-
Notifications
You must be signed in to change notification settings - Fork 54
/
Copy pathgradio_hpo.py
207 lines (178 loc) · 10 KB
/
gradio_hpo.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
import gradio as gr
from train_sl import TrainClassifier
from cords.utils.config_utils import load_config_data
from cords.utils.data.datasets.SL import gen_dataset
import os, sys
from dotmap import DotMap
import pandas as pd
from matplotlib import pyplot as plt
from fastapi import FastAPI
# CUSTOM_PATH = "/gradio_hpo"
# app = FastAPI()
def return_search_space(dset):
if dset == 'TREC6':
return {
'num_layers': [1],
'hidden_size': [64, 128],
'learning_rate': [0.001, 0.01, 0.1]
}
def return_full_results():
search_space = {
'num_layers': [1, 2],
'hidden_size': [64, 128, 256],
'learning_rate': [0.001, 0.01, 0.1]
}
search_configs = []
key1, key2, key3 = list(search_space.keys())[0], list(search_space.keys())[1], list(search_space.keys())[2]
for i in range(len(search_space[key1])):
for j in range(len(search_space[key2])):
for k in range(len(search_space[key3])):
search_configs.append({key1: search_space[key1][i], key2: search_space[key2][j], key3: search_space[key3][k]})
full_results = [[0.866, 287.729421377182],
[0.874, 287.24425220489502], [0.276, 287.74825239181519], [0.868, 287.75174808502197], [0.858, 287.92841958999634], [0.374, 287.94037413597107], [0.856, 287.2798364162445], [0.874, 287.7269721031189], [0.276, 287.57475852966309], [0.878, 287.33644080162048], [0.864, 287.51423025131226], [0.278, 287.54987382888794], [0.86, 287.80005693435669], [0.888, 287.19879388809204], [0.286, 287.0971245765686], [0.886, 287.03492903709412], [0.87, 287.60258483886719], [0.412, 287.89436173439026]]
full_results_dict = {}
for i in range(len(full_results)):
full_results_dict[tuple(sorted(search_configs[i].items()))] = full_results[i]
return full_results_dict
def update_model_text(choice):
if choice == "MNIST":
return gr.update(choices=['LeNet'], value="LeNet"), return_search_space(choice)
elif choice == "TREC6":
return gr.update(choices=['LSTM'], value="LSTM"), return_search_space(choice)
elif choice == 'CIFAR10':
return gr.update(choices=['ResNet18'], value="ResNet18"), return_search_space(choice)
def update_out_text(choice):
# print(choice)
return gr.update(label='Best Test Accuracy obtained by '+ str(choice)), gr.update(label="Time taken by " + str(choice) + " in seconds")
def hpo(dset, ml_model, strategy, budget):
metric = 'cossim'
kw = 0.01
search_space = return_search_space(dset)
dset = dset.lower()
if dset in ['mnist', 'cifar10']:
feat_model = 'dino_cls'
elif dset in ['trec6']:
feat_model = 'all-distilroberta-v1'
temperature = 1
strategy = strategy.lower()
per_class = True
if ml_model == 'LeNet':
ml_model = 'MnistNet'
else:
ml_model = ml_model
run_cnt = 0
if strategy in ['gradmatchpb', 'craigpb', 'glister']:
if dset in ['trec6']:
select_every = 3
else:
select_every = 3
else:
select_every = 1
submod_function = 'disp_min_pc'
data_dir = '../data'
stochastic_subsets_file = os.path.join(os.path.abspath(data_dir), dset + '_' + feat_model + '_' + metric + '_' + 'gc_pc' + '_' + str(kw) + '_' + str(budget/100) + '_stochastic_subsets.pkl')
gc_stochastic_subsets_file = os.path.join(os.path.abspath(data_dir), dset + '_' + feat_model + '_' + metric + '_' + 'gc_pc' + '_' + str(kw) + '_' + str(budget/100) + '_stochastic_subsets.pkl')
global_order_file = os.path.join(os.path.abspath(data_dir), dset + '_' + feat_model + '_' + metric + '_' + submod_function + '_' + str(kw) + '_global_order.pkl')
#Pre-trained HPO full results
full_results_dict = return_full_results()
#Naive code for now! Will have an improved version soon
search_configs = []
key1, key2, key3 = list(search_space.keys())[0], list(search_space.keys())[1], list(search_space.keys())[2]
for i in range(len(search_space[key1])):
for j in range(len(search_space[key2])):
for k in range(len(search_space[key3])):
search_configs.append({key1: search_space[key1][i], key2: search_space[key2][j], key3: search_space[key3][k]})
trainset, validset, testset, num_cls = gen_dataset('../data/TREC6',
'hf_trec6',
'dss',
dataset=DotMap(dict(name="hf_trec6",
datadir="../data/TREC6/",
feature="dss",
type="text",
wordvec_dim=300,
weight_path='../data/glove.6B/',)))
results_dict = {}
