extra_scripts

Helpful scripts for data preparation and model conversion

We provide several helpful scripts to prepare data, to convert VISSL models to detectron2 compatible models or to convert caffe2 models to VISSL compatible models.

Data preparation

VISSL supports benchmarks inspired by the VTAB and CLIP papers, for which the datasets do not directly exist but are transformations of existing dataset.

To run these benchmarks, the following data preparation scripts are mandatory:

extra_scripts/datasets/create_clevr_count_data_files.py: to create a disk_filelist dataset from CLEVR where the goal is to count the number of object in the scene
extra_scripts/datasets/create_clevr_dist_data_files.py: to create a disk_filelist dataset from CLEVR where the goal is to estimate the distance to the closest object in the scene
extra_scripts/datasets/create_dsprites_location_data_files.py: to create a disk_folder dataset from dSprites where the goal is to estimate the x coordinate of the sprite on the scene
extra_scripts/datasets/create_dsprites_orientation_data_files.py: to create a disk_folder dataset from dSprites where the goal is to estimate the orientation of the sprite on the scene
extra_scripts/datasets/create_euro_sat_data_files.py: to transform the EUROSAT dataset to the disk_folder format
extra_scripts/datasets/create_food101_data_files.py: to transform the FOOD101 dataset to the disk_folder format
extra_scripts/datasets/create_imagenet_ood_data_files.py: to create test sets in disk_filelist format for Imagenet based on Imagenet-A and Imagenet-R
extra_scripts/datasets/create_kitti_dist_data_files.py: to create a disk_folder dataset from KITTI where the goal is to estimate the distance of the closest car, van or truck
extra_scripts/datasets/create_patch_camelyon_data_files.py: to transform the PatchCamelyon dataset to the disk_folder format
extra_scripts/datasets/create_small_norb_azimuth_data_files.py to create a disk_folder dataset from Small NORB where the goal is to find the azimuth or the photographed object
extra_scripts/datasets/create_small_norb_elevation_data_files.py to create a disk_folder dataset from Small NORB where the goal is to predict the elevation in the image
extra_scripts/datasets/create_sun397_data_files.py to transform the SUN397 dataset to the disk_filelist format
extra_scripts/datasets/create_ucf101_data_files.py: to create a disk_folder image action recognition dataset from the video action recognition dataset UCF101 by extracting the middle frame

Unified data preparation interface

All of these scripts follow the same easy to use interface:

python create_[***]_data_files.py -i /path/to/input_datset -o /path/to/tranformed/dataset -d

-i gives the path to the official dataset format
-o gives the path to the output transformed dataset (the one to feed to VISSL)
-d (optional) automatically downloads the dataset in the input path

Scripts producing a disk_filelist format will create the following structure:

output_folder/
    train_images.npy  # Paths to the train images
    train_labels.npy  # Labels for each of the train images
    val_images.npy    # Paths to the val images
    val_labels.npy    # Labels for each of the val images

These files should be referenced in the dataset_catalog.json like so:

"dataset_filelist": {
    "train": ["/path/to/train_images.npy", "/path/to/train_labels.npy"],
    "val": ["/path/to/val_images.npy", "/path/to/val_labels.npy"]
},

Scripts producing a disk_folder format will create the following structure:

train/
    label1/
        image_1.jpeg
        image_2.jpeg
        ...
    label2/
        image_x.jpeg
        image_y.jpeg
        ...
    ...
val/
    label1/
        image_1.jpeg
        image_2.jpeg
        ...
    label2/
        image_x.jpeg
        image_y.jpeg
        ...
    ...

These files should be referenced in the dataset_catalog.json like so:

"dataset_folder": {
    "train": ["/path/to/dataset/train", "<ignored>"],
    "val": ["/path/to/dataset/val", "<ignored>"]
},

The following sections will describe each of these data preparation scripts in detail.

