Spaces:
Runtime error
Runtime error
File size: 1,918 Bytes
98e527d 0c48c4a 98e527d |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 |
from typing import Dict
import numpy as np
import tensorflow as tf
from PIL import Image
from tensorflow import keras
RESOLUTION = 224
PATCH_SIZE = 16
crop_layer = keras.layers.CenterCrop(RESOLUTION, RESOLUTION)
norm_layer = keras.layers.Normalization(
mean=[0.485 * 255, 0.456 * 255, 0.406 * 255],
variance=[(0.229 * 255) ** 2, (0.224 * 255) ** 2, (0.225 * 255) ** 2],
)
def preprocess_image(orig_image: Image, size: int):
"""Image preprocessing utility."""
image = np.array(orig_image)
image_resized = tf.expand_dims(image, 0)
resize_size = int((256 / 224) * size)
image_resized = tf.image.resize(
image_resized, (resize_size, resize_size), method="bicubic"
)
image_resized = crop_layer(image_resized)
return image_resized.numpy().squeeze(), norm_layer(image_resized).numpy()
# Reference:
# https://github.com/facebookresearch/dino/blob/main/visualize_attention.py
def get_cls_attention_map(
preprocessed_image: np.ndarray,
attn_score_dict: Dict[str, np.ndarray],
block_key="ca_ffn_block_0_att",
):
"""Utility to generate class-attention map modeling spatial-class relationships."""
w_featmap = preprocessed_image.shape[2] // PATCH_SIZE
h_featmap = preprocessed_image.shape[1] // PATCH_SIZE
attention_scores = attn_score_dict[block_key]
nh = attention_scores.shape[1] # Number of attention heads.
# Taking the representations from CLS token.
attentions = attention_scores[0, :, 0, 1:].reshape(nh, -1)
# Reshape the attention scores to resemble mini patches.
attentions = attentions.reshape(nh, w_featmap, h_featmap)
attentions = attentions.transpose((1, 2, 0))
# Resize the attention patches to 224x224 (224: 14x16)
attentions = tf.image.resize(
attentions,
size=(h_featmap * PATCH_SIZE, w_featmap * PATCH_SIZE),
method="bicubic",
)
return attentions.numpy()
|