2_WEIGHT_DECAY),
140
+ name=conv_name_base + '2b')(x)
141
+ x = layers.BatchNormalization(
142
+ axis=bn_axis, momentum=BATCH_NORM_DECAY, epsilon=BATCH_NORM_EPSILON,
143
+ name=bn_name_base + '2b')(x, training=training)
144
+
145
+ shortcut = layers.Conv2D(filters2, (1, 1), strides=strides, use_bias=False,
146
+ kernel_initializer='he_normal',
147
+ kernel_regularizer=regularizers.l2(L2_WEIGHT_DECAY),
148
+ name=conv_name_base + '1')(input_tensor)
149
+ shortcut = layers.BatchNormalization(
150
+ axis=bn_axis, momentum=BATCH_NORM_DECAY, epsilon=BATCH_NORM_EPSILON,
151
+ name=bn_name_base + '1')(shortcut, training=training)
152
+
153
+ x = layers.add([x, shortcut])
154
+ x = layers.Activation('relu')(x)
155
+ return x
156
+
157
+
158
+ def resnet_block(input_tensor,
159
+ size,
160
+ kernel_size,
161
+ filters,
162
+ stage,
163
+ conv_strides=(2, 2),
164
+ training=None):
165
+ """A block which applies conv followed by multiple identity blocks.
166
+
167
+ Arguments:
168
+ input_tensor: input tensor
169
+ size: integer, number of constituent conv/identity building blocks.
170
+ A conv block is applied once, followed by (size - 1) identity blocks.
171
+ kernel_size: default 3, the kernel size of
172
+ middle conv layer at main path
173
+ filters: list of integers, the filters of 3 conv layer at main path
174
+ stage: integer, current stage label, used for generating layer names
175
+ conv_strides: Strides for the first conv layer in the block.
176
+ training: Only used if training keras model with Estimator. In other
177
+ scenarios it is handled automatically.
178
+
179
+ Returns:
180
+ Output tensor after applying conv and identity blocks.
181
+ """
182
+
183
+ x = conv_building_block(input_tensor, kernel_size, filters, stage=stage,
184
+ strides=conv_strides, block='block_0',
185
+ training=training)
186
+ for i in range(size - 1):
187
+ x = identity_building_block(x, kernel_size, filters, stage=stage,
188
+ block='block_%d' % (i + 1), training=training)
189
+ return x
190
+
191
+
192
+ def resnet(num_blocks, classes=10, training=None):
193
+ """Instantiates the ResNet architecture.
194
+
195
+ Arguments:
196
+ num_blocks: integer, the number of conv/identity blocks in each block.
197
+ The ResNet contains 3 blocks with each block containing one conv block
198
+ followed by (layers_per_block - 1) number of idenity blocks. Each
199
+ conv/idenity block has 2 convolutional layers. With the input
200
+ convolutional layer and the pooling layer towards the end, this brings
201
+ the total size of the network to (6*num_blocks + 2)
202
+ classes: optional number of classes to classify images into
203
+ training: Only used if training keras model with Estimator. In other
204
+ scenarios it is handled automatically.
205
+
206
+ Returns:
207
+ A Keras model instance.
208
+ """
209
+
210
+ input_shape = (32, 32, 3)
211
+ img_input = layers.Input(shape=input_shape)
212
+
213
+ if backend.image_data_format() == 'channels_first':
214
+ x = layers.Lambda(lambda x: backend.permute_dimensions(x, (0, 3, 1, 2)),
215
+ name='transpose')(img_input)
216
+ bn_axis = 1
217
+ else: # channel_last
218
+ x = img_input
219
+ bn_axis = 3
220
+
221
+ x = layers.ZeroPadding2D(padding=(1, 1), name='conv1_pad')(x)
222
+ x = layers.Conv2D(16, (3, 3),
223
+ strides=(1, 1),
224
+ padding='valid', use_bias=False,
225
+ kernel_initializer='he_normal',
226
+ kernel_regularizer=regularizers.l2(L2_WEIGHT_DECAY),
227
+ name='conv1')(x)
228
+ x = layers.BatchNormalization(axis=bn_axis,
229
+ momentum=BATCH_NORM_DECAY,
230
+ epsilon=BATCH_NORM_EPSILON,
231
+ name='bn_conv1',)(x, training=training)
232
+ x = layers.Activation('relu')(x)
233
+
234
+ x = resnet_block(x, size=num_blocks, kernel_size=3, filters=[16, 16],
235
+ stage=2, conv_strides=(1, 1), training=training)
236
+
237
+ x = resnet_block(x, size=num_blocks, kernel_size=3, filters=[32, 32],
238
+ stage=3, conv_strides=(2, 2), training=training)
239
+
240
+ x = resnet_block(x, size=num_blocks, kernel_size=3, filters=[64, 64],
241
+ stage=4, conv_strides=(2, 2), training=training)
242
+
243
+ rm_axes = [1, 2] if backend.image_data_format() == 'channels_last' else [2, 3]
244
+ x = layers.Lambda(lambda x: backend.mean(x, rm_axes), name='reduce_mean')(x)
245
+ x = layers.Dense(classes,
246
+ activation='softmax',
247
+ kernel_initializer=initializers.RandomNormal(stddev=0.01),
248
+ kernel_regularizer=regularizers.l2(L2_WEIGHT_DECAY),
249
+ bias_regularizer=regularizers.l2(L2_WEIGHT_DECAY),
250
+ name='fc10')(x)
251
+
252
+ inputs = img_input
253
+ # Create model.
