utils.py

# utils.py

# Code adapted from https://github.com/facebookresearch/mae/tree/efb2a8062c206524e35e47d04501ed4f544c0ae8

# Define the utils

import matplotlib.pyplot as plt
import numpy as np
import torch

imgs_mean = np.array([0.5, 0.5, 0.5])
imgs_std = np.array([0.5, 0.5, 0.5])

def show_image(image, title=''):
    # image is [H, W, 3]
    assert image.shape[2] == 3
    plt.imshow(torch.clip((image * imgs_std + imgs_mean) * 255, 0, 255).int())
    plt.title(title, fontsize=16)
    plt.axis('off')
    return

def run_one_image(img, model,filename=None):
    
    x = torch.tensor(img)

    # make it a batch-like
    x = x.unsqueeze(dim=0)
    x = torch.einsum('nhwc->nchw', x)

    # run MAE
    loss, y, mask = model(x.float(), mask_ratio=0.75)
    y = model.unpatchify(y)
    y = torch.einsum('nchw->nhwc', y).detach().cpu()

    # visualize the mask
    mask = mask.detach()
    mask = mask.unsqueeze(-1).repeat(1, 1, model.patch_embed.patch_size[0]**2 *3)  # (N, H*W, p*p*3)
    mask = model.unpatchify(mask)  # 1 is removing, 0 is keeping
    mask = torch.einsum('nchw->nhwc', mask).detach().cpu()

    x = torch.einsum('nchw->nhwc', x)

    # masked image
    im_masked = x * (1 - mask)

    # MAE reconstruction pasted with visible patches
    im_paste = x * (1 - mask) + y * mask

    # make the plt figure larger
    plt.rcParams['figure.figsize'] = [24, 24]

    plt.subplot(1, 4, 1)
    show_image(x[0], "original")

    plt.subplot(1, 4, 2)
    show_image(im_masked[0], "masked")

    plt.subplot(1, 4, 3)
    show_image(y[0], "reconstruction")

    plt.subplot(1, 4, 4)
    show_image(im_paste[0], "reconstruction + visible")

    # Save the figure to a file
    if filename:
      plt.savefig(filename)

    plt.show()


# --------------------------------------------------------
# Position embedding utils
# --------------------------------------------------------


# --------------------------------------------------------
# 2D sine-cosine position embedding
# References:
# Transformer: https://github.com/tensorflow/models/blob/master/official/nlp/transformer/model_utils.py
# MoCo v3: https://github.com/facebookresearch/moco-v3
# --------------------------------------------------------
def get_2d_sincos_pos_embed(embed_dim, grid_size, cls_token=False):
    """
    grid_size: int of the grid height and width
    return:
    pos_embed: [grid_size*grid_size, embed_dim] or [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token)
    """
    grid_h = np.arange(grid_size, dtype=np.float64)
    grid_w = np.arange(grid_size, dtype=np.float64)
    grid = np.meshgrid(grid_w, grid_h)  # here w goes first
    grid = np.stack(grid, axis=0)

    grid = grid.reshape([2, 1, grid_size, grid_size])
    pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
    if cls_token:
        pos_embed = np.concatenate([np.zeros([1, embed_dim]), pos_embed], axis=0)
    return pos_embed


def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):
    assert embed_dim % 2 == 0

    # use half of dimensions to encode grid_h
    emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0])  # (H*W, D/2)
    emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1])  # (H*W, D/2)

    emb = np.concatenate([emb_h, emb_w], axis=1) # (H*W, D)
    return emb


def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
    """
    embed_dim: output dimension for each position
    pos: a list of positions to be encoded: size (M,)
    out: (M, D)
    """
    assert embed_dim % 2 == 0
    omega = np.arange(embed_dim // 2, dtype=np.float64)
    omega /= embed_dim / 2.
    omega = 1. / 10000**omega  # (D/2,)

    pos = pos.reshape(-1)  # (M,)
    out = np.einsum('m,d->md', pos, omega)  # (M, D/2), outer product

    emb_sin = np.sin(out) # (M, D/2)
    emb_cos = np.cos(out) # (M, D/2)

    emb = np.concatenate([emb_sin, emb_cos], axis=1)  # (M, D)
    return emb


# --------------------------------------------------------
# Interpolate position embeddings for high-resolution
# References:
# DeiT: https://github.com/facebookresearch/deit
# --------------------------------------------------------
def interpolate_pos_embed(model, checkpoint_model):
    if 'pos_embed' in checkpoint_model:
        pos_embed_checkpoint = checkpoint_model['pos_embed']
        embedding_size = pos_embed_checkpoint.shape[-1]
        num_patches = model.patch_embed.num_patches
        num_extra_tokens = model.pos_embed.shape[-2] - num_patches
        # height (== width) for the checkpoint position embedding
        orig_size = int((pos_embed_checkpoint.shape[-2] - num_extra_tokens) ** 0.5)
        # height (== width) for the new position embedding
        new_size = int(num_patches ** 0.5)
        # class_token and dist_token are kept unchanged
        if orig_size != new_size:
            print("Position interpolate from %dx%d to %dx%d" % (orig_size, orig_size, new_size, new_size))
            extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens]
            # only the position tokens are interpolated
            pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:]
            pos_tokens = pos_tokens.reshape(-1, orig_size, orig_size, embedding_size).permute(0, 3, 1, 2)
            pos_tokens = torch.nn.functional.interpolate(
                pos_tokens, size=(new_size, new_size), mode='bicubic', align_corners=False)
            pos_tokens = pos_tokens.permute(0, 2, 3, 1).flatten(1, 2)
            new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1)
            checkpoint_model['pos_embed'] = new_pos_embed