init

README.md

@@ -0,0 +1 @@
+# BLIP

configs/bert_config.json

@@ -0,0 +1,21 @@
+{
+  "architectures": [
+    "BertForMaskedLM"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 768,
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "layer_norm_eps": 1e-12,
+  "max_position_embeddings": 512,
+  "model_type": "bert",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 12,
+  "pad_token_id": 0,
+  "type_vocab_size": 2,
+  "vocab_size": 30522,
+  "encoder_width": 768,
+  "add_cross_attention": true   
+}

configs/caption_coco.yaml

@@ -0,0 +1,33 @@
+image_root: '/export/share/datasets/vision/coco/images/'
+ann_root: 'annotation'
+coco_gt_root: 'annotation/coco_gt'
+
+# set pretrained as a file path or an url
+pretrained: 'https://storage.googleapis.com/sfr-vision-language-research/BLIP/models/model*_base_caption.pth'
+
+# size of vit model; base or large
+vit: 'base'
+vit_grad_ckpt: False
+vit_ckpt_layer: 0
+batch_size: 32
+init_lr: 1e-5
+
+# vit: 'large'
+# vit_grad_ckpt: True
+# vit_ckpt_layer: 5
+# batch_size: 16
+# init_lr: 2e-6
+
+image_size: 384
+
+# generation configs
+max_length: 20  
+min_length: 5
+num_beams: 3
+prompt: 'a picture of '
+
+# optimizer
+weight_decay: 0.05
+min_lr: 0
+max_epoch: 5
+

configs/med_config.json

@@ -0,0 +1,21 @@
+{
+  "architectures": [
+    "BertModel"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 768,
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "layer_norm_eps": 1e-12,
+  "max_position_embeddings": 512,
+  "model_type": "bert",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 12,
+  "pad_token_id": 0,
+  "type_vocab_size": 2,
+  "vocab_size": 30524,
+  "encoder_width": 768,
+  "add_cross_attention": true   
+}

configs/nlvr.yaml

@@ -0,0 +1,21 @@
+image_root: '/export/share/datasets/vision/NLVR2/' 
+ann_root: 'annotation'
+
+# set pretrained as a file path or an url
+pretrained: 'https://storage.googleapis.com/sfr-vision-language-research/BLIP/models/model_base_nlvr.pth'
+
+#size of vit model; base or large
+vit: 'base'
+batch_size_train: 16 
+batch_size_test: 64 
+vit_grad_ckpt: False
+vit_ckpt_layer: 0
+max_epoch: 15
+
+image_size: 384
+
+# optimizer
+weight_decay: 0.05
+init_lr: 3e-5
+min_lr: 0
+

configs/nocaps.yaml

@@ -0,0 +1,15 @@
+image_root: '/export/share/datasets/vision/nocaps/'
+ann_root: 'annotation'
+
+# set pretrained as a file path or an url
+pretrained: 'https://storage.googleapis.com/sfr-vision-language-research/BLIP/models/model*_base_caption.pth'
+
+vit: 'base'
+batch_size: 32
+
+image_size: 384
+
+max_length: 20
+min_length: 5
+num_beams: 3
+prompt: 'a picture of '

configs/pretrain.yaml

@@ -0,0 +1,27 @@
+train_file: ['/export/share/junnan-li/VL_pretrain/annotation/coco_karpathy_train.json',
+             '/export/share/junnan-li/VL_pretrain/annotation/vg_caption.json',
+             ]
+laion_path: ''   
+
+# size of vit model; base or large
+vit: 'base'
+vit_grad_ckpt: False
+vit_ckpt_layer: 0
+
+image_size: 224
+batch_size: 75
+
+queue_size: 57600
+alpha: 0.4
+
+# optimizer
+weight_decay: 0.05
+init_lr: 3e-4
+min_lr: 1e-6
+warmup_lr: 1e-6
+lr_decay_rate: 0.9
+max_epoch: 20
+warmup_steps: 3000
+
+
+

configs/retrieval_coco.yaml

@@ -0,0 +1,34 @@
+image_root: '/export/share/datasets/vision/coco/images/'
+ann_root: 'annotation'
+dataset: 'coco'
+
+# set pretrained as a file path or an url
+pretrained: 'https://storage.googleapis.com/sfr-vision-language-research/BLIP/models/model_base_retrieval_coco.pth'
+
+# size of vit model; base or large
+
+vit: 'base'
+batch_size_train: 32
+batch_size_test: 64
+vit_grad_ckpt: True
+vit_ckpt_layer: 4
+init_lr: 1e-5
+
+# vit: 'large'
+# batch_size_train: 16
+# batch_size_test: 32
+# vit_grad_ckpt: True
+# vit_ckpt_layer: 12
+# init_lr: 5e-6
+
+image_size: 384
+queue_size: 57600
+alpha: 0.4
+k_test: 256
+negative_all_rank: True
+
+# optimizer
+weight_decay: 0.05
+min_lr: 0
+max_epoch: 6
+

configs/retrieval_flickr.yaml

@@ -0,0 +1,34 @@
+image_root: '/export/share/datasets/vision/flickr30k/'
+ann_root: 'annotation'
+dataset: 'flickr'
+
+# set pretrained as a file path or an url
+pretrained: 'https://storage.googleapis.com/sfr-vision-language-research/BLIP/models/model_base_retrieval_flickr.pth'
+
+# size of vit model; base or large
+
+vit: 'base'
+batch_size_train: 32
+batch_size_test: 64
+vit_grad_ckpt: True
+vit_ckpt_layer: 4
+init_lr: 1e-5
+
+# vit: 'large'
+# batch_size_train: 16
+# batch_size_test: 32
+# vit_grad_ckpt: True
+# vit_ckpt_layer: 10
+# init_lr: 5e-6
+
+image_size: 384
+queue_size: 57600
+alpha: 0.4
+k_test: 128
+negative_all_rank: False
+
+# optimizer
+weight_decay: 0.05
+min_lr: 0
+max_epoch: 6
+

