Merge pull request #599 from artemisp/main

X-InstructBLIP Code

.gitignore

@@ -152,5 +152,8 @@ debug*/
 *.dat
 *.tsv
 *.gz
+*.csv
+*.p
+*.pdf
 
 cache/

README.md

@@ -27,6 +27,9 @@
 # LAVIS - A Library for Language-Vision Intelligence
 
 ## What's New: 🎉 
+  * [Model Release] November 2023, released implementation of **X-InstructBLIP** <br>
+  [Paper](https://arxiv.org/pdf/2311.18799.pdf), [Project Page](https://github.com/salesforce/LAVIS/tree/main/projects/xinstructblip), [Website](https://artemisp.github.io/X-InstructBLIP-page/), [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/salesforce/LAVIS/blob/main/projects/xinstructblip/demo/run_demo.ipynb)
+  > A simple, yet effective, cross-modality framework built atop frozen LLMs that allows the integration of various modalities (image, video, audio, 3D) without extensive modality-specific customization.
   * [Model Release] July 2023, released implementation of **BLIP-Diffusion** <br>
   [Paper](https://arxiv.org/abs/2305.06500), [Project Page](https://github.com/salesforce/LAVIS/tree/main/projects/blip-diffusion), [Website](https://dxli94.github.io/BLIP-Diffusion-website/)
   > A text-to-image generation model that trains 20x than DreamBooth. Also facilitates zero-shot subject-driven generation and editing.

lavis/common/utils.py

@@ -1,5 +1,5 @@
 """
- Copyright (c) 2022, salesforce.com, inc.
+ Copyright (c) 2023, 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
@@ -440,3 +440,16 @@ def get_file_size(filename):
     """
     size_in_mb = os.path.getsize(filename) / float(1024**2)
     return size_in_mb
+
+def is_serializable(value):
+    """
+    This function checks if the provided value can be serialized into a JSON string.
+    """
+    try:
+        json.dumps(value)
+        return True
+    except (TypeError, OverflowError):
+        return False
+
+def is_convertible_to_int(value):
+    return bool(re.match(r'^-?\d+$', str(value)))

lavis/configs/datasets/aokvqa/defaults_instruct.yaml

@@ -0,0 +1,52 @@
+ # Copyright (c) 2023, 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
+
+datasets:
+  aok_vqa_instruct:
+    # data_dir: ${env.data_dir}/datasets
+    data_type: images # [images|videos|features]
+
+    vis_processor:
+        train:
+          name: "clip_image_train"
+          image_size: 224
+        eval:
+          name: "clip_image_eval"
+          image_size: 224
+
+    text_processor:
+        train:
+          name: blip_instruction
+          modality: image
+          task: qa
+        eval:
+          name: blip_question
+
+    build_info:
+      # Be careful not to append minus sign (-) before split to avoid itemizing
+      annotations:
+        train:
+          url:
+              - https://storage.googleapis.com/sfr-vision-language-research/LAVIS/datasets/aokvqa/aokvqa_v1p0_train.json
+          storage:
+              - aokvqa/annotations/aokvqa_v1p0_train.json
+        # val:
+        #   url:
+        #       - https://storage.googleapis.com/sfr-vision-language-research/LAVIS/datasets/aokvqa/aokvqa_v1p0_val.json
+        #       - https://storage.googleapis.com/sfr-vision-language-research/LAVIS/datasets/aokvqa/specialized_vocab_train.json
+        #   storage:
+        #       - aokvqa/annotations/aokvqa_v1p0_val.json
+        #       - aokvqa/annotations/specialized_vocab_train_lavis.json
+        #       # - aokvqa/annotations/large_vocab_train_lavis.json
+        # test:
+        #   url:
+        #       - https://storage.googleapis.com/sfr-vision-language-research/LAVIS/datasets/aokvqa/aokvqa_v1p0_test.json
+        #       - https://storage.googleapis.com/sfr-vision-language-research/LAVIS/datasets/aokvqa/specialized_vocab_train.json
+        #   storage:
+        #       - aokvqa/annotations/aokvqa_v1p0_test.json
+        #       - aokvqa/annotations/specialized_vocab_train_lavis.json
+      images:
+          # storage: /coco/images
+          storage: /export/share/datasets/vision/coco/images

