[feat] Add UNITER yaml config

Add configs for UNITER vqa2 finetuning.

ghstack-source-id: a0301cb
Pull Request resolved: #1128

Author: Ryan-Qiyu-Jiang

mmf/configs/zoo/models.yaml

@@ -546,3 +546,21 @@ unit:
       - url: mmf://models/unit_models/unit.all_8_datasets.shared_dec_with_coco_init_without_task_embedding.tar.gz
         file_name: unit.all_8_datasets.shared_dec_with_coco_init_without_task_embedding.tar.gz
         hashcode: c4a3c1072f07e605de76402ff7c0e2bc8d4514561c52dc94b3d28b3d4d7f698e
+
+uniter:
+  defaults: ${uniter.pretrained}
+  pretrained:
+    version: 1.0_2020_11_29
+    resources:
+    - url: mmf://models/uniter/uniter.pretrained.tar.gz
+      file_name: uniter.pretrained.tar.gz
+      hashcode: e843e65a3d8b4ed353779b96d654dde0440f5f04a79b3b4740309d40a3542664
+
+villa:
+  defaults: ${villa.pretrained}
+  pretrained:
+    version: 1.0_2020_11_29
+    resources:
+    - url: mmf://models/uniter/villa.pretrained.tar.gz
+      file_name: villa.pretrained.tar.gz
+      hashcode: 7a8f31421ef644fddc99bd142a0090660573dd526a779d025253c3fd996754fc

projects/uniter/README.md

@@ -0,0 +1,18 @@
+# UNITER
+
+This repository contains the code for pytorch implementation of UNITER model, released originally under this ([repo](https://github.com/ChenRocks/UNITER/)). Please cite the following papers if you are using UNITER model from mmf:
+
+* Chen, Y.-C., Li, L., Yu, L., Kholy, A. E., Ahmed, F., Gan,
+Z., Cheng, Y., and jing Liu, J. *Uniter: Universal imagetext representation learning.* In European Conference on
+Computer Vision, 2020b. ([arXiV](https://arxiv.org/pdf/1909.11740))
+```
+@inproceedings{chen2020uniter,
+  title={Uniter: Universal image-text representation learning},
+  author={Chen, Yen-Chun and Li, Linjie and Yu, Licheng and Kholy, Ahmed El and Ahmed, Faisal and Gan, Zhe and Cheng, Yu and Liu, Jingjing},
+  booktitle={ECCV},
+  year={2020}
+}
+```
+
+
+Please see [https://mmf.sh/docs/projects/uniter](https://mmf.sh/docs/projects/uniter) for more details on how to use the UNITER model.

projects/uniter/configs/masked_coco/defaults.yaml

@@ -0,0 +1,74 @@
+model_config:
+  uniter:
+    do_pretraining: true
+    mask_probability: 0.15
+    tasks:
+      - mlm
+      - itm
+      - mrc
+      - mrfr
+      - wra
+    heads:
+      mlm:
+        type: mlm
+        vocab_size: 28996
+      itm:
+        type: itm
+      mrc:
+        type: mrc
+      mrfr:
+        type: mrfr
+      wra:
+        type: wra
+    text_embeddings:
+      vocab_size: 28996
+
+dataset_config:
+  masked_coco:
+    false_caption: true
+    false_caption_probability: 0.1
+    zoo_requirements:
+    - coco.defaults
+    return_features_info: true
+    use_features: true
+    use_images: false
+    processors:
+      masked_token_processor:
+        type: uniter_text_tokenizer
+        params:
+          from_pretrained: bert-base-cased
+          tokenizer_config:
+            type: bert-base-cased
+            params:
+              do_lower_case: false
+          mask_probability: 0.15
+          max_seq_length: 25
+    images:
+      train:
+      - coco/defaults/images/train2014
+      - coco/defaults/images/val2014
+      - coco/defaults/images/train2014
+      - coco/defaults/images/val2014
+      - coco/defaults/images/val2014
+      val:
+      - coco/defaults/images/val2014
+      test:
+      - coco/defaults/images/test2015
+
+optimizer:
+  type: adam_w
+  params:
+    lr: 5e-5
+    eps: 1e-8
+
+scheduler:
+  type: warmup_linear
+  params:
+    num_warmup_steps: 6000
+    num_training_steps: 60000
+
+training:
+  batch_size: 480
+  lr_scheduler: true
+  # Don't forget to update schedule_attributes if you update this
+  max_updates: 60000

projects/uniter/configs/vqa2/defaults.yaml

@@ -0,0 +1,59 @@
+# includes:
+# - ../projects/unit/configs/vqa2_dataset_cfg.yaml
+
+model_config:
+  uniter:
+    do_pretraining: false
+    tasks: vqa2
+    heads:
+      vqa2:
+        type: mlp
+        num_labels: 3129
+        lr_multiplier: 10
+    losses:
+      vqa2: logit_bce
+    text_embeddings:
+      vocab_size: 28996
+
+dataset_config:
+  vqa2:
+    return_features_info: true
+    use_features: true
+    use_images: false
+    processors:
+      text_processor:
+        type: uniter_text_tokenizer
+        params:
+          from_pretrained: bert-base-cased
+          tokenizer_config:
+            type: bert-base-cased
+            params:
+              do_lower_case: false
+          mask_probability: 0
+          max_seq_length: 60
+
+optimizer:
+  type: adam_w
+  params:
+    lr: 8e-5
+    eps: 1e-8
+    weight_decay: 0.01
+
+scheduler:
+  type: warmup_linear
+  params:
+    num_warmup_steps: 600
+    num_training_steps: 5000
+
+evaluation:
+  metrics:
+  - vqa_accuracy
+
+training:
+  batch_size: 5120
+  lr_scheduler: true
+  # Don't forget to update schedule_attributes if you update this
+  max_updates: 5000
+  early_stop:
+    criteria: vqa2/vqa_accuracy
+    minimize: false