git clone https://github.com/dmarx/Multi-Modal-Comparators
cd 'Multi-Modal-Comparators'
cp pyproject.toml.INSTALL-ALL pyproject.toml
pip install poetry
poetry build
pip install dist/mmc*.whl

# optional final step:
#poe napm_installs
python src/mmc/napm_installs/__init__.py

To see which models are immediately available, run:

python -m mmc.loaders

For the most convenient experience, it is recommended that you perform the final poe napm_installs step. Omitting this step will make your one-time setup faster, but will make certain use cases more complex.

If you did not perform the optional poe napm_installs step, you likely received several warnings about models whose loaders could not be registered. These are models whose codebases depend on python code which is not trivially installable. You will still have access to all of the models supported by the library as if you had run the last step, but their loaders will not be queryable from the registry (see below) and will need to be loaded via the appropriate mmc.loader directly, which may be non-trivial to identify without the ability to query it from mmc's registry.

As a concrete example, if the napm step is skipped, the model [cloob - corwsonkb - cloob_laion_400m_vit_b_16_32_epochs] will not appear in the list of registered loaders, but can still be loaded like this:

from mmc.loaders import KatCloobLoader

model = KatCloobLoader(id='cloob_laion_400m_vit_b_16_32_epochs').load()

Invoking the load() method on an unregistered loader will invoke napm to prepare any uninstallable dependencies required to load the model. Next time you run python -m mmc.loaders, the CLOOB loader will show as registered and spinning up the registry will longer emit a warning for that model.

TLDR

# spin up the registry
from mmc import loaders

## Using the 'mocked' openai API to load supported CLIP models

from mmc.ez.CLIP import clip
clip.available_models()

# requesting a tokenizer before loading the model
# returns the openai clip SimpleTokenizer
#tokenize = clip.tokenize

# either of these works
model, preprocessor = clip.load('RN50')
model, preprocessor = clip.load('[clip - openai - RN50]')

# if we request the tokenizer *after* a model has been loaded, 
# the tokenizer appropriate to the loaded model is returned 
tokenize = clip.tokenize

###############################

## Slightly "closer to the metal" usage

from mmc.mock.openai import MockOpenaiClip
from mmc.registry import REGISTRY

cloob_query = {architecture='cloob'}
cloob_loaders = REGISTRY.find(**cloob_query)

# loader repl prints attributes for uniquely querying
print(cloob_loaders)

# loader returns a perceptor whose API is standardized across mmc
cloob_model = cloob_loaders[0].load()

# wrapper classes are provided for mocking popular implementations
# to facilitate drop-in compatibility with existing code
drop_in_replacement__cloob_model = MockOpenaiClip(cloob_model)

Spin up the model registry by importing the loaders module:

from mmc import loaders

To see which models are available:

from mmc.registry import REGISTRY

for loader in REGISTRY.find():
    print(loader)

You can constrain the result set by querying the registry for specific metadata attributes

# all CLIP models
clip_loaders = REGISTRY.find(architecture='clip')

# CLIP models published by openai
openai_clip_loaders = REGISTRY.find(architecture='clip', publisher='openai')

# All models published by MLFoundations (openCLIP)
mlf_loaders = REGISTRY.find(publisher='mlfoundations)'

# A specific model
rn50_loader = REGISTRY.find(architecture='clip', publisher='openai', id='RN50')
# NB: there may be multiple models matching a particular "id". the 'id' field
# only needs to be unique for a given architecture-publisher pair.

All pretrained checkpoints are uniquely identifiable by a combination of architecture, publisher, and id.

The above queries return lists of loader objects. If model artifacts (checkpoints, config) need to be downloaded, they will only be downloaded after the load() method on the loader is invoked.

loaders = REGISTRY.find(...)
loader = loaders[0] # just picking an arbitrary return value here, remember: loaders is a *list* of loaders
model = loader.load()

The load() method returns an instance of an mmc.MultiModalComparator. The MultiModalComparator class is a modality-agnostic abstraction. I'll get to the ins and outs of that another time.

You want something you can just drop into your code and it'll work. We got you. This library provides wrapper classes to mock the APIs of commonly used CLIP implementations (at present, OpenAI's CLIP is the only API which can be mocked). Individual loaders can be wrapped after instantiation (see below), but we also provide an "easy mode" API for best user experience.

Let's consider a codebase that already has openai/CLIP installed, via e.g. pip install git+https://github.com/openai/CLIP or pip install clip-anytorch.

All we have to do to integrate MMC is change

Step 1:

# let's call this the "normal clip object"
import clip

to

#from mmc import loaders ## optional, populates the mmc registry with all supported loaders
from mmc.ez.CLIP import clip

Step 2. (optional but strongly advised)

Make sure an references to clip.tokenize appear after clip.load() has already been invoked

And that's it!

Here's what this change gives us:

• clip.available_models()

  • In addition to all of the values normally returned when this method is invoked on the normal clip object, also returns all currently available models known to the MMC registry.

• clip.load()

  • Supports any model aliases returned by clip.available_models()
  • Invoking clip.load() with arguments like 'RN50' or 'ViT-B/16' will return the expected OopenAI clip model.
  • Additionally supports loading models using the mmc alias convention, i.e. clip.load('RN50') is equivalent to clip.load('[clip - openai - RN50]')

• clip.tokenize

  • Returns the OpenAI tokenizer by default, exactly as if we had invoked clip.tokenize on the original clip object rather than mmc.ez.clip.
  • if a model has already been loaded using the clip.load() method above, then clip.tokenize returns the text preprocessor appropriate to that model. At present, most CLIP implementations are published with Openai's tokenizer so you might not experience any issues if you don't reorganize this part of your code.

