diff --git a/.circleci/test.yml b/.circleci/test.yml
index 2d5713cf1a..169bba2778 100644
--- a/.circleci/test.yml
+++ b/.circleci/test.yml
@@ -73,6 +73,10 @@ jobs:
name: Install timm
command: |
pip install timm
+ - run:
+ name: Install transformers
+ command: |
+ pip install transformers
- when:
condition:
equal: [ "0.10.0", << parameters.torchvision >> ]
@@ -118,6 +122,10 @@ jobs:
command: |
docker exec mmaction pip install timm
docker exec mmaction python -m pip install pytorchvideo
+ - run:
+ name: Install transformers
+ command: |
+ docker exec mmaction pip install transformers
- run:
name: Install mmaction dependencies
command: |
diff --git a/.github/workflows/merge_stage_test.yml b/.github/workflows/merge_stage_test.yml
index de01615037..0a0222903a 100644
--- a/.github/workflows/merge_stage_test.yml
+++ b/.github/workflows/merge_stage_test.yml
@@ -69,6 +69,8 @@ jobs:
if: ${{matrix.torchvision == '0.10.0'}}
- name: Install timm
run: pip install timm
+ - name: Install transformers
+ run: pip install transformers
- name: Build and install
run: rm -rf .eggs && pip install -e .
- name: Run unittests and generate coverage report
@@ -110,6 +112,8 @@ jobs:
run: pip install lmdb
- name: Install timm
run: pip install timm
+ - name: Install transformers
+ run: pip install transformers
- name: Install TurboJpeg lib
run: sudo apt-get install -y libturbojpeg
- name: Install PyTorch
@@ -183,6 +187,8 @@ jobs:
run: pip install librosa soundfile
- name: Install lmdb
run: pip install lmdb
+ - name: Install transformers
+ run: pip install transformers
- name: Install mmaction dependencies
run: |
pip install git+https://github.com/open-mmlab/mmengine.git@main
@@ -240,6 +246,8 @@ jobs:
run: python -m pip install pytorchvideo
- name: Install timm
run: python -m pip install timm
+ - name: Install transformers
+ run: python -m pip install transformers
- name: Build and install
run: |
pip install -e . -v
diff --git a/configs/multimodal/vindlu/README.md b/configs/multimodal/vindlu/README.md
new file mode 100644
index 0000000000..c49fed61fa
--- /dev/null
+++ b/configs/multimodal/vindlu/README.md
@@ -0,0 +1,87 @@
+# VindLU
+
+[VindLU: A Recipe for Effective Video-and-Language Pretraining](https://arxiv.org/abs/2212.05051)
+
+
+
+## Abstract
+
+
+
+The last several years have witnessed remarkable progress in video-and-language (VidL) understanding. However, most modern VidL approaches use complex and specialized model architectures and sophisticated pretraining protocols, making the reproducibility, analysis and comparisons of these frameworks difficult. Hence, instead of proposing yet another new VidL model, this paper conducts a thorough empirical study demystifying the most important factors in the VidL model design. Among the factors that we investigate are (i) the spatiotemporal architecture design, (ii) the multimodal fusion schemes, (iii) the pretraining objectives, (iv) the choice of pretraining data, (v) pretraining and finetuning protocols, and (vi) dataset and model scaling. Our empirical study reveals that the most important design factors include: temporal modeling, video-to-text multimodal fusion, masked modeling objectives, and joint training on images and videos. Using these empirical insights, we then develop a step-by-step recipe, dubbed VindLU, for effective VidL pretraining. Our final model trained using our recipe achieves comparable or better than state-of-the-art results on several VidL tasks without relying on external CLIP pretraining. In particular, on the text-to-video retrieval task, our approach obtains 61.2% on DiDeMo, and 55.0% on ActivityNet, outperforming current SOTA by 7.8% and 6.1% respectively. Furthermore, our model also obtains state-of-the-art video question-answering results on ActivityNet-QA, MSRVTT-QA, MSRVTT-MC and TVQA. Our code and pretrained models are publicly available at: https://github.com/klauscc/VindLU.
+
+
+
+
+
+