Delete lxmert/run

lxmert/run/README.md

@@ -1,49 +0,0 @@
-# Running Script Arguments
-
-```
-Data Splits: 
-    --train [str,str,...]: use the splits (separated by comma) in training.
-    --valid [str,str,...]: use the splits (separated by comma) in validation.
-    --test [str,str,...]: use the splits (separated by comma) in testing.
-Model Architecture:
-    --llayers [int]: number of layers in language encoder.
-    --xlayers [int]: number of layers in cross-modality encoder.
-    --rlayers [int]: number of layers in object relationship encoder.
-Load Weights:
-    --load [str='path/to/saved_model']: load fine-tuned model path/to/saved_model.pth.
-    --loadLXMERT [str='path/to/saved_model']: load pre-trained model without answer heads from path/to/saved_model_LXRT.pth.
-    --loadLXMERTQA [str='path/to/saved_model']: load pre-trained model with answer head path/to/saved_model_LXRT.pth.
-    --fromScratch: If none of the above loading parameters are set, the default mode would 
-      load the pre-trained BERT weights.
-      As we promised to EMNLP reviewers, the language encoder would be re-initialized with this one-line argument to test the performance without BERT weights.
-Training Hyper Parameters:
-    --batchSize [int]: batch size.
-    --optim [str]: optimizers.
-    --lr [float]: peak learning rate.
-    --epochs [int]: training epochs.
-Debugging:
-    --tiny: Load 512 images for each data split. (Note: number of images might be changed due to dataset specification)
-    --fast: Load 5000 images for each data split. (Note: number of images might be changed due to dataset specification)
-```
-
-# Pre-training-Specific Arguments
-```
-Pre-training Tasks:
-    --taskMaskLM: use the masked language model task.
-    --taskObjPredict: use the masked object prediction task.
-    --taskMatched: use the cross-modality matched task.
-    --taskQA: use the image QA task.
-Visual Pre-training Losses (Tasks):
-    --visualLosses [str,str,...]: The sub-tasks in pre-training visual modality. Each one is from 'obj,attr,feat'. 
-      obj: detected-object-label classification. 
-      attr: detected-object-attribute classification. 
-      feat: RoI-feature regression.
-Mask Rate in Pre-training:
-    --wordMaskRate [float]: The prob of masking a word.
-    --objMaskRate [float]: The prob of masking an object.
-Initialization:
-    --fromScratch: The default mode would load the pre-trained BERT weights into the model. 
-      As we promised to EMNLP reviewers, this option would re-initialize the language encoder.
-```
-
-

lxmert/run/gqa_finetune.bash

@@ -1,17 +0,0 @@
-# The name of this experiment.
-name=$2
-
-# Save logs and models under snap/gqa; make backup.
-output=snap/gqa/$name
-mkdir -p $output/src
-cp -r src/* $output/src/
-cp $0 $output/run.bash
-
-# See Readme.md for option details.
-CUDA_VISIBLE_DEVICES=$1 PYTHONPATH=$PYTHONPATH:./src \
-    python src/tasks/gqa.py \
-    --train train,valid --valid testdev \
-    --llayers 9 --xlayers 5 --rlayers 5 \
-    --loadLXMERTQA snap/pretrained/model \
-    --batchSize 32 --optim bert --lr 1e-5 --epochs 4 \
-    --tqdm --output $output ${@:3}

lxmert/run/gqa_test.bash

@@ -1,15 +0,0 @@
-# The name of this experiment.
-name=$2
-
-# Save logs and models under snap/gqa; make backup.
-output=snap/gqa/$name
-mkdir -p $output/src
-cp -r src/* $output/src/
-cp $0 $output/run.bash
-
-# See Readme.md for option details.
-CUDA_VISIBLE_DEVICES=$1 PYTHONPATH=$PYTHONPATH:./src \
-    python src/tasks/gqa.py \
-    --tiny --train train --valid "" \
-    --llayers 9 --xlayers 5 --rlayers 5 \
-    --tqdm --output $output ${@:3}