Preparing CLEVR/Counts data files

Automatic download

Run the create_clevr_count_data_files.py script with the -d option as follows:

python extra_scripts/datasets/create_clevr_count_data_files.py \
    -i /path/to/clevr/ \
    -o /output_path/to/clevr_count
    -d

The folder /output_path/clevr_count now contains the CLEVR/Counts disk_filelist dataset. The last step is to set this path in dataset_catalog.json and you are good to go:

"clevr_count_filelist": {
    "train": ["/output_path/to/clevr_count/train_images.npy", "/output_path/to/clevr_count/train_labels.npy"],
    "val": ["/output_path/to/clevr_count/val_images.npy", "/output_path/to/clevr_count/val_labels.npy"]
},

Manual download

Download the full dataset by visiting CLEVR website and clicking on Download CLEVR v1.0 (18 GB) dataset. Expand the archive.

The resulting folder should have the following structure:

/path/to/clevr/
    CLEVR_v1.0/
        COPYRIGHT.txt
        LICENSE.txt
        README.txt
        images/
            train/
                ... 75000 images ...
            val/
                ... 15000 images ...
            test/
                ... 15000 images ...
        questions/
            CLEVR_test_questions.json
            CLEVR_train_questions.json
            CLEVR_val_questions.json
        scenes/
            CLEVR_train_scenes.json
            CLEVR_val_scenes.json

Run the script where /path/to/clevr/ is the path of the folder containing the CLEVR_v1.0 folder:

python extra_scripts/datasets/create_clevr_count_data_files.py \
    -i /path/to/clevr/ \
    -o /output_path/to/clevr_count

The folder /output_path/clevr_count now contains the CLEVR/Counts dataset.

Preparing CLEVR/Dist data files

Follow the exact same steps as for the preparation of the CLEVR/Count dataset described above, but use create_clevr_dist_data_files.py instead of create_clevr_count_data_files.py.

Once the dataset is prepared and available at /path/to/clevr_dist, the last step is to set this path in dataset_catalog.json and you are good to go:

"clevr_dist_filelist": {
    "train": ["/path/to/clevr_dist/train_images.npy", "/path/to/clevr_dist/train_labels.npy"],
    "val": ["/path/to/clevr_dist/val_images.npy", "/path/to/clevr_dist/val_labels.npy"]
},

Preparing the dSprites/location data files

Run the create_dsprites_location_data_files.py script with the -d option as follows:

python extra_scripts/datasets/create_dsprites_location_data_files.py \
    -i /path/to/dsprites/ \
    -o /output_path/to/dsprites_loc
    -d

The folder /output_path/to/dsprites_loc now contains the dSprites/location disk_folder dataset. The last step is to set this path in dataset_catalog.json and you are good to go:

"dsprites_loc_folder": {
    "train": ["/output_path/to/dsprites_loc/train", "<ignored>"],
    "val": ["/output_path/to/dsprites_loc/val", "<ignored>"]
},

Preparing the dSprites/orientation data files

Run the create_dsprites_orientation_data_files.py script with the -d option as follows:

python extra_scripts/datasets/create_dsprites_orientation_data_files.py \
    -i /path/to/dsprites/ \
    -o /output_path/to/dsprites_orient
    -d

The folder /output_path/to/dsprites_orient now contains the dSprites/orientation disk_folder dataset. The last step is to set this path in dataset_catalog.json and you are good to go:

"dsprites_orient_folder": {
    "train": ["/output_path/to/dsprites_orient/train", "<ignored>"],
    "val": ["/output_path/to/dsprites_orient/val", "<ignored>"]
},

Preparing the Caltech101 data files

Run the create_caltech101_data_files.py script with the -d option as follows:

python extra_scripts/datasets/create_caltech101_data_files.py \
    -i /path/to/caltech101/ \
    -o /output_path/to/caltech101
    -d

The folder /output_path/to/caltech101 now contains the Caltech101 disk_folder dataset. The last step is to set this path in dataset_catalog.json and you are good to go:

