import torch
import torch.nn as nn
from skimage import io
import numpy as np
import albumentations as A
import matplotlib.pyplot as pltVisualize Salincy Maps in Timm Models (ViT) using PyTorch Hooks!!
Transformer
Deep Learning
PyTorch
Tutorial code of how I Visualize Attention Maps and Feature Maps (Salieancy Maps) in PyTorch timm using PyTorch Hooks as timm does not provide attention intermediates
This Blogs is a Part 2 of my Vision Transformer From Scratch Series, where I implement ViT from scratch. This Blog explores more on the Exaplainibility and visualizing the Saliency maps of any Transformer based Vision Model, It also delves into the concept of PyTorch Hooks which is used here to get the attention outputs from timm models. Where my basic understanding comes from Official Pytorch Hooks Documentation.
also credits : - https://github.com/huggingface/pytorch-image-models/blob/main/timm/models/vision_transformer.py
also available as youtube Video: - https://youtu.be/7q3NGMkEtjI?si=zRnBBIo5LEwrQM7e
img_path = "/kaggle/input/image-net-visualization/bird_1.png"
model_path = "vit_small_patch16_224.dino"
image_size = 1024def get_image(im_path, shape=image_size):
img = io.imread(im_path)[:,:,:3] ## HXWXC
H,W,C = img.shape
org_img = img.copy()
img = A.Normalize()(image=img)["image"] ## HxWxC
norm_img = img.copy()
img = np.transpose(img,(2,0,1)) ## CxHxW
img = np.expand_dims(img,0) ## B=1,CxHxW
img = torch.tensor(img)
img = nn.Upsample((shape,shape),mode='bilinear')(img)
return img, norm_img, org_imgimg,norm_img,org_img = get_image(img_path)io.imshow(org_img)
io.imshow(norm_img)
%%capture
! pip install timmimport timm
model = timm.create_model(model_path,pretrained=True,
img_size=image_size,
dynamic_img_size=True)model = model.cuda()outputs = {}
def get_outputs(name:str):
def hook(model, input, output):
outputs[name] = output.detach()
return hookmodel.blocks[-1].attn.q_norm.register_forward_hook(get_outputs('Q'))
model.blocks[-1].attn.k_norm.register_forward_hook(get_outputs('K'))
model.blocks[-1].register_forward_hook(get_outputs('features'))
model(img.cuda())tensor([[-1.6873e+00, 3.3118e+00, 1.6206e-02, 7.5276e+00, -3.8783e+00,
7.9405e-01, -2.5275e-01, -9.2937e-01, -7.6219e+00, 1.8461e+00,
-3.7896e+00, 6.9062e+00, 4.9580e+00, 7.0582e+00, -1.4598e+00,
1.8604e+00, 3.1002e+00, -2.3646e-01, 9.7898e-01, 1.4716e-01,
4.8010e+00, 9.8696e+00, -6.2256e+00, -4.0314e+00, 4.8572e-01,
4.6386e+00, 5.2622e+00, 8.2102e+00, 6.6765e+00, -1.0041e+01,
9.1688e-01, 2.6036e+00, 1.0652e+00, -1.0549e+01, -1.5771e+00,
-2.4766e+00, -4.6636e+00, 5.6768e+00, -1.7559e+00, -4.9503e-01,
-2.5628e+00, 1.0277e+01, -4.6592e+00, -3.6269e+00, 6.4073e+00,
-2.8656e+00, -8.2731e+00, 9.1952e+00, -2.4074e+00, -5.6380e-01,
