RobustBind/train.py at main · TensorNeural/RobustBind · GitHub

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
import torch
from torch import nn
import torch.nn.functional as F
from torch.utils.tensorboard import SummaryWriter
import models.PointBind_models as models
from imagebind.imagebind_model import ModalityType
import argparse
import json
import pickle
import logging
import os
from tqdm import tqdm
from data.process_data import TrainDataset, EvalVisionDataset
from torch.utils.data import Dataset, DataLoader, SequentialSampler, RandomSampler
from transformers import (
    AdamW,
    get_linear_schedule_with_warmup,
)
import random
import numpy as np
from model import UniBind
from utils.utils import set_env, gen_label, loss_fun, load_centre_embeddings

logger = logging.getLogger(__name__)

def gen_visual_embeddings(model, data_loader):
    all_embeddings = []
    all_visual_labels = []
    for _, batch in enumerate(tqdm(data_loader)):
        with torch.no_grad():
            all_visual_labels.extend(batch['labels'])
            embeddings = model.encode_vision_with_mlp(batch['inputs'])
            all_embeddings.append(embeddings)
    all_embeddings = torch.cat(all_embeddings, dim=0)
    return all_embeddings, all_visual_labels


def evaluate(args, model, val_data_loader, device):
    model.eval()
    logger.info('Start to generate visual embeddings!')
    visual_embeddings, visual_labels = gen_visual_embeddings(model, val_data_loader)
    logger.info('visual embedding generation done!')
    logger.info('---------------------------------')
    logger.info('Start to load centre embeddings!')
    centre_embeddings, centre_labels = load_centre_embeddings(args.centre_embeddings_path, device)
    logger.info('centre embedding load done!')
    logger.info('---------------------------------')
    centre_embeddings /= centre_embeddings.norm(dim=-1, keepdim=True)
    visual_embeddings /= visual_embeddings.norm(dim=-1, keepdim=True)
    logic = (visual_embeddings.to(device) @ centre_embeddings.to(device).t()).softmax(dim=-1)
    acc = 0.0
    for i in range(logic.shape[0]):
        _, index = logic[i].topk(1)
        if visual_labels[i] == centre_labels[int(index[0])]:
            acc = acc + 1
    return acc/logic.shape[0]

def train(args, model, train_dataloader, val_dataloader, device):
    real_batch_size = args.train_batch_size * args.gradient_accumulation_steps
    t_total = train_dataloader.dataset.__len__() // real_batch_size * args.num_train_epochs
    param_optimizer = list(model.named_parameters())
    optimizer_grouped_parameters = [
            {'params': [p for n, p in param_optimizer], 'weight_decay': 0.01}
        ]

    optimizer = AdamW(optimizer_grouped_parameters,
                      lr=args.learning_rate, eps=args.adam_epsilon)
    scheduler = get_linear_schedule_with_warmup(
        optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
    logger.info("***** Running training *****")
    logger.info("  Num Epochs = %d", args.num_train_epochs)
    logger.info("  Instantaneous batch size per GPU = %d",
                args.train_batch_size)
    logger.info("  Gradient Accumulation steps = %d",
                args.gradient_accumulation_steps)
    logger.info("  Total optimization steps = %d", t_total)
    global_step = 0
    tr_loss = 0.0
    best_acc = 0.0
    writer = SummaryWriter(log_dir=(args.output_dir + '/tb_loss'))
    model.zero_grad()
    for epoch in range(int(args.num_train_epochs)):
        epoch_steps = len(train_dataloader)
        for step, batch in enumerate(train_dataloader):
            model.train()
            text_embeddings, vision_embeddings = model(batch['inputs'])
            logic = vision_embeddings @ text_embeddings.t()
            labels = gen_label(logic, device)
            loss = loss_fun(logic, labels)
            if args.gradient_accumulation_steps > 1:
                loss = loss / args.gradient_accumulation_steps
            loss.backward()
            tr_loss += loss.item()
            if (step + 1) % args.gradient_accumulation_steps == 0:
                global_step += 1
                torch.nn.utils.clip_grad_norm_(
                    model.parameters(), args.max_grad_norm)
                logger.info(
                    'Epoch: {}, Step: {}, Loss: {:.8f}, lr: {:.6f}'.format(epoch, global_step, (tr_loss / global_step),
                                                                           optimizer.param_groups[0]["lr"]))
                writer.add_scalar(tag="cl_loss", scalar_value= tr_loss / global_step, global_step=step+epoch_steps*epoch)
                optimizer.step()
                scheduler.step()
                model.zero_grad()
                if global_step % args.eval_steps == 0:
                    logger.info('Start eval!')
                    acc = evaluate(args, model, val_dataloader, device)
                    logger.info('Dev acc: {0}'.format(acc))
                    if acc >= best_acc:
                        best_acc = acc
                        torch.save(
                            model.state_dict(),
                            os.path.join(args.output_dir, f"{args.modality}_model_best.pt")
                        )
    writer.close()
    torch.save(model.state_dict(), os.path.join(args.output_dir, f"{args.modality}_model_{args.num_train_epochs}.pt"))

