-
Notifications
You must be signed in to change notification settings - Fork 2
/
train.py
200 lines (164 loc) · 8.59 KB
/
train.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
# train.py -pn RecipeMind_CIKM2022 -sn ablated_cpmx --model_struct recipemind_cpmx_sum_cat --random_seed 1001 --batch_size 1024 --dataset_index ver1 --dataset_name recipemind_mixed_sPMId02
from env_config import *
from trainer import *
from models import *
import wandb
import torch
import numpy as np
import os
import random
import json
import setproctitle
torch.set_num_threads(1)
os.environ["MKL_NUM_THREADS"] = "20"
os.environ["NUMEXPR_NUM_THREADS"] = "20"
os.environ["OMP_NUM_THREADS"] = "20"
os.environ['OPENBLAS_NUM_THREADS'] = "20"
os.environ['CUDA_DEVICE_ORDER'] = 'PCI_BUS_ID'
# torch.autograd.set_detect_anomaly(True)
def setup_seed(seed):
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
np.random.seed(seed)
random.seed(seed)
torch.backends.cudnn.deterministic = True
def get_vector_dimensions(args):
lang_dim = dict()
dim_dict = {
'reciptor': 600,
'bert-base-uncased': 768,
'flavorgraph': 300,
'im2recipe': 300,
'binary': 630
}
lang_dim['J'] = dim_dict[args.initial_vectors_J]
lang_dim['T'] = dim_dict[args.initial_vectors_T]
lang_dim['R'] = dim_dict[args.initial_vectors_R]
args.lang_dim = lang_dim
return args
def baseline_arguments(args):
if args.model_struct == 'kitchenette':
print("Kitchenette Baseline Model")
args.dataset_name = 'recipemind_doublets'
args.initial_vectors_J = 'im2recipe'
args.hidden_dim = 1024
args.dropout_rate = 0.2
args.learning_rate = 1e-4
args.weight_decay = 1e-5
args.num_epochs = 60
args.batch_size = 32
# elif 'recipebowl' in args.model_struct:
# print("RecipeBowl Pretraining Model")
# args.dataset_name = 'recipebowl_original'
# args.initial_vectors_J = 'flavorgraph'
# args.initial_vectors_T = 'binary'
# args.initial_vectors_R = 'reciptor'
# args.hidden_dim = 1024
# args.weight_decay = 0.0
# args.dropout_rate = 0.2
# args.learning_rate = 0.0003
# args.num_epochs = 60
# args.batch_size = 64
else:
pass
return args
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--project_name', '-pn', default='Test SonyAI', type=str)
# parser.add_argument('--group_name', '-gn', default='Test SonyAI', type=str)
parser.add_argument('--session_name', '-sn', default='Test SonyAI', type=str)
parser.add_argument('--random_seed', default=911012, type=int)
# parser.add_argument('--test_mode', default='none', type=str)
parser.add_argument('--fine_tuning', default=None, type=str)
parser.add_argument('--debug_mode', '-dm', default=False, action='store_true')
parser.add_argument('--dataset_index', default='ver1', type=str)
parser.add_argument('--dataset_version', default='211210', type=str)
parser.add_argument('--dataset_name', default='recipemind_mixed_sPMId02', type=str)
parser.add_argument('--initial_vectors_J', default='flavorgraph', type=str)
parser.add_argument('--initial_vectors_T', default='bert-base-uncased', type=str)
parser.add_argument('--initial_vectors_R', default='bert-base-uncased', type=str)
parser.add_argument('--model_struct', default='recipemind', type=str)
parser.add_argument('--model_analysis', default=False, action='store_true')
parser.add_argument('--hidden_dim', default=128, type=int) # 1024
parser.add_argument('--dropout_rate', default=0.025, type=float) # 0.2
parser.add_argument('--learning_rate', default=1e-4, type=float)
parser.add_argument('--weight_decay', default=1e-5, type=float)
parser.add_argument('--num_epochs', default=30, type=int)
parser.add_argument('--batch_size', default=512, type=int)
parser.add_argument('--loss_function', default='rmse', type=str)
parser.add_argument('--grad_update', default='default', type=str)
parser.add_argument('--train_eval', default=False, action='store_true')
# parser.add_argument('--pretrained_recipebowl', default='none', type=str)
parser.add_argument('--hybrid_coef', default=0.5, type=float)
parser.add_argument('--mc_dropout', default=False, action='store_true')