To wrap a MultiModalComparator so it can be used as a drop-in replacement with code compatible with OpenAI's CLIP:

from mmc.mock.openai import MockOpenaiClip

my_model = my_model_loader.load()
model = MockOpenaiClip(my_model)

(WIP, behavior likely to change in near future)

The MultiMMC class can be used to run inference against multiple mmc models in parallel. This form of ensemble is sometimes referred to as a "multi-perceptor".

To ensure that all models loaded into the MultiMMC are compatible, the MultiMMC instance is initialized by specifying the modalities it supports. We'll discuss modality objects in a bit.

from mmc.multimmc import MultiMMC
from mmc.modalities import TEXT, IMAGE

perceptor = MultiMMC(TEXT, IMAGE)

To load and use a model:

perceptor.load_model(
    architecture='clip', 
    publisher='openai', 
    id='RN50',
)

score = perceptor.compare(
    image=PIL.Image.open(...), 
    text=text_pos),
)

Additional models can be added to the ensemble via the load_model() method.

The MultiMMC does not support API mocking because of its reliance on the compare method.

Some model comparisons here

# [<architecture> - <publisher> - <id>]
[clip - openai - RN50]
[clip - openai - RN101]
[clip - openai - RN50x4]
[clip - openai - RN50x16]
[clip - openai - RN50x64]
[clip - openai - ViT-B/32]
[clip - openai - ViT-B/16]
[clip - openai - ViT-L/14]
[clip - openai - ViT-L/14@336px]
[clip - mlfoundations - RN50--openai]
[clip - mlfoundations - RN50--yfcc15m]
[clip - mlfoundations - RN50--cc12m]
[clip - mlfoundations - RN50-quickgelu--openai]
[clip - mlfoundations - RN50-quickgelu--yfcc15m]
[clip - mlfoundations - RN50-quickgelu--cc12m]
[clip - mlfoundations - RN101--openai]
[clip - mlfoundations - RN101--yfcc15m]
[clip - mlfoundations - RN101-quickgelu--openai]
[clip - mlfoundations - RN101-quickgelu--yfcc15m]
[clip - mlfoundations - RN50x4--openai]
[clip - mlfoundations - RN50x16--openai]
[clip - mlfoundations - ViT-B-32--openai]
[clip - mlfoundations - ViT-B-32--laion400m_e31]
[clip - mlfoundations - ViT-B-32--laion400m_e32]
[clip - mlfoundations - ViT-B-32--laion400m_avg]
[clip - mlfoundations - ViT-B-32-quickgelu--openai]
[clip - mlfoundations - ViT-B-32-quickgelu--laion400m_e31]
[clip - mlfoundations - ViT-B-32-quickgelu--laion400m_e32]
[clip - mlfoundations - ViT-B-32-quickgelu--laion400m_avg]
[clip - mlfoundations - ViT-B-16--openai]
[clip - mlfoundations - ViT-L-14--openai]
[clip - sbert - ViT-B-32-multilingual-v1]
[clip - sajjjadayobi - clipfa]

# The following models depend on napm for setup
[clip - navervision - kelip_ViT-B/32]
[cloob - crowsonkb - cloob_laion_400m_vit_b_16_16_epochs]
[cloob - crowsonkb - cloob_laion_400m_vit_b_16_32_epochs]
[clip - facebookresearch - clip_small_25ep]
[clip - facebookresearch - clip_base_25ep]
[clip - facebookresearch - clip_large_25ep]
[slip - facebookresearch - slip_small_25ep]
[slip - facebookresearch - slip_small_50ep]
[slip - facebookresearch - slip_small_100ep]
[slip - facebookresearch - slip_base_25ep]
[slip - facebookresearch - slip_base_50ep]
[slip - facebookresearch - slip_base_100ep]
[slip - facebookresearch - slip_large_25ep]
[slip - facebookresearch - slip_large_50ep]
[slip - facebookresearch - slip_large_100ep]
[simclr - facebookresearch - simclr_small_25ep]
[simclr - facebookresearch - simclr_base_25ep]
[simclr - facebookresearch - simclr_large_25ep]
[clip - facebookresearch - clip_base_cc3m_40ep]
[clip - facebookresearch - clip_base_cc12m_35ep]
[slip - facebookresearch - slip_base_cc3m_40ep]
[slip - facebookresearch - slip_base_cc12m_35ep]

The following is an estimate of the amount of space the loaded model occupies in memory:

If you would like to suggest a pre-trained model for future addition, you can add a comment to this issue

1. Create a loader class that encapsulates the logic for importing the model, loading weights, preprocessing inputs, and performing projections.
2. At the bottom of the file defining the loader class should be a code snippet that adds each respective checkpoint's loader to the registry.
3. Add an import for the new file to mmc/loaders/__init__.py. The imports in this file are the reason import mmc.loaders "spins up" the registry.
4. If the codebase on which the model depends can be installed, update pytproject.toml to install it.
5. Otherwise, add napm preparation at the top of the loaders load method (see cloob or kelip for examples), and also add napm setup to mmc/napm_installs/__init__.py
6. Add a test case to tests/test_mmc_loaders.py
7. Add a test script for the loader (see test_mmc_katcloob as an example)