def v1_9_pos_log(output, eye_mask, eps, neg_freq, pos_freq):
output_len = len(output)
output_mat = torch.cat(output)
output_mat = normalize(output_mat, dim=1)
output_mat_sim = output_mat.mm(output_mat.t())
dt_size = output_mat.shape[0]
pred = (output_mat_sim * eye_mask[:dt_size, :dt_size]).argmax(1)
label_range = (torch.from_numpy(np.cumsum(
[i.shape[0] for i in output])).to("cuda").repeat([dt_size, 1]).t())
pred = (pred > label_range).sum(0)
label = torch.cat([
torch.ones(o.shape[0], device="cuda") * l
for l, o in zip(range(output_len), output)
])
acc = (pred == label).float().mean()
# concept_mask = torch.stack([label != i for i in label]).float()
concept_mask = torch.stack([label != i for i in label])
concept_mask_not = torch.logical_not(concept_mask)
pos_score = output_mat_sim * concept_mask_not + concept_mask
pos_loss = -((1 - pos_score) * torch.log((1 + pos_score) / 2)).sum(0)
pos_loss = pos_loss / (concept_mask_not.sum(0) - 1 + eps)
pos_loss = pos_loss.sum()
neg_score = output_mat_sim * concept_mask
neg_score = torch.cos_(neg_score * neg_freq)
neg_loss = -(0.5 * (2 - neg_score) * torch.log(
(1 + neg_score) / 2)).sum(0) / (concept_mask.sum(0) + eps)
neg_loss = neg_loss.sum()
loss = pos_loss + neg_loss
result = {
"loss": loss,
"pos_loss": pos_loss,
"neg_loss": neg_loss,
"acc": acc,
}
return result
print ("El valor de b[1, 1] es de ", b[1, 1])
c = b[1:, 1:]
print ("El tamaño de c es de", c.size())
print ("El stride de c es de ", c.stride())
print ("El offset de c es de ", c.storage_offset())
# In[6]:
#Ejercicio 2
torchcos = torch.Tensor([0,1,2])
torchcos_1 = torch.cos(torchcos)
print(torchcos_1)
#In place
torch.cos_(torch.Tensor([0,1,2]))
# In[20]:
#Ejercicio 3
d = torch.Tensor([[0,1,2],[0,1,2]])
d.unsqueeze_(0).shape
# In[15]:
#Ejercicio 4
d.squeeze_(0).shape