"caltech101_folder": {
    "train": ["/output_path/to/caltech101/train", "<ignored>"],
    "val": ["/output_path/to/caltech101/test", "<ignored>"]
},

Preparing the DTD data files

Run the create_dtd_data_files.py script with the -d option as follows:

python extra_scripts/datasets/create_dtd_data_files.py \
    -i /path/to/dtd/ \
    -o /output_path/to/dtd
    -d

The folder /output_path/to/dtd now contains the DTD disk_folder dataset. The last step is to set this path in dataset_catalog.json and you are good to go:

"dtd_folder": {
    "train": ["/output_path/to/dtd/train", "<ignored>"],
    "val": ["/output_path/to/dtd/test", "<ignored>"]
},

Preparing the EuroSAT data files

Run the create_euro_sat_data_files.py script with the -d option as follows:

python extra_scripts/datasets/create_euro_sat_data_files.py \
    -i /path/to/euro_sat/ \
    -o /output_path/to/euro_sat
    -d

The folder /output_path/to/euro_sat now contains the EuroSAT disk_folder dataset. The last step is to set this path in dataset_catalog.json and you are good to go:

"euro_sat_folder": {
    "train": ["/output_path/to/euro_sat/train", "<ignored>"],
    "val": ["/output_path/to/euro_sat/val", "<ignored>"]
},

Preparing the Food-101 data files

Run the create_food101_data_files.py script with the -d option as follows:

python extra_scripts/datasets/create_food101_data_files.py \
    -i /path/to/food101/ \
    -o /output_path/to/food101
    -d

The folder /output_path/to/food101 now contains the Food101 disk_folder dataset. The last step is to set this path in dataset_catalog.json and you are good to go:

"food101_folder": {
    "train": ["/output_path/to/food101/train", "<ignored>"],
    "val": ["/output_path/to/food101/val", "<ignored>"]
},

Preparing the FGVC Aircrafts data files

Run the create_fgvc_aircraft_data_files.py script with the -d option as follows:

python extra_scripts/datasets/create_fgvc_aircraft_data_files.py \
    -i /path/to/aircrafts/ \
    -o /output_path/to/aircrafts
    -d

The folder /output_path/to/aircrafts now contains the FGVC Aircrafts disk_folder dataset. The last step is to set this path in dataset_catalog.json and you are good to go:

"aircrafts_folder": {
    "train": ["/output_path/to/aircrafts/trainval", "<ignored>"],
    "val": ["/output_path/to/aircrafts/test", "<ignored>"]
},

Preparing the GTSRB data files

Run the create_gtsrb_data_files.py script with the -d option as follows:

python extra_scripts/datasets/create_gtsrb_data_files.py \
    -i /path/to/gtsrb/ \
    -o /output_path/to/gtsrb
    -d

The folder /output_path/to/gtsrb now contains the GTSRB disk_folder dataset. The last step is to set this path in dataset_catalog.json and you are good to go:

"gtsrb_folder": {
    "train": ["/output_path/to/gtsrb/train", "<ignored>"],
    "val": ["/output_path/to/gtsrb/test", "<ignored>"]
},

Preparing the Imagenet-A and Imagenet-R data files

Run the create_imagenet_ood_data_files.py script with the -d option as follows:

python extra_scripts/datasets/create_imagenet_ood_data_files.py \
    -i /path/to/input_folder/ \
    -o /path/to/output/
    -d

After running this script:

The folder /path/to/input_folder/ will contain the expanded imagenet-a and imagenet-r dataset in their original format
The folder /path/to/output/imagenet-a will contain the Imagenet-A disk_filelist to provide to VISSL
The folder /path/to/output/imagenet-r will contain the Imagenet-R disk_filelist to provide to VISSL

Note: all these folders are necessary as the disk_filelist format references images and does not copy them. Deleting /path/to/input_folder/ will result in an error during training.