254
+ model = tf.keras.models.Model(inputs, x, name='resnet56')
255
+
256
+ return model
257
+
258
+
259
+ resnet20 = functools.partial(resnet, num_blocks=3)
260
+ resnet32 = functools.partial(resnet, num_blocks=5)
261
+ resnet56 = functools.partial(resnet, num_blocks=9)
262
+ resnet10 = functools.partial(resnet, num_blocks=110)
models/official/benchmark/models/resnet_cifar_test.py ADDED
@@ -0,0 +1,180 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2018 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
+ """Test the keras ResNet model with Cifar data."""
16
+
17
+ from __future__ import absolute_import
18
+ from __future__ import division
19
+ from __future__ import print_function
20
+
21
+ import tempfile
22
+
23
+ import tensorflow as tf
24
+
25
+ from tensorflow.python.eager import context
26
+ from tensorflow.python.platform import googletest
27
+ from official.benchmark.models import cifar_preprocessing
28
+ from official.benchmark.models import resnet_cifar_main
29
+ from official.utils.testing import integration
30
+
31
+
32
+ class KerasCifarTest(googletest.TestCase):
33
+ """Unit tests for Keras ResNet with Cifar."""
34
+
35
+ _extra_flags = [
36
+ "-batch_size", "4",
37
+ "-train_steps", "1",
38
+ "-use_synthetic_data", "true"
39
+ ]
40
+ _tempdir = None
41
+
42
+ def get_temp_dir(self):
43
+ if not self._tempdir:
44
+ self._tempdir = tempfile.mkdtemp(dir=googletest.GetTempDir())
45
+ return self._tempdir
46
+
47
+ @classmethod
48
+ def setUpClass(cls): # pylint: disable=invalid-name
49
+ super(KerasCifarTest, cls).setUpClass()
50
+ resnet_cifar_main.define_cifar_flags()
51
+
52
+ def setUp(self):
53
+ super(KerasCifarTest, self).setUp()
54
+ cifar_preprocessing.NUM_IMAGES["validation"] = 4
55
+
56
+ def tearDown(self):
57
+ super(KerasCifarTest, self).tearDown()
58
+ tf.io.gfile.rmtree(self.get_temp_dir())
59
+
60
+ def test_end_to_end_no_dist_strat(self):
61
+ """Test Keras model with 1 GPU, no distribution strategy."""
62
+
63
+ extra_flags = [
64
+ "-distribution_strategy", "off",
65
+ "-model_dir", "keras_cifar_no_dist_strat",
66
+ "-data_format", "channels_last",
67
+ ]
68
+ extra_flags = extra_flags + self._extra_flags
69
+
70
+ integration.run_synthetic(
71
+ main=resnet_cifar_main.run,
72
+ tmp_root=self.get_temp_dir(),
73
+ extra_flags=extra_flags
74
+ )
75
+
76
+ def test_end_to_end_graph_no_dist_strat(self):
77
+ """Test Keras model in legacy graph mode with 1 GPU, no dist strat."""
78
+ extra_flags = [
79
+ "-enable_eager", "false",
80
+ "-distribution_strategy", "off",
81
+ "-model_dir", "keras_cifar_graph_no_dist_strat",
82
+ "-data_format", "channels_last",
83
+ ]
84
+ extra_flags = extra_flags + self._extra_flags
85
+
86
+ integration.run_synthetic(
87
+ main=resnet_cifar_main.run,
88
+ tmp_root=self.get_temp_dir(),
89
+ extra_flags=extra_flags
90
+ )
91
+
92
+ def test_end_to_end_1_gpu(self):
93
+ """Test Keras model with 1 GPU."""
