import torch
import torch.nn as nn
import torch.nn.functional as F
import src.data.data as data
import src.data.config as cfg
import src.models.utils as model_utils
import src.evaluate.utils as eval_utils
import src.train.batch as batch_utils
def make_sampler(sampler_type, opt, *args, **kwargs):
print("Initializing Greedy Sampler")
return GreedySampler(opt, *args, **kwargs)
class Sampler():
def __init__(self, opt, data_loader, batch_mode=False):
# Token on which to end sampling
self.end_token = data_loader.vocab_encoder[data.end_token]
self.opt = opt
def generate_sequence(self, batch, model):
raise
class GreedySampler(Sampler):
def __init__(self, opt, data_loader, batch_mode=True):
super(GreedySampler, self).__init__(opt, data_loader)
def append_batch(self, X, next_idx, mask):
next_pos = X[:, -1:, 1] + 1
next_x = torch.cat((next_idx, next_pos), -1).unsqueeze(1)
next_mask = torch.cat([mask, torch.ones(X.size(0), 1, device=mask.device)], 1)
return torch.cat((X, next_x), 1), next_mask
def generate_sequence(self, batch, model, data_loader, start_idx, end_len):
XMB = batch["sequences"][:, :start_idx]
MMB = batch["attention_mask"][:, :start_idx]
XMB = model_utils.prepare_position_embeddings(
self.opt, data_loader.vocab_encoder, XMB.unsqueeze(-1))
lm_probs = F.log_softmax(model(
XMB.unsqueeze(1), sequence_mask=MMB), dim=-1)
values, indices = lm_probs[:, -1, :].max(dim=-1)
seqs = indices.clone().unsqueeze(1)
loss = values
counts = 1
next_pos = XMB[:, -1:, 1] + 1
next_x = torch.cat((indices.view(-1, 1), next_pos), -1).unsqueeze(1)
XMB = torch.cat((XMB, next_x), 1)
MMB = torch.cat([MMB, torch.ones(XMB.size(0), 1, device=MMB.device)], 1)
# Sample from top k
for _ in range(self.opt.eval.smax):
lm_probs = F.log_softmax(model(
XMB.unsqueeze(1), sequence_mask=MMB), dim=-1)
# Sample from top k
values, next_idx = lm_probs[:, -1, :].max(dim=-1)
loss += values
counts += 1
next_idx = next_idx.unsqueeze(1)
seqs = torch.cat([seqs, next_idx], 1)
if (next_idx.item() == self.end_token) or (_ == end_len - 1):
break
XMB, MMB = self.append_batch(XMB, next_idx, MMB)
beams = []
for beam in seqs:
beams.append(" ".join("".join(
[data_loader.vocab_decoder[tok.item()].replace(
'</w>', ' ').replace('\n', '')
for tok in beam if tok != self.end_token]).split()))
sampling_result = {
"sequence": beams[0],
"beams": beams,
"beam_losses": [loss.item()],
"loss": loss.item(),
"beam_lengths": [counts],
"length": counts
}
return sampling_result
class TopKSampler(Sampler):
def __init__(self, opt, data_loader, batch_mode=True):
super(TopKSampler, self).__init__(opt, data_loader)
def append_batch(self, X, next_idx, mask):
next_pos = X[:, -1:, 1] + 1
next_x = torch.cat((next_idx, next_pos), -1).unsqueeze(1)
next_mask = torch.cat([mask, torch.ones(X.size(0), 1, device=mask.device)], 1)
return torch.cat((X, next_x), 1), next_mask
def generate_sequence(self, batch, model, data_loader, start_idx, end_len):
# start_idx = context_size_event + 1
# start_idx = max_e1 + max_r
# end_idx = context_size_effect - 1
# end_idx = max_e2
XMB = batch["sequences"][:, :start_idx]
MMB = batch["attention_mask"][:, :start_idx]
XMB = model_utils.prepare_position_embeddings(
