from fastai.vision.all import *Problem setup
path = untar_data(URLs.IMAGEWOOF_320)model = resnet50(pretrained=True).cuda()/home/molly/miniconda3/envs/fastai/lib/python3.10/site-packages/torchvision/models/_utils.py:208: UserWarning: The parameter 'pretrained' is deprecated since 0.13 and will be removed in 0.15, please use 'weights' instead.
warnings.warn(
/home/molly/miniconda3/envs/fastai/lib/python3.10/site-packages/torchvision/models/_utils.py:223: UserWarning: Arguments other than a weight enum or `None` for 'weights' are deprecated since 0.13 and will be removed in 0.15. The current behavior is equivalent to passing `weights=ResNet50_Weights.IMAGENET1K_V1`. You can also use `weights=ResNet50_Weights.DEFAULT` to get the most up-to-date weights.
warnings.warn(msg)dls = DataBlock((ImageBlock,CategoryBlock),
splitter=IndexSplitter([0,2]),
item_tfms=Resize(320),
batch_tfms=aug_transforms(),
get_items=get_image_files).dataloaders(path)def model_call(model,x):
#same as model.forward
x = model.conv1(x)
x = model.bn1(x)
x = model.relu(x)
x = model.maxpool(x)
x = model.layer1(x)
x = model.layer2(x)
x = model.layer3(x)
x = model.layer4(x)
#x = model.avgpool(x)
#x = torch.flatten(x, 1)
#x = model.fc(x)
return xmodel.fcLinear(in_features=2048, out_features=1000, bias=True)#model_call(model,dls.one_batch()[0])Going to use a relatively small number of buckets, could use much more if we wanted to have less total collisions.
nbits=8
buckets=2**nbits
bits=torch.arange(0,8,device='cuda')
nbits,buckets,2.**bits(8,
256,
tensor([ 1., 2., 4., 8., 16., 32., 64., 128.], device='cuda:0'))bitstensor([0, 1, 2, 3, 4, 5, 6, 7], device='cuda:0')torch.sum(2**bits)tensor(255, device='cuda:0')How would we calculate the maximum value that we can represent?
[64, 2048, 10, 10][64, 2048, 10, 10][64, 8, 10, 10][64, 8, 10, 10](2.**bits).sum()tensor(255., device='cuda:0')#dividing by nbits gets us close to 1
projector=torch.randn(2048,nbits).cuda()/nbits**0.5
torch.linalg.norm(projector,dim=1)tensor([1.0337, 1.0348, 1.0620, ..., 1.0139, 1.1947, 1.1155], device='cuda:0')nbits8[0,1,0,0,0], [0,0,1,0,0],[1,0,0,0,0]([0, 1, 0, 0, 0], [0, 0, 1, 0, 0], [1, 0, 0, 0, 0])b=dls.one_batch()[0]dls.show_batch()
ys=model_call(model,b)(ys[:,:,None]*projector[None,...,None,None]).shapetorch.Size([64, 2048, 8, 10, 10])ys.shape,projector.shape(torch.Size([64, 2048, 10, 10]), torch.Size([2048, 8]))torch.einsum('ijkl,jm->imkl',ys,projector).shapetorch.Size([64, 8, 10, 10])(torch.einsum('ijkl,jm->ijmkl',ys,projector)==(ys[:,:,None]*projector[None,...,None,None])).any()TensorImage(True, device='cuda:0')How would you calculate the dot product for the second dim and first dim? (2048,10,10)@(2048,8)=(8,10,10)
Using Einstein notation?
projection=torch.einsum('bcxy,cz->bzxy',ys,projector)projectionTensorImage([[[[-6.2792e+00, 1.3167e-01, -9.4779e-02, ..., -2.4656e+00,
-5.0584e+00, -1.3296e+00],
[ 8.1754e+00, 9.6036e+00, -4.4357e+00, ..., -4.6396e+00,
-7.8996e+00, -2.8259e+00],
[ 4.4467e+00, -2.1758e+00, -6.7790e+00, ..., 4.6684e+00,
1.3650e+01, 1.0253e+01],
...,
[ 6.3236e-01, -3.8342e+00, -1.3020e+01, ..., -4.5836e+00,
-3.6719e+00, 4.8875e+00],
[-9.5392e+00, -4.3794e+00, -4.9069e+00, ..., 4.2450e+00,
-7.0259e+00, -6.6205e+00],
[-2.8886e+00, -2.2777e+00, -3.5640e+00, ..., -1.4655e+00,
-5.7932e+00, -2.0072e+00]],
[[ 9.2835e+00, 7.7462e+00, 6.1908e+00, ..., 2.9067e+00,
8.0073e+00, 1.1350e+01],
[ 1.1556e+01, 1.3565e+01, 9.3358e+00, ..., 3.8952e+00,
8.9265e+00, 1.1539e+01],
[ 9.3274e+00, 7.3811e+00, 1.1513e+01, ..., 1.6263e+01,
2.4063e+01, 1.9001e+01],
...,
