File size: 8,120 Bytes
9157432
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import gc
import logging
import os
import time
from pathlib import Path
import numpy as np
import pandas as pd
import torch
from config import get_config_regression
from data_loader import MMDataLoader
from trains import ATIO
from utils import assign_gpu, setup_seed
from trains.singleTask.model import DLF
from trains.singleTask.distillnets import get_distillation_kernel, get_distillation_kernel_homo
from trains.singleTask.misc import softmax
import sys

from datetime import datetime      
now = datetime.now()
format = "%Y/%m/%d %H:%M:%S"
formatted_now = now.strftime(format)
formatted_now = str(formatted_now)+" - "

os.environ["CUDA_DEVICE_ORDER"]="PCI_BUS_ID"
os.environ["CUBLAS_WORKSPACE_CONFIG"] = ":4096:2"
logger = logging.getLogger('MMSA')

def _set_logger(log_dir, model_name, dataset_name, verbose_level):

    # base logger
    log_file_path = Path(log_dir) / f"{model_name}-{dataset_name}.log"
    logger = logging.getLogger('MMSA')
    logger.setLevel(logging.DEBUG)

    # file handler
    fh = logging.FileHandler(log_file_path)
    fh_formatter = logging.Formatter('%(asctime)s - %(name)s [%(levelname)s] - %(message)s')
    fh.setLevel(logging.DEBUG)
    fh.setFormatter(fh_formatter)
    logger.addHandler(fh)

    # stream handler
    stream_level = {0: logging.ERROR, 1: logging.INFO, 2: logging.DEBUG}
    ch = logging.StreamHandler()
    ch.setLevel(stream_level[verbose_level])
    ch_formatter = logging.Formatter('%(name)s - %(message)s')
    ch.setFormatter(ch_formatter)
    logger.addHandler(ch)

    return logger


def DLF_run(
    model_name, dataset_name, config=None, config_file="", seeds=[], is_tune=False,
    tune_times=500, feature_T="", feature_A="", feature_V="",
    model_save_dir="", res_save_dir="", log_dir="",
    gpu_ids=[0], num_workers=1, verbose_level=1, mode = '', is_training = False 
):
    # Initialization
    model_name = model_name.upper()
    dataset_name = dataset_name.lower()
    
    if config_file != "":
        config_file = Path(config_file)
    else: # use default config files
        config_file = Path(__file__).parent / "config" / "config.json"
    if not config_file.is_file():
        raise ValueError(f"Config file {str(config_file)} not found.")
    if model_save_dir == "":
        model_save_dir = Path.home() / "MMSA" / "saved_models"
    Path(model_save_dir).mkdir(parents=True, exist_ok=True)
    if res_save_dir == "":
        res_save_dir = Path.home() / "MMSA" / "results"
    Path(res_save_dir).mkdir(parents=True, exist_ok=True)
    if log_dir == "":
        log_dir = Path.home() / "MMSA" / "logs"
    Path(log_dir).mkdir(parents=True, exist_ok=True)
    seeds = seeds if seeds != [] else [1111, 1112, 1113, 1114, 1115]
    logger = _set_logger(log_dir, model_name, dataset_name, verbose_level)
    

    args = get_config_regression(model_name, dataset_name, config_file)
    args.is_training = is_training  
    args.mode = mode # train or test
    args['model_save_path'] = Path(model_save_dir) / f"{args['model_name']}-{args['dataset_name']}.pth"
    args['device'] = assign_gpu(gpu_ids)
    args['train_mode'] = 'regression'
    args['feature_T'] = feature_T
    args['feature_A'] = feature_A
    args['feature_V'] = feature_V
    if config:
        args.update(config)


    res_save_dir = Path(res_save_dir) / "normal"
    res_save_dir.mkdir(parents=True, exist_ok=True)
    model_results = []
    for i, seed in enumerate(seeds):
        setup_seed(seed)
        args['cur_seed'] = i + 1
        result = _run(args, num_workers, is_tune)
        model_results.append(result)
    if args.is_training:
        criterions = list(model_results[0].keys())
        # save result to csv
        csv_file = res_save_dir / f"{dataset_name}.csv"
        if csv_file.is_file():
            df = pd.read_csv(csv_file)
        else:
            df = pd.DataFrame(columns=["Time"]+["Model"] + criterions)
        # save results
        res = [model_name]
        for c in criterions:
            values = [r[c] for r in model_results]
            mean = round(np.mean(values)*100, 2)
            std = round(np.std(values)*100, 2)
            res.append((mean, std))
        
        res = [formatted_now]+res 
        df.loc[len(df)] = res    
        df.to_csv(csv_file, index=None)
        logger.info(f"Results saved to {csv_file}.")


def _run(args, num_workers=4, is_tune=False, from_sena=False): 

    dataloader = MMDataLoader(args, num_workers)

    if args.is_training:
        print("training for DLF")

        
        args.gd_size_low = 64  
        args.w_losses_low = [1, 10]  
        args.metric_low = 'l1'  

        
        args.gd_size_high = 32  
        args.w_losses_high = [1, 10]  
        args.metric_high = 'l1'  

        to_idx = [0, 1, 2]  
        from_idx = [0, 1, 2]  
        assert len(from_idx) >= 1

        model = []
        model_DLF = getattr(DLF, 'DLF')(args)

        model_distill_homo = getattr(get_distillation_kernel_homo, 'DistillationKernel')(n_classes=1,
                                                                               hidden_size=
                                                                               args.dst_feature_dim_nheads[0],
                                                                               gd_size=args.gd_size_low,
                                                                               to_idx=to_idx, from_idx=from_idx,
                                                                               gd_prior=softmax([0, 0, 1, 0, 1, 0], 0.25),
                                                                               gd_reg=10,
                                                                               w_losses=args.w_losses_low,
                                                                               metric=args.metric_low,
                                                                               alpha=1 / 8,
                                                                               hyp_params=args)

        model_distill_hetero = getattr(get_distillation_kernel, 'DistillationKernel')(n_classes=1,
                                                                                   hidden_size=
                                                                                   args.dst_feature_dim_nheads[0] * 2,
                                                                                   gd_size=args.gd_size_high,
                                                                                   to_idx=to_idx, from_idx=from_idx,
                                                                                   gd_prior=softmax([0, 0, 1, 0, 1, 1], 0.25),
                                                                                   gd_reg=10,
                                                                                   w_losses=args.w_losses_high,
                                                                                   metric=args.metric_high,
                                                                                   alpha=1 / 8,
                                                                                   hyp_params=args)

        model_DLF = model_DLF.cuda()

        model = [model_DLF]         
    else:
        print("testing phase for DLF")
        model = getattr(DLF, 'DLF')(args)
        model = model.cuda()

    trainer = ATIO().getTrain(args)


    #test
    if args.mode == 'test':
        model.load_state_dict(torch.load('./pt/DLF'+str(args.dataset_name)+'.pth'),strict=False) 
        results = trainer.do_test(model, dataloader['test'], mode="TEST")
        sys.stdout.flush()
        input('[Press Any Key to start another run]')
    #train
    else:
        epoch_results = trainer.do_train(model, dataloader, return_epoch_results=from_sena)
        model[0].load_state_dict(torch.load('./pt/DLF'+str(args.dataset_name)+'.pth'))

        results = trainer.do_test(model[0], dataloader['test'], mode="TEST")

        del model
        torch.cuda.empty_cache()
        gc.collect()
        time.sleep(1)
    return results