code/clip/clip.py

# Code ported from https://github.com/openai/CLIP

import hashlib
import os
import urllib
import warnings
from typing import Union, List

import torch
from PIL import Image
from torchvision.transforms import Compose, Resize, CenterCrop, ToTensor, Normalize, RandomResizedCrop, RandomAffine
from tqdm import tqdm

from clip.model import build_model
from clip.tokenizer import SimpleTokenizer as _Tokenizer

__all__ = ["available_models", "load", "tokenize"]
_tokenizer = _Tokenizer()

_MODELS = {
    "RN50": "https://openaipublic.azureedge.net/clip/models/afeb0e10f9e5a86da6080e35cf09123aca3b358a0c3e3b6c78a7b63bc04b6762/RN50.pt",
    "RN101": "https://openaipublic.azureedge.net/clip/models/8fa8567bab74a42d41c5915025a8e4538c3bdbe8804a470a72f30b0d94fab599/RN101.pt",
    "RN50x4": "https://openaipublic.azureedge.net/clip/models/7e526bd135e493cef0776de27d5f42653e6b4c8bf9e0f653bb11773263205fdd/RN50x4.pt",
    "ViT-B/32": "https://openaipublic.azureedge.net/clip/models/40d365715913c9da98579312b702a82c18be219cc2a73407c4526f58eba950af/ViT-B-32.pt",
    "ViT-B/16": "https://openaipublic.azureedge.net/clip/models/5806e77cd80f8b59890b7e101eabd078d9fb84e6937f9e85e4ecb61988df416f/ViT-B-16.pt",
}


def _download(url: str, root: str = os.path.expanduser("~/.cache/clip")):
    os.makedirs(root, exist_ok=True)
    filename = os.path.basename(url)

    expected_sha256 = url.split("/")[-2]
    download_target = os.path.join(root, filename)

    if os.path.exists(download_target) and not os.path.isfile(download_target):
        raise RuntimeError(f"{download_target} exists and is not a regular file")

    if os.path.isfile(download_target):
        if hashlib.sha256(open(download_target, "rb").read()).hexdigest() == expected_sha256:
            return download_target
        else:
            warnings.warn(f"{download_target} exists, but the SHA256 checksum does not match; re-downloading the file")

    with urllib.request.urlopen(url) as source, open(download_target, "wb") as output:
        with tqdm(total=int(source.info().get("Content-Length")), ncols=80, unit='iB', unit_scale=True) as loop:
            while True:
                buffer = source.read(8192)
                if not buffer:
                    break

                output.write(buffer)
                loop.update(len(buffer))

    if hashlib.sha256(open(download_target, "rb").read()).hexdigest() != expected_sha256:
        raise RuntimeError(f"Model has been downloaded but the SHA256 checksum does not not match")

    return download_target

def _convert_to_rgb(image):
    return image.convert('RGB')

def _transform(n_px: int, is_train: bool, affine: bool = False):
    normalize = Normalize((0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711))
    if is_train:
        if affine:
            return Compose([
                RandomAffine(degrees=30,translate=(0.3,0.3),shear =[-30,30,-30,30], scale=(1,2), fill=255, interpolation=Image.BICUBIC),
                RandomResizedCrop(n_px, scale=(0.8, 1.0), interpolation=Image.BICUBIC),
                _convert_to_rgb,
                ToTensor(),
                normalize,
            ])
        else:
            return Compose([
                RandomResizedCrop(n_px, scale=(0.9, 1.0), interpolation=Image.BICUBIC),
                _convert_to_rgb,
                ToTensor(),
                normalize,
            ])
    else:
        return Compose([
            Resize(n_px, interpolation=Image.BICUBIC),
            CenterCrop(n_px),
            _convert_to_rgb,
            ToTensor(),
            normalize,
        ])



def available_models() -> List[str]:
    """Returns the names of available CLIP models"""
    return list(_MODELS.keys())


def load(name: str, device: Union[str, torch.device] = "cuda" if torch.cuda.is_available() else "cpu", jit=True, is_train=False, pretrained=True, weight_sharing=False, feature_fusion='avg', affine_transformation=False, num_class=None):
    """Load a CLIP model
    Parameters
    ----------
    name : str
        A model name listed by `clip.available_models()`, or the path to a model checkpoint containing the state_dict
    device : Union[str, torch.device]
        The device to put the loaded model
    jit : bool
        Whether to load the optimized JIT model (default) or more hackable non-JIT model.
    Returns
    -------
    model : torch.nn.Module
        The CLIP model
    preprocess : Callable[[PIL.Image], torch.Tensor]
        A torchvision transform that converts a PIL image into a tensor that the returned model can take as its input
    """
    assert num_class is not None
    if name in _MODELS:
        model_path = _download(_MODELS[name])
    elif os.path.isfile(name):
        model_path = name
    else:
        raise RuntimeError(f"Model {name} not found; available models = {available_models()}")