[ 6.8415e+00, 5.9298e+00, 7.7832e+00, ..., 2.8651e+01,
4.1376e+00, -4.5388e+00],
[ 5.3508e+00, 1.1588e+01, 1.2607e+01, ..., 1.5089e+01,
7.7975e+00, 5.8521e+00],
[ 9.5215e+00, 6.9118e+00, 1.0100e+01, ..., 8.9737e+00,
6.7271e+00, 8.5211e+00]],
[[-3.1688e+00, -6.4222e+00, -2.8567e+00, ..., -3.6577e+00,
-5.9065e+00, -2.9208e+00],
[-6.0068e+00, -9.4032e+00, -2.8770e+00, ..., -6.6112e-01,
-5.0611e+00, -8.4874e+00],
[-4.4515e+00, -1.0154e+01, -1.2417e+00, ..., -1.5574e-01,
-1.1492e+00, 7.6063e+00],
...,
[ 1.5988e+01, 1.3046e+01, 6.4920e+00, ..., -4.4126e+00,
1.4175e+00, 1.2447e+01],
[ 1.9409e+01, 1.7188e+01, 6.8644e+00, ..., -3.7321e+00,
-2.8769e+00, 1.6388e+00],
[ 1.6584e+01, 1.4844e+01, 9.3730e+00, ..., -8.5079e-01,
-6.0974e+00, -3.7285e+00]],
...,
[[-7.1310e+00, 6.5308e-01, 2.6390e+00, ..., -4.4481e+00,
-7.0668e+00, -9.1038e+00],
[ 2.2496e+00, 5.2074e+00, 1.0821e+01, ..., -3.1807e+00,
8.5318e-01, -1.1862e+00],
[-5.3933e+00, 3.0677e+00, 6.6602e+00, ..., -4.9860e+00,
1.2720e+00, 1.7180e+01],
...,
[-3.8841e+00, -1.0031e+00, 1.4554e+01, ..., 1.7657e+01,
8.1946e+00, 8.0810e+00],
[-7.9001e+00, -9.9230e+00, 1.9062e+00, ..., 2.0380e+00,
-1.1232e+00, 2.4761e+00],
[-1.1327e+01, -1.5734e+01, -1.2547e+01, ..., -1.1905e+00,
3.4819e+00, 1.7690e+00]],
[[-5.7783e+00, 1.5963e-01, 3.8980e+00, ..., 1.9761e+00,
1.8847e+00, 2.4780e-01],
[-1.7248e+00, -3.0383e+00, -6.9456e+00, ..., 8.2464e+00,
-2.3109e+00, -2.6400e+00],
[ 3.1076e-01, 1.4344e+00, -1.1225e+01, ..., 5.7751e+00,
1.9494e+00, 3.9672e+00],
...,
[-1.1729e+01, -1.0031e+01, -2.1186e+01, ..., -7.7848e+00,
-1.1644e+01, -1.2071e+01],
[-1.5569e+01, -2.1363e+01, -2.6422e+01, ..., -1.6087e+01,
-1.9381e+01, -2.5736e+01],
[-2.8877e+01, -2.3298e+01, -1.9927e+01, ..., -1.1429e+01,
-1.3280e+01, -2.2741e+01]],
[[-9.2639e+00, -3.3929e+00, 9.7048e+00, ..., -3.8852e+00,
8.4170e-01, 1.7012e+00],
[-2.6492e+00, 3.2599e+00, 3.2146e+01, ..., 2.9210e+00,
2.3555e-01, 3.9248e-01],
[ 1.1340e+01, 1.2774e+01, 2.0421e+01, ..., -7.9307e+00,
-1.6897e+01, -4.7462e+00],
...,
[ 9.6164e+00, 1.3184e+01, 1.3542e+01, ..., 7.6629e+00,
1.3888e+01, 1.5293e+01],
[ 1.0481e+01, 1.4385e+01, 1.1060e+01, ..., 4.2098e+00,
1.1673e+01, 1.5041e+01],
[ 4.5687e+00, 1.4530e+01, 1.3015e+01, ..., 9.6600e+00,
1.2996e+01, 1.3536e+01]]],
[[[ 9.5124e+00, 7.0540e+00, 1.0731e+01, ..., -1.1898e-01,
-4.8526e+00, -1.6273e+01],
[ 1.6990e-01, -1.3911e-01, 6.2190e+00, ..., -6.3489e-01,
-1.0237e+01, -8.5378e+00],
[-3.7613e+00, -4.9234e+00, -2.4806e+00, ..., -2.5140e+00,
-7.3274e+00, -7.1875e+00],
...,
[-6.5095e+00, 1.7462e+00, -1.9302e-01, ..., -9.5685e+00,
-7.6806e+00, -8.3060e-01],
[-9.2985e-01, -6.4332e+00, 6.1271e+00, ..., -1.7758e+01,
-1.5903e+01, -1.8803e+01],
[ 9.8972e+00, 1.4579e+01, 8.9024e+00, ..., -2.1343e+01,
-3.3487e+01, -2.7978e+01]],
[[ 9.7840e+00, 1.3461e+01, 2.7833e+01, ..., 5.9922e+00,
1.1757e+01, 2.2186e+01],
[ 6.5224e+00, 6.7404e+00, 1.3893e+01, ..., 9.1900e+00,
1.8095e+01, 1.9578e+01],
[ 8.7155e-01, 1.9546e+00, 8.3535e+00, ..., 1.5843e+01,
1.7649e+01, 1.1515e+01],
...,
[-1.0384e+01, -1.4986e+01, 5.2173e+00, ..., 3.0871e+01,
4.3548e+01, 4.0713e+01],
[ 1.8919e+01, 1.2526e+01, 2.4128e+00, ..., 2.9831e+01,
4.2634e+01, 3.9587e+01],
[ 3.8178e+01, 3.2932e+01, 4.6763e+00, ..., 3.2836e+01,
3.7258e+01, 4.4807e+01]],
[[-6.7811e+00, -1.3298e+01, -2.0355e+00, ..., -4.4833e+00,
-6.0731e+00, -5.2660e+00],
[-4.5190e+00, -4.7397e+00, -4.3055e+00, ..., -8.3795e+00,
-8.6657e+00, -3.1813e+00],
[-4.0083e+00, -5.4050e+00, -8.3738e+00, ..., -8.1771e+00,
-5.7712e+00, 3.3252e+00],
...,
[-2.0961e+01, -1.7522e+01, -1.9436e+01, ..., -3.9490e+00,
3.5565e+00, 4.4478e+00],