for strat in [strategy]:
if dset in ['trec6']:
config_file = "configs/SL/config_" + strat + "_glove_" + dset + ".py"
else:
config_file = "configs/SL/config_" + strat + "_" + dset + ".py"
config_data = load_config_data(config_file)
config_data.train_args.device = 'cuda'
config_data.train_args.run = run_cnt
config_data.train_args.wandb = False
if dset in ['trec6']:
config_data.train_args.num_epochs = 20
config_data.train_args.print_every = 5
config_data.dataloader.batch_size = 16
else:
config_data.train_args.num_epochs = 10
config_data.train_args.print_every = 1
config_data.scheduler.T_max = 10
config_data.scheduler.type = "cosine_annealing"
config_data.optimizer.type = 'sgd'
config_data.optimizer.lr = 5e-2
config_data.dataloader.batch_size = 128
config_data.train_args.print_args=["tst_loss", "tst_acc", "time"]
config_data.dss_args.fraction = (budget/100)
config_data.dss_args.global_order_file = global_order_file
config_data.dss_args.gc_stochastic_subsets_file = gc_stochastic_subsets_file
config_data.dss_args.stochastic_subsets_file = stochastic_subsets_file
config_data.dss_args.gc_ratio = 0.1
config_data.dss_args.kw = kw
config_data.dss_args.per_class = per_class
config_data.dss_args.temperature = temperature
config_data.dss_args.submod_function = submod_function
config_data.model.architecture = ml_model
config_data.dss_args.select_every = select_every
classifier = TrainClassifier(config_data)
results_dict[strat] = []
for search_config in search_configs:
if dset in ['trec6']:
if "num_layers" in search_config:
classifier.cfg.model.num_layes = search_config["num_layers"]
if "hidden_size" in search_config:
classifier.cfg.model.hidden_size = search_config["hidden_size"]
if "learning_rate" in search_config:
classifier.cfg.optimizer.lr = search_config["learning_rate"]
trn_acc, val_acc, tst_acc, best_acc, omp_cum_timing = classifier.train(trainset=trainset, validset=validset, testset=testset, num_cls=num_cls)
results_dict[strat].append([best_acc[-1], omp_cum_timing[-1]])
best_strat_acc = -1
best_full_acc = -1
best_strat_idx = -1
strat_tuning_time = 0
full_tuning_time = 0
for i in range(len(search_configs)):
strat_tuning_time += results_dict[strategy][i][1]
full_tuning_time += full_results_dict[tuple(sorted(search_configs[i].items()))][1]
if results_dict[strategy][i][0] > best_strat_acc:
best_strat_acc = results_dict[strategy][i][0]
best_strat_config = search_configs[i]
if full_results_dict[tuple(sorted(search_configs[i].items()))][0] > best_full_acc:
best_full_acc = full_results_dict[tuple(sorted(search_configs[i].items()))][0]
best_full_config = search_configs[i]
return best_strat_config, full_results_dict[tuple(sorted(best_strat_config.items()))][0], strat_tuning_time , best_full_config, best_full_acc, full_tuning_time
with gr.Blocks(title = "Hyper-parameter Optimization") as demo:
with gr.Row():
with gr.Column():
dset = gr.Dropdown(choices=['TREC6'], label='Dataset Name')
model = gr.Radio(["LSTM"], label="Model Architecture")
strategy = gr.Dropdown(choices=['Random', 'AdaptiveRandom', 'MILO', 'WRE', 'SGE', 'MILOFixed', 'GradMatchPB', 'CraigPB', 'GLISTER'], label='Subset Selection Strategy')
budget = gr.Slider(minimum=1, maximum=100, label='Budget (in %)')
search_space = gr.JSON(label='Hyper-parameter Search Space')
submit = gr.Button(value="Perform Grid Search")
with gr.Column():
# df1 = gr.DataFrame(label='Hyper-parameter Optimization Results')
with gr.Row():
strat_config = gr.JSON(label='Best Configuration obtained using selected strategy')
strat_acc = gr.Number(label="Best Validation Accuracy using selected strategy")
strat_timing = gr.Number(label="Time taken for tuning using selected strategy in seconds")
with gr.Row():
full_config = gr.JSON(label='Best Configuration obtained using Full')
full_acc = gr.Number(label="Best Validation Accuracy using Full")
full_timing = gr.Number(label="Time taken for tuning using Full")
dset.change(fn=update_model_text, inputs=dset, outputs=[model, search_space])
# strategy.change(fn=update_out_text, inputs=strategy, outputs=[strat_acc, strat_timing])
submit.click(fn=hpo, inputs=[dset, model, strategy, budget], outputs=[strat_config, strat_acc, strat_timing, full_config, full_acc, full_timing])
demo.launch()
# app = gr.mount_gradio_app(app, demo, path=CUSTOM_PATH)