How to interpret the miner output?

[celsofranssa]

How to interpret the following miner output:

import torch
from pytorch_metric_learning import miners, losses

# miner and loss function
miner = miners.MultiSimilarityMiner()
loss_func = losses.TripletMarginLoss()

# embeddings and labels
embeddings = torch.tensor([
                [.91, .70, .82],
                [.89, .71, .79],
                [.09, .07, .09],
                [.05, .03, .08],
])
labels = torch.tensor([1,1,0,0])

miner_outs = miner(embeddings,labels)
# (tensor([0, 1, 2, 3]),
# tensor([1, 0, 3, 2]),
# tensor([0, 0, 1, 1, 2, 2, 3, 3]),
# tensor([3, 2, 3, 2, 1, 0, 1, 0]))

loss_func(embeddings, labels, miner_outs) 
# tensor(0.1183)

[KevinMusgrave]

MultiSimilarityMiner returns pairs, so the output format is (a1, p, a2, n).

• (a1, p) is the indices of positive pairs
• (a2, n) is the indices the negative pairs

In your case:

• a1 = tensor([0, 1, 2, 3])
• p = tensor([1, 0, 3, 2])
• a2 = tensor([0, 0, 1, 1, 2, 2, 3, 3])
• n = tensor([3, 2, 3, 2, 1, 0, 1, 0]))

So there are 4 positive pairs and 8 negative pairs.

For example:

The 1st positive pair is (a1[0], p[0]), which is (0, 1), which corresponds with (embeddings[0], embeddings[1]).

The 5th negative pair is (a2[4], n[4]), which is (2, 1), which corresponds with (embeddings[2], embeddings[1]).

[celsofranssa]

MultiSimilarityMiner returns pairs, so the output format is (a1, p, a2, n).

• (a1, p) is the indices of positive pairs
• (a2, n) is the indices the negative pairs

In your case:

• a1 = tensor([0, 1, 2, 3])
• p = tensor([1, 0, 3, 2])
• a2 = tensor([0, 0, 1, 1, 2, 2, 3, 3])
• n = tensor([3, 2, 3, 2, 1, 0, 1, 0]))

So there are 4 positive pairs and 8 negative pairs.

For example:

The 1st positive pair is (a1[0], p[0]), which is (0, 1), which corresponds with (embeddings[0], embeddings[1]).

The 5th negative pair is (a2[4], n[4]), which is (2, 1), which corresponds with (embeddings[2], embeddings[1]).

Thank you, @KevinMusgrave,

Is all more clear explained now!