self.opt, data_loader.vocab_encoder, XMB.unsqueeze(-1))
lm_probs = F.log_softmax(model(
XMB.unsqueeze(1), sequence_mask=MMB), dim=-1)
values, indices = lm_probs[:, -1, :].topk(self.opt.eval.k)
seqs = indices.t().clone()
losses = - values.view(-1, 1)
ended = (seqs == self.end_token).float()
counts = (1 - ended)
XMB = XMB.repeat(self.opt.eval.k, 1, 1)
MMB = MMB.repeat(self.opt.eval.k, 1)
next_pos = XMB[:, -1:, 1] + 1
next_x = torch.cat((indices.view(self.opt.eval.k, -1), next_pos), -1).unsqueeze(1)
XMB = torch.cat((XMB, next_x), 1)
MMB = torch.cat([MMB, torch.ones(XMB.size(0), 1, device=MMB.device)], 1)
# Sample from top k
for _ in range(end_len):
lm_probs = F.log_softmax(model(
XMB.unsqueeze(1), sequence_mask=MMB), dim=-1)
# Sample from top k
values, indices = lm_probs[:, -1, :].topk(self.opt.eval.k)
choice = torch.multinomial(values.exp(), 1)
next_idx = indices.gather(-1, choice)
ended = ended + (next_idx == self.end_token).float() * (1 - ended)
next_idx = next_idx * (1 - ended).long() + ended.long() * self.end_token
counts += (1 - ended)
seqs = torch.cat([seqs, next_idx], 1)
if ended.sum().item() == self.opt.eval.k:
break
losses -= values.gather(-1, choice) * (1 - ended)
XMB, MMB = self.append_batch(XMB, next_idx, MMB)
beams = []
for beam in seqs:
beams.append(" ".join("".join(
[data_loader.vocab_decoder[tok.item()].replace(
'</w>', ' ').replace('\n', '')
for tok in beam if tok != self.end_token]).split()))
sampling_result = {
"sequence": beams[0],
"beams": beams,
"beam_losses": losses.squeeze().tolist(),
"loss": losses[0].item(),
"beam_lengths": counts.long().squeeze().tolist(),
"length": counts[0].long().item()
}
return sampling_result
class BeamSampler(TopKSampler):
def __init__(self, opt, data_loader, batch_mode=True, scorer=None):
super(BeamSampler, self).__init__(opt, data_loader, batch_mode)
self.kill_mask = torch.ones(opt.eval.bs, opt.eval.bs).to(cfg.device) * 9000
self.kill_mask[:, 0] = 0
def make_batch(self, X):
X = np.array(X)
assert X.ndim in [1, 2]
if X.ndim == 1:
X = np.expand_dims(X, axis=0)
pos_enc = np.arange(n_vocab + n_special, n_vocab + n_special + X.shape[-1])
pos_enc = np.expand_dims(pos_enc, axis=0)
batch = np.stack([X, pos_enc], axis=-1)
batch = torch.tensor(batch, dtype=torch.long).to(device)
return batch
def append_batch(self, X, beam_toks, mask):
next_pos = X[:, -1:, 1] + 1
next_x = torch.cat((beam_toks.unsqueeze(1), next_pos), -1).unsqueeze(1)
next_mask = torch.cat([mask, torch.ones(X.size(0), 1, device=mask.device)], 1)
return torch.cat((X, next_x), 1), next_mask
def generate_sequence(self, batch, model, data_loader, start_idx, end_len):
# start_idx = context_size_event + 1
# start_idx = max_e1 + max_r
# end_idx = context_size_effect - 1
# end_idx = max_e2
XMB = batch["sequences"][:, :start_idx]
MMB = batch["attention_mask"][:, :start_idx]
XMB = model_utils.prepare_position_embeddings(
self.opt, data_loader.vocab_encoder, XMB.unsqueeze(-1))
tokens = []
beam_losses = []
# Beam Search
beam_lls, beam_toks, beam_seqs = None, None, None
lm_probs = F.log_softmax(model(
XMB.unsqueeze(1), sequence_mask=MMB), dim=-1)
dist = lm_probs[:, -1, :].squeeze()
beam_lls, beam_toks = dist.topk(self.opt.eval.bs)