[-2.4070e+01, -2.2667e+01, -7.1590e+00, ..., -3.1348e+00,
5.0762e+00, 1.2960e+01],
[-7.6899e+00, -8.6212e+00, 6.7536e-01, ..., -2.6530e-01,
1.1745e+01, 1.5273e+01]],
...,
[[ 9.2374e-01, 3.2793e+00, -6.5183e+00, ..., 4.7281e-01,
-2.2870e+00, 1.3027e+00],
[ 4.0763e+00, 3.0901e+00, 4.1876e+00, ..., 2.0954e+00,
-9.8625e-01, -3.1275e+00],
[ 9.6910e-01, 2.8828e-01, 7.4595e-01, ..., 1.0479e+01,
8.2900e+00, 4.0798e+00],
...,
[ 1.2553e+01, 1.4372e+01, 1.3448e+01, ..., 1.2941e+01,
7.6967e+00, 1.1426e+01],
[ 1.1466e+01, 1.4855e+01, 2.0851e+01, ..., 3.2110e+01,
4.5123e+01, 4.3206e+01],
[ 1.0600e+01, 7.3636e+00, 1.0523e+01, ..., 1.6592e+01,
2.7320e+01, 5.5086e+01]],
[[ 2.4515e+00, 1.0092e+01, 1.2336e+01, ..., 6.2933e+00,
4.9184e+00, 9.3574e+00],
[ 1.2444e+00, 3.0530e+00, 7.1421e+00, ..., 5.0260e+00,
6.4202e+00, -1.7965e+00],
[-2.0972e+00, 3.1371e-01, 8.9068e-01, ..., 2.6291e+00,
3.4665e+00, -1.9193e+00],
...,
[ 1.5958e+00, -6.7022e+00, -5.6058e+00, ..., 6.6774e+00,
4.8079e+00, 3.4565e+00],
[-4.5736e-02, -1.8359e+01, -7.8809e+00, ..., 6.7820e+00,
3.7120e+00, 4.0987e+00],
[ 1.4360e+00, -5.5607e+00, 6.2634e-01, ..., -1.5635e+00,
-1.0811e+01, -2.0711e+00]],
[[ 2.8068e+00, 3.3115e+00, 7.0982e+00, ..., 2.5611e+00,
-1.5446e+00, -7.0554e+00],
[ 3.2848e+00, -2.6821e+00, 4.3762e+00, ..., 1.6894e+00,
-1.6904e+00, -6.2212e+00],
[-4.2415e-01, -3.8746e+00, -4.9796e-01, ..., -3.5530e-01,
2.7174e+00, 1.3833e+01],
...,
[ 6.1781e+00, 3.1638e+00, 7.9841e+00, ..., -9.4526e+00,
-3.6841e+00, -1.4651e+01],
[-6.8839e+00, -3.6912e+00, -3.6544e+00, ..., -1.6174e+01,
-2.0845e+01, -2.3711e+01],
[-3.4264e+00, -5.4144e+00, -1.1610e-01, ..., -1.2922e+01,
-1.8294e+01, -2.1884e+01]]],
[[[ 1.6726e+01, 1.7428e+01, 4.2822e+00, ..., 1.3308e+01,
7.7680e+00, 1.4643e-01],
[ 1.0972e+01, 1.4134e+01, 4.4387e+00, ..., 1.5983e+01,
7.5606e+00, 2.1118e+00],
[ 8.8190e+00, 7.4110e+00, 1.1246e+01, ..., 1.3696e+01,
3.8566e+00, 6.2899e+00],
...,
[ 1.2071e+01, 1.0601e+01, 1.6731e+00, ..., -1.3564e+01,
-1.3250e+01, 3.3810e-01],
[ 1.8464e+00, 7.9721e+00, 3.5149e+00, ..., -1.6235e+01,
-1.0623e+01, 8.7157e+00],
[ 8.7276e+00, -9.2056e-01, 7.5868e-01, ..., -1.1893e+01,
2.7289e+00, 2.7667e+01]],
[[ 6.6531e+00, 1.0010e+01, 2.2753e+01, ..., 1.4925e+01,
1.3476e+01, 1.6543e+01],
[ 4.5088e+00, 1.9536e+01, 3.0127e+01, ..., 1.1684e+01,
9.8105e+00, 3.2875e+00],
[ 4.2340e+00, 5.5834e+00, 1.8127e+01, ..., -3.0256e+00,
2.6275e+00, -7.1714e-01],
...,
[ 1.6779e+01, 2.9478e+00, 1.5335e+01, ..., 3.1999e+01,
1.1919e+01, 8.4015e+00],
[ 1.5926e+01, 2.0441e+00, 7.7098e+00, ..., 1.2695e+01,
1.0315e+01, 5.9256e+00],
[ 4.5396e+00, -3.9510e+00, 3.0082e+00, ..., 4.2423e+00,
1.4620e+00, -8.1835e-01]],
[[-1.7676e+01, -1.7258e+01, -1.1684e+01, ..., -1.0038e+01,
-7.8505e-02, 5.6030e-01],
[-1.2038e+01, -1.0079e+01, 3.9319e-01, ..., -1.3313e+00,
-2.6005e+00, 6.5192e+00],
[-5.7970e+00, -1.0169e+01, -1.4580e+01, ..., 3.7920e+00,
2.1885e+00, -2.9092e+00],
...,
[-9.1141e+00, -1.7944e+01, -2.5657e+00, ..., -2.0504e+01,
-8.4545e+00, 3.7092e+00],
[-4.9958e+00, -9.3027e+00, -7.6127e+00, ..., -1.3733e+01,
-7.3914e+00, 1.0248e+01],
[-2.4956e+00, -4.2638e+00, -3.6001e+00, ..., -5.5006e-01,
2.5708e+00, 5.7056e+00]],
...,
[[ 4.2114e-01, -7.9980e+00, -8.6686e+00, ..., -4.2628e+00,
8.8063e+00, -9.2941e-02],
[-8.4785e+00, -1.7714e+01, -5.5977e-01, ..., 9.8289e-01,
1.2428e+01, 9.1162e+00],
[-9.5216e+00, -2.6300e+01, -1.7297e+01, ..., 4.7637e+00,
6.2845e+00, 4.5398e+00],
...,
[ 1.2768e+00, -6.3151e+00, -1.0408e+00, ..., 7.4808e-01,
-1.0534e+00, 4.4392e+00],
[ 3.0409e+00, -1.3620e+00, 3.0350e+00, ..., -9.6885e+00,