def test(args, val_dataloader, device):
    model = UniBind(args)
    model.to(device)
    ckpt_path = os.path.join(args.output_dir, f"{args.modality}_model_best.pt")
    model.load_state_dict(torch.load(ckpt_path))
    acc = evaluate(args, model, val_dataloader, device)
    return acc


def main():
    torch.multiprocessing.set_start_method('spawn')
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    parser = argparse.ArgumentParser()
    parser.add_argument("--train_dataset_dir", type=str, default='', required=True)
    parser.add_argument("--eval_dataset_dir", type=str, default='', required=True)
    parser.add_argument("--test_dataset_dir", type=str, default='', required=True)
    parser.add_argument("--train_data_path", type=str, default='', required=True)
    parser.add_argument("--eval_data_path", type=str, default='', required=True)
    parser.add_argument("--test_data_path", type=str, default='', required=True)
    parser.add_argument("--centre_embeddings_path", type=str, default='', required=True)
    parser.add_argument("--pretrain_weights", type=str, default='', required=True)
    parser.add_argument("--output_dir", type=str, default='', required=True)
    parser.add_argument("--modality", type=str, default='vision', required=True)
    parser.add_argument("--train_batch_size", type=int, default=8, required=True)
    parser.add_argument("--val_batch_size", type=int, default=8, required=True)
    parser.add_argument("--num_workers", type=int, default=0, required=True)
    parser.add_argument("--gradient_accumulation_steps", type=int, default=1)
    parser.add_argument("--learning_rate", default=1e-5, type=float, required=True, help="The initial learning rate for Adam.")
    parser.add_argument("--adam_epsilon", type=float, default=1e-8)
    parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
    parser.add_argument("--num_train_epochs", default=10, type=float, required=True, help="Total number of training epochs to perform.")
    parser.add_argument("--warmup_steps", default=0, type=int, help="Linear warmup over warmup_steps.")
    parser.add_argument("--eval_steps", type=int, default=10, required=True, help="eval model every X updates steps.")
    parser.add_argument("--seed", type=int, default=1234, required=True, help="random seed for initialization")

    args = parser.parse_args()
    log_name = args.modality + '_finetune'
    set_env(args, log_name)
    model = UniBind(args)
    model.to(device)
    logger.info("Loading training set.")
    train_data = TrainDataset(args, device)
    train_sampler = RandomSampler(train_data)
    train_reader = DataLoader(dataset=train_data, sampler=train_sampler, num_workers=args.num_workers,
                              batch_size=args.train_batch_size, collate_fn=train_data.Collector)
    eval_data = EvalVisionDataset(args, device, infer_type="eval")
    eval_sampler = SequentialSampler(eval_data)
    eval_reader = DataLoader(dataset=eval_data, sampler=eval_sampler, num_workers=args.num_workers,
                              batch_size=args.val_batch_size, collate_fn=eval_data.Collector, drop_last=False)
    test_data = EvalVisionDataset(args, device, infer_type="test")
    test_sampler = SequentialSampler(test_data)
    test_reader = DataLoader(dataset=test_data, sampler=test_sampler, num_workers=args.num_workers,
                              batch_size=args.val_batch_size, collate_fn=eval_data.Collector, drop_last=False)

    train(args, model, train_reader, eval_reader, device)

    test_acc = test(args, test_reader, device)
    logger.info(f"Acc on test set: {test_acc}")


if __name__ == "__main__":
    main()