The last step is to set these paths in dataset_catalog.json and you are good to go:

"imagenet_a_filelist": {
    "train": ["<not_used>", "<not_used>"],
    "val": ["/path/to/output/imagenet-a/test_images.npy", "/path/to/output/imagenet-a/test_labels.npy"]
},
"imagenet_r_filelist": {
    "train": ["<not_used>", "<not_used>"],
    "val": ["/path/to/output/imagenet-r/test_images.npy", "/path/to/output/imagenet-r/test_labels.npy"]
},

Preparing iNaturalist2018 data files

Automatic download

Run the create_inaturalist2018_data_files.py script with the -d option as follows:

python extra_scripts/datasets/create_inaturalist2018_data_files.py \
    -i /path/to/inaturalist2018/ \
    -o /output_path/to/inaturalist2018
    -d

The folder /output_path/to/inaturalist2018 now contains the inaturalist2018 disk_filelist dataset. The last step is to set this path in dataset_catalog.json and you are good to go:

"inaturalist2018_filelist": {
    "train": ["/output_path/to/inaturalist2018/train_images.npy", "/output_path/to/inaturalist2018/train_labels.npy"],
    "val": ["/output_path/to/inaturalist2018/val_images.npy", "/output_path/to/inaturalist2018/val_labels.npy"]
},

Manual download

Download the full dataset by visiting the Inaturalist competion Github and downloading the "All training and validation images [120GB]", "Training annotations [26MB]", and "Validation annotations [26MB]" into the same directory, for example: "/path/to/inaturalist2018/". Expand each .tar archive.

The resulting folder should have the following structure:

/path/to/inaturalist2018/
    train_val2018/
        Actinopterygii/
            2229
                ... Images for class 2229 ...
            2230
                ... Images for class 2230 ...
            ...
        Amphibia
        ... All 14 "super categories" ...
    train2018.json
    val2018.json

Run the script where /path/to/inaturalist2018/ is the path of the folder containing the expanded tars:

python extra_scripts/datasets/create_inaturalist2018_data_files.py \
    -i /path/to/inaturalist2018/ \
    -o /output_path/to/inaturalist2018

The folder /output_path/to/inaturalist2018 now contains the inaturalist2018 disk_filelist .

Preparing the KITTI/distance data files

Run the create_kitti_dist_data_files.py script with the -d option as follows:

python extra_scripts/datasets/create_kitti_dist_data_files.py \
    -i /path/to/kitti/ \
    -o /output_path/to/kitti_distance
    -d

The folder /output_path/to/kitti_distance now contains the KITTI/distance disk_folder dataset. The last step is to set this path in dataset_catalog.json and you are good to go:

"kitti_dist_folder": {
    "train": ["/output_path/to/kitti_distance/train", "<ignored>"],
    "val": ["/output_path/to/kitti_distance/val", "<ignored>"]
},

Preparing the Oxford Flowers data files

Run the create_oxford_flowers_data_files.py script with the -d option as follows:

python extra_scripts/datasets/create_oxford_flowers_data_files.py \
    -i /path/to/flowers/ \
    -o /output_path/to/flowers
    -d

The folder /output_path/to/flowers now contains the Oxford Pets disk_folder dataset. The last step is to set this path in dataset_catalog.json and you are good to go:

"flowers_folder": {
    "train": ["/output_path/to/flowers/train", "<ignored>"],
    "val": ["/output_path/to/flowers/test", "<ignored>"]
},

Preparing the Oxford Pets data files

Run the create_oxford_pets_data_files.py script with the -d option as follows:

python extra_scripts/datasets/create_oxford_pets_data_files.py \
    -i /path/to/pets/ \
    -o /output_path/to/pets
    -d

The folder /output_path/to/pets now contains the Oxford Pets disk_folder dataset. The last step is to set this path in dataset_catalog.json and you are good to go:

"oxford_pets_folder": {
    "train": ["/output_path/to/pets/train", "<ignored>"],
    "val": ["/output_path/to/pets/test", "<ignored>"]
},

