File size: 3,720 Bytes
561c629
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# -*- coding: utf-8 -*-
import sys
import os
import torch


# Import files from the local folder
root_path = os.path.abspath('.')
sys.path.append(root_path)
from opt import opt
from architecture.grl import GRL               # This place need to adjust for different models
from train_code.train_master import train_master



# Mixed precision training
scaler = torch.cuda.amp.GradScaler()


class train_grl(train_master):
    def __init__(self, options, args) -> None:
        super().__init__(options, args, "grl")      # Pass a model name unique code 


    def loss_init(self):
        # Prepare pixel loss
        self.pixel_loss_load()
        

    def call_model(self):
        patch_size = 144
        window_size = 8

        if opt['model_size'] == "small":
            # GRL small model 
            self.generator = GRL(
                upscale = opt['scale'],              
                img_size = patch_size,
                window_size = 8,
                depths = [4, 4, 4, 4],
                embed_dim = 128,
                num_heads_window = [2, 2, 2, 2],
                num_heads_stripe = [2, 2, 2, 2],
                mlp_ratio = 2,
                qkv_proj_type = "linear",
                anchor_proj_type = "avgpool",
                anchor_window_down_factor = 2,
                out_proj_type = "linear",
                conv_type = "1conv",
                upsampler = "pixelshuffle",
            ).cuda()

        elif opt['model_size'] == "tiny":
            # GRL tiny model
            self.generator = GRL(
                upscale = opt['scale'],
                img_size = 64,
                window_size = 8,
                depths = [4, 4, 4, 4],
                embed_dim = 64,
                num_heads_window = [2, 2, 2, 2],
                num_heads_stripe = [2, 2, 2, 2],
                mlp_ratio = 2,
                qkv_proj_type = "linear",
                anchor_proj_type = "avgpool",
                anchor_window_down_factor = 2,
                out_proj_type = "linear",
                conv_type = "1conv",
                upsampler = "pixelshuffledirect",
            ).cuda()


        elif opt['model_size'] == "tiny2":
            # GRL tiny model
            self.generator = GRL(
                upscale = opt['scale'],
                img_size = 64,
                window_size = 8,
                depths = [4, 4, 4, 4],
                embed_dim = 64,
                num_heads_window = [2, 2, 2, 2],
                num_heads_stripe = [2, 2, 2, 2],
                mlp_ratio = 2,
                qkv_proj_type = "linear",
                anchor_proj_type = "avgpool",
                anchor_window_down_factor = 2,
                out_proj_type = "linear",
                conv_type = "1conv",
                upsampler = "nearest+conv",     # Change
            ).cuda()

        else:
            raise NotImplementedError("We don't support such model size in GRL model")
        
        # self.generator = torch.compile(self.generator).cuda()     # Don't use this for 3090Ti
        self.generator.train()

    
    def run(self):
        self.master_run()
                        

    
    def calculate_loss(self, gen_hr, imgs_hr):
        # Define the loss function here

        # Generator pixel loss (l1 loss):  generated vs. GT
        l_g_pix = self.cri_pix(gen_hr, imgs_hr, self.batch_idx)
        self.weight_store["pixel_loss"] = l_g_pix
        self.generator_loss += l_g_pix


    def tensorboard_report(self, iteration):
        # self.writer.add_scalar('Loss/train-Generator_Loss-Iteration', self.generator_loss, iteration)
        self.writer.add_scalar('Loss/train-Pixel_Loss-Iteration', self.weight_store["pixel_loss"], iteration)