File size: 6,698 Bytes
b84549f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from typing import Any, Dict
from schema import Schema
from data import Scenario, MergedDataset
from methods.base.alg import BaseAlg
from data import build_dataloader
from ..model import ElasticDNN_OfflineFMModel
from ...model.base import ElasticDNNUtil
import torch.optim
import tqdm
from torch import nn
from torchvision.transforms import Compose
from utils.dl.common.env import create_tbwriter
import os
import random
import numpy as np
from copy import deepcopy
from utils.dl.common.model import get_module
from utils.common.log import logger


class ElasticDNN_FMLoRAAlg(BaseAlg):
    def get_required_models_schema(self) -> Schema:
        return Schema({
            'fm': ElasticDNN_OfflineFMModel
        })
        
    def get_required_hyp_schema(self) -> Schema:
        from schema import Optional
        return Schema({
            'launch_tbboard': bool,
            
            'samples_size': object,
            'ab_r': int,
            
            'train_batch_size': int,
            'val_batch_size': int,
            'num_workers': int,
            'optimizer': str,
            'optimizer_args': dict,
            'scheduler': str,
            'scheduler_args': dict,
            'num_iters': int,
            'val_freq': int,
            
            Optional('fm_lora_ckpt_path'): str,
            Optional('transform'): Compose,
        })
        
    def run(self, scenario: Scenario, hyps: Dict, collate_fn=None) -> Dict[str, Any]:
        super().run(scenario, hyps)
        
        assert isinstance(self.models['fm'], ElasticDNN_OfflineFMModel) # for auto completion
        
        # 1. add LoRA
        lora_util = self.models['fm'].get_lora_util()
        device = self.models['fm'].device
        
        sample = hyps['samples_size']
        if isinstance(sample, (tuple, list)) and isinstance(sample[0], int):
            sample = torch.rand(hyps['samples_size']).to(device)
        lora_util.add_lora_ab_to_fm(self.models['fm'].models_dict['main'], hyps['ab_r'], sample)
        
        if 'fm_lora_ckpt_path' in hyps.keys() and hyps['fm_lora_ckpt_path'] != '' and hyps['fm_lora_ckpt_path'] is not None:
            _ckpt = torch.load(hyps['fm_lora_ckpt_path'])['main']

            new_state_dict = deepcopy(self.models['fm'].models_dict['main'].state_dict())
            
            for n, p in _ckpt.named_parameters():
                if 'qkv.abs' not in n:
                    continue
                
                new_state_dict[n] = p
                logger.info(f'use {n} from ckpt')
            
            self.models['fm'].models_dict['main'].load_state_dict(new_state_dict)
            
        
        # 2. train (knowledge distillation, index relationship)
        if 'transform' in hyps.keys():
            offline_datasets = scenario.get_offline_datasets(transform=hyps['transform'])
        else:
            offline_datasets = scenario.get_offline_datasets()
        train_dataset = MergedDataset([d['train'] for d in offline_datasets.values()])
        
        # debug
        # from data.visualize import visualize_classes_in_object_detection
        # d = offline_datasets['GTA5Det']['val']
        # class_to_idx_map = {c: d.idx_map[i] for i, c in enumerate(d.classes)}
        # print(class_to_idx_map)
        # visualize_classes_in_object_detection(d, class_to_idx_map,
        #                                       {}, os.path.join(self.res_save_dir, 'debug.png'))
        # exit()
        
        val_dataset = MergedDataset([d['val'] for d in offline_datasets.values()])
        train_loader = iter(build_dataloader(train_dataset, hyps['train_batch_size'], hyps['num_workers'],
                                        True, None, collate_fn=collate_fn))
        
        # if hyps['use_train_loader_for_val']:
        #     val_loader = build_dataloader(train_dataset, hyps['val_batch_size'], hyps['num_workers'],
        #                               False, False)
        #     logger.warn('use train loader for val!!!')
        # else:
        val_loader = build_dataloader(val_dataset, hyps['val_batch_size'], hyps['num_workers'],
                                    False, False, collate_fn=collate_fn)
        
        lora_params = lora_util.train_only_lora(self.models['fm'].models_dict['main'])
        head_params = self.models['fm'].get_task_head_params()
        
        num_lora_params = sum([np.prod(p.size()) for p in lora_params])
        total_params = sum([np.prod(p.size()) for p in self.models['fm'].models_dict['main'].parameters()])
        logger.info(f'num lora params: {num_lora_params}, total params: {total_params}, ratio: {num_lora_params / total_params}')
        
        optimizer = torch.optim.__dict__[hyps['optimizer']](lora_params + head_params, **hyps['optimizer_args'])
        scheduler = torch.optim.lr_scheduler.__dict__[hyps['scheduler']](optimizer, **hyps['scheduler_args'])
        
        fbs_tb_writer = create_tbwriter(os.path.join(self.res_save_dir, 'tb_log'), launch_tbboard=hyps['launch_tbboard'])
        pbar = tqdm.tqdm(range(hyps['num_iters']), dynamic_ncols=True)
        
        best_val_acc = 0
        val_acc = 0
        
        for iter_index in pbar:
            self.models['fm'].to_train_mode()
            
            x, y = next(train_loader)
            
            if isinstance(x, dict):
                for k, v in x.items():
                    if isinstance(v, torch.Tensor):
                        x[k] = v.to(device)
                y = y.to(device)
            else:
                x, y = x.to(device), y.to(device)
            task_loss = self.models['fm'].forward_to_get_task_loss(x, y)
            optimizer.zero_grad()
            task_loss.backward()
            optimizer.step()
            scheduler.step()
            
            if (iter_index + 1) % hyps['val_freq'] == 0:
                # logger.warn('use train loader for val!!!')
                
                self.models['fm'].to_eval_mode()
                val_acc = self.models['fm'].get_accuracy(val_loader)
                
                self.models['fm'].save_model(os.path.join(self.res_save_dir, 'models/fm_last.pt'))
                if val_acc > best_val_acc:
                    best_val_acc = val_acc
                    self.models['fm'].save_model(os.path.join(self.res_save_dir, 'models/fm_best.pt'))
                
            fbs_tb_writer.add_scalar(f'losses/task_loss', task_loss, iter_index)
            fbs_tb_writer.add_scalar(f'accs/val_acc', val_acc, iter_index)
            fbs_tb_writer.add_scalar(f'lr', optimizer.param_groups[0]['lr'], iter_index)
            pbar.set_description(f'loss: {task_loss:.6f}, val_acc: {val_acc:.4f}')