-2.5018e+00, -3.8327e+00],
[-1.6695e+01, -1.1960e+01, -6.9895e+00, ..., -8.2594e+00,
-1.4068e+01, -2.0667e+01]],
[[ 5.1652e+00, 1.0395e+01, 7.4535e+00, ..., 8.6804e+00,
4.5424e+00, 1.4589e+01],
[ 4.9551e+00, 1.8278e+01, 2.8135e+01, ..., 2.0059e+01,
1.2688e+01, 1.1655e+01],
[ 8.6485e+00, 2.4904e+01, 4.3337e+01, ..., 1.4686e+01,
1.6079e+01, 2.0526e+01],
...,
[ 3.6101e+00, -3.5364e+00, 7.2125e+00, ..., -8.2165e-01,
-7.4522e+00, 6.1776e+00],
[-3.6109e+00, -6.7193e+00, 1.2664e+01, ..., -5.5758e+00,
-7.2992e+00, 3.7080e+00],
[-9.0229e-01, -2.1786e+00, 7.6342e+00, ..., -7.0221e+00,
-9.6212e+00, 5.1720e+00]],
[[-1.1960e+01, -1.1723e+01, -2.1839e+01, ..., -7.9723e+00,
9.2299e+00, 5.2175e+00],
[-1.4355e+01, -2.1408e+01, -3.0463e+01, ..., 3.5703e+00,
4.7135e+00, 7.2011e+00],
[-1.3361e+01, -2.6255e+01, -3.8794e+01, ..., 2.9431e+00,
-4.4089e+00, 1.5832e-01],
...,
[-4.0332e+00, -1.0877e+01, -1.0693e+01, ..., -7.4630e+00,
-1.3602e-01, 1.2742e+00],
[-3.7040e+00, -2.8553e+00, -8.7539e+00, ..., 6.7736e-01,
2.3848e+00, -1.8044e-01],
[-8.3037e+00, -1.0215e+01, -8.9308e+00, ..., 1.6606e+01,
1.4776e+00, -7.1054e+00]]],
...,
[[[ 3.0742e+01, 2.0617e+01, 3.4883e+00, ..., -9.0017e+00,
-1.6916e+00, 3.9473e-01],
[ 3.5029e+01, 2.8505e+01, 1.3211e+01, ..., -5.5471e+00,
-6.9257e+00, -2.2475e+00],
[ 1.4213e+01, 2.3229e+01, 1.6051e+01, ..., -1.8002e-01,
-3.7035e-01, 6.3750e+00],
...,
[ 8.4374e-01, 4.3312e+00, -3.1439e-01, ..., -4.4065e+00,
-1.1842e+00, 3.8927e+00],
[-1.3196e+00, -4.4678e-01, -2.1221e+00, ..., -2.8409e+00,
-4.0868e+00, 2.5064e+00],
[-3.6652e-01, -9.4469e-01, -9.2601e+00, ..., -3.7837e+00,
-1.7367e+00, 8.7911e+00]],
[[ 2.1170e+01, 1.9605e+01, 8.4797e+00, ..., 1.7911e+01,
8.9692e+00, 6.8814e+00],
[ 1.5301e+01, 1.3120e+01, 1.9084e+00, ..., 1.5672e+01,
1.0492e+01, 9.9011e+00],
[ 4.6515e+00, -1.0032e+00, 1.7698e+00, ..., 9.0264e+00,
7.3372e+00, 2.9377e+00],
...,
[ 9.4733e+00, 1.7173e+01, 1.2714e+01, ..., 2.1929e+01,
1.6661e+01, 1.2190e+01],
[ 4.9162e+00, 1.5067e+01, 1.7427e+01, ..., 1.6764e+01,
1.1263e+01, 1.2861e+01],
[ 3.4370e+00, 1.0124e+01, 1.2676e+01, ..., 2.0788e+01,
1.6364e+01, 2.2770e+01]],
[[ 1.7073e+01, 1.2411e+01, 3.9940e+00, ..., -2.5094e-01,
1.0745e+00, 2.0860e+00],
[ 9.7372e+00, 2.9299e+00, -5.5394e+00, ..., 1.9889e+00,
3.3939e+00, -1.5617e+00],
[-2.5642e-01, -1.1352e+00, -8.8535e+00, ..., -4.5452e+00,
-1.9397e+00, -7.2823e+00],
...,
[-2.0316e+00, -2.9228e+00, -4.4708e+00, ..., -3.9330e+00,
-1.1184e+01, -9.1596e+00],
[ 4.0511e-01, -5.1108e-01, -5.8130e+00, ..., -3.4118e+00,
-4.2866e+00, -2.3001e+00],
[-3.2900e+00, -7.9968e+00, -4.4886e+00, ..., -4.5846e+00,
-5.1580e+00, -5.1870e+00]],
...,
[[-1.0441e+01, 3.6222e-01, 5.7083e+00, ..., -8.7860e-01,
-5.9626e+00, -2.4280e+00],
[-3.2910e-01, 3.1008e+00, 1.1946e+01, ..., 2.7110e+00,
-5.6768e+00, -6.1909e+00],
[-5.4021e+00, 1.0762e+01, 2.7438e+01, ..., 5.8677e+00,
-1.0364e-01, -2.4942e+00],
...,
[ 7.0567e-01, -3.1743e+00, -1.8512e+00, ..., 4.5055e+00,
1.3033e+00, -8.2645e+00],
[-3.0940e+00, 2.9549e+00, 1.0386e+01, ..., -4.2258e+00,
-8.3496e+00, -6.9864e+00],
[-5.8167e+00, 5.5948e+00, 2.3419e+01, ..., -6.6175e+00,
-1.2134e+01, -7.1784e+00]],
[[-1.5039e+01, -5.6994e+00, 6.0646e+00, ..., 4.8761e+00,
-1.6647e+00, -3.3653e+00],
[-1.2232e+01, 3.2679e-01, 2.1025e+00, ..., 4.5207e+00,
-6.3115e-01, 4.4318e-01],
[-1.6360e+01, -4.2183e-01, -2.9451e+00, ..., -3.2819e+00,
-7.8153e+00, -3.9869e+00],
...,
[-9.9416e+00, 1.5021e+01, 1.2003e+01, ..., -5.2349e+00,
2.6565e+00, -2.1975e+00],
[-1.3382e+01, 2.7150e+00, 7.4891e+00, ..., -5.7495e+00,
8.4802e+00, 6.6035e+00],