lxmert/run/lxmert_pretrain.bash

@@ -1,21 +0,0 @@
-# The name of experiment
-name=lxmert
-
-# Create dirs and make backup
-output=snap/pretrain/$name
-mkdir -p $output/src
-cp -r src/* $output/src/
-cp $0 $output/run.bash
-
-# Pre-training
-CUDA_VISIBLE_DEVICES=$1 PYTHONPATH=$PYTHONPATH:./src \
-    python src/pretrain/lxmert_pretrain.py \
-    --taskMaskLM --taskObjPredict --taskMatched --taskQA \
-    --visualLosses obj,attr,feat \
-    --wordMaskRate 0.15 --objMaskRate 0.15 \
-    --train mscoco_train,mscoco_nominival,vgnococo --valid mscoco_minival \
-    --llayers 9 --xlayers 5 --rlayers 5 \
-    --fromScratch \
-    --batchSize 256 --optim bert --lr 1e-4 --epochs 20 \
-    --tqdm --output $output ${@:2}
-

lxmert/run/nlvr2_finetune.bash

@@ -1,18 +0,0 @@
-# The name of this experiment.
-name=$2
-
-# Save logs and models under snap/nlvr2; Make backup.
-output=snap/nlvr2/$name
-mkdir -p $output/src
-cp -r src/* $output/src/
-cp $0 $output/run.bash
-
-# See run/Readme.md for option details.
-CUDA_VISIBLE_DEVICES=$1 PYTHONPATH=$PYTHONPATH:./src \
-    python src/tasks/nlvr2.py \
-    --train train --valid valid \
-    --llayers 9 --xlayers 5 --rlayers 5 \
-    --loadLXMERT snap/pretrained/model \
-    --batchSize 32 --optim bert --lr 5e-5 --epochs 4 \
-    --tqdm --output $output ${@:3}
-

lxmert/run/nlvr2_test.bash

@@ -1,14 +0,0 @@
-# The name of this experiment.
-name=$2
-
-# Save logs and models under snap/nlvr2; make backup.
-output=snap/nlvr2/$name
-mkdir -p $output/src
-cp -r src/* $output/src/
-cp $0 $output/run.bash
-
-# See Readme.md for option details.
-CUDA_VISIBLE_DEVICES=$1 PYTHONPATH=$PYTHONPATH:./src \
-    python src/tasks/nlvr2.py \
-    --tiny --llayers 9 --xlayers 5 --rlayers 5 \
-    --tqdm --output $output ${@:3}

lxmert/run/vqa_finetune.bash

@@ -1,17 +0,0 @@
-# The name of this experiment.
-name=$2
-
-# Save logs and models under snap/vqa; make backup.
-output=snap/vqa/$name
-mkdir -p $output/src
-cp -r src/* $output/src/
-cp $0 $output/run.bash
-
-# See Readme.md for option details.
-CUDA_VISIBLE_DEVICES=$1 PYTHONPATH=$PYTHONPATH:./src \
-    python src/tasks/vqa.py \
-    --train train,nominival --valid minival  \
-    --llayers 9 --xlayers 5 --rlayers 5 \
-    --loadLXMERTQA snap/pretrained/model \
-    --batchSize 32 --optim bert --lr 5e-5 --epochs 4 \
-    --tqdm --output $output ${@:3}

lxmert/run/vqa_test.bash

@@ -1,16 +0,0 @@
-# The name of this experiment.
-name=$2
-
-# Save logs and models under snap/vqa; make backup.
-output=snap/vqa/$name
-mkdir -p $output/src
-cp -r src/* $output/src/
-cp $0 $output/run.bash
-
-# See Readme.md for option details.
-CUDA_VISIBLE_DEVICES=$1 PYTHONPATH=$PYTHONPATH:./src \
-    python src/tasks/vqa.py \
-    --tiny --train train --valid ""  \
-    --llayers 9 --xlayers 5 --rlayers 5 \
-    --batchSize 32 --optim bert --lr 5e-5 --epochs 4 \
-    --tqdm --output $output ${@:3}