94
+
95
+ if context.num_gpus() < 1:
96
+ self.skipTest(
97
+ "{} GPUs are not available for this test. {} GPUs are available".
98
+ format(1, context.num_gpus()))
99
+
100
+ extra_flags = [
101
+ "-num_gpus", "1",
102
+ "-distribution_strategy", "mirrored",
103
+ "-model_dir", "keras_cifar_1_gpu",
104
+ "-data_format", "channels_last",
105
+ ]
106
+ extra_flags = extra_flags + self._extra_flags
107
+
108
+ integration.run_synthetic(
109
+ main=resnet_cifar_main.run,
110
+ tmp_root=self.get_temp_dir(),
111
+ extra_flags=extra_flags
112
+ )
113
+
114
+ def test_end_to_end_graph_1_gpu(self):
115
+ """Test Keras model in legacy graph mode with 1 GPU."""
116
+ if context.num_gpus() < 1:
117
+ self.skipTest(
118
+ "{} GPUs are not available for this test. {} GPUs are available".
119
+ format(1, context.num_gpus()))
120
+
121
+ extra_flags = [
122
+ "-num_gpus", "1",
123
+ "-noenable_eager",
124
+ "-distribution_strategy", "mirrored",
125
+ "-model_dir", "keras_cifar_graph_1_gpu",
126
+ "-data_format", "channels_last",
127
+ ]
128
+ extra_flags = extra_flags + self._extra_flags
129
+
130
+ integration.run_synthetic(
131
+ main=resnet_cifar_main.run,
132
+ tmp_root=self.get_temp_dir(),
133
+ extra_flags=extra_flags
134
+ )
135
+
136
+ def test_end_to_end_2_gpu(self):
137
+ """Test Keras model with 2 GPUs."""
138
+
139
+ if context.num_gpus() < 2:
140
+ self.skipTest(
141
+ "{} GPUs are not available for this test. {} GPUs are available".
142
+ format(2, context.num_gpus()))
143
+
144
+ extra_flags = [
145
+ "-num_gpus", "2",
146
+ "-distribution_strategy", "mirrored",
147
+ "-model_dir", "keras_cifar_2_gpu",
148
+ ]
149
+ extra_flags = extra_flags + self._extra_flags
150
+
151
+ integration.run_synthetic(
152
+ main=resnet_cifar_main.run,
153
+ tmp_root=self.get_temp_dir(),
154
+ extra_flags=extra_flags
155
+ )
156
+
157
+ def test_end_to_end_graph_2_gpu(self):
158
+ """Test Keras model in legacy graph mode with 2 GPUs."""
159
+ if context.num_gpus() < 2:
160
+ self.skipTest(
161
+ "{} GPUs are not available for this test. {} GPUs are available".
162
+ format(2, context.num_gpus()))
163
+
164
+ extra_flags = [
165
+ "-num_gpus", "2",
166
+ "-enable_eager", "false",
167
+ "-distribution_strategy", "mirrored",
168
+ "-model_dir", "keras_cifar_graph_2_gpu",
169
+ ]
170
+ extra_flags = extra_flags + self._extra_flags
171
+
172
+ integration.run_synthetic(
173
+ main=resnet_cifar_main.run,
174
+ tmp_root=self.get_temp_dir(),
175
+ extra_flags=extra_flags
176
+ )
177
+
178
+
179
+ if __name__ == "__main__":
180
+ googletest.main()
models/official/benchmark/models/resnet_imagenet_main.py ADDED
@@ -0,0 +1,301 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2019 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
+ """Runs a ResNet model on the ImageNet dataset."""
16
+
17
+ from __future__ import absolute_import
18
+ from __future__ import division
19
+ from __future__ import print_function
20
+
21
+ import os
22
+
23
+ from absl import app
24
+ from absl import flags
25
+ from absl import logging
26
+ import tensorflow as tf
27
+
28
+ import tensorflow_model_optimization as tfmot
29
+ from official.modeling import performance
30
+ from official.utils.flags import core as flags_core
31
+ from official.utils.misc import distribution_utils
32
+ from official.utils.misc import keras_utils
33
+ from official.utils.misc import model_helpers
34
+ from official.vision.image_classification import test_utils
35
+ from official.vision.image_classification.resnet import common
36
+ from official.vision.image_classification.resnet import imagenet_preprocessing
37
+ from official.vision.image_classification.resnet import resnet_model
38
+
39
+
40
+ def run(flags_obj):
41
+ """Run ResNet ImageNet training and eval loop using native Keras APIs.
