English
File size: 11,322 Bytes
6eff0e6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
from pathlib import Path
import comet_ml
import datasets
import evaluate
import lightning as L
import torch
from timm import create_model, data
from tokenizers import Tokenizer
from torch import nn
from torch.utils.data import DataLoader
from transformers import (
    GPT2LMHeadModel,
)
from lightning.pytorch.loggers import TensorBoardLogger
from lightning.pytorch.callbacks import ModelCheckpoint


eos_token_id = 50256  # obtained from gpt model


class Projection(nn.Module):
    def __init__(self, in_features, out_features):
        super().__init__()
        self.network = nn.Sequential(
            nn.Linear(in_features, in_features * 3),
            nn.GELU(),
            nn.Linear(in_features * 3, out_features),
        )

    def forward(self, input):
        return self.network(input)


class ImageNetCaptionModel(L.LightningModule):
    def __init__(self):
        super().__init__()
        # backbone model to extract image feature token
        self.backbone = create_model(
            "vit_mediumd_patch16_reg4_gap_384", pretrained=True
        )
        self.llm = GPT2LMHeadModel.from_pretrained("gpt2")

        self.image_start_token = "<image_start>"
        self.image_end_token = "<image_end>"
        self.tokenizer = Tokenizer.from_pretrained("gpt2")
        self.tokenizer.add_special_tokens(
            [self.image_start_token, self.image_end_token]
        )
        self.image_start_token_id = self.tokenizer.token_to_id(self.image_start_token)
        self.image_end_token_id = self.tokenizer.token_to_id(self.image_end_token)
        self.eos_token = eos_token_id

        self.llm.resize_token_embeddings(self.tokenizer.get_vocab_size())
        self.embedding = self.llm.get_input_embeddings()

        self.projection = Projection(
            in_features=512, out_features=self.llm.config.hidden_size
        )

        self.bleu_metric = evaluate.load("bleu")
        self.meteor_metric = evaluate.load("meteor")

        ## freeze backbone and gpt models.
        for param in self.backbone.parameters():
            param.requires_grad = False

        for param in self.llm.parameters():
            param.requires_grad = True

    def get_tokenizer(self):
        return self.tokenizer

    def forward(self, image=None, input_caption=None, **kwargs):
        image_feature = self.backbone.forward_features(image)
        projection = self.projection(image_feature)
        input_caption_embedding = self.embedding(input=input_caption)

        # concat start_image_token + projection + end_image_token + input_caption
        image_start_token, image_end_token = self.get_image_seperation_token(
            image=image
        )
        input_embedding = torch.cat(
            [image_start_token, projection, image_end_token, input_caption_embedding],
            dim=1,
        )
        attention_mask = torch.ones(
            input_embedding.size()[:-1], dtype=torch.long, device=image.device
        )

        labels = torch.full(
            (input_embedding.size(0), input_embedding.size(1)),
            -100,
            dtype=torch.long,
            device=image.device,
        )
        labels[:, projection.size(1) + 2 :] = input_caption  # align text labels

        llm_output = self.llm(
            inputs_embeds=input_embedding, attention_mask=attention_mask, labels=labels
        )
        return llm_output

    def training_step(self, batch, batch_idx):
        output = self.forward(**batch)
        self.log("loss", output.loss.item())
        return output.loss

    def validation_step(self, batch, batch_idx):
        if batch_idx < 5:
            pred = self.predict_step(batch=batch, batch_idx=batch_idx)
            print(
                "evaluation ",
                "pred",
                pred,
                "original caption",
                batch["original_caption_enriched"],
            )
            bleu = self.bleu_metric.compute(
                predictions=pred, references=batch["original_caption_enriched"]
            )
            self.log("bleu", bleu["bleu"])
            self.log("precision", bleu["brevity_penalty"])
            metor = self.meteor_metric.compute(
                predictions=pred, references=batch["original_caption_enriched"]
            )
            print(metor)
            self.log_dict(metor)

    def get_image_seperation_token(self, image):
        image_start_embedding = self.embedding(
            torch.tensor([self.image_start_token_id], device=image.device)
        )
        image_end_embedding = self.embedding(
            torch.tensor([self.image_end_token_id], device=image.device)
        )
        image_start_token = image_start_embedding.unsqueeze(0).repeat(len(image), 1, 1)
        image_end_token = image_end_embedding.unsqueeze(0).repeat(len(image), 1, 1)

        return image_start_token, image_end_token

    def configure_optimizers(self):
        proj_params = [p for p in self.projection.parameters() if p.requires_grad]
        llm_params = [p for p in self.llm.parameters() if p.requires_grad]

        optimizer = torch.optim.AdamW(
            [
                {"params": proj_params, "lr": 1e-4, "weight_decay": 0.01},
                {"params": llm_params, "lr": 5e-6, "weight_decay": 0.01},
            ]
        )

        return optimizer

    def predict_step(self, batch, batch_idx, dataloader_idx=0):
        image = batch["image"]
        image_feature = self.backbone.forward_features(image)
        projection = self.projection(image_feature)

        image_start_embedding = self.embedding(
            torch.tensor([self.image_start_token_id], device=image.device)
        )
        image_end_embedding = self.embedding(
            torch.tensor([self.image_end_token_id], device=image.device)
        )
        input_start_image_embedding_batch = image_start_embedding.unsqueeze(0).repeat(
            len(image), 1, 1
        )
        input_end_image_embedding_batch = image_end_embedding.unsqueeze(0).repeat(
            len(image), 1, 1
        )

        input_embedding = torch.cat(
            [
                input_start_image_embedding_batch,
                projection,
                input_end_image_embedding_batch,
            ],
            dim=1,
        )
        attention_mask = torch.ones(
            input_embedding.size()[:-1], dtype=torch.long, device=image.device
        )

        outputs = self.llm.generate(
            inputs_embeds=input_embedding,
            attention_mask=attention_mask,
            eos_token_id=0,
            max_new_tokens=30,
            do_sample=True,  # add randomness
            top_p=0.9,  # nucleus sampling
            temperature=0.7,
        )

