def evaluate_fixed(self, test_iter, n_cl_sample):
cum_loss = 0
cum_prec = 0
cum_rec = 0
cum_size = 0
for t in range(test_iter):
words, context, target, cluster_iden, ixs = self.generate(
n_cl_sample)
input_x = torch.cat([words, context], dim=1)
kernel = self.kernel_net(input_x)
vals, vecs = custom_decomp()(kernel)
subset = DPP()(vals, vecs)
pick = subset.diag().mm(words).sum(0, keepdim=True)
self.pred = self.pred_net(pick).squeeze()
loss = nn.MSELoss()(self.pred, target)
# Subset Statistics
precision, recall, set_size = self.assess(subset.data, ixs)
cum_loss += loss.data[0]
cum_prec += precision
cum_rec += recall
cum_size += set_size
print(cum_loss / test_iter, cum_prec / test_iter, cum_rec / test_iter,
cum_size / test_iter)
return (cum_loss / test_iter, cum_prec / test_iter,
cum_rec / test_iter, cum_size / test_iter)
def sample(self):
words, context, ixs, target = self.generate()
input_x = torch.cat([words, context], dim=1)
kernel = self.kernel_net(input_x)
vals, vecs = custom_decomp()(kernel)
subset = DPP()(vals, vecs)
pick = subset.diag().mm(words).sum(0, keepdim=True)
self.pred = self.pred_net(pick).squeeze()
loss = nn.BCELoss()(self.pred, target)
# Classification Accuracy
label = torch.round(self.pred.data)
acc = ((label == target.data).sum() / self.aspects_n)
# Subset Statistics
precision, recall, set_size = self.assess(subset.data, ixs)
# Print
print('Target is: ', target.data)
print('Pred is: ', self.pred.data)
print('Loss is:', loss.data[0])
print('Acc is:', ((label == target.data).sum() / self.aspects_n))
print('Subset is:', subset.data)
print('Ix is:', ixs)
print('Subset statistics are:', precision, recall, set_size)
return words, context, ixs, target, self.pred, loss, subset
def evaluate(self, test_iter):
cum_loss = 0
cum_acc = 0
cum_prec = 0
cum_rec = 0
cum_size = 0
for t in range(test_iter):
words, context, ixs, target = self.generate()
input_x = torch.cat([words, context], dim=1)
kernel = self.kernel_net(input_x)
vals, vecs = custom_decomp()(kernel)
subset = DPP()(vals, vecs)
pick = subset.diag().mm(words).sum(0, keepdim=True)
self.pred = self.pred_net(pick).squeeze()
loss = nn.BCELoss()(self.pred, target)
# Classification Accuracy
label = torch.round(self.pred.data)
cum_acc += ((label == target.data).sum() / self.aspects_n)
# Subset Statistics
precision, recall, set_size = self.assess(subset.data, ixs)
cum_loss += loss.data[0]
cum_prec += precision
cum_rec += recall
cum_size += set_size
print('Loss:', cum_loss / test_iter, 'Pred Acc:', cum_acc / test_iter,
'Precision:', cum_prec / test_iter, 'Recall:',
cum_rec / test_iter, 'Set Size:', cum_size / test_iter)
def forward(self, kernels, words):
# need to set beforehand!
s_ix = self.s_ix
e_ix = self.e_ix
self.exp_sizes = []
assert s_ix != None and e_ix != None
alpha_iter = self.alpha_iter
batch_size = len(s_ix)
embd_dim = words.size(1)
output = []
lengths = []
actions = self.saved_subsets = [[] for i in range(batch_size)]
for i, (s, e) in enumerate(zip(s_ix, e_ix)):
V = kernels[s:e]
word = words[s:e]
vals, vecs = custom_decomp()(
V) # check if gradients work for non-square matrix
exp_size = (vals / (1 + vals)).pow(2).sum()
for j in range(alpha_iter):
while True: # to avoid zero subsets, problematic as not unbiased anymore, temp fix.