beam_losses.append(beam_lls)
ended = (beam_toks == self.end_token).float()
counts = (2 - ended)
beam_toks = beam_toks.unsqueeze(1)
beam_seqs = beam_toks.clone()
XMB = XMB.repeat(self.opt.eval.bs, 1, 1)
MMB = MMB.repeat(self.opt.eval.bs, 1)
next_pos = XMB[:, -1:, 1] + 1
next_x = torch.cat((beam_toks, next_pos), -1).unsqueeze(1)
XMB = torch.cat((XMB, next_x), 1)
MMB = torch.cat([MMB, torch.ones(XMB.size(0), 1, device=MMB.device)], 1)
for _ in range(end_len):
# Compute distribution for current beam
lm_probs = F.log_softmax(model(
XMB.unsqueeze(1), sequence_mask=MMB), dim=-1)
dist = lm_probs[:, -1, :].squeeze()
# get hypothesis tokens for distribution
hyp_beam_lls, hyp_beam_toks = dist.topk(self.opt.eval.bs)
# Compute masks and expand beam
expanded_ended = ended.unsqueeze(1).repeat(1, self.opt.eval.bs)
hypothesis_mask = expanded_ended * self.kill_mask + (1 - expanded_ended)
paper_results = False
if paper_results:
# Results from paper with slightly buggy beam search
current_beam_lls = beam_lls.unsqueeze(1).repeat(
1, self.opt.eval.bs).view(self.opt.eval.bs**2)
else:
# Current beam search implementation
current_beam_lls = beam_losses[-1].unsqueeze(1).repeat(
1, self.opt.eval.bs).view(self.opt.eval.bs**2)
# Compute losses of hypotheses, masking those that have ended
hyp_beam_lls = (hyp_beam_lls.view(self.opt.eval.bs**2) *
hypothesis_mask.view(-1)) + current_beam_lls
# Get normalizer for sequences
temp_counts = counts.unsqueeze(1).repeat(1, self.opt.eval.bs).view(
self.opt.eval.bs ** 2)
# Select best beams with lowest aggregate loss
beam_lls, top_beam_idxs = (hyp_beam_lls / temp_counts).topk(self.opt.eval.bs)
# Update placements in beam based on selecetion
beam_losses = [i.index_select(0, top_beam_idxs // self.opt.eval.bs)
for i in beam_losses]
ended = ended.index_select(0, top_beam_idxs // self.opt.eval.bs)
counts = temp_counts.index_select(0, top_beam_idxs)
# Save beam losses
beam_losses.append(beam_lls * counts)
# Update beam tokens
ended_mask = (1 - ended).long()
end_replacement = (self.end_token * ended).long()
next_toks = hyp_beam_toks.view(-1)[top_beam_idxs]
beam_toks = next_toks * ended_mask + end_replacement
# Update ended and counts
ended = ended + (beam_toks == self.end_token).float() * (1 - ended)
counts = counts + (1 - ended)
# Update beam sequences
beam_seqs = beam_seqs.t().repeat(self.opt.eval.bs, 1).t().contiguous().view(
self.opt.eval.bs**2, -1)[top_beam_idxs]
beam_seqs = torch.cat((beam_seqs, beam_toks.unsqueeze(1)), dim=1)
# I have no idea what's going on but Ari's on point with it
XMB = XMB.transpose(0, 1).transpose(1, 2).repeat(
self.opt.eval.bs, 1, 1).transpose(2, 1).transpose(
1, 0).contiguous().view(
self.opt.eval.bs**2, XMB.size(1), XMB.size(2))[top_beam_idxs]
XMB, MMB = self.append_batch(XMB, beam_toks, MMB)
if (beam_toks == self.end_token).sum().item() == self.opt.eval.bs:
break
beams = []
for beam in beam_seqs:
beams.append(" ".join("".join(
[data_loader.vocab_decoder[tok.item()].replace(
'</w>', ' ').replace('\n', '')
for tok in beam if tok != self.end_token]).split()))
sampling_result = {
"sequence": beams[0],
"beams": beams,
"beam_losses": beam_lls.tolist(),
"loss": beam_lls[0].item