# Set Attention Blocks as Element Encoder
parser.add_argument('--sab_num_aheads', default=8, type=int)
parser.add_argument('--sab_num_blocks', default=3, type=int)
# ApproxRepSet as Element Encoder
parser.add_argument('--ars_num_hsets', default=256, type=int)
parser.add_argument('--ars_num_helms', default=8, type=int)
# Pooling By Multihead Attention as Set Encoder
parser.add_argument('--pma_num_aheads', default=8, type=int)
parser.add_argument('--pma_num_sdvecs', default=4, type=int)
parser.add_argument('--pma_num_blocks', default=2, type=int)
# Multihead Attention related parameters
parser.add_argument('--multihead_sim', default='general_dot', type=str)
parser.add_argument('--multihead_big', default=False, action='store_true')
args = parser.parse_args()
args = baseline_arguments(args)
if 'wnd' in args.model_struct: args.batch_size = 32
print(f"[1] ======================================= Setting Random Seed {args.random_seed}")
setup_seed(args.random_seed)
print(f"[2] ======================================= Getting Vector Dimensions")
args = get_vector_dimensions(args)
print(f"[3] ======================================= Setting Up Wandb.AI")
wandb_init_args = {'project': args.project_name,
'group' : args.session_name,
'name' : f'training_{args.random_seed}',
'config' : args}
for k ,v in wandb_init_args.items(): print(k, v)
wandb.init(**wandb_init_args)
setproctitle.setproctitle(f'{args.session_name}')
wandb.define_metric('train/step'); wandb.define_metric('train/*', step_metric='train/step')
wandb.define_metric('valid/step'); wandb.define_metric('valid/*', step_metric='valid/step')
print(f"[4] ======================================= Loading Model, Trainer and CollateFn")
trainer = load_recipe_trainer(args)
collate = CollateFn(args)
# args.model_analysis = True
model = load_recipe_model(args).cuda()
num_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
print(f"[5] ======================================= Number of Trainable Parameters for {args.model_struct}: {num_params}")
wandb.watch(model, log='gradients', log_freq=1000)
pickle.dump(args, open(trainer.checkpoint_path+'model_config.pkl', 'wb'))
print(f"[6] ======================================= Loading Train/Valid Dataset and Dataloader")
train = get_train_loader(args, collate)
valid = get_valid_loader(args, collate)
print(f"[7] ======================================= Training the Model {trainer.checkpoint_path}")
model = trainer.train_model(model, train, valid, args.fine_tuning)
del train; torch.cuda.empty_cache()
if not args.debug_mode:
print(f"[8] ======================================= Evaluating the Model on Full Validation Set")
trainer.test_model(model, valid, True)
# if args.fine_tuning:
# print(f"[6] ======================================= Loading the Pretrained Model {args.fine_tuning}")
# checkpoint = torch.load(f'{OUT_PATH}{args.fine_tuning}/epoch_final.mdl')
# model.load_state_dict(checkpoint['model_state_dict'])
# if args.test_mode == 'none':
# print(f"[7] ======================================= Loading Train/Valid Dataset and Dataloader")
# train = get_train_loader(args, collate)
# valid = get_valid_loader(args, collate)
# print(f"[8] ======================================= Training the Model {trainer.checkpoint_path}")
# model = trainer.train_model(model, train, valid, args.fine_tuning)
# del train; torch.cuda.empty_cache()
# if not args.debug_mode:
# print(f"[9] ======================================= Evaluating the Model on Full Validation Set")
# trainer.test_model(model, valid, True)
# else:
# print(f"[7] ======================================= Loading Test Dataset and Dataloader")
# checkpoint = torch.load(trainer.checkpoint_path+f'{args.test_mode}.mdl')
# model.load_state_dict(checkpoint['model_state_dict'])
# test = get_test_loader(args, collate)
# print(f"[8] ======================================= Evaluating the Model on Full Test Set")
# test_session = f'final_eval_{args.dataset_index}_{args.dataset_name}'
# trainer.test_model(model, test, True)
wandb.finish()