File size: 14,911 Bytes
cacafc1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
'''
LinCIR
Copyright (c) 2023-present NAVER Corp.
CC BY-NC-4.0 (https://creativecommons.org/licenses/by-nc/4.0/)
'''
import json
import os
import pickle
import random
import math
from argparse import ArgumentParser
from pathlib import Path
from typing import Literal, Tuple, Dict, List, Set
import logging

import numpy as np
import torch
import torch.nn.functional as F
from tqdm import tqdm

from loader import build_loader, CIRRDataset
from encode_with_pseudo_tokens import encode_with_pseudo_tokens_HF
from models import build_text_encoder, Phi, EMAModel
from utils import extract_image_features, extract_pseudo_tokens_with_phi
from validate import cirr_compute_val_metrics

import transformers
from transformers import get_scheduler
from accelerate import Accelerator, DeepSpeedPlugin
from accelerate.logging import get_logger
from accelerate.utils import set_seed
from accelerate.state import AcceleratorState
from accelerate.logging import get_logger


logger = get_logger(__name__)


def parse_args():
    parser = ArgumentParser()

    parser.add_argument("--output_dir", default="trained_models", type=str,
                        help="The output directory where the model predictions and checkpoints will be written")
    parser.add_argument("--logging_dir", default="logs", type=str, help="tensorboard logs will saved here")
    parser.add_argument("--cache_dir", default="./hf_models", type=str,
                        help="Path to model cache folder")
    parser.add_argument("--report_to", default="tensorboard", type=str, help="")
    parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")

    parser.add_argument("--clip_model_name", default="giga", type=str,
                        help="CLIP model to use, e.g 'large', 'giga'")
    parser.add_argument("--cirr_dataset_path", type=str, help="Path to CIRR dataset", required=True)
    parser.add_argument("--keywords_path", type=str, help="Path to keywords json file")
    parser.add_argument("--resume", default=None, type=str, help="Path to pretrained ckpt")

    parser.add_argument("--use_8bit_adam", action="store_true", help="Whether or not to use 8-bit Adam from bitsandbytes.")
    parser.add_argument("--lr_scheduler", type=str, default="constant",
                        choices=["linear", "cosine", "cosine_with_restarts", "polynomial", "constant", "constant_with_warmup"],
                        help="")
    parser.add_argument("--lr_warmup_steps", type=int, default=500, help="Number of steps for the warmup in the lr scheduler.")
    parser.add_argument("--max_train_steps", type=int, default=50000, help="Total number of training steps to perform")
    parser.add_argument("--phi_dropout", default=0.5, type=float, help="Dropout probability for the phi network")
    parser.add_argument("--l2_normalize", action="store_true", help="Whether or not to use l2 normalization")
    parser.add_argument("--batch_size", default=256, type=int, help="Phi training batch size")
    parser.add_argument("--num_workers", default=10, type=int, help="Number of workers")
    parser.add_argument("--learning_rate", default=1e-4, type=float, help="Learning rate")
    parser.add_argument("--weight_decay", type=float, default=0.01, help="Weight decay")
    parser.add_argument("--gradient_accumulation_steps", default=1, type=int, help="Number of updates steps to accumulate before performing a backward/update pass")
    parser.add_argument("--max_grad_norm", default=None, type=float, help="Max gradient norm.")
    parser.add_argument("--mixed_precision", default=None, type=str, choices=["no", "fp16", "bf16"], help="mixed precision")
    parser.add_argument("--validation_steps", default=1, type=int, help="Validation frequency expressed in epochs")
    parser.add_argument("--checkpointing_steps", default=None, type=int, help="Save a checkpoint of the training state every X updates")
    parser.add_argument("--use_ema", action="store_true", help="Whether to use EMA model.")

    parser.add_argument("--seed", type=int, default=None, help="seed for reproducibility")

    args = parser.parse_args()

    env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
    if env_local_rank != -1 and env_local_rank != args.local_rank:
        args.local_rank = env_local_rank

    return args


def save_phi(name: str, cur_epoch: int, model_to_save: Phi, training_path: Path) -> None:
    """
    Save the weights of Phi during training
    """
    models_path = os.path.join(training_path, "checkpoints")
    os.makedirs(models_path, exist_ok=True)
    model_name = model_to_save.__class__.__name__
    torch.save({
        'epoch': cur_epoch,
        model_name: model_to_save.state_dict(),
    }, os.path.join(models_path, f'{name}.pt'))


def train_phi(args):
    # We are going to use the pre-extracted clip image features. so we do not need image_encoder anymore.