# subset = AllInOne()(V) # scrap this after gradient check!
subset = DPP()(vals, vecs)
if subset.data.sum() >= 1:
break
else:
print("Zero Subset was produced. Re-sample")
continue
actions[i].append(subset)
pick = subset.diag().mm(
word
) # but this creates zero rows, however it links the two graphs :)
pick = pick.masked_select(
Variable(subset.data.byte().expand_as(pick.t())).t())
pick = pick.view(-1, embd_dim)
output.append(pick)
lengths.append(pick.size(0))
self.exp_sizes.append(exp_size)
output = torch.cat(output, dim=0)
cum_lengths = list(accumulate(lengths))
self.s_ix = [ix1 - ix2 for (ix1, ix2) in zip(cum_lengths, lengths)]
self.e_ix = cum_lengths
return output
def run(self, words, context, batch_size, alpha_iter):
"""
This may be used by sample and by evaluate.
Samples once from DPP.
Can be used with any batch_size.
Returns a tensor of many subsets
"""
set_size = self.set_size
kernel_in = self.kernel_in
kernel_out = self.kernel_out
actions = self.saved_subsets = [[] for i in range(batch_size)]
rewards = self.saved_losses = [[] for i in range(batch_size)]
cum_loss = 0.
# Concatenate individual words and set context
# Dimensions are batch_size x set_size x kernel_in
batch_x = Variable(torch.cat([words, context], dim = 2))
# Compute embedding of DPP kernel
batch_kernel = self.kernel_net(batch_x.view(-1, kernel_in))
batch_kernel = batch_kernel.view(-1, set_size, kernel_out)
for i, kernel in enumerate(batch_kernel):
vals, vecs = custom_decomp()(kernel)
for j in range(alpha_iter):
subset = DPP()(vals, vecs)
actions[i].append(subset)
subset_tensor = torch.stack([torch.stack(subsets) for subsets in actions])
return subset_tensor
def __call__(self, kernels, batched_words):
batch_size, set_size, kernel_dim = kernels.size()
batch_size, set_size, embd_dim = batched_words.size()
alpha_iter = self.alpha_iter
self.exp_sizes = exp_sizes = []
self.saved_subsets = actions = [[] for i in range(batch_size)]
self.saved_picks = picks = [[] for i in range(batch_size)]
for i, (kernel, words) in enumerate(zip(kernels, batched_words)):
vals, vecs = custom_decomp()(kernel)
exp_size = (vals / (1 + vals)).sum()
exp_sizes.append(exp_size)
for j in range(alpha_iter):
while True: # to avoid zero subsets, problematic as not unbiased anymore, temp fix.
subset = DPP()(vals, vecs)
if subset.data.sum() >= 1:
break
else:
print("Zero Subset was produced. Re-sample")
continue
actions[i].append(subset)
pick = words.masked_select(
Variable(subset.data.byte().expand_as(words.t())).t())
pick = pick.view(-1, embd_dim)
picks[i].append(pick)
return picks
def train(self, train_iter, batch_size, lr, alpha_iter=1, baseline=True):
"""
Training the model.
Doesn't use the forward pass as want to sample repeatedly!
"""
set_size = self.set_size
kernel_in = self.kernel_in
kernel_out = self.kernel_out
loss_log = 100
optimizer = optim.Adam(self.kernel_net.parameters(), lr=lr)
for t in range(train_iter):
actions = self.saved_subsets = [[] for i in range(batch_size)]
rewards = self.saved_losses = [[] for i in range(batch_size)]
cum_loss = 0.