Preparing the Patch Camelyon data files

Run the create_patch_camelyon_data_files.py script with the -d option as follows:

python extra_scripts/datasets/create_patch_camelyon_data_files.py \
    -i /path/to/pcam/ \
    -o /output_path/to/pcam
    -d

The folder /output_path/to/pcam now contains the Patch Camelyon disk_folder dataset. The last step is to set this path in dataset_catalog.json and you are good to go:

"pcam_folder": {
    "train": ["/output_path/to/pcam/train", "<ignored>"],
    "val": ["/output_path/to/pcam/val", "<ignored>"]
},

Preparing the SmallNORB/azimuth data files

Run the create_small_norb_azimuth_data_files.py script with the -d option as follows:

python extra_scripts/datasets/create_small_norb_azimuth_data_files.py \
    -i /path/to/snorb/ \
    -o /output_path/to/snorb_azimuth
    -d

The folder /output_path/to/snorb_azimuth now contains the SmallNORB/azimuth disk_folder dataset. The last step is to set this path in dataset_catalog.json and you are good to go:

"small_norb_azimuth_folder": {
    "train": ["/output_path/to/snorb_azimuth/train", "<ignored>"],
    "val": ["/output_path/to/snorb_azimuth/val", "<ignored>"]
},

Preparing the SmallNORB/elevation data files

Run the create_small_norb_elevation_data_files.py script with the -d option as follows:

python extra_scripts/datasets/create_small_norb_elevation_data_files.py \
    -i /path/to/snorb/ \
    -o /output_path/to/snorb_elevation
    -d

The folder /output_path/to/snorb_elevation now contains the SmallNORB/elevation disk_folder dataset. The last step is to set this path in dataset_catalog.json and you are good to go:

"small_norb_elevation_folder": {
    "train": ["/output_path/to/snorb_elevation/train", "<ignored>"],
    "val": ["/output_path/to/snorb_elevation/val", "<ignored>"]
},

Preparing the Stanford Cars data files

Run the create_stanford_cars_data_files.py script with the -d option as follows:

python extra_scripts/datasets/create_stanford_cars_data_files.py \
    -i /path/to/cars/ \
    -o /output_path/to/cars
    -d

The folder /output_path/to/cars now contains the Stanford Cars disk_folder dataset. The last step is to set this path in dataset_catalog.json and you are good to go:

"stanford_cars_folder": {
    "train": ["/output_path/to/cars/train", "<ignored>"],
    "val": ["/output_path/to/cars/val", "<ignored>"]
},

Preparing the SUN397 data files

Run the create_sun397_data_files.py script with the -d option as follows:

python extra_scripts/datasets/create_sun397_data_files.py \
    -i /path/to/sun397/ \
    -o /path/to/sun397/ \
    -d

The folder /path/to/sun397/ now contains the Stanford Cars disk_filelist dataset. The last step is to set this path in dataset_catalog.json and you are good to go:

"sun397_filelist": {
    "train": ["/checkpoint/qduval/datasets/sun397/train_images.npy", "/checkpoint/qduval/datasets/sun397/train_labels.npy"],
    "val": ["/checkpoint/qduval/datasets/sun397/val_images.npy", "/checkpoint/qduval/datasets/sun397/val_labels.npy"]
},

Preparing UCF101/image data files

Automatic download

Run the create_ucf101_data_files.py script with the -d option as follows:

python extra_scripts/datasets/create_ucf101_data_files.py \
    -i /path/to/ucf101/ \
    -o /output_path/to/ucf101
    -d

The folder /output_path/ucf101 now contains the UCF101 image action recognition disk_folder dataset. The last step is to set this path in dataset_catalog.json and you are good to go:

"ucf101_folder": {
    "train": ["/output_path/to/ucf101/train", "<ignored>"],
    "val": ["/output_path/to/ucf101/val", "<ignored>"]
},

Manual download

Download the full dataset by visiting the UCF101 website:

Click on The UCF101 data set can be downloaded by "clicking here" to retrieve the data (all the videos).
Click on The Train/Test Splits for Action Recognition on UCF101 data set can be downloaded by clicking here to retrieve the splits.

