ImgSegRefExp/train.py at master · ramithp/ImgSegRefExp · 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
import sys
import json
import os
import numpy as np
from data import ImageSegmentationDataset
from torch.utils.data import DataLoader
from model.baseline import ImgSegRefExpModel
from model.resnet_exp import ResImgSeg
import config
from torch.optim import SGD
import torch
from torch import nn
import time

def get_lr(optimizer):
    for param_group in optimizer.param_groups:
        return param_group['lr']
def train_epoch(model, train_loader, criterion, optimizer, scheduler, epoch_num, device=config.device):
    print("Training\n Number of batches: {}\tBatch Size: {}\tDataset size: {}".format(len(train_loader),train_loader.batch_size,len(train_loader.dataset)))
    start = time.time()
    model.train()
    cls_loss_avg = 0.0
    end_time = 0
    batch_start = time.time()

    for batchId, (image_sizes, processed_ims, ground_truth_masks, texts) in enumerate(train_loader):
        optimizer.zero_grad()
        texts = texts.long()

        output_mask = model((processed_ims.to(device), texts.to(device)))
        output_mask = output_mask.squeeze(1)
        ground_truth_masks = ground_truth_masks.to(device).squeeze(1).float()

        # Adapted from https://pytorch.org/docs/stable/nn.html#bcewithlogitsloss
        # and https://github.com/ramithp/text_objseg/blob/tensorflow-1.x-compatibility/util/loss.py
        pos_loss_mult = 1.
        neg_loss_mult = 1. /3.
        loss_mult = (ground_truth_masks * (pos_loss_mult-neg_loss_mult)) +  neg_loss_mult

        # Classification loss as the average of weighed per-score loss
        cls_loss = (criterion(output_mask, ground_truth_masks) * loss_mult).mean()

        cls_loss.backward()
        cls_loss_avg = config.decay * cls_loss + (1 - config.decay) * cls_loss.item()
        optimizer.step()
        scheduler.step()

        if batchId % 20 == 0:
            print("Batch Time with data loading = {}s, Batch #{}: Loss = {}\tAvg Loss: {}\tTime: {}s".format(time.time() - end_time,
                                                                                                            batchId,
                                                                                                            cls_loss.item(),
                                                                                                            cls_loss_avg,time.time() - batch_start))
        if batchId % 500 == 0:
            torch.save(model.state_dict(), 'model_dict_ep_'+ str(epoch_num) + '_iter_' + str(batchId) + '.pt')

        end_time = time.time()

    print('Batch Loss (avg) = {}, lr = {}, time = {}s'.format(cls_loss_avg,
                                                             get_lr(optimizer),
                                                             time.time()-start))

def eval_model(model, val_loader, criterion, epoch_num, device):
    print("Validating\nNumber of batches: {}\tBatch Size: {}\tDataset size: {}".format(len(val_loader),
                                                                                        val_loader.batch_size,
                                                                                        len(val_loader.dataset)))
    cls_loss_avg = 0.0
    model.eval()
    with torch.no_grad():
        batch_start = time.time()
        for batchId, (image_sizes, processed_ims, ground_truth_masks, texts) in enumerate(val_loader):
            texts = texts.long()
            output_masks = model((processed_ims.to(device), texts.to(device)))
            output_masks = output_masks.squeeze(1)
            ground_truth_masks = ground_truth_masks.to(device).squeeze(1).float()

            cls_loss_val = criterion(output_masks, ground_truth_masks)
            cls_loss_avg = config.decay * cls_loss_avg + (1 - config.decay) * cls_loss_val.item()

            if batchId % 100 == 0:
                print("Batch #{}: Loss = {}\tAvg Loss: {}\tTime: {}s".format(batchId,
                                                                            cls_loss_val.item(),
                                                                            cls_loss_avg,time.time() - batch_start))
        print('Batch Loss (avg) = {}, lr = {}, time = {}s'.format(cls_loss_avg,
                                                                 get_lr(optimiser),
                                                                 time.time()))

def main():
    print("Starting training")
    # Load model and weights
    # model = ImgSegRefExpModel(mlp_hidden=500, vocab_size=8803, emb_size=1000, lstm_hidden_size=1000)
    model = ResImgSeg(mlp_hidden=500, vocab_size=8803, emb_size=1000, lstm_hidden_size=1000)

    # if config.pretrained_wts:
        # pre_trained = torch.load(config.pretrained_model_file)
        # model.load_state_dict(pre_trained)

    model.to(config.device)
    print(model)

    # Combine weight decay regularisation with optimiser
    optimizer = torch.optim.Adam(model.parameters(),lr=1e-4)

    # optimizer = torch.optim.SGD(model.parameters(),lr=config.start_lr, momentum=config.momentum, weight_decay=config.weight_decay)
    scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=config.lr_decay_step, gamma=config.lr_decay_rate)
    criterion = nn.BCEWithLogitsLoss(reduction="none")

    train_dataset = ImageSegmentationDataset(config.query_file, config.image_dir, config.mask_dir)
    val_dataset = ImageSegmentationDataset(config.query_file_val, config.image_dir, config.mask_dir)

    train_loader = DataLoader(train_dataset, batch_size=32, shuffle=True)
    val_loader = DataLoader(val_dataset,batch_size=32, shuffle=False)

    model_parameters = filter(lambda p: p.requires_grad, model.img_features.parameters())
    params = sum([np.prod(p.size()) for p in model_parameters])
    print("Total params:", params)

    model_parameters = filter(lambda p: p.requires_grad, model.parameters())
    params = sum([np.prod(p.size()) for p in model_parameters])
    print("Total params:", params)

    for i in range(0, config.n_epochs):
        print("Training for epoch %d" % (i))
        train_loss = train_epoch(model, train_loader, criterion, optimizer, scheduler, i, config.device)
        test_loss = eval_model(model, val_loader, criterion, i, config.device)

        print('='*20)

if __name__ == '__main__':
    main()