-3.4731e+00, -4.7151e+00, 1.8822e-01, -3.8197e+00, 2.5455e+00,
-1.3325e-01, -4.3335e+00, 7.5765e+00, -3.5712e+00, -4.4365e+00,
4.9343e-01, 9.1502e-01, 1.0584e+01, 3.9611e+00, 3.8866e+00,
-4.6030e+00, 1.4714e+00, 4.2984e-01, -3.4557e+00, 1.0801e+00,
-4.1478e+00, 5.2226e+00, 2.8259e+00, 9.7525e-01, 6.9041e+00,
1.2095e-01, 2.4013e+00, 1.6391e+00, -6.0108e+00, -9.3346e+00,
-4.0121e+00, 1.8157e+00, -7.2864e-01, 8.5839e+00, -1.0082e+01,
1.2923e+00, -1.8654e+00, -1.0347e+01, 4.6857e+00, -2.7471e+00,
3.4723e+00, 4.6032e+00, -2.5349e+00, -2.0960e+00, 1.6807e-01,
1.9170e+00, -6.1443e+00, 1.0321e+00, -1.6014e+00, 5.8918e+00,
9.2721e+00, -1.2786e+00, -6.3613e+00, 5.5546e+00, -1.2483e+01,
3.0828e+00, -6.0050e-02, -1.3473e+01, 6.6318e-01, -5.3011e+00,
-3.5312e+00, -1.5129e+00, 4.1372e+00, 1.4356e+00, 1.0351e+01,
-1.7940e+00, -1.2211e+01, 2.8142e+00, -1.5071e+00, -6.9505e-01,
-3.1786e+00, 7.8665e+00, 6.7557e-01, 2.6826e+00, -2.4210e+00,
-2.8962e+00, 5.5639e+00, -1.8775e+00, -1.2859e+01, -7.5849e-01,
1.6650e+00, 2.6447e+00, -6.3936e-01, 1.5060e+00, 6.0126e-01,
3.9601e+00, -1.8724e+00, -3.1135e+00, -2.8265e+00, 1.6418e+00,
-5.6146e+00, -9.4629e-02, -1.3282e-01, 7.6737e+00, -1.5090e+00,
5.6526e+00, -9.7128e-01, -3.6167e+00, 1.7872e+00, -2.7470e+00,
-2.4107e+00, -7.7301e-03, 5.4260e+00, -2.6830e-01, -6.0379e-01,
2.4405e+00, -3.0422e+00, -6.1176e+00, -1.3285e+00, -1.1136e+01,
-3.1303e+00, 2.8008e+00, 2.7368e+00, -2.8170e+00, -4.1028e+00,
5.4076e+00, 1.6082e+00, -9.9726e-01, 5.4990e+00, 1.8774e+00,
-5.4387e-01, -2.9433e-01, 3.6973e+00, -4.5119e+00, -3.5505e+00,
2.8779e+00, 3.0147e-01, -2.3486e+00, -1.5237e+00, 1.1389e+00,
2.2250e+00, 6.8706e+00, 3.7546e+00, -1.6722e+00, -6.4871e-01,
3.2394e+00, 6.6144e-01, 1.0893e+00, 1.1054e+00, 4.0759e+00,
-4.9504e+00, -4.0154e+00, -1.4490e+00, 1.8689e+00, -3.9057e+00,
2.2533e+00, -6.3996e+00, -5.5835e-01, 2.4417e+00, 2.1308e+00,
3.9925e+00, 4.4550e-01, 1.1531e+00, 4.7416e+00, 2.0096e+00,
-1.2605e+00, -9.5855e-01, -5.2844e+00, -4.0194e+00, 7.0700e-01,
5.4294e+00, 3.3919e+00, 4.9089e+00, 4.1371e-01, 2.2920e+00,
-1.7367e-01, -4.7534e+00, 5.0581e+00, 8.1151e+00, -1.5215e+00,
3.7808e+00, -2.9530e+00, -9.3015e+00, 6.1947e-01, -3.9256e-01,
-4.1782e+00, 1.9628e+00, 6.7975e+00, 1.2619e+00, 1.7882e+00,
-9.8904e-01, -3.4642e-01, -2.1700e+00, 8.9316e+00, 4.0332e+00,
-3.3242e+00, 3.7758e+00, 2.3272e+00, -2.6639e+00, 1.8211e+00,
2.0843e+00, 9.2511e+00, 7.0950e-01, 2.3567e+00, -3.4504e+00,
-1.2123e+01, -2.2370e-01, -4.0639e+00, -7.6606e-01, -5.7467e-01,
6.2722e+00, -1.1846e+00, 7.4569e-01, -2.3022e+00, -7.6787e+00,
1.5879e+00, -5.4248e+00, 4.4352e+00, -3.0992e-01, -9.8239e+00,
4.7596e+00, -7.8933e+00, -1.5991e+00, 3.5556e+00, -8.2829e+00,
