models/eva_clip/utils_qformer.py

import importlib
import math
import os
import random

import cv2
import numpy as np
import torch
import torch.nn.functional as F
from torchvision.utils import make_grid
from transformers import PretrainedConfig


def seed_everything(seed):
    os.environ["PL_GLOBAL_SEED"] = str(seed)
    random.seed(seed)
    np.random.seed(seed)
    torch.manual_seed(seed)
    torch.cuda.manual_seed_all(seed)


def is_torch2_available():
    return hasattr(F, "scaled_dot_product_attention")


def instantiate_from_config(config):
    if "target" not in config:
        if config == '__is_first_stage__' or config == "__is_unconditional__":
            return None
        raise KeyError("Expected key `target` to instantiate.")
    return get_obj_from_str(config["target"])(**config.get("params", {}))


def get_obj_from_str(string, reload=False):
    module, cls = string.rsplit(".", 1)
    if reload:
        module_imp = importlib.import_module(module)
        importlib.reload(module_imp)
    return getattr(importlib.import_module(module, package=None), cls)


def drop_seq_token(seq, drop_rate=0.5):
    idx = torch.randperm(seq.size(1))
    num_keep_tokens = int(len(idx) * (1 - drop_rate))
    idx = idx[:num_keep_tokens]
    seq = seq[:, idx]
    return seq


def import_model_class_from_model_name_or_path(
    pretrained_model_name_or_path: str, revision: str, subfolder: str = "text_encoder"
):
    text_encoder_config = PretrainedConfig.from_pretrained(
        pretrained_model_name_or_path, subfolder=subfolder, revision=revision
    )
    model_class = text_encoder_config.architectures[0]

    if model_class == "CLIPTextModel":
        from transformers import CLIPTextModel

        return CLIPTextModel
    elif model_class == "CLIPTextModelWithProjection":  # noqa RET505
        from transformers import CLIPTextModelWithProjection

        return CLIPTextModelWithProjection
    else:
        raise ValueError(f"{model_class} is not supported.")


def resize_numpy_image_long(image, resize_long_edge=768):
    h, w = image.shape[:2]
    if max(h, w) <= resize_long_edge:
        return image
    k = resize_long_edge / max(h, w)
    h = int(h * k)
    w = int(w * k)
    image = cv2.resize(image, (w, h), interpolation=cv2.INTER_LANCZOS4)
    return image


# from basicsr
def img2tensor(imgs, bgr2rgb=True, float32=True):
    """Numpy array to tensor.

    Args:
        imgs (list[ndarray] | ndarray): Input images.
        bgr2rgb (bool): Whether to change bgr to rgb.
        float32 (bool): Whether to change to float32.

    Returns:
        list[tensor] | tensor: Tensor images. If returned results only have
            one element, just return tensor.
    """

    def _totensor(img, bgr2rgb, float32):
        if img.shape[2] == 3 and bgr2rgb:
            if img.dtype == 'float64':
                img = img.astype('float32')
            img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
        img = torch.from_numpy(img.transpose(2, 0, 1))
        if float32:
            img = img.float()
        return img

    if isinstance(imgs, list):
        return [_totensor(img, bgr2rgb, float32) for img in imgs]
    return _totensor(imgs, bgr2rgb, float32)


def tensor2img(tensor, rgb2bgr=True, out_type=np.uint8, min_max=(0, 1)):
    """Convert torch Tensors into image numpy arrays.

    After clamping to [min, max], values will be normalized to [0, 1].

    Args:
        tensor (Tensor or list[Tensor]): Accept shapes:
            1) 4D mini-batch Tensor of shape (B x 3/1 x H x W);
            2) 3D Tensor of shape (3/1 x H x W);
            3) 2D Tensor of shape (H x W).
            Tensor channel should be in RGB order.
        rgb2bgr (bool): Whether to change rgb to bgr.
        out_type (numpy type): output types. If ``np.uint8``, transform outputs
            to uint8 type with range [0, 255]; otherwise, float type with
            range [0, 1]. Default: ``np.uint8``.
        min_max (tuple[int]): min and max values for clamp.

    Returns:
        (Tensor or list): 3D ndarray of shape (H x W x C) OR 2D ndarray of
        shape (H x W). The channel order is BGR.
    """
    if not (torch.is_tensor(tensor) or (isinstance(tensor, list) and all(torch.is_tensor(t) for t in tensor))):
        raise TypeError(f'tensor or list of tensors expected, got {type(tensor)}')

    if torch.is_tensor(tensor):
        tensor = [tensor]
    result = []
    for _tensor in tensor:
        _tensor = _tensor.squeeze(0).float().detach().cpu().clamp_(*min_max)
        _tensor = (_tensor - min_max[0]) / (min_max[1] - min_max[0])

        n_dim = _tensor.dim()
        if n_dim == 4:
            img_np = make_grid(_tensor, nrow=int(math.sqrt(_tensor.size(0))), normalize=False).numpy()
            img_np = img_np.transpose(1, 2, 0)
            if rgb2bgr:
                img_np = cv2.cvtColor(img_np, cv2.COLOR_RGB2BGR)
        elif n_dim == 3:
            img_np = _tensor.numpy()
            img_np = img_np.transpose(1, 2, 0)
            if img_np.shape[2] == 1:  # gray image
                img_np = np.squeeze(img_np, axis=2)
            else:
                if rgb2bgr:
                    img_np = cv2.cvtColor(img_np, cv2.COLOR_RGB2BGR)
        elif n_dim == 2:
            img_np = _tensor.numpy()
        else:
            raise TypeError(f'Only support 4D, 3D or 2D tensor. But received with dimension: {n_dim}')
        if out_type == np.uint8:
            # Unlike MATLAB, numpy.unit8() WILL NOT round by default.
            img_np = (img_np * 255.0).round()
        img_np = img_np.astype(out_type)
        result.append(img_np)
    if len(result) == 1:
        result = result[0]
    return result