I have a 2d tensor of the shape (batch_size, 500), 500 is no of voice frames taken at a time and each of these frames has two labels i.e. 0 and 1, 0 denoting silence while 1 denotes speech.
tensor([[0., 0., 0., ..., 0., 0., 0.],
[0., 0., 0., ..., 0., 0., 0.],
[0., 0., 0., ..., 0., 0., 0.],
...,
[0., 0., 0., ..., 1., 0., 0.],
[0., 0., 0., ..., 0., 0., 0.],
[0., 0., 1., ..., 0., 0., 0.]], device='cuda:0')
After passing the input features into my BiLSTM model from this is the output tensor
tensor([[0.5058, 0.5088, 0.5075, ..., 0.5071, 0.5079, 0.5057],
[0.5010, 0.4988, 0.4984, ..., 0.5046, 0.5041, 0.5022],
[0.5081, 0.5079, 0.5069, ..., 0.4985, 0.4982, 0.4992],
...,
[0.5064, 0.5104, 0.5117, ..., 0.5039, 0.5040, 0.5041],
[0.5049, 0.5075, 0.5079, ..., 0.5178, 0.5174, 0.5162],
[0.4936, 0.4948, 0.4970, ..., 0.5033, 0.5038, 0.5041]],
device='cuda:0', grad_fn=<SqueezeBackward0>)
Now, from the above tensor i am assigning 1 if the value is above 0.5 and 0 otherwise, so that each of the 500 frames has 1 or 0 assigned to it. After that i am calculating the BCEloss and then back propagating. But the issue is the input labels are unbalanced, so i want to assign class_weight to the labels during calculating the BCEloss.
After obtaining the class_weight from compute_class_weight module in sklearn, i am getting class_weights as array([0.59432247, 3.15048184]). But when i pass this as a tensor to the BCEloss i am getting an error.
RuntimeError: The size of tensor a (500) must match the size of tensor b (2) at non-singleton dimension 1
This is probably happening because the output tensor size is (batch_size, 500) while i have class_weights for two labels [0, 1].
Can anyone help me as to how i can assign the class_weights during training. Any kind of help would be greatly appreciated.
Thank You