[-1.3810e+01, 7.6381e+00, 5.1958e+00, ..., -1.8646e+00,
3.0005e+00, 9.3771e-01]],
[[ 1.3810e+01, 1.7808e+01, 8.6061e+00, ..., 1.1546e+01,
8.8377e+00, 8.1417e+00],
[ 1.1785e+01, 1.7363e+01, -4.3577e+00, ..., 3.6378e+00,
1.6090e+01, 1.5553e+01],
[ 1.0965e+01, 2.3075e+00, -2.3853e+00, ..., -8.7184e+00,
1.7130e+01, 1.8968e+01],
...,
[ 2.2361e+00, 2.7280e+00, 6.5013e+00, ..., 5.6273e+00,
8.9686e+00, 9.0201e+00],
[ 5.5273e+00, 3.4220e+00, -9.4277e+00, ..., 8.6742e-01,
7.3521e+00, 1.1186e+01],
[ 4.3832e+00, -6.7473e-01, -1.1554e+01, ..., 6.7014e+00,
8.0967e+00, 7.0652e+00]]],
[[[-2.6913e+01, -2.9568e+01, -8.9120e+00, ..., -3.6777e+00,
-9.5993e+00, -1.3907e+01],
[-3.3114e+01, -4.0868e+01, -3.1856e+01, ..., -2.1573e+00,
-7.5814e+00, -1.1474e+01],
[-2.7869e+01, -3.9983e+01, -2.6464e+01, ..., 6.0545e+00,
-6.4322e+00, -6.3355e+00],
...,
[-1.6259e+01, -1.7102e+01, -9.8119e+00, ..., -2.6094e+00,
-9.1046e+00, -4.5698e+00],
[-4.2792e+00, -1.4512e+01, -7.5014e+00, ..., -3.2948e-01,
-8.0459e+00, -5.2389e+00],
[-2.2665e-03, -1.0965e+01, -7.7879e+00, ..., -9.6266e+00,
-1.0545e+01, -7.5697e-01]],
[[ 1.5671e+00, 1.2056e+01, 1.2069e+01, ..., 2.4591e+01,
3.2034e+01, 3.6939e+01],
[-3.8125e+00, 3.2298e+00, 1.2571e+01, ..., 2.2087e+01,
2.9005e+01, 1.8920e+01],
[ 7.9042e-01, 2.6120e+01, 3.5221e+01, ..., 3.8826e+00,
1.6314e+01, 2.7712e+01],
...,
[-2.7849e+00, -1.0391e+01, -5.8262e+00, ..., 6.0313e+00,
6.1854e+00, 7.3942e+00],
[ 8.9106e+00, 6.6072e+00, 3.7478e+00, ..., 4.4233e+00,
-1.8716e-01, -3.8388e-01],
[ 2.1626e+01, 1.3852e+01, 1.0201e+01, ..., 3.2672e+00,
-1.1369e+00, 3.3530e+00]],
[[ 8.4911e+00, 8.4341e+00, 4.6288e+00, ..., 1.3576e+01,
-7.2182e+00, -1.2758e+01],
[ 2.1120e+01, 1.2484e+01, 8.6090e-02, ..., -1.1351e+00,
-1.4208e+01, -1.1975e+01],
[-3.1700e+00, -4.0954e+00, -1.6492e+00, ..., -1.0604e+01,
-9.4700e+00, 6.6750e-01],
...,
[ 2.9052e+00, 6.3958e+00, 1.3706e-01, ..., 5.7390e+00,
3.1581e+00, -6.0765e-01],
[-3.8417e+00, -4.2816e+00, -5.9917e-01, ..., 6.7638e+00,
5.0139e+00, 5.6882e+00],
[-1.8144e+01, -1.2465e+01, -4.2393e+00, ..., 4.2922e+00,
6.0590e+00, 9.7644e+00]],
...,
[[-7.3576e+00, 1.0356e+01, 3.1724e-01, ..., 1.8843e+01,
1.8283e+01, 1.0810e+01],
[-7.3645e+00, 5.2899e+00, 1.3292e+00, ..., 1.3994e+01,
8.5479e+00, -4.9072e+00],
[-9.9293e+00, 2.7770e+00, -8.8405e+00, ..., 3.1427e+00,
1.9661e+00, 8.4572e+00],
...,
[-8.4440e+00, -5.7747e+00, -7.7229e+00, ..., 8.1245e+00,
-3.8143e+00, -9.9999e+00],
[ 6.4093e-01, 5.9969e+00, -6.3158e+00, ..., 1.4593e+00,
-5.8362e+00, -1.3630e+01],
[ 4.7849e+00, 4.6974e+00, 3.1298e+00, ..., -4.9041e+00,
-1.0216e+01, -1.2580e+01]],
[[-5.8014e+00, -1.0372e+01, 1.0254e+01, ..., -4.6294e-01,
1.8217e+01, 1.0038e+01],
[ 5.7797e+00, 9.2750e+00, 2.1096e+01, ..., 8.8768e+00,
2.2236e+01, 2.0851e+01],
[ 1.4236e+01, 3.3656e+01, 3.3081e+01, ..., 2.2807e+01,
3.4478e+01, 1.7018e+01],
...,
[ 1.9373e+01, 1.7202e+01, 7.0146e-01, ..., 9.4769e+00,
5.3170e+00, 7.5832e+00],
[ 2.3346e+01, 2.6482e+01, 1.0779e+01, ..., 1.0170e+00,
9.9077e+00, 6.9110e+00],
[ 1.7564e+01, 2.1025e+01, 1.2639e+01, ..., 4.6908e+00,
5.8073e+00, 3.1757e+00]],
[[-5.2906e+00, -2.7127e+00, 1.5336e+01, ..., -3.3235e+00,
3.0431e+00, 4.4982e+00],
[-3.8740e+00, -2.7825e-01, 1.8444e+01, ..., -6.8304e+00,
-9.7458e+00, -2.2539e+00],
[-2.3341e+00, 1.1108e+00, 9.1996e+00, ..., 1.4541e+00,
4.0044e+00, 5.3693e+00],
...,
[-5.5670e+00, -3.2424e+00, -1.0724e+01, ..., -2.3431e+00,
3.1732e+00, 2.2250e+00],
[ 4.7924e+00, 2.9135e+00, 1.0695e+00, ..., -3.1337e+00,
3.6863e+00, 9.4232e+00],
[ 1.3064e+01, 1.4679e+00, 4.2512e+00, ..., 2.0431e+00,
5.5197e+00, 9.0792e+00]]],