Expand both archives in the same folder, say /path/to/ucf101. The resulting folder should have the following structure:

ucf101/
    UCF-101/
        ApplyEyeMakeup/
            ... videos ...
        ApplyLipstick/
            ... videos ...
    ucfTrainTestlist/
        classInd.txt
        testlist01.txt
        testlist02.txt
        testlist03.txt
        trainlist01.txt
        trainlist02.txt
        trainlist03.txt

Run the following commands (where /path/to/ucf101 is the path of the folder above):

python extra_scripts/datasets/create_ucf101_data_files.py \
    -i /path/to/ucf101/ \
    -o /output_path/to/ucf101

The folder /output_path/ucf101 now contains the UCF101 image action recognition dataset.

Data preparation (optional)

The following scripts are optional as VISSL's dataset_catalog.py supports reading the downloaded data directly for most of these. For all the datasets below, we assume that the datasets are in the format as described here.

Preparing COCO data files

python extra_scripts/datasets/create_coco_data_files.py \
    --json_annotations_dir /path/to/coco/annotations/ \
    --output_dir /tmp/vissl/datasets/coco/ \
    --train_imgs_path /path/to/coco/train2014 \
    --val_imgs_path /path/to/coco/val2014

Preparing VOC data files

for VOC2007 data_source_dir='/mnt/fair/VOC2007/' output_dir='/path/to/my/output/dir/voc2007/'
for VOC2012 data_source_dir='/mnt/fair/VOC2012/' output_dir='/path/to/my/output/dir/voc2012'

For VOC2007 dataset:

python extra_scripts/datasets/create_voc_data_files.py \
    --data_source_dir /path/to/VOC2007/ \
    --output_dir /tmp/vissl/datasets/voc07/

For VOC2012 dataset:

python extra_scripts/datasets/create_voc_data_files.py \
    --data_source_dir /path/to/VOC2012/ \
    --output_dir /tmp/vissl/datasets/voc12/

Preparing ImageNet and Places{205, 365} data files

python extra_scripts/datasets/create_imagenet_data_files.py \
    --data_source_dir /path/to/imagenet_full_size/ \
    --output_dir /tmp/vissl/datasets/imagenet1k/

python extra_scripts/datasets/create_imagenet_data_files.py \
    --data_source_dir /path/to/places205/ \
    --output_dir /tmp/vissl/datasets/places205/

python extra_scripts/datasets/create_imagenet_data_files.py \
    --data_source_dir /path/to/places365/ \
    --output_dir /tmp/vissl/datasets/places365/

Low-shot data sampling

Low-shot image classification is one of the benchmark tasks in the paper. VISSL support low-shot sampling and benchmarking on the PASCAL VOC dataset only.

We train on trainval split of VOC2007 dataset which has 5011 images and 20 classes. Hence the labels are of shape 5011 x 20. We generate 5 independent samples (for a given low-shot value k) by essentially generating 5 independent target files. For each class, we randomly pick the positive k samples and 19 * k negatives. Rest of the samples are ignored. We perform low-shot image classification on various different layers on the model (AlexNet, ResNet-50). The targets targets_data_file is usually obtained by extracting features for a given layer. Below command generates 5 samples for various k values:

python extra_scripts/datasets/create_voc_low_shot_samples.py \
    --targets_data_file /path/to/voc/numpy_targets.npy \
    --output_path /tmp/vissl/datasets/voc07/low_shot/labels/ \
    --k_values "1,2,4,8,16,32,64,96" \
    --num_samples 5

Generating Jigsaw Permutations for varying problem complexity

We provide scripts to change problem complexity of Jigsaw approach (as an axis of scaling in paper).