42
+
43
+ Args:
44
+ flags_obj: An object containing parsed flag values.
45
+
46
+ Raises:
47
+ ValueError: If fp16 is passed as it is not currently supported.
48
+ NotImplementedError: If some features are not currently supported.
49
+
50
+ Returns:
51
+ Dictionary of training and eval stats.
52
+ """
53
+ keras_utils.set_session_config(
54
+ enable_xla=flags_obj.enable_xla)
55
+
56
+ # Execute flag override logic for better model performance
57
+ if flags_obj.tf_gpu_thread_mode:
58
+ keras_utils.set_gpu_thread_mode_and_count(
59
+ per_gpu_thread_count=flags_obj.per_gpu_thread_count,
60
+ gpu_thread_mode=flags_obj.tf_gpu_thread_mode,
61
+ num_gpus=flags_obj.num_gpus,
62
+ datasets_num_private_threads=flags_obj.datasets_num_private_threads)
63
+ common.set_cudnn_batchnorm_mode()
64
+
65
+ dtype = flags_core.get_tf_dtype(flags_obj)
66
+ performance.set_mixed_precision_policy(
67
+ flags_core.get_tf_dtype(flags_obj),
68
+ flags_core.get_loss_scale(flags_obj, default_for_fp16=128))
69
+
70
+ data_format = flags_obj.data_format
71
+ if data_format is None:
72
+ data_format = ('channels_first' if tf.config.list_physical_devices('GPU')
73
+ else 'channels_last')
74
+ tf.keras.backend.set_image_data_format(data_format)
75
+
76
+ # Configures cluster spec for distribution strategy.
77
+ _ = distribution_utils.configure_cluster(flags_obj.worker_hosts,
78
+ flags_obj.task_index)
79
+
80
+ strategy = distribution_utils.get_distribution_strategy(
81
+ distribution_strategy=flags_obj.distribution_strategy,
82
+ num_gpus=flags_obj.num_gpus,
83
+ all_reduce_alg=flags_obj.all_reduce_alg,
84
+ num_packs=flags_obj.num_packs,
85
+ tpu_address=flags_obj.tpu)
86
+
87
+ if strategy:
88
+ # flags_obj.enable_get_next_as_optional controls whether enabling
89
+ # get_next_as_optional behavior in DistributedIterator. If true, last
90
+ # partial batch can be supported.
91
+ strategy.extended.experimental_enable_get_next_as_optional = (
92
+ flags_obj.enable_get_next_as_optional
93
+ )
94
+
95
+ strategy_scope = distribution_utils.get_strategy_scope(strategy)
96
+
97
+ # pylint: disable=protected-access
98
+ if flags_obj.use_synthetic_data:
99
+ input_fn = common.get_synth_input_fn(
100
+ height=imagenet_preprocessing.DEFAULT_IMAGE_SIZE,
101
+ width=imagenet_preprocessing.DEFAULT_IMAGE_SIZE,
102
+ num_channels=imagenet_preprocessing.NUM_CHANNELS,
103
+ num_classes=imagenet_preprocessing.NUM_CLASSES,
104
+ dtype=dtype,
105
+ drop_remainder=True)
106
+ else:
107
+ input_fn = imagenet_preprocessing.input_fn
108
+
109
+ # When `enable_xla` is True, we always drop the remainder of the batches
110
+ # in the dataset, as XLA-GPU doesn't support dynamic shapes.
111
+ drop_remainder = flags_obj.enable_xla
112
+
113
+ # Current resnet_model.resnet50 input format is always channel-last.
114
+ # We use keras_application mobilenet model which input format is depends on
115
+ # the keras beckend image data format.
116
+ # This use_keras_image_data_format flags indicates whether image preprocessor
117
+ # output format should be same as the keras backend image data format or just
118
+ # channel-last format.
