upload code

LICENSE.txt

@@ -0,0 +1,14 @@
+BSD 3-Clause License
+
+Copyright (c) 2022 Salesforce, Inc.
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met:
+
+1. Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer.
+
+2. Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution.
+
+3. Neither the name of Salesforce.com nor the names of its contributors may be used to endorse or promote products derived from this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

MANIFEST.in

@@ -0,0 +1,7 @@
+recursive-include lavis/configs *.yaml *.json
+recursive-include lavis/projects *.yaml *.json
+
+recursive-exclude lavis/datasets/download_scripts *
+recursive-exclude lavis/output *
+
+include requirements.txt

README.md

@@ -1,12 +1,9 @@
 # Bootstrapping Vision-Language Learning with Decoupled Language Pre-training
 
-
-This repo covers implementations of BLIP2 + Pformer in **Bootstrapping Vision-Language Learning with Decoupled Language Pre-training** by [Yiren Jian](https://yiren-jian.github.io/), [Chongyang Gao](https://gcyzsl.github.io/) and [Soroush Vosoughi](https://www.cs.dartmouth.edu/~soroush/).
+This repo covers implementations of BLIP2 with Pformer in **[Bootstrapping Vision-Language Learning with Decoupled Language Pre-training](https://arxiv.org/abs/2307.07063)**. The paper is accepted to NeurIPS 2023. The code is developed based on [LAVIS](https://github.com/salesforce/LAVIS/) project (cloned on Feb 23, 2023).
 
 <img src="overview.png" width="800">
 
-The code is developed based on [LAVIS](https://github.com/salesforce/LAVIS/) project (cloned on Feb 23, 2023).
-
 We mainly add following files in `lavis/models/blip2_models`:
 
 - [x] `blip2_pformer_opt.py`
@@ -29,7 +26,53 @@ pip install -e .
 Please follow instructions from [LAVIS](https://github.com/salesforce/LAVIS/) to download pre-training datasets.
 
 ## Training
-The code will be released soon.
+stage 0 (training P-Former)
+```bash
+bash run_scripts/blip-T/train/pretrain_stage0.sh
+```
+
+stage 1 (training Q-Former with pre-trained P-Former)
+```bash
+bash run_scripts/blip-T/train/pretrain_stage1.sh
+```
+
+stage 2 (End-to-end BLIP2 with pre-trained P-Former)
+```bash
+bash run_scripts/blip-T/train/pretrain_stage2.sh
+```
+
+finetuning on MSCOCO
+```bash
+bash run_scripts/blip-T/train/train_caption_coco.sh
+```
+
+## Pretrained Models
+- [ ] [our_stage0]()
+- [ ] [our_stage1]()
+- [ ] [our_stage2]()
 
 ## Evaluation
-The code will be released soon.
+```bash
+bash run_scripts/blip-T/eval/eval_gqa_zeroshot_opt2.7b.sh
+bash run_scripts/blip-T/eval/eval_okvqa_zeroshot_opt2.7b.sh
+bash run_scripts/blip-T/eval/validate_vqa_zeroshot_opt2.7b.sh
+bash run_scripts/blip-T/eval/eval_cap_coco_opt2.7b.sh
+```
+
+## Training and Evaluation Logs
+You can find our training and evaluation logs [here](training_logs/).
+
+## Acknowlegements
+The code is developed based on [BLIP2](https://openreview.net/forum?id=KU9UojoX7U) and [LAVIS](https://github.com/salesforce/LAVIS/) project.
+
+## Citation
+```bibtex
+@inproceedings{
+ jian2023bootstrapping,
+ title={Bootstrapping Vision-Language Learning with Decoupled Language Pre-training},
+ author = {Jian, Yiren and Gao, Chongyang and Vosoughi, Soroush},
+ booktitle={Thirty-seventh Conference on Neural Information Processing Systems},
+ year={2023},
+ url={https://openreview.net/forum?id=8Kch0ILfQH}
+}
+```

app/__init__.py

@@ -0,0 +1,26 @@
+"""
+ # Copyright (c) 2022, salesforce.com, inc.
+ # All rights reserved.
+ # SPDX-License-Identifier: BSD-3-Clause
+ # For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause
+"""
+
+from PIL import Image
+import requests
+
+import streamlit as st
+import torch
+
+
+@st.cache()
+def load_demo_image():
+ img_url = (
+ "https://storage.googleapis.com/sfr-vision-language-research/BLIP/demo.jpg"
+ )
+ raw_image = Image.open(requests.get(img_url, stream=True).raw).convert("RGB")
+ return raw_image
+
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+cache_root = "/export/home/.cache/lavis/"