[[[-4.1837e+00, 4.4228e+00, 7.5466e+00, ..., 9.0294e+00,
1.8985e+00, 1.3472e+00],
[ 3.4190e+00, 7.7751e+00, 1.4765e+01, ..., 1.0512e+01,
1.2134e+01, 1.3603e+01],
[ 1.1120e+01, 1.9633e+01, 1.5074e+01, ..., 7.3059e+00,
5.4364e+00, 1.3671e+01],
...,
[ 7.7573e+00, 6.5959e+00, 2.4744e+00, ..., 1.3797e+01,
2.7515e+00, 2.0304e+00],
[ 5.0850e+00, 9.8228e+00, 7.2073e+00, ..., 3.0873e-01,
-2.5682e+01, -1.1913e+01],
[-1.9177e+00, 1.0239e+00, 2.1603e+00, ..., 7.6998e-02,
-1.0977e+01, -1.0928e+01]],
[[ 9.6061e+00, 1.5370e+00, 2.8735e+00, ..., 5.0786e+00,
1.1083e+01, 9.2034e+00],
[ 3.0860e+00, 7.1579e+00, 6.0781e+00, ..., 7.9867e+00,
3.1649e+00, 5.9685e+00],
[ 2.0483e+00, 1.4068e+01, 9.0135e+00, ..., 4.9285e+00,
6.5073e+00, -3.5669e+00],
...,
[-1.0659e+00, 4.2513e+00, 3.0158e+00, ..., 2.1836e+00,
-7.4681e+00, -1.0047e+01],
[-6.4479e+00, -1.6330e+00, 7.5435e+00, ..., 1.3588e+01,
1.1536e+00, -1.1312e+01],
[-4.2848e+00, 4.4101e+00, 9.2249e+00, ..., 1.7328e+01,
1.4831e+01, -7.3449e-01]],
[[-1.3682e+00, -5.1753e+00, -4.2238e+00, ..., -1.3184e+01,
-1.0707e+01, -9.5371e+00],
[-1.6860e+00, -3.3730e+00, 3.3757e+00, ..., -9.6586e+00,
-4.3802e+00, -4.1719e-02],
[-4.7046e-01, -4.2123e-01, -3.8035e+00, ..., -9.6757e+00,
-1.5742e+01, -1.8032e+00],
...,
[-4.2253e+00, -3.7856e+00, -4.7338e+00, ..., 6.0688e+00,
-2.2821e+00, 6.7935e+00],
[-8.7673e+00, -2.8977e-01, -1.8588e+00, ..., 7.1490e+00,
-4.2642e+00, -4.3403e-01],
[-1.3027e+01, -9.8119e+00, -4.1386e+00, ..., -3.9124e+00,
-1.8475e+01, -5.6218e+00]],
...,
[[-1.3048e+01, -6.4990e+00, 3.7771e+00, ..., -3.5662e-01,
-2.1649e+01, -1.9933e+01],
[-1.3308e+01, 1.0137e+00, 4.0163e+00, ..., 1.5178e+01,
-1.8908e+01, -2.6649e+01],
[-6.0838e+00, 4.4509e+00, 9.2461e+00, ..., 1.2517e+01,
-6.9830e+00, -1.0261e+01],
...,
[ 3.8149e+00, 1.0815e+01, 3.1839e+00, ..., 9.5991e+00,
3.8434e+00, 2.1540e+00],
[-2.1296e+00, -1.8112e+00, -2.8965e+00, ..., -6.4432e+00,
-6.9710e+00, -5.0703e-01],
[ 7.2261e+00, -2.4904e+00, -6.8260e+00, ..., -8.1620e+00,
-4.7305e+00, -1.3990e+00]],
[[ 6.4019e+00, 1.4235e+01, 1.1706e+01, ..., 8.1696e+00,
1.3590e+01, 1.1112e+01],
[ 9.0817e+00, 1.8450e+01, 1.8520e+00, ..., 2.9886e+00,
1.0353e+01, 1.9502e+01],
[ 1.2688e+01, 1.2391e+01, 7.0384e+00, ..., -4.8993e+00,
4.6831e+00, 1.3654e+01],
...,
[ 6.8772e+00, 4.8467e+00, 3.6678e+00, ..., -3.3305e+00,
2.7749e+01, 3.7785e+01],
[ 1.4010e+01, 1.1408e+01, 5.3943e+00, ..., -2.1537e+01,
2.1972e+00, 3.4320e+01],
[ 1.4877e+01, 1.7063e+01, 6.9275e+00, ..., -1.4477e+01,
-3.4017e+00, 3.0101e+01]],
[[-3.0831e+00, 3.0010e+00, 6.9832e+00, ..., -1.8217e+00,
-7.1681e+00, -3.4423e+00],
[ 8.3796e+00, 3.6513e+00, -2.5611e+00, ..., -6.8060e+00,
-1.4611e+01, -8.5084e+00],
[-6.5332e-01, -1.1911e+01, -1.2325e+01, ..., -4.0894e+00,
-1.0735e+01, 1.7551e+00],
...,
[ 3.4406e+00, 1.4224e+00, -3.6680e+00, ..., -1.0623e+01,
-2.8591e+01, -1.4722e+01],
[ 3.2616e+00, 2.6747e+00, -1.9402e+00, ..., -6.7384e+00,
-1.2895e+01, -1.5971e+01],
[ 1.2498e+01, 2.2431e+00, -3.8494e+00, ..., -9.0071e+00,
-2.1719e+01, -1.8405e+01]]]], device='cuda:0',
grad_fn=<AliasBackward0>)2**bits, 0-255(tensor([ 1, 2, 4, 8, 16, 32, 64, 128], device='cuda:0'), -255)(1*(projection>0)).shape,(2**bits[...,None,None]).shape(torch.Size([64, 8, 10, 10]), torch.Size([8, 1, 1]))torch.Size([64, 10, 10])projection.shape,(2.**bits).shape(torch.Size([64, 8, 10, 10]), torch.Size([8]))(1*(projection>0)).shape,(2**bits).shape(torch.Size([64, 8, 10, 10]), torch.Size([8]))2**bits,(1*(projection>0)).shape(tensor([ 1, 2, 4, 8, 16, 32, 64, 128], device='cuda:0'),
torch.Size([64, 8, 10, 10]))((1*(projection>0)[:,:,0,0])*(2**bits))TensorImage([[ 0, 2, 0, 8, 0, 0, 0, 0],