configs/vqa.yaml

@@ -0,0 +1,25 @@
+vqa_root: '/export/share/datasets/vision/VQA/Images/mscoco/' #followed by train2014/
+vg_root: '/export/share/datasets/vision/visual-genome/'  #followed by image/
+train_files: ['vqa_train','vqa_val','vg_qa']
+ann_root: 'annotation'
+
+# set pretrained as a file path or an url
+pretrained: 'https://storage.googleapis.com/sfr-vision-language-research/BLIP/models/model*_vqa.pth'
+
+# size of vit model; base or large
+vit: 'base'
+batch_size_train: 16 
+batch_size_test: 32 
+vit_grad_ckpt: False
+vit_ckpt_layer: 0
+init_lr: 2e-5
+
+image_size: 480
+
+k_test: 128
+inference: 'rank'
+
+# optimizer
+weight_decay: 0.05
+min_lr: 0
+max_epoch: 10

eval_nocaps.py

@@ -0,0 +1,118 @@
+'''
+ * Copyright (c) 2022, salesforce.com, inc.
+ * All rights reserved.
+ * SPDX-License-Identifier: BSD-3-Clause
+ * For full license text, see LICENSE.txt file in the repo root or https://opensource.org/licenses/BSD-3-Clause
+ * By Junnan Li
+'''
+import argparse
+import os
+import ruamel_yaml as yaml
+import numpy as np
+import random
+import time
+import datetime
+import json
+from pathlib import Path
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.backends.cudnn as cudnn
+import torch.distributed as dist
+from torch.utils.data import DataLoader
+
+from models.blip import blip_decoder
+import utils
+from data import create_dataset, create_sampler, create_loader
+from data.utils import save_result
+
+@torch.no_grad()
+def evaluate(model, data_loader, device, config):
+    # evaluate
+    model.eval() 
+
+    metric_logger = utils.MetricLogger(delimiter="  ")
+    header = 'Evaluation:'
+    print_freq = 10
+
+    result = []
+    for image, image_id in metric_logger.log_every(data_loader, print_freq, header): 
+
+        image = image.to(device)       
+
+        captions = model.generate(image, sample=False, num_beams=config['num_beams'], max_length=config['max_length'], 
+                                  min_length=config['min_length'], repetition_penalty=1.1)
+
+        for caption, img_id in zip(captions, image_id):
+            result.append({"image_id": img_id.item(), "caption": caption})
+
+    return result
+
+
+def main(args, config):
+    utils.init_distributed_mode(args)    
+
+    device = torch.device(args.device)
+
+    # fix the seed for reproducibility
+    seed = args.seed + utils.get_rank()
+    torch.manual_seed(seed)
+    np.random.seed(seed)
+    random.seed(seed)
+    cudnn.benchmark = True
+
+    #### Dataset #### 
+    print("Creating captioning dataset")
+    val_dataset, test_dataset = create_dataset('nocaps', config)  
+
+    if args.distributed:
+        num_tasks = utils.get_world_size()
+        global_rank = utils.get_rank()            
+        samplers = create_sampler([val_dataset,test_dataset], [False,False], num_tasks, global_rank)         
+    else:
+        samplers = [None,None]
+
+    val_loader, test_loader = create_loader([val_dataset, test_dataset],samplers, 
+                                            batch_size=[config['batch_size']]*2,num_workers=[4,4],
+                                            is_trains=[False, False], collate_fns=[None,None])        
+
+    #### Model #### 
+    print("Creating model")
+    model = blip_decoder(pretrained=config['pretrained'], image_size=config['image_size'], vit=config['vit'], 
+                           prompt=config['prompt'])
+
+    model = model.to(device)   
+
+    model_without_ddp = model
+    if args.distributed:
+        model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.gpu])
+        model_without_ddp = model.module    
+
+    val_result = evaluate(model_without_ddp, val_loader, device, config)  
+    val_result_file = save_result(val_result, args.result_dir, 'val', remove_duplicate='image_id')   
+    test_result = evaluate(model_without_ddp, test_loader, device, config)  
+    test_result_file = save_result(test_result, args.result_dir, 'test', remove_duplicate='image_id') 
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--config', default='./configs/nocaps.yaml')
+    parser.add_argument('--output_dir', default='output/NoCaps')        
+    parser.add_argument('--device', default='cuda')
+    parser.add_argument('--seed', default=42, type=int)
+    parser.add_argument('--world_size', default=1, type=int, help='number of distributed processes')    
+    parser.add_argument('--dist_url', default='env://', help='url used to set up distributed training')
+    parser.add_argument('--distributed', default=True, type=bool)
+    args = parser.parse_args()
+
+    config = yaml.load(open(args.config, 'r'), Loader=yaml.Loader)
+
+    args.result_dir = os.path.join(args.output_dir, 'result')
+
+    Path(args.output_dir).mkdir(parents=True, exist_ok=True)
+    Path(args.result_dir).mkdir(parents=True, exist_ok=True)
+
+    yaml.dump(config, open(os.path.join(args.output_dir, 'config.yaml'), 'w'))    
+
+    main(args, config)