    try:
        # loading JIT archive
        model = torch.jit.load(model_path, map_location=device if jit else "cpu").eval()
        state_dict = None
    except RuntimeError:
        # loading saved state dict
        if jit:
            warnings.warn(f"File {model_path} is not a JIT archive. Loading as a state dict instead")
            jit = False
        state_dict = torch.load(model_path, map_location="cpu")

    if not jit:
        try:
            model = build_model(state_dict or model.state_dict(), weight_sharing, feature_fusion, num_class = num_class).to(device)
        except KeyError:
            sd = {k[7:]: v for k,v in state_dict["state_dict"].items()}
            model = build_model(sd, weight_sharing, feature_fusion, num_class=num_class).to(device)

        if str(device) == "cpu":
            model.float()
        return model, \
               _transform(model.visual.input_resolution, is_train=True, affine=affine_transformation), \
               _transform(model.visual.input_resolution, is_train=False)
    #sanity check to make sure we are not loading up old version of networks directly
    assert model.visual2 is not None
    # patch the device names
    device_holder = torch.jit.trace(lambda: torch.ones([]).to(torch.device(device)), example_inputs=[])
    device_node = [n for n in device_holder.graph.findAllNodes("prim::Constant") if "Device" in repr(n)][-1]

    def patch_device(module):
        graphs = [module.graph] if hasattr(module, "graph") else []
        if hasattr(module, "forward1"):
            graphs.append(module.forward1.graph)

        for graph in graphs:
            for node in graph.findAllNodes("prim::Constant"):
                if "value" in node.attributeNames() and str(node["value"]).startswith("cuda"):
                    node.copyAttributes(device_node)
                    
    #load sketch branch
    weight_sharing = model.weight_sharing
    if weight_sharing:
        model.visual2 = model.visual
    else:
        #copy weight from image branch
        sd1 = model.visual.state_dict()
        sd2 = model.visual2.state_dict()
        for name, param in sd1.items():
            assert name in sd2
            sd2[name].copy_(param)

    model.apply(patch_device)
    patch_device(model.encode_image)
    patch_device(model.encode_text)

    # patch dtype to float32 on CPU
    if str(device) == "cpu":
        float_holder = torch.jit.trace(lambda: torch.ones([]).float(), example_inputs=[])
        float_input = list(float_holder.graph.findNode("aten::to").inputs())[1]
        float_node = float_input.node()

        def patch_float(module):
            graphs = [module.graph] if hasattr(module, "graph") else []
            if hasattr(module, "forward1"):
                graphs.append(module.forward1.graph)

            for graph in graphs:
                for node in graph.findAllNodes("aten::to"):
                    inputs = list(node.inputs())
                    for i in [1, 2]:  # dtype can be the second or third argument to aten::to()
                        if inputs[i].node()["value"] == 5:
                            inputs[i].node().copyAttributes(float_node)

        model.apply(patch_float)
        patch_float(model.encode_image)
        patch_float(model.encode_text)

        model.float()

    return model, \
           _transform(model.input_resolution.item(), is_train=True, affine=affine_transformation), \
           _transform(model.input_resolution.item(), is_train=False)


def tokenize(texts: Union[str, List[str]], context_length: int = 77) -> torch.LongTensor:
    """
    Returns the tokenized representation of given input string(s)
    Parameters
    ----------
    texts : Union[str, List[str]]
        An input string or a list of input strings to tokenize
    context_length : int
        The context length to use; all CLIP models use 77 as the context length
    Returns
    -------
    A two-dimensional tensor containing the resulting tokens, shape = [number of input strings, context_length]
    """
    if isinstance(texts, str):
        texts = [texts]

    sot_token = _tokenizer.encoder["<start_of_text>"]
    eot_token = _tokenizer.encoder["<end_of_text>"]
    all_tokens = [[sot_token] + _tokenizer.encode(text) + [eot_token] for text in texts]
    result = torch.zeros(len(all_tokens), context_length, dtype=torch.long)

    for i, tokens in enumerate(all_tokens):
        if len(tokens) > context_length: # Truncate
            tokens = tokens[:context_length]
        result[i, :len(tokens)] = torch.tensor(tokens)

    return result