File size: 10,243 Bytes
9ec3d0b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
import os, math, json, time, random
from dataclasses import dataclass
import unicodedata

import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.amp import autocast, GradScaler
from torch.utils.data import Dataset, DataLoader
from PIL import Image, ImageDraw, ImageFont
import textwrap
import numpy as np
from .tokenizer import Tokenizer

def _load_img(path, imgsz):
    if not isinstance(path, str): # PIL
        img = path.convert("RGB").resize((imgsz, imgsz))
        data = torch.from_numpy(np.array(img)).permute(2,0,1).unsqueeze(0).float() / 255.0            
        return data

    assert os.path.exists(path)
    img = np.array(Image.open(path).convert("RGB").resize((imgsz, imgsz)))    
    data = torch.from_numpy(img).permute(2,0,1).unsqueeze(0).float() / 255.0
    return data # [B, C, H, W]

def load_images(paths, imgsz, device=torch.device('cpu')):
    if isinstance(paths, dict): # For huggingface datasets my example "ituperceptron/image-captioning-turkish"
        try:
            return torch.cat([_load_img(p, imgsz) for p in paths["image"]], dim=0).to(device=device)
        except Exception as e:
            print(e)
            print("If you are using Huggingface dataset or something else, please implement your own Trainer.load_images and Trainer._load_img function.")
            exit(1)
    elif isinstance(paths[0], dict):
        return torch.cat([_load_img(p["image"], imgsz) for p in paths], dim=0).to(device=device)

    return torch.cat([_load_img(p, imgsz) for p in paths], dim=0).to(device=device)

def load_captions(captions:list[list[int]], max_len, device=torch.device('cpu')):
    """
    Used to apply teacher forcing for training
    """
    tin = []
    tout = []
    for cap in captions:
        cap = cap[:max_len]
        tin.append(cap[:-1])
        tout.append(cap[1:])
    
    tin = torch.tensor(tin).long().to(device=device)
    tout = torch.tensor(tout).long().to(device=device)
    return tin, tout

def remove_accents(text: str) -> str:    
    normalized = unicodedata.normalize("NFKD", text)    
    return "".join(ch for ch in normalized if not unicodedata.combining(ch))

def turkish_ascii(text: str) -> str:
    """
    Used to prediction captions visualization in Turkish sentences.
    """
    text = remove_accents(text)
    return text.replace("ı","i").replace("İ","I")

def save_pred(img: torch.Tensor, caption: str, save_path: str = "pred.png"):
    """
    This function is used to save image and caption for validation.
    """
    if img.ndim != 3 or img.shape[0] not in (1,3):
        raise ValueError(f"img shape must be [C,H,W] with C=1 or 3, got {tuple(img.shape)}")

    C, H, W = img.shape
    imgsz = W
    caption = turkish_ascii(caption)
    if C == 1:
        img = img.repeat(3, 1, 1)
    img_np = (img.clamp(0,1).permute(1,2,0) * 255.0).cpu().numpy().astype(np.uint8)

    panel_pil = Image.new("RGB", (imgsz, imgsz), (0,0,0))
    draw = ImageDraw.Draw(panel_pil)
    font = ImageFont.load_default(size=26)
    wrapped = textwrap.fill(caption, width=38)

    x0, y0 = 24, 24
    bbox = draw.multiline_textbbox((x0, y0), wrapped, font=font, spacing=6)
    pad = 12
    
    draw.rectangle([bbox[0]-pad, bbox[1]-pad, bbox[2]+pad, bbox[3]+pad], fill=(0,0,0))
    draw.multiline_text((x0, y0), wrapped, font=font, fill=(255,255,255),
                        spacing=6, stroke_width=2, stroke_fill=(0,0,0))
    
    panel_np = np.array(panel_pil)
    frame = np.concatenate([img_np, panel_np], axis=1)

    Image.fromarray(frame).save(save_path)
    

@dataclass
class TrainConfig:
    epoch:int = 20
    batch_size:int = 64
    lr:float = 2e-4
    weight_decay:float = 1e-2
    grad_clip:float = 1.0
    save_dir:str = "checkpoints"
    sentence_lim:int = 128
    max_seq_len:int = 128
    min_seq_len:int = 1
    avg_seq_len:int = 0
    std_seq_len:int = 0
    recommended_seq_len:int = 0

    def __str__(self):
        return f"""
        epoch: {self.epoch}
        batch_size: {self.batch_size}
        lr: {self.lr}
        weight_decay: {self.weight_decay}
        grad_clip: {self.grad_clip}
        save_dir: {self.save_dir}
        min sentece length: {self.min_seq_len}
        max sentece length: {self.max_seq_len}
        avg sentece length: {self.avg_seq_len:.2f}
        std sentece length: {self.std_seq_len:.2f}
        recommended sentece length: {self.recommended_seq_len}
        Training sentece limit: {self.sentence_lim}
        """


class Trainer:
    def __init__(self, model:nn.Module, device:torch.device=torch.device('cpu')):
        self.model  = model
        self.device = device    