lavis/configs/datasets/audiocaps/defaults_mm_cap.yaml

@@ -0,0 +1,49 @@
+ # Copyright (c) 2023, 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
+
+datasets:
+  audiocaps_mm_caption: # name of the dataset builder
+    audio_processor:
+        train:
+          name: beats_audio
+          sampling_rate: 16000
+        eval:
+          name: beats_audio
+          sampling_rate: 16000
+
+    text_processor:
+        train:
+          name: "blip_instruction"
+          modality: audio
+          task: caption
+        eval:
+          name: "blip_caption"
+
+    data_type: [audio]
+
+    build_info:
+      kwargs:
+            missing_ids: [2sh7ZkazyO8, 966jA2-z0mQ, 52RlolYyjAE, HVAc9hm4jjk, 8lPjqvYWNyM, eXgPnnE3TuQ]
+      annotations:
+        train:
+          url: 
+            - https://raw.githubusercontent.com/cdjkim/audiocaps/master/dataset/train.csv
+          storage: 
+            - audiocaps/annotations/train.csv
+
+        val:
+          url: 
+            - https://raw.githubusercontent.com/cdjkim/audiocaps/master/dataset/val.csv
+          storage: 
+            - audiocaps/annotations/val.csv
+
+        test:
+          url: 
+            - https://raw.githubusercontent.com/cdjkim/audiocaps/master/dataset/test.csv
+          storage: 
+            - audiocaps/annotations/test.csv
+
+      audio:
+        storage: /export/einstein-vision/audio_datasets/audiocaps/AUDIOCAPS_32000Hz/audio

lavis/configs/datasets/audiocaps/defaults_mm_cap_instruct.yaml

@@ -0,0 +1,52 @@
+ # Copyright (c) 2023, 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
+
+datasets:
+  audiocaps_mm_caption_instruct: # name of the dataset builder
+    audio_processor:
+        train:
+          name: beats_audio
+          sampling_rate: 16000
+        eval:
+          name: beats_audio
+          sampling_rate: 16000
+
+    text_processor:
+        train:
+          name: "blip_instruction"
+          modality: audio
+          task: caption
+        eval:
+          name: "blip_caption"
+
+    data_type: [audio]
+
+    missing_ids: [2sh7ZkazyO8, 966jA2-z0mQ, 52RlolYyjAE, HVAc9hm4jjk, 8lPjqvYWNyM, eXgPnnE3TuQ]
+
+    build_info:
+      kwargs:
+        cached: False
+        cached_dir: /export/einstein-vision/audio_datasets/audiocaps/beats_features
+      annotations:
+        train:
+          url: 
+            - https://raw.githubusercontent.com/cdjkim/audiocaps/master/dataset/train.csv
+          storage: 
+            - audiocaps/annotations/train.csv
+
+        # val:
+        #   url: 
+        #     - https://raw.githubusercontent.com/cdjkim/audiocaps/master/dataset/val.csv
+        #   storage: 
+        #     - audiocaps/annotation/val.csv
+
+        # test:
+        #   url: 
+        #     - https://raw.githubusercontent.com/cdjkim/audiocaps/master/dataset/test.csv
+        #   storage: 
+        #     - /export/einstein-vision/audio_datasets/audiocaps/dataset/test.csv
+
+      audio:
+        storage: /export/einstein-vision/audio_datasets/audiocaps/AUDIOCAPS_32000Hz/audio

lavis/configs/datasets/audiocaps/defaults_mm_qa.yaml

@@ -0,0 +1,51 @@
+ # Copyright (c) 2023, 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
+
+datasets:
+  audiocaps_mm_qa: # name of the dataset builder
+    audio_processor:
+        train:
+          name: beats_audio
+          sampling_rate: 16000
+        eval:
+          name: beats_audio
+          sampling_rate: 16000
+          is_eval: True
+
+    text_processor:
+        train:
+          name: "blip_instruction"
+          modality: audio
+          task: qa
+        eval:
+          name: "blip_question"
+
+    data_type: [audio] 
+
+    build_info:
+      kwargs:
+        cached: False
+        # add_binary: True
+        cached_dir: /export/einstein-vision/audio_datasets/audiocaps/beats_features
+        missing_ids: [2sh7ZkazyO8, 966jA2-z0mQ, 52RlolYyjAE, HVAc9hm4jjk, 8lPjqvYWNyM, eXgPnnE3TuQ]
+      annotations:
+        train:
+          url: 
+            - https://storage.googleapis.com/sfr-xinstructblip-data-research/data/audiocaps/audio_qa_final_train.csv
+            # - /export/home/LAVIS-xgen_mm/projects/xinstructblip/data_aug/audio_qa_data/audio_qa_final_train.csv
+          storage: 
+            - audiocaps_qa/annotations/train.csv
+            # - /export/home/LAVIS-xgen_mm/projects/xinstructblip/data_aug/audio_qa_data/audio_qa_final_train.csv
+
+        # val:
+        #   url: 
+        #     # - https://storage.googleapis.com/sfr-xinstructblip-data-research/data/audiocaps/audio_qa_final_val.csv
+        #     - /export/home/LAVIS-xgen_mm/projects/xinstructblip/data_aug/audio_qa_data/audio_qa_final_val.csv
+        #   storage: 
+        #     # - audiocaps_qa/annotations/val.csv
+        #     - /export/home/LAVIS-xgen_mm/projects/xinstructblip/data_aug/audio_qa_data/audio_qa_final_val.csv
+
+      audio:
+        storage: /export/einstein-vision/audio_datasets/audiocaps/AUDIOCAPS_32000Hz/audio