119
+ use_keras_image_data_format = (flags_obj.model == 'mobilenet')
120
+ train_input_dataset = input_fn(
121
+ is_training=True,
122
+ data_dir=flags_obj.data_dir,
123
+ batch_size=flags_obj.batch_size,
124
+ parse_record_fn=imagenet_preprocessing.get_parse_record_fn(
125
+ use_keras_image_data_format=use_keras_image_data_format),
126
+ datasets_num_private_threads=flags_obj.datasets_num_private_threads,
127
+ dtype=dtype,
128
+ drop_remainder=drop_remainder,
129
+ tf_data_experimental_slack=flags_obj.tf_data_experimental_slack,
130
+ training_dataset_cache=flags_obj.training_dataset_cache,
131
+ )
132
+
133
+ eval_input_dataset = None
134
+ if not flags_obj.skip_eval:
135
+ eval_input_dataset = input_fn(
136
+ is_training=False,
137
+ data_dir=flags_obj.data_dir,
138
+ batch_size=flags_obj.batch_size,
139
+ parse_record_fn=imagenet_preprocessing.get_parse_record_fn(
140
+ use_keras_image_data_format=use_keras_image_data_format),
141
+ dtype=dtype,
142
+ drop_remainder=drop_remainder)
143
+
144
+ lr_schedule = common.PiecewiseConstantDecayWithWarmup(
145
+ batch_size=flags_obj.batch_size,
146
+ epoch_size=imagenet_preprocessing.NUM_IMAGES['train'],
147
+ warmup_epochs=common.LR_SCHEDULE[0][1],
148
+ boundaries=list(p[1] for p in common.LR_SCHEDULE[1:]),
149
+ multipliers=list(p[0] for p in common.LR_SCHEDULE),
150
+ compute_lr_on_cpu=True)
151
+ steps_per_epoch = (
152
+ imagenet_preprocessing.NUM_IMAGES['train'] // flags_obj.batch_size)
153
+
154
+ with strategy_scope:
155
+ if flags_obj.optimizer == 'resnet50_default':
156
+ optimizer = common.get_optimizer(lr_schedule)
157
+ elif flags_obj.optimizer == 'mobilenet_default':
158
+ initial_learning_rate = \
159
+ flags_obj.initial_learning_rate_per_sample * flags_obj.batch_size
160
+ optimizer = tf.keras.optimizers.SGD(
161
+ learning_rate=tf.keras.optimizers.schedules.ExponentialDecay(
162
+ initial_learning_rate,
163
+ decay_steps=steps_per_epoch * flags_obj.num_epochs_per_decay,
164
+ decay_rate=flags_obj.lr_decay_factor,
165
+ staircase=True),
166
+ momentum=0.9)
167
+ if flags_obj.fp16_implementation == 'graph_rewrite':
168
+ # Note: when flags_obj.fp16_implementation == "graph_rewrite", dtype as
169
+ # determined by flags_core.get_tf_dtype(flags_obj) would be 'float32'
170
+ # which will ensure tf.compat.v2.keras.mixed_precision and
171
+ # tf.train.experimental.enable_mixed_precision_graph_rewrite do not double
172
+ # up.
173
+ optimizer = tf.train.experimental.enable_mixed_precision_graph_rewrite(
174
+ optimizer)
175
+
176
+ # TODO(hongkuny): Remove trivial model usage and move it to benchmark.
177
+ if flags_obj.use_trivial_model:
178
+ model = test_utils.trivial_model(imagenet_preprocessing.NUM_CLASSES)
179
+ elif flags_obj.model == 'resnet50_v1.5':
180
+ model = resnet_model.resnet50(
181
+ num_classes=imagenet_preprocessing.NUM_CLASSES)
182
+ elif flags_obj.model == 'mobilenet':
183
+ # TODO(kimjaehong): Remove layers attribute when minimum TF version
184
+ # support 2.0 layers by default.
185
+ model = tf.keras.applications.mobilenet.MobileNet(
186
+ weights=None,
187
+ classes=imagenet_preprocessing.NUM_CLASSES,
188
+ layers=tf.keras.layers)
189
+ if flags_obj.pretrained_filepath:
190
+ model.load_weights(flags_obj.pretrained_filepath)
191
+
192
+ if flags_obj.pruning_method == 'polynomial_decay':
193
+ if dtype != tf.float32:
194
+ raise NotImplementedError(
195
+ 'Pruning is currently only supported on dtype=tf.float32.')