words, context, target = self.generate(batch_size)
# Concatenate individual words and set context
# Dimensions are batch_size x set_size x kernel_in
batch_x = Variable(torch.cat([words, context], dim = 2))
# Compute embedding of DPP kernel
batch_kernel = self.kernel_net(batch_x.view(-1, kernel_in))
batch_kernel = batch_kernel.view(-1, set_size, kernel_out)
for i, kernel in enumerate(batch_kernel):
vals, vecs = custom_decomp()(kernel)
for j in range(alpha_iter):
subset = DPP()(vals, vecs)
actions[i].append(subset)
loss, _, _, _, _ = self._assess(target[i], subset.data)
rewards[i].append(loss)
cum_loss += loss
if baseline:
self.saved_baselines = [compute_baseline(i) for i in self.saved_losses]
else:
self.saved_baselines = self.saved_losses
# Register the baselines
for actions, rewards in zip(self.saved_subsets, self.saved_baselines):
for action, reward in zip(actions, rewards):
action.reinforce(reward)
pseudo_loss = torch.stack([torch.stack(subsets) for subsets in self.saved_subsets]).sum()
pseudo_loss.backward(None)
optimizer.step()
optimizer.zero_grad()
self.loss_dict[t].append(cum_loss / (batch_size * alpha_iter))
if not ((t + 1) % loss_log):
print("Loss at it ", t+1, " is: ", cum_loss / (batch_size * alpha_iter))
def sample(self):
words, context, target, cluster_iden, ixs = self.generate()
input_x = torch.cat([words, context], dim=1)
kernel = self.kernel_net(input_x)
vals, vecs = custom_decomp()(kernel)
subset = DPP()(vals, vecs)
pick = subset.diag().mm(words).sum(0, keepdim=True)
self.pred = self.pred_net(pick).squeeze()
loss = nn.MSELoss()(self.pred, target)
# Subset Statistics
precision, recall, set_size = self.assess(subset.data, ixs)
# Print
print('Target is: ', target.data)
print('Pred is: ', self.pred.data)
print('Loss is:', loss.data[0])
print('Subset is:', subset.data)
print('Ix is:', ixs)
print('Subset statistics are:', precision, recall, set_size)
def evaluate(self, test_iter):
set_size = self.set_size
kernel_in = self.kernel_in
kernel_out = self.kernel_out
set_size = self.set_size
n_clusters = self.n_clusters
kernel_in = self.kernel_in
kernel_out = self.kernel_out
embd_dim = self.pred_in
dtype = self.dtype
criterion = self.criterion
cum_loss = 0.
cum_prec = 0.
cum_rec = 0.
cum_ssize = 0.
n_missed = 0.
n_one = 0
n_many = 0.
n_perfect = 0.
mean = 0.
temp = 0.
var = 0.
criterion = self.criterion
batch_size = 1
embd_dim = self.pred_in
for t in range(test_iter):
picks = [[] for i in range(batch_size)]
words, context, target, index = self.generate(batch_size)
# Concatenate individual words and set context
# Dimensions are batch_size x set_size x kernel_in
batch_x = Variable(torch.cat([words, context], dim = 2))
# Compute embedding of DPP kernel
batch_kernel = self.kernel_net(batch_x.view(-1, kernel_in))
batch_kernel = batch_kernel.view(-1, set_size, kernel_out)
for i, kernel in enumerate(batch_kernel):
vals, vecs = custom_decomp()(kernel)
subset = DPP()(vals, vecs)
pick = subset.diag().mm(Variable(words[i])).sum(0, keepdim=True)
picks[i].append(pick)
_, prec, rec, ssize = self._big_assess(index[i], subset.data)
missed, one, many, perfect = self._eval_assess(index, subset.data)
cum_prec += prec
cum_rec += rec
cum_ssize += ssize
n_missed += missed
n_one += one
n_many += many
n_perfect += perfect
picks = torch.stack([torch.stack(pick) for pick in picks]).view(-1, embd_dim)
preds = self.pred_net(picks).view(1, -1)
targets = target.unsqueeze(0).expand_as(preds)