-2.8837e+00, 8.0843e+00, 4.4499e+00, -7.8938e+00, 1.1607e+01,
3.3790e+00, -1.0773e+01, 4.5276e+00, 2.0844e+00, 2.4876e+00,
2.4884e+00, 2.1223e-01, -3.0398e+00, -6.3939e+00, -2.2402e+00,
-9.1261e-01, 2.2716e+00, 7.5233e+00, 3.2600e+00, 1.3579e+00,
3.0965e+00, -1.2645e+00, 3.8311e-01, -3.8111e+00, -1.9618e+00,
2.7574e+00, 1.0933e+01, -8.1753e+00, -5.7976e-01, 4.3519e+00,
-3.1721e+00, 3.5473e+00, 8.9204e-01, -5.8862e+00, 5.9441e+00,
-4.1584e+00, 2.6372e+00, 5.9613e-01, -3.5491e+00, -7.7043e+00,
-5.9683e+00, -1.9161e-01, -3.4110e+00, -4.9706e+00, 3.6422e+00,
2.0380e+00, -1.7211e-01, 2.2398e+00, -8.1027e-01, -1.7448e+00,
-8.5672e-01, -3.8151e-02, -1.5993e+00, 3.6409e+00, 3.0737e+00,
1.5809e-01, -2.5176e-01, 3.4933e+00, -6.0633e+00, 6.0766e+00,
6.8786e+00, -2.1457e+00, 4.7027e-01, -5.3296e-01, -9.8374e+00,
-4.7503e+00, -1.6798e+01, 5.6179e+00, 1.0797e+00, 9.5722e-01,
1.0259e+01, 2.4940e+00, 5.1237e+00, -3.9690e+00, 6.3755e+00,
-2.4782e+00, 5.0692e+00, 1.1048e+00, -3.4466e+00, -8.6174e+00,
-1.3207e+00, -1.0411e+01, -1.8757e+00, -1.1057e+01, -5.9575e+00,
1.0189e+00, 2.0985e+00, -3.2184e+00, -3.6003e+00, -4.3049e+00,
3.8259e+00, -4.0951e+00, -1.8868e+00, -4.7576e+00, -6.7396e+00,
6.9816e-01, -2.2439e+00, -4.5493e+00, 8.9310e+00, -1.6068e-01,
-2.5645e+00, 7.0637e+00, 1.7403e+01, 2.9385e+00, 1.5559e+00,
5.3462e+00, 1.3728e+01, 3.9456e+00, -5.9510e+00, 3.1551e+00,
4.6709e+00, -3.6971e+00, -1.2921e+00, 2.9455e+00, 4.7313e+00,
-2.5174e+00, 6.0789e+00, -7.3726e+00, -2.1013e+00]], device='cuda:0',
grad_fn=<SelectBackward0>)scale = model.blocks[-1].attn.scaleoutputs["attention"] = (outputs["Q"]@ outputs["K"].transpose(-2, -1))#*scale
# outputs["attention"] = outputs["attention"].softmax(dim=-1)outputs["attention"].shapetorch.Size([1, 6, 4097, 4097])b,num_heads,num_patches_1,num_patches_1 = outputs["attention"].shape
map_size = int(np.sqrt(num_patches_1-1))
for attention_head in range(num_heads):
attention_map = outputs["attention"][:,attention_head,0,1:] ## 1,4096
attention_map = attention_map.view(1,1,map_size,map_size)
attention_map = nn.Upsample(size=(image_size,image_size))(attention_map)
attention_map = attention_map[0,0,:,:].detach().cpu().numpy()
io.imshow(attention_map)
plt.show()
outputs["features"].shapetorch.Size([1, 4097, 384])features = outputs["features"].mean(-1)
features = features[:,1:]
features = features.view(1,1,map_size,map_size)
features = nn.Upsample(size=(image_size,image_size))(features)
features = features[0,0,:,:].detach().cpu().numpy()
io.imshow(features)/opt/conda/lib/python3.10/site-packages/skimage/io/_plugins/matplotlib_plugin.py:149: UserWarning: Low image data range; displaying image with stretched contrast.
lo, hi, cmap = _get_display_range(image)