app/calculate_coco_features.py

@@ -0,0 +1,87 @@
+"""
+ # Copyright (c) 2022, salesforce.com, inc.
+ # All rights reserved.
+ # SPDX-License-Identifier: BSD-3-Clause
+ # For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause
+"""
+
+from PIL import Image
+import requests
+import torch
+
+import os
+
+from lavis.common.registry import registry
+from lavis.processors import *
+from lavis.models import *
+from lavis.common.utils import build_default_model
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+
+def load_demo_image():
+ img_url = (
+ "https://storage.googleapis.com/sfr-vision-language-research/BLIP/demo.jpg"
+ )
+ raw_image = Image.open(requests.get(img_url, stream=True).raw).convert("RGB")
+
+ return raw_image
+
+
+def read_img(filepath):
+ raw_image = Image.open(filepath).convert("RGB")
+
+ return raw_image
+
+
+# model
+model_url = "https://storage.googleapis.com/sfr-vision-language-research/BLIP/models/model_base.pth"
+feature_extractor = BlipFeatureExtractor(pretrained=model_url)
+
+feature_extractor.eval()
+feature_extractor = feature_extractor.to(device)
+
+# preprocessors
+vis_processor = BlipImageEvalProcessor(image_size=224)
+text_processor = BlipCaptionProcessor()
+
+# files to process
+# file_root = "/export/home/.cache/lavis/coco/images/val2014"
+file_root = "/export/home/.cache/lavis/coco/images/train2014"
+filepaths = os.listdir(file_root)
+
+print(len(filepaths))
+
+caption = "dummy"
+
+path2feat = dict()
+bsz = 256
+
+images_in_batch = []
+filepaths_in_batch = []
+
+for i, filename in enumerate(filepaths):
+ if i % bsz == 0 and i > 0:
+ images_in_batch = torch.cat(images_in_batch, dim=0).to(device)
+ with torch.no_grad():
+ image_features = feature_extractor(
+ images_in_batch, caption, mode="image", normalized=True
+ )[:, 0]
+
+ for filepath, image_feat in zip(filepaths_in_batch, image_features):
+ path2feat[os.path.basename(filepath)] = image_feat.detach().cpu()
+
+ images_in_batch = []
+ filepaths_in_batch = []
+
+ print(len(path2feat), image_features.shape)
+ else:
+ filepath = os.path.join(file_root, filename)
+
+ image = read_img(filepath)
+ image = vis_processor(image).unsqueeze(0)
+
+ images_in_batch.append(image)
+ filepaths_in_batch.append(filepath)
+
+torch.save(path2feat, "path2feat_coco_train2014.pth")

app/caption.py

@@ -0,0 +1,98 @@
+"""
+ # Copyright (c) 2022, salesforce.com, inc.
+ # All rights reserved.
+ # SPDX-License-Identifier: BSD-3-Clause
+ # For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause
+"""
+
+import streamlit as st
+from app import device, load_demo_image
+from app.utils import load_model_cache
+from lavis.processors import load_processor
+from PIL import Image
+
+
+def app():
+ # ===== layout =====
+ model_type = st.sidebar.selectbox("Model:", ["BLIP_base", "BLIP_large"])
+
+ sampling_method = st.sidebar.selectbox(
+ "Sampling method:", ["Beam search", "Nucleus sampling"]
+ )
+
+ st.markdown(
+ "<h1 style='text-align: center;'>Image Description Generation</h1>",
+ unsafe_allow_html=True,
+ )
+
+ instructions = """Try the provided image or upload your own:"""
+ file = st.file_uploader(instructions)
+
+ use_beam = sampling_method == "Beam search"
+
+ col1, col2 = st.columns(2)
+
+ if file:
+ raw_img = Image.open(file).convert("RGB")
+ else:
+ raw_img = load_demo_image()
+
+ col1.header("Image")
+
+ w, h = raw_img.size
+ scaling_factor = 720 / w
+ resized_image = raw_img.resize((int(w * scaling_factor), int(h * scaling_factor)))
+
+ col1.image(resized_image, use_column_width=True)
+ col2.header("Description")
+
+ cap_button = st.button("Generate")
+
+ # ==== event ====
+ vis_processor = load_processor("blip_image_eval").build(image_size=384)
+
+ if cap_button:
+ if model_type.startswith("BLIP"):
+ blip_type = model_type.split("_")[1].lower()
+ model = load_model_cache(
+ "blip_caption",
+ model_type=f"{blip_type}_coco",
+ is_eval=True,
+ device=device,
+ )
+
+ img = vis_processor(raw_img).unsqueeze(0).to(device)
+ captions = generate_caption(
+ model=model, image=img, use_nucleus_sampling=not use_beam
+ )
+
+ col2.write("\n\n".join(captions), use_column_width=True)
+
+
+def generate_caption(
+ model, image, use_nucleus_sampling=False, num_beams=3, max_length=40, min_length=5
+):
+ samples = {"image": image}
+
+ captions = []
+ if use_nucleus_sampling:
+ for _ in range(5):
+ caption = model.generate(
+ samples,
+ use_nucleus_sampling=True,
+ max_length=max_length,
+ min_length=min_length,
+ top_p=0.9,
+ )
+ captions.append(caption[0])
+ else:
+ caption = model.generate(
+ samples,
+ use_nucleus_sampling=False,
+ num_beams=num_beams,
+ max_length=max_length,
+ min_length=min_length,
+ )
+ captions.append(caption[0])
+
+ return captions