loss = criterion(preds, Variable(targets, volatile=True))
cum_loss += loss.data[0]
delta = ssize - mean
mean += delta / (t+1)
delta2 = ssize - mean
temp += delta * delta2
var = temp / test_iter
print("Average Subset Size: ", mean)
print("Subset Variance: ", var)
print("Average Loss", cum_loss / test_iter)
print("n_missed share", n_missed / (test_iter * self.n_clusters))
print("n_one share", n_one / (test_iter * self.n_clusters))
print("n_many share", n_many / (test_iter * self.n_clusters))
print("n_perfect share", n_perfect/ test_iter)
# Concatenate individual words and set context
# Dimensions are batch_size x set_size x kernel_in
batch_x = Variable(torch.cat([words, context], dim = 2))
# Compute embedding of DPP kernel
batch_kernel = self.kernel_net(batch_x.view(-1, kernel_in))
batch_kernel = batch_kernel.view(-1, set_size, kernel_out)
if reg:
reg_loss = 0
for i, kernel in enumerate(batch_kernel):
vals, vecs = custom_decomp()(kernel)
for j in range(alpha_iter):
subset = DPP()(vals, vecs)
pick = subset.diag().mm(.a(words[i])).sum(0, keepdim=True)
actions[i].append(subset)
picks[i].append(pick)
_, prec, rec, ssize = self._big_assess(index[i], subset.data)
cum_prec += prec
cum_rec += rec
cum_ssize += ssize
if reg:
exp_ssize = (vals / (1 + vals)).sum()
reg_loss += reg * (exp_ssize - reg_mean)**2
picks = torch.stack([torch.stack(pick) for pick in picks]).view(-1, embd_dim)
preds = self.pred_net(picks).view(batch_size, alpha_iter, -1)
def train(self,
train_steps,
batch_size=1,
sample_iter=1,
lr=1e-3,
baseline=False,
reg=0,
reg_mean=0):
if baseline:
assert sample_iter > 1
params = [{
'params': self.kernel_net.parameters()
}, {
'params': self.pred_net.parameters()
}]
optimizer = optim.Adam(params, lr=lr)
train_iter = train_steps * batch_size
cum_loss = 0
cum_prec = 0
cum_rec = 0
cum_size = 0
for t in range(train_iter):
actions = self.saved_subsets = []
rewards = self.saved_losses = []
picks = []
words, context, ixs, target = self.generate()
input_x = torch.cat([words, context], dim=1)
kernel = self.kernel_net(input_x)
vals, vecs = custom_decomp()(kernel)
pred_loss = 0
for j in range(sample_iter):
subset = DPP()(vals, vecs)
actions.append(subset)
pick = subset.diag().mm(words).sum(0, keepdim=True)
self.pred = self.pred_net(pick).squeeze()
loss = nn.BCELoss()(self.pred, target)
rewards.append(loss.data[0])
pred_loss += loss
# For the statistics
precision, recall, set_size = self.assess(subset.data, ixs)
cum_loss += loss.data[0]
cum_prec += precision
cum_rec += recall
cum_size += set_size
# Compute baselines
if baseline:
self.saved_baselines = compute_baseline(self.saved_losses)
else:
self.saved_baselines = self.saved_losses
# Register rewards
for action, reward in zip(self.saved_subsets,
self.saved_baselines):
action.reinforce(reward)
# Apply Regularization
total_loss = pred_loss
if reg:
card = (vals / (1 + vals)).sum()
reg_loss = sample_iter * reg * ((card - reg_mean)**2)
total_loss += reg_loss
total_loss.backward()
if not ((t + 1) % batch_size):
optimizer.step()
optimizer.zero_grad()
if not ((t + 1) % (batch_size * 100)):
print(cum_loss / (batch_size * sample_iter))
self.loss_dict[self.counter].append(cum_loss /
(batch_size * sample_iter))
self.prec_dict[self.counter].append(cum_prec /
(batch_size * sample_iter))
self.rec_dict[self.counter].append(cum_rec /
(batch_size * sample_iter))
self.ssize_dict[self.counter].append(
cum_size / (batch_size * sample_iter))
self.counter += 1
cum_loss = 0
cum_prec = 0
cum_rec = 0
cum_size = 0