        # Convert tensor to list of lists for decode_batch
        if outputs.dim() == 2:
            # outputs is [batch_size, sequence_length], convert to list of lists
            outputs_list = outputs.tolist()
        else:
            # outputs is already a list/sequence
            outputs_list = outputs

        return self.tokenizer.decode_batch(outputs_list, skip_special_tokens=True)
    
    def generate(self, image):
        data_config = data.resolve_model_data_config(
        create_model("vit_mediumd_patch16_reg4_gap_384", pretrained=True)
        )
        transforms = data.create_transform(**data_config, is_training=False)
        image = transforms(image)
        
        return self.predict_step(batch={"image":image.unsqueeze(0)},batch_idx=0)[0]
        


def collate_fn(batch):
    collected = {"image": [], "input_caption": [], "original_caption_enriched": []}

    for data in batch:
        collected["image"].append(torch.tensor(data["image"], dtype=torch.float))
        collected["input_caption"].append(
            torch.tensor(data["input_caption"], dtype=torch.long)
        )
        collected["original_caption_enriched"].append(data["original_caption_enriched"])

    return {
        "image": torch.stack(collected["image"], dim=0),
        "input_caption": torch.stack(collected["input_caption"], dim=0),
        "original_caption_enriched": collected["original_caption_enriched"],
    }


def agument(tokenizer: Tokenizer):
    data_config = data.resolve_model_data_config(
        create_model("vit_mediumd_patch16_reg4_gap_384", pretrained=True)
    )
    transforms = data.create_transform(**data_config, is_training=False)

    def transform(data):
        ids = tokenizer.encode(data["caption_enriched"])

        # Handle sequences based on length
        if len(ids.ids) <= 59:
            # For short sequences, just append EOS
            ids.ids.append(eos_token_id)
        else:
            # For long sequences, truncate to 59 tokens and append EOS
            ids.ids = ids.ids[:59]
            ids.ids.append(eos_token_id)

            # Pad to exactly 60 tokens
            ids.ids = ids.ids[:60]  # Ensure we don't exceed 60
        ids.pad(60)

        decoded = tokenizer.decode(ids.ids, skip_special_tokens=True)
        print("original", data["caption_enriched"], "decoded", decoded)

        data["input_caption"] = torch.tensor(ids.ids, dtype=torch.long)

        data["original_caption_enriched"] = data["caption_enriched"]
        data["image"] = transforms(data["image"])
        return data

    return transform


def is_valid_image(example):
    try:
        # Try opening the image
        if example["image"].mode == "RGB":
            return True

        return False
    except Exception as e:
        # ValueError will catch the MAX_TEXT_CHUNK error
        print("false", example["image"])
        print("Exception:", e)
        return False


def train(
    root_path: Path,
    dataset: datasets.Dataset,
    num_loader_worker: int = 0,
    batch_size=16,
    logger=None,
):
    # dataset = datasets.load_dataset("visual-layer/imagenet-1k-vl-enriched", split="validation").shuffle(seed=42).select(range(20000)).train_test_split(test_size=0.1)
    test_ds = dataset["test"]
    train_ds = dataset["train"]

    model = ImageNetCaptionModel()

    tokenizer = model.get_tokenizer()

    # Apply transformation to both datasets
    train_ds = train_ds.filter(is_valid_image)
    train_ds = train_ds.map(agument(tokenizer=tokenizer))

    test_ds = test_ds.filter(is_valid_image)
    test_ds = test_ds.map(agument(tokenizer=tokenizer))

    train_data_loader = DataLoader(
        dataset=train_ds,
        drop_last=True,
        batch_size=batch_size,
        collate_fn=collate_fn,
        num_workers=num_loader_worker,
    )
    evaluation_data_loader = DataLoader(
        dataset=test_ds,
        drop_last=True,
        batch_size=batch_size,
        collate_fn=collate_fn,
        num_workers=num_loader_worker,
    )

    if logger is None:
        logger = TensorBoardLogger(save_dir=str(root_path), version=1, name="logs")
    checkpoint_callback = ModelCheckpoint(
        dirpath=root_path / "checkpoint",
        filename="checkpoint-{epoch:02d}-{loss:.2f}",
        every_n_epochs=1,
        save_top_k=-1,
    )
    print("path", root_path)
    trainer = L.Trainer(
        logger=logger,
        max_epochs=2,
        default_root_dir=root_path,
        callbacks=[checkpoint_callback],
    )
    trainer.fit(
        model=model,
        train_dataloaders=train_data_loader,
        val_dataloaders=evaluation_data_loader,
    )