For the problem of Jigsaw, we vary the number of permutations used to solve the jigsaw task. In the paper, the permutations used ∈ [100, 2000, 10000]. We provide these permutations files for download here. To generate the permutations, use the command below:

python extra_scripts/generate_jigsaw_permutations.py \
    --output_dir /tmp/vissl//jigsaw_perms/ \
    -- N 2000

Converting FSDP checkpoints

VISSL allows to pretrain algorithms such as SwAV with FullyShardedDataParallel (FSDP) instead of DistributedDataParallel (DDP) to reduce the amount of GPU memory used during training and therefore train much bigger models.

FSDP models will save sharded checkpoints (one parameter file for each model shard, i.e. one parameter file for each GPU) instead of consolidated checkpoints (which contain all the weights).

While this allow faster training, it also couples the evaluation with the pre-training by forcing the benchmarking code to use exactly as many GPUs as the pre-training did. In some cases (linear evaluation), we would like to reduce the number of GPUs used during evaluation.

The convert_sharded_checkpoint.py script allow to transform sharded training checkpoints, containing parameters and optimizer states, into evaluation checkpoints containing only the parameter of the trunk to evaluate them, and compatible with any number of GPUs.

You can use the script like so:

python extra_scripts/convert_sharded_checkpoint.py \
    -i path/to/fsdp_checkpoint.torch \
    -o path/to/eval_checkpoint.torch \
    -t consolidated

The resulting eval_checkpoint.torch will be usable in DDP mode or FSDP mode, with any number of GPU.

Converting Models VISSL -> {Detectron2, ClassyVision, TorchVision}

We provide scripts to convert VISSL models to Detectron2 and ClassyVision compatible models.

Converting to Detectron2

All the ResNe(X)t models in VISSL can be converted to Detectron2 weights using following command:

python extra_scripts/convert_vissl_to_detectron2.py \
    --input_model_file <input_model>.pth  \
    --output_model <d2_model>.torch \
    --weights_type torch \
    --state_dict_key_name classy_state_dict

Converting to ClassyVision

All the ResNe(X)t models in VISSL can be converted to Detectron2 weights using following command:

python extra_scripts/convert_vissl_to_classy_vision.py \
    --input_model_file <input_model>.pth  \
    --output_model <d2_model>.torch \
    --state_dict_key_name classy_state_dict

Converting to TorchVision

All the ResNe(X)t models in VISSL can be converted to Torchvision weights using following command:

python extra_scripts/convert_vissl_to_torchvision.py \
    --model_url_or_file <input_model>.pth  \
    --output_dir /path/to/output/dir/ \
    --output_name <my_converted_model>.torch

Converting Caffe2 models -> VISSL

We provide conversion of all the caffe2 models in the paper

ResNet-50 models to VISSL

All the models have been added to ICCV19_MODEL_ZOO_FB.md.

Jigsaw model:

python extra_scripts/convert_caffe2_to_torchvision_resnet.py \
    --c2_model <model>.pkl \
    --output_model <pth_model>.torch \
    --jigsaw True --bgr2rgb True

Colorization model:

python extra_scripts/convert_caffe2_to_torchvision_resnet.py \
    --c2_model <model>.pkl \
    --output_model <pth_model>.torch \
    --bgr2rgb False

Supervised model:

python extra_scripts/convert_caffe2_to_pytorch_rn50.py \
    --c2_model <model>.pkl \
    --output_model <pth_model>.torch \
    --bgr2rgb True

AlexNet models to VISSL

AlexNet Jigsaw models:

python extra_scripts/convert_caffe2_to_vissl_alexnet.py \
    --weights_type caffe2 \
    --model_name jigsaw \
    --bgr2rgb True \
    --input_model_weights <model.pkl> \
    --output_model <pth_model>.torch

AlexNet Colorization models:

python extra_scripts/convert_caffe2_to_vissl_alexnet.py \
    --weights_type caffe2 \
    --model_name colorization \
    --input_model_weights <model.pkl> \
    --output_model <pth_model>.torch