[ 1, 2, 0, 0, 0, 32, 64, 128],
[ 1, 2, 0, 0, 0, 32, 64, 0],
[ 0, 2, 0, 0, 0, 0, 0, 0],
[ 0, 2, 4, 8, 16, 0, 0, 128],
[ 0, 2, 0, 8, 0, 0, 0, 128],
[ 0, 2, 4, 0, 16, 0, 64, 128],
[ 0, 2, 4, 0, 0, 32, 64, 128],
[ 0, 2, 0, 0, 0, 32, 0, 0],
[ 1, 2, 0, 0, 0, 0, 64, 128],
[ 1, 2, 0, 0, 0, 32, 0, 128],
[ 0, 2, 0, 0, 0, 0, 64, 128],
[ 1, 2, 0, 0, 0, 0, 64, 0],
[ 0, 2, 0, 0, 0, 32, 64, 128],
[ 1, 2, 0, 0, 0, 32, 0, 0],
[ 0, 2, 4, 8, 0, 0, 0, 0],
[ 1, 2, 0, 0, 0, 32, 64, 128],
[ 0, 2, 4, 8, 16, 0, 64, 128],
[ 0, 2, 4, 0, 0, 32, 0, 0],
[ 0, 2, 0, 8, 0, 0, 0, 128],
[ 1, 2, 4, 0, 0, 32, 64, 0],
[ 0, 2, 0, 0, 0, 0, 64, 0],
[ 1, 2, 4, 0, 0, 32, 0, 0],
[ 0, 2, 0, 0, 16, 0, 0, 128],
[ 1, 2, 0, 0, 16, 0, 0, 128],
[ 0, 2, 4, 0, 16, 0, 64, 128],
[ 0, 2, 4, 8, 16, 0, 64, 128],
[ 0, 2, 4, 8, 16, 0, 64, 128],
[ 0, 2, 0, 0, 16, 32, 0, 0],
[ 1, 2, 0, 0, 0, 0, 64, 128],
[ 1, 2, 0, 0, 0, 0, 0, 0],
[ 0, 2, 4, 8, 0, 0, 0, 128],
[ 1, 2, 0, 0, 0, 32, 0, 128],
[ 1, 2, 0, 0, 0, 0, 0, 128],
[ 0, 2, 4, 8, 0, 0, 64, 0],
[ 0, 2, 4, 8, 0, 0, 64, 0],
[ 0, 2, 0, 8, 0, 0, 0, 128],
[ 1, 2, 0, 8, 0, 32, 64, 0],
[ 0, 0, 4, 8, 16, 0, 64, 128],
[ 0, 0, 4, 8, 16, 32, 0, 128],
[ 1, 0, 0, 0, 0, 32, 64, 0],
[ 0, 2, 4, 8, 16, 0, 64, 128],
[ 0, 2, 0, 0, 16, 0, 0, 128],
[ 1, 2, 4, 8, 0, 0, 0, 128],
[ 1, 2, 0, 0, 0, 32, 64, 0],
[ 0, 2, 0, 8, 16, 0, 64, 128],
[ 0, 2, 0, 0, 0, 32, 0, 0],
[ 0, 2, 4, 0, 0, 0, 0, 128],
[ 1, 0, 0, 0, 0, 32, 64, 0],
[ 1, 2, 0, 0, 0, 32, 0, 128],
[ 0, 2, 4, 0, 16, 32, 0, 128],
[ 1, 2, 0, 0, 16, 0, 64, 128],
[ 0, 0, 0, 0, 16, 32, 64, 0],
[ 1, 2, 4, 0, 16, 0, 64, 128],
[ 0, 2, 0, 0, 0, 0, 64, 0],
[ 0, 2, 0, 8, 0, 0, 64, 0],
[ 0, 2, 0, 0, 0, 0, 0, 128],
[ 0, 2, 4, 8, 0, 0, 64, 0],
[ 1, 2, 0, 8, 0, 0, 64, 0],
[ 1, 2, 0, 0, 0, 32, 64, 0],
[ 0, 2, 4, 8, 16, 0, 64, 0],
[ 1, 2, 4, 8, 16, 0, 0, 128],
[ 0, 2, 4, 0, 16, 0, 0, 0],
[ 0, 2, 0, 8, 16, 0, 64, 0]], device='cuda:0')torch.einsum('ijkl,j->ikl',1.*(projection>0),(2.**bits))TensorImage([[[ 10., 99., 226., ..., 66., 194., 194.],
[ 35., 163., 162., ..., 194., 162., 130.],
[211., 242., 178., ..., 67., 115., 119.],
...,
[151., 134., 166., ..., 162., 166., 165.],
[134., 134., 166., ..., 163., 146., 182.],
[150., 150., 150., ..., 146., 178., 178.]],
[[227., 243., 211., ..., 226., 66., 106.],
[227., 98., 227., ..., 226., 82., 18.],
[ 34., 98., 98., ..., 114., 242., 166.],
...,
[232., 169., 170., ..., 98., 102., 110.],
[ 42., 42., 43., ..., 98., 102., 102.],
[ 99., 35., 111., ..., 42., 38., 38.]],
[[ 99., 83., 67., ..., 67., 227., 215.],
[ 67., 67., 87., ..., 227., 243., 231.],
[ 67., 67., 83., ..., 229., 119., 241.],
...,
[ 99., 3., 67., ..., 34., 2., 231.],
[ 35., 3., 107., ..., 130., 130., 71.],
[ 3., 8., 75., ..., 130., 135., 69.]],
...,
[[159., 183., 231., ..., 194., 134., 151.],
[151., 247., 99., ..., 246., 150., 210.],
[147., 161., 35., ..., 34., 146., 147.],
...,
[179., 195., 194., ..., 162., 226., 131.],
[150., 226., 98., ..., 130., 194., 195.],
[146., 98., 98., ..., 130., 194., 195.]],
[[ 22., 54., 246., ..., 38., 226., 226.],
[ 68., 102., 246., ..., 98., 98., 66.],
[ 82., 226., 210., ..., 227., 226., 230.],
...,
[ 68., 68., 68., ..., 102., 198., 202.],
[226., 226., 194., ..., 102., 196., 204.],
[226., 226., 226., ..., 222., 212., 214.]],
[[ 90., 195., 227., ..., 91., 83., 83.],
[203., 235., 119., ..., 115., 75., 75.],
[ 75., 115., 115., ..., 51., 83., 201.],
...,
[249., 243., 115., ..., 63., 105., 109.],
[193., 193., 91., ..., 31., 90., 64.],
[232., 203., 91., ..., 27., 18., 64.]]], device='cuda:0')How would we calculate the bucket for each x,y location?