    ### init accelerator here
    logging_dir = os.path.join(args.output_dir, args.logging_dir)
    accelerator = Accelerator(
            gradient_accumulation_steps=args.gradient_accumulation_steps,
            mixed_precision=args.mixed_precision,
            log_with=args.report_to,
            project_dir=logging_dir,
    )

    os.makedirs(args.output_dir, exist_ok=True)

    logging.basicConfig(
            format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
            datefmt="%m/%d/%Y %H:%M:%S",
            level=logging.INFO,
    )
    logger.info(accelerator.state, main_process_only=False)

    if accelerator.is_local_main_process:
        transformers.utils.logging.set_verbosity_info()
    else:
        transformers.utils.logging.set_verbosity_error()

    if args.seed is not None:
        set_seed(args.seed)

    ### Define the text encoder from clip
    image_encoder, clip_preprocess, text_encoder, tokenizer = build_text_encoder(args)

    ### Define the phi model
    phi = Phi(input_dim=text_encoder.config.projection_dim,
                    hidden_dim=text_encoder.config.projection_dim * 4,
                    output_dim=text_encoder.config.hidden_size, dropout=args.phi_dropout)

    if args.resume:
        phi.load_state_dict(
                torch.load(args.resume, map_location=accelerator.device)[
                phi.__class__.__name__])


    ### GPU handling
    weight_dtype = torch.float32
    if accelerator.mixed_precision == "fp16":
        weight_dtype = torch.float16
    elif accelerator.mixed_precision == "bf16":
        weight_dtype = torch.bfloat16

    image_encoder.to(accelerator.device, dtype=weight_dtype)
    text_encoder.to(accelerator.device, dtype=weight_dtype)

    image_encoder.requires_grad_(False)
    text_encoder.requires_grad_(False)

    if args.use_ema:
        import copy
        ema_phi = copy.deepcopy(phi)
        ema_phi = EMAModel(ema_phi.parameters())
        ema_phi.to(accelerator.device, dtype=weight_dtype)

    ### Define the train datasets
    print('pytorch loader')
    train_dataset = build_loader(args, tokenizer, accelerator)

    ## evaluator
    if accelerator.is_main_process:
        ## Define CIRR validation set
        cirr_relative_val_dataset = CIRRDataset(args.cirr_dataset_path, 'val', 'relative', clip_preprocess)
        cirr_classic_val_dataset = CIRRDataset(args.cirr_dataset_path, 'val', 'classic', clip_preprocess)

        # Extract the features for the CIRR validation set
        cirr_val_index_features, cirr_val_index_names = extract_image_features(cirr_classic_val_dataset, image_encoder)

    # Define the optimizer, the loss and the grad scaler
    if args.use_8bit_adam:
        try:
            import bitsandbytes as bnb
        except ImportError:
            raise ImportError(
                    "Please install bitsandbytes to use 8-bit Adam. You can do so by running `pip install bitsandbytes`"
            )
        optimizer_cls = bnb.optim.AdamW8bit
    else:
        optimizer_cls = torch.optim.AdamW

    optimizer = optimizer_cls(phi.parameters(),
                              lr=args.learning_rate,
                              weight_decay=args.weight_decay)

    lr_scheduler = get_scheduler(
            args.lr_scheduler,
            optimizer=optimizer,
            num_warmup_steps=args.lr_warmup_steps * args.gradient_accumulation_steps * accelerator.num_processes,
            num_training_steps=args.max_train_steps * args.gradient_accumulation_steps * accelerator.num_processes,
    )

    phi, optimizer, lr_scheduler, train_dataset = accelerator.prepare(
            phi, optimizer, lr_scheduler, train_dataset
    )

    if accelerator.is_main_process:
        accelerator.init_trackers("zeroshot-cir", config=vars(args))

    # Start with the training loop
    total_batch_size = args.batch_size * accelerator.num_processes * args.gradient_accumulation_steps

    logger.info("***** Running training *****")
    logger.info(f"  Instantaneous batch size per device = {args.batch_size}")
    logger.info(f"  Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
    logger.info(f"  Gradient Accumulation steps = {args.gradient_accumulation_steps}")
    logger.info(f"  Total steps = {args.max_train_steps}")

    phi.train()

    train_loss = 0.0
    global_step = 0
    best_recall = -1

    progress_bar = tqdm(range(global_step, args.max_train_steps), disable=not accelerator.is_local_main_process)
    progress_bar.set_description("Steps")

    while True:
        for idx, (original_tokens, replaced_tokens, indicators) in enumerate(train_dataset):
            original_tokens = original_tokens.to(accelerator.device)
            replaced_tokens = replaced_tokens.to(accelerator.device)

            org = text_encoder(input_ids=original_tokens)
            original_text_embeddings, original_last_hidden_states = org.text_embeds, org.last_hidden_state
            input_features = original_text_embeddings.clone()
            input_features += 1.0 * torch.rand(input_features.shape[0], device=input_features.device).unsqueeze(-1) * torch.randn(input_features.shape, device=input_features.device)