196
+ pruning_params = {
197
+ 'pruning_schedule':
198
+ tfmot.sparsity.keras.PolynomialDecay(
199
+ initial_sparsity=flags_obj.pruning_initial_sparsity,
200
+ final_sparsity=flags_obj.pruning_final_sparsity,
201
+ begin_step=flags_obj.pruning_begin_step,
202
+ end_step=flags_obj.pruning_end_step,
203
+ frequency=flags_obj.pruning_frequency),
204
+ }
205
+ model = tfmot.sparsity.keras.prune_low_magnitude(model, **pruning_params)
206
+ elif flags_obj.pruning_method:
207
+ raise NotImplementedError(
208
+ 'Only polynomial_decay is currently supported.')
209
+
210
+ model.compile(
211
+ loss='sparse_categorical_crossentropy',
212
+ optimizer=optimizer,
213
+ metrics=(['sparse_categorical_accuracy']
214
+ if flags_obj.report_accuracy_metrics else None),
215
+ run_eagerly=flags_obj.run_eagerly)
216
+
217
+ train_epochs = flags_obj.train_epochs
218
+
219
+ callbacks = common.get_callbacks(
220
+ pruning_method=flags_obj.pruning_method,
221
+ enable_checkpoint_and_export=flags_obj.enable_checkpoint_and_export,
222
+ model_dir=flags_obj.model_dir)
223
+
224
+ # if mutliple epochs, ignore the train_steps flag.
225
+ if train_epochs <= 1 and flags_obj.train_steps:
226
+ steps_per_epoch = min(flags_obj.train_steps, steps_per_epoch)
227
+ train_epochs = 1
228
+
229
+ num_eval_steps = (
230
+ imagenet_preprocessing.NUM_IMAGES['validation'] // flags_obj.batch_size)
231
+
232
+ validation_data = eval_input_dataset
233
+ if flags_obj.skip_eval:
234
+ # Only build the training graph. This reduces memory usage introduced by
235
+ # control flow ops in layers that have different implementations for
236
+ # training and inference (e.g., batch norm).
237
+ if flags_obj.set_learning_phase_to_train:
238
+ # TODO(haoyuzhang): Understand slowdown of setting learning phase when
239
+ # not using distribution strategy.
240
+ tf.keras.backend.set_learning_phase(1)
241
+ num_eval_steps = None
242
+ validation_data = None
243
+
244
+ if not strategy and flags_obj.explicit_gpu_placement:
245
+ # TODO(b/135607227): Add device scope automatically in Keras training loop
246
+ # when not using distribition strategy.
247
+ no_dist_strat_device = tf.device('/device:GPU:0')
248
+ no_dist_strat_device.__enter__()
249
+
250
+ history = model.fit(train_input_dataset,
251
+ epochs=train_epochs,
252
+ steps_per_epoch=steps_per_epoch,
253
+ callbacks=callbacks,
254
+ validation_steps=num_eval_steps,
255
+ validation_data=validation_data,
256
+ validation_freq=flags_obj.epochs_between_evals,
257
+ verbose=2)
258
+
259
+ eval_output = None
260
+ if not flags_obj.skip_eval:
261
+ eval_output = model.evaluate(eval_input_dataset,
262
+ steps=num_eval_steps,
263
+ verbose=2)
264
+
265
+ if flags_obj.pruning_method:
266
+ model = tfmot.sparsity.keras.strip_pruning(model)
267
+ if flags_obj.enable_checkpoint_and_export:
268
+ if dtype == tf.bfloat16:
269
+ logging.warning('Keras model.save does not support bfloat16 dtype.')
270
+ else:
271
+ # Keras model.save assumes a float32 input designature.
272
+ export_path = os.path.join(flags_obj.model_dir, 'saved_model')
273
+ model.save(export_path, include_optimizer=False)
274
+
275
+ if not strategy and flags_obj.explicit_gpu_placement:
276
+ no_dist_strat_device.__exit__()
277
+
278
+ stats = common.build_stats(history, eval_output, callbacks)
279
+ return stats
280
+
281
+
282
+ def define_imagenet_keras_flags():
283
+ common.define_keras_flags(
284
+ model=True,
285
+ optimizer=True,
286
+ pretrained_filepath=True)
287
+ common.define_pruning_flags()
288
+ flags_core.set_defaults()
289
+ flags.adopt_module_key_flags(common)
290
+
291
+
292
+ def main(_):
293
+ model_helpers.apply_clean(flags.FLAGS)
294
+ stats = run(flags.FLAGS)
295
+ logging.info('Run stats:\n%s', stats)
296
+
297
+
298
+ if __name__ == '__main__':
299
+ logging.set_verbosity(logging.INFO)
300
+ define_imagenet_keras_flags()
301
+ app.run(main)