hash_loc=torch.einsum('bchw,c->bhw',(projection>0).float(),2.**bits).int()
hash_loc.shape,hash_loc(torch.Size([64, 10, 10]),
TensorImage([[[ 10, 99, 226, ..., 66, 194, 194],
[ 35, 163, 162, ..., 194, 162, 130],
[211, 242, 178, ..., 67, 115, 119],
...,
[151, 134, 166, ..., 162, 166, 165],
[134, 134, 166, ..., 163, 146, 182],
[150, 150, 150, ..., 146, 178, 178]],
[[227, 243, 211, ..., 226, 66, 106],
[227, 98, 227, ..., 226, 82, 18],
[ 34, 98, 98, ..., 114, 242, 166],
...,
[232, 169, 170, ..., 98, 102, 110],
[ 42, 42, 43, ..., 98, 102, 102],
[ 99, 35, 111, ..., 42, 38, 38]],
[[ 99, 83, 67, ..., 67, 227, 215],
[ 67, 67, 87, ..., 227, 243, 231],
[ 67, 67, 83, ..., 229, 119, 241],
...,
[ 99, 3, 67, ..., 34, 2, 231],
[ 35, 3, 107, ..., 130, 130, 71],
[ 3, 8, 75, ..., 130, 135, 69]],
...,
[[159, 183, 231, ..., 194, 134, 151],
[151, 247, 99, ..., 246, 150, 210],
[147, 161, 35, ..., 34, 146, 147],
...,
[179, 195, 194, ..., 162, 226, 131],
[150, 226, 98, ..., 130, 194, 195],
[146, 98, 98, ..., 130, 194, 195]],
[[ 22, 54, 246, ..., 38, 226, 226],
[ 68, 102, 246, ..., 98, 98, 66],
[ 82, 226, 210, ..., 227, 226, 230],
...,
[ 68, 68, 68, ..., 102, 198, 202],
[226, 226, 194, ..., 102, 196, 204],
[226, 226, 226, ..., 222, 212, 214]],
[[ 90, 195, 227, ..., 91, 83, 83],
[203, 235, 119, ..., 115, 75, 75],
[ 75, 115, 115, ..., 51, 83, 201],
...,
[249, 243, 115, ..., 63, 105, 109],
[193, 193, 91, ..., 31, 90, 64],
[232, 203, 91, ..., 27, 18, 64]]], device='cuda:0',
dtype=torch.int32))hash_loc.flatten().mode()torch.return_types.mode(
values=TensorImage(99, device='cuda:0', dtype=torch.int32),
indices=TensorImage(1, device='cuda:0'))hash_loc==99TensorImage([[[False, True, False, ..., False, False, False],
[False, False, False, ..., False, False, False],
[False, False, False, ..., False, False, False],
...,
[False, False, False, ..., False, False, False],
[False, False, False, ..., False, False, False],
[False, False, False, ..., False, False, False]],
[[False, False, False, ..., False, False, False],
[False, False, False, ..., False, False, False],
[False, False, False, ..., False, False, False],
...,
[False, False, False, ..., False, False, False],
[False, False, False, ..., False, False, False],
[ True, False, False, ..., False, False, False]],
[[ True, False, False, ..., False, False, False],
[False, False, False, ..., False, False, False],
[False, False, False, ..., False, False, False],
...,
[ True, False, False, ..., False, False, False],
[False, False, False, ..., False, False, False],
[False, False, False, ..., False, False, False]],
...,
[[False, False, False, ..., False, False, False],
[False, False, True, ..., False, False, False],
[False, False, False, ..., False, False, False],
...,
[False, False, False, ..., False, False, False],
[False, False, False, ..., False, False, False],
[False, False, False, ..., False, False, False]],
[[False, False, False, ..., False, False, False],
[False, False, False, ..., False, False, False],
[False, False, False, ..., False, False, False],
...,
[False, False, False, ..., False, False, False],
[False, False, False, ..., False, False, False],
[False, False, False, ..., False, False, False]],
[[False, False, False, ..., False, False, False],
[False, False, False, ..., False, False, False],
[False, False, False, ..., False, False, False],
...,
[False, False, False, ..., False, False, False],
[False, False, False, ..., False, False, False],
[False, False, False, ..., False, False, False]]],
device='cuda:0')How would we generate a mask for the locations that equal the mode(199)?
(hash_loc==199).shape,(hash_loc==199)(torch.Size([64, 10, 10]),
TensorImage([[[False, False, False, ..., False, False, False],
[False, False, False, ..., False, False, False],
[False, False, False, ..., False, False, False],
...,
[False, False, False, ..., False, False, False],
[False, False, False, ..., True, True, True],
[False, False, False, ..., True, False, False]],
[[False, False, False, ..., False, False, False],
[False, False, False, ..., False, False, False],
[False, False, False, ..., False, False, False],
...,
[False, False, False, ..., False, False, False],
[False, False, False, ..., False, False, False],
[False, False, False, ..., False, False, False]],
[[False, False, False, ..., True, True, True],
[False, False, False, ..., False, False, False],
[False, False, False, ..., False, False, False],
...,
[False, False, False, ..., False, True, True],
[False, False, False, ..., False, False, True],
[False, False, False, ..., True, False, False]],
...,
[[False, False, False, ..., False, False, False],
[False, False, False, ..., True, False, False],
[ True, False, False, ..., False, False, False],
...,
[False, False, False, ..., False, False, False],
[False, False, False, ..., False, False, False],
[False, False, False, ..., False, False, False]],
[[False, False, False, ..., True, False, False],
[False, False, False, ..., True, False, False],
[False, False, False, ..., True, False, False],
...,
[False, False, True, ..., False, False, False],
[False, False, True, ..., False, False, False],
[False, False, True, ..., False, False, False]],
[[False, False, False, ..., False, False, False],
[False, False, False, ..., False, False, False],
[False, False, False, ..., False, False, False],
...,
[False, False, False, ..., False, False, False],
[False, False, False, ..., False, False, False],
[False, False, False, ..., False, False, False]]],
device='cuda:0'))b.shape,hash_loc.shape(torch.Size([64, 3, 320, 320]), torch.Size([64, 10, 10]))%timeit b.view(64, 3, 10,320//10, 10,320//10)15.4 µs ± 2.08 µs per loop (mean ± std. dev. of 7 runs, 10,000 loops each)tiled_img=b.view(64, 3, 10,320//10, 10,320//10).permute(0,2,4,1,3,5)
b.shape,tiled_img.shape(torch.Size([64, 3, 320, 320]), torch.Size([64, 10, 10, 3, 32, 32]))show_images(b.view(64, 10, 10, 3, 32, 32)[0].flatten(end_dim=1),nrows=10)
b.view(64, 3, 10,320//10, 10,320//10).shapetorch.Size([64, 3, 10, 32, 10, 32])show_images(tiled_img[0].flatten(end_dim=1),nrows=10)
tiled_img[1].shape,(hash_loc==99)[0].shape(torch.Size([10, 10, 3, 32, 32]), torch.Size([10, 10]))tiled_img[8][(hash_loc==99)[8]].shapetorch.Size([15, 3, 32, 32])How would select select the patches of the image that fall into our mode bucket?
tiled_img[1][hash_loc[1]==199].shapetorch.Size([11, 3, 32, 32])show_image(b[0])<AxesSubplot:>
Seems to detect “objects.” Other images included scales for weight things, and sea shells. Dog nose, might be due to that space being close.
show_images(tiled_img[0][hash_loc[0]==199][:64],nrows=4)
show_images(tiled_img[8][(hash_loc==99)[8]],nrows=4)
show_image(b[8])<AxesSubplot:>
len(dls.dataset)*10*1012952001295200/2565059.375