    @torch.no_grad()
    def validate(self, 
                 imagepaths:list[str], 
                 tokenizer:Tokenizer
                 ) -> str:
        images = load_images(imagepaths, self.model.imgsz, device=self.device) # [B, C, H, W]
        preds = []
        for img in images:
            p = self.model.forward(img.unsqueeze(0)) # [B, S]
            preds.append(p.squeeze(0))
        
        for i,c in enumerate(preds):
            cap = tokenizer.decode(c.tolist())
            os.makedirs(f"{self.cfg.save_dir}/preds", exist_ok=True)
            save_pred(images[i], cap, os.path.join(f"{self.cfg.save_dir}/preds", f"{i}.png"))

    def fit(self,
            tokenizer:Tokenizer,
            imagepaths:list[str],
            epoch:int=20,
            batch_size:int=64,
            lr:float=2e-4,
            weight_decay:float=1e-2,
            grad_clip:float=1.0,
            save_dir:str="checkpoints",
            max_len = None
            ):
        len_limit = max_len if max_len is not None else tokenizer.recommended_seq_len
        self.cfg = TrainConfig(epoch=epoch,
                          batch_size=batch_size,
                          lr=lr,
                          weight_decay=weight_decay,
                          grad_clip=grad_clip,
                          save_dir=save_dir,
                          sentence_lim=len_limit,
                          max_seq_len=tokenizer.max_seq_len,
                          min_seq_len=tokenizer.min_seq_len,
                          avg_seq_len=tokenizer.avg_len,
                          std_seq_len=tokenizer.leng_std,
                          recommended_seq_len=tokenizer.recommended_seq_len)
        cfg = self.cfg
        print(cfg)
        
        os.makedirs(cfg.save_dir, exist_ok=True)

        self.captions = tokenizer.tokenized_captions

        global_step = 0
        total_step = cfg.epoch*math.ceil(len(self.captions)/cfg.batch_size)
        opt = torch.optim.AdamW([
            {'params': self.model.backbone.parameters(), 'lr': 1e-5},
            {'params': self.model.encoder.parameters()},
            {'params': self.model.decoder.parameters()},
        ], lr=cfg.lr, weight_decay=cfg.weight_decay)
        sched = torch.optim.lr_scheduler.CosineAnnealingLR(opt, T_max=total_step)
        criterion = nn.CrossEntropyLoss(ignore_index=int(self.model.pad_id))

        best_loss = float("inf")
        self.model.to(self.device)
        self.model.train(True)
        scaler = GradScaler(device=self.device.type, enabled=True)

        for ep in range(1, cfg.epoch+1):
            global_loss = 0.0
            seen = 0
            t0 = time.time()

            for batch_start_idx in range(0, len(self.captions), cfg.batch_size):
                batch_size = min(cfg.batch_size, len(self.captions)-batch_start_idx)

                paths = imagepaths[batch_start_idx:batch_start_idx+batch_size]
                captions = self.captions[batch_start_idx:batch_start_idx+batch_size]

                imgs = load_images(paths, imgsz=self.model.imgsz, device=self.device)
                tins, touts = load_captions(captions, len_limit, device=self.device) # 2x [B, T, V]

                with autocast(device_type="cuda", dtype=torch.float16):
                    logits = self.model.forward(imgs, tins)
                    loss = criterion.forward(logits.reshape(-1, logits.size(-1)), touts.reshape(-1))

                opt.zero_grad()
                scaler.scale(loss).backward()
                nn.utils.clip_grad_norm_(self.model.parameters(), cfg.grad_clip)
                old_scale = scaler.get_scale()
                scaler.step(opt)
                scaler.update()
                stepped = scaler.get_scale() >= old_scale

                if stepped:
                    sched.step()

                bsz = imgs.size(0)
                global_loss += loss.item() * bsz
                seen += bsz
                global_step += 1
                print(f"[%{100*((batch_start_idx+batch_size)/len(self.captions)):>6.2f}]Epoch {ep:03d} loss={global_loss/seen:<6.3f} lr={sched.get_last_lr()[-1]:<.6f}", end="\r")

            random_idx:list = np.random.randint(0, len(self.captions), cfg.batch_size).tolist()
            paths = []
            for idx in random_idx:
                paths.append(imagepaths[idx])
            
            self.model.train(False)
            self.validate(paths, tokenizer)
            self.model.train(True)

            ep_loss = global_loss/seen
            dt = time.time()-t0
            print(f"[EPOCH {ep:03d}] loss={ep_loss:.4f}  time={dt:.1f}s")

            ckpt = {
                "model_name": self.model.model_name if self.model is not None else "ViT Transformer",
                "model_state_dict": self.model.state_dict(),
                "vocab": tokenizer.vocap,
                "imgsz": self.model.imgsz,
                "dim": self.model.dim,
                "encoder_depth": self.model.encoder_depth,
                "decoder_depth": self.model.decoder_depth,
                "encoder_num_heads": self.model.encoder_num_heads,
                "decoder_num_heads": self.model.decoder_num_heads,
                "last_avg_loss": ep_loss
            }

            last_path = os.path.join(cfg.save_dir, "last.pt")
            torch.save(ckpt, last_path)
            if ep_loss < best_loss:
                best_loss = ep_loss
                torch.save(ckpt, os.path.join(cfg.save_dir, "best.pt"))

        print(f"[TRAIN DONE] best_loss={best_loss:.4f}  ckpts => {cfg.save_dir}")
        return {"best_loss": best_loss, "save_dir": cfg.save_dir}