            # normalize test
            if args.l2_normalize:
                input_features = F.normalize(input_features, dim=-1)
            #################

            estimated_token_embeddings = phi(input_features)

            replaced_text_embeddings, replaced_last_hidden_states = encode_with_pseudo_tokens_HF(text_encoder, replaced_tokens, estimated_token_embeddings, return_last_states=True)

            loss = F.mse_loss(replaced_text_embeddings.float(), original_text_embeddings.float(), reduction="mean")

            avg_loss = accelerator.gather(loss.repeat(args.batch_size)).mean()
            train_loss += avg_loss.item() / args.gradient_accumulation_steps

            # Backpropagation
            accelerator.backward(loss)
            if accelerator.sync_gradients and args.max_grad_norm is not None:
                accelerator.clip_grad_norm_(phi.parameters(), arg.max_grad_norm)
            optimizer.step()
            lr_scheduler.step()
            optimizer.zero_grad()
            
            if accelerator.sync_gradients:
                if args.use_ema:
                    ema_phi.step(phi.module.parameters())
                progress_bar.update(1)
                global_step += 1
                accelerator.log({"train/train_loss": train_loss}, step=global_step)
                train_loss = 0.0

                accelerator.log({'train/lr': lr_scheduler.get_last_lr()[0]}, step=global_step)
                accelerator.log({'train/preproc_rate': torch.sum(indicators).item() / len(indicators)}, step=global_step)
                if args.checkpointing_steps and global_step % args.checkpointing_steps == 0:
                    if accelerator.is_main_process:
                        logger.info(f"model saving... step: {global_step}")
                        save_phi(f"phi_{global_step:09}", global_step, accelerator.unwrap_model(phi), args.output_dir)
                        save_phi(f"phi_latest", global_step, accelerator.unwrap_model(phi), args.output_dir)
                    if args.use_ema:
                        phi_for_saving = copy.deepcopy(accelerator.unwrap_model(phi))
                        ema_phi.copy_to(phi_for_saving.parameters())
                        save_phi(f"ema_phi_{global_step:09}", global_step, phi_for_saving, args.output_dir)
                        save_phi(f"ema_phi_latest", global_step, phi_for_saving, args.output_dir)

                if global_step % args.validation_steps == 0 or global_step == 50:
                    if accelerator.is_main_process:
                        logger.info(f"evaluate model... step: {global_step}")

                        if args.use_ema:
                            phi_for_eval = copy.deepcopy(accelerator.unwrap_model(phi))
                            ema_phi.copy_to(phi_for_eval.parameters())
                        else:
                            phi_for_eval = phi

                        phi_for_eval.eval()

                        # Extract the pseudo tokens for the CIRR validation set using Phi
                        cirr_val_pseudo_tokens, cirr_val_ref_names_list = extract_pseudo_tokens_with_phi(image_encoder, phi_for_eval,
                                                                                                         cirr_relative_val_dataset, args)
                        cirr_val_pseudo_tokens = cirr_val_pseudo_tokens.to(accelerator.device)

                        # Compute the CIRR validation metrics
                        cirr_results_dict = cirr_compute_val_metrics(cirr_relative_val_dataset, text_encoder,
                                                                     cirr_val_index_features, cirr_val_index_names,
                                                                     cirr_val_ref_names_list, cirr_val_pseudo_tokens)
                        check_list = ['cirr_recall_at1', 'cirr_recall_at5', 'cirr_recall_at10', 'cirr_recall_at50']
                        for check_key in check_list:
                            accelerator.log({f"validate/{check_key}": cirr_results_dict[check_key]}, step=global_step)
                        print(json.dumps(cirr_results_dict, indent=4))

                        # Save the best model.
                        if args.checkpointing_steps:
                            if cirr_results_dict['cirr_recall_at1'] > best_recall:
                                best_recall = cirr_results_dict['cirr_recall_at1']
                                logger.info(f"best model saving... step: {global_step}")
                                save_phi("phi_best", global_step, accelerator.unwrap_model(phi), args.output_dir)

                        phi.train()

            if global_step >= args.max_train_steps:
                break


if __name__ == '__main__':
    args = parse_args()

    train_phi(args)