AlexNet Supervised models:

python extra_scripts/convert_caffe2_to_vissl_alexnet.py \
    --weights_type caffe2 \
    --model_name supervised \
    --bgr2rgb True \
    --input_model_weights <model.pkl> \
    --output_model <pth_model>.torch

Converting Models ClassyVision -> VISSL

We provide scripts to convert ClassyVision models to VISSL compatible models.

python extra_scripts/convert_classy_vision_to_vissl_resnet.py \
    --input_model_file <input_model>.pth  \
    --output_model <d2_model>.torch \
    --depth 50

Converting Official RotNet and DeepCluster models -> VISSL

AlexNet RotNet model:

python extra_scripts/convert_caffe2_to_vissl_alexnet.py \
    --weights_type torch \
    --model_name rotnet \
    --input_model_weights <model> \
    --output_model <pth_model>.torch

AlexNet DeepCluster model:

python extra_scripts/convert_alexnet_models.py \
    --weights_type torch \
    --model_name deepcluster \
    --input_model_weights <model> \
    --output_model <pth_model>.torch

Name		Name	Last commit message	Last commit date
parent directory ..
datasets		datasets
README.md		README.md
__init__.py		__init__.py
convert_caffe2_to_torchvision_resnet.py		convert_caffe2_to_torchvision_resnet.py
convert_caffe2_to_vissl_alexnet.py		convert_caffe2_to_vissl_alexnet.py
convert_classy_vision_to_vissl_resnet.py		convert_classy_vision_to_vissl_resnet.py
convert_folder_to_filelist.py		convert_folder_to_filelist.py
convert_sharded_checkpoint.py		convert_sharded_checkpoint.py
convert_vissl_to_classy_vision.py		convert_vissl_to_classy_vision.py
convert_vissl_to_detectron2.py		convert_vissl_to_detectron2.py
convert_vissl_to_torchvision.py		convert_vissl_to_torchvision.py
create_low_shot_samples.py		create_low_shot_samples.py
experiment_spreadsheet_from_logs.py		experiment_spreadsheet_from_logs.py
generate_jigsaw_permutations.py		generate_jigsaw_permutations.py

Files

extra_scripts

Directory actions

More options

Directory actions

More options

Latest commit

History

extra_scripts

Folders and files

parent directory

README.md

Helpful scripts for data preparation and model conversion

Data preparation

Unified data preparation interface

Preparing CLEVR/Counts data files

Automatic download

Manual download

Preparing CLEVR/Dist data files

Preparing the dSprites/location data files

Preparing the dSprites/orientation data files

Preparing the Caltech101 data files

Preparing the DTD data files

Preparing the EuroSAT data files

Preparing the Food-101 data files

Preparing the FGVC Aircrafts data files

Preparing the GTSRB data files

Preparing the Imagenet-A and Imagenet-R data files

Preparing iNaturalist2018 data files

Automatic download

Manual download

Preparing the KITTI/distance data files

Preparing the Oxford Flowers data files

Preparing the Oxford Pets data files

Preparing the Patch Camelyon data files

Preparing the SmallNORB/azimuth data files

Preparing the SmallNORB/elevation data files

Preparing the Stanford Cars data files

Preparing the SUN397 data files

Preparing UCF101/image data files

Automatic download

Manual download

Data preparation (optional)

Preparing COCO data files

Preparing VOC data files

Preparing ImageNet and Places{205, 365} data files

Low-shot data sampling

Generating Jigsaw Permutations for varying problem complexity

Converting FSDP checkpoints

Converting Models VISSL -> {Detectron2, ClassyVision, TorchVision}

Converting to Detectron2

Converting to ClassyVision

Converting to TorchVision

Converting Caffe2 models -> VISSL

ResNet-50 models to VISSL

AlexNet models to VISSL

Converting Models ClassyVision -> VISSL

Converting Official RotNet and DeepCluster models -> VISSL