lavis/configs/datasets/audioset/defaults_mm_cap.yaml

@@ -0,0 +1,47 @@
+ # Copyright (c) 2023, 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
+
+datasets:
+  audioset_mm_caption: # 14141
+    audio_processor:
+        train:
+          name: beats_audio
+          sampling_rate: 16000
+        eval:
+          name: beats_audio
+          sampling_rate: 16000
+          is_eval: False
+
+    text_processor:
+        train:
+          name: blip_instruction
+          modality: audio
+          task: classification
+        eval:
+          name: blip_caption
+
+    data_type: [audio] 
+
+    build_info:
+      annotations:
+        train:
+          url: 
+            - https://storage.googleapis.com/sfr-xinstructblip-data-research/data//audioset/balanced_train_clean.csv
+            # - /export/home/LAVIS-xgen_mm/lavis/configs/datasets/audioset/balanced_train_clean.csv
+            - http://storage.googleapis.com/us_audioset/youtube_corpus/v1/csv/class_labels_indices.csv
+          storage:
+            - audioset/balanced_train_clean.csv
+            # - /export/home/LAVIS-xgen_mm/lavis/configs/datasets/audioset/balanced_train_clean.csv
+            - audioset/annotations/class_labels_indices.csv
+
+        # val:
+        #   url: 
+        #     - http://storage.googleapis.com/us_audioset/youtube_corpus/v1/csv/eval_segments.csv
+        #     - http://storage.googleapis.com/us_audioset/youtube_corpus/v1/csv/class_labels_indices.csv
+        #   storage: 
+        #     - audioset/annotations/eval_segments.csv
+        #     - audioset/annotations/class_labels_indices.csv
+      audio:
+        storage: /export/einstein-vision/audio_datasets/AudioSet/all_audio

lavis/configs/datasets/audioset/defaults_mm_cap_instruct.yaml

@@ -0,0 +1,48 @@
+ # Copyright (c) 2023, 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
+
+datasets:
+  audioset_mm_caption_instruct: # 14141
+    audio_processor:
+        train:
+          name: beats_audio
+          sampling_rate: 16000
+        eval:
+          name: beats_audio
+          sampling_rate: 16000
+          is_eval: False
+
+    text_processor:
+        train:
+          name: blip_instruction
+          modality: audio
+          task: classification
+        eval:
+          name: blip_caption
+
+    data_type: [audio] 
+
+    build_info:
+      annotations:
+        train:
+          url: 
+            # - https://storage.googleapis.com/sfr-xinstructblip-data-research/data//audioset/balanced_train_clean.csv
+            - /export/home/LAVIS-xgen_mm/lavis/configs/datasets/audioset/balanced_train_clean.csv
+            - http://storage.googleapis.com/us_audioset/youtube_corpus/v1/csv/class_labels_indices.csv
+          storage:
+            - audioset/annotations/balanced_train_clean.csv
+            # - /export/home/LAVIS-xgen_mm/lavis/configs/datasets/audioset/balanced_train_clean.csv
+            - audioset/annotations/class_labels_indices.csv
+
+        # val:
+        #   url: 
+        #     - http://storage.googleapis.com/us_audioset/youtube_corpus/v1/csv/eval_segments.csv
+        #     - http://storage.googleapis.com/us_audioset/youtube_corpus/v1/csv/class_labels_indices.csv
+        #   storage: 
+        #     - audioset/annotations/eval_segments.csv
+        #     - audioset/annotations/class_labels_indices.csv
+
+      audio:
+        storage: /export/einstein-vision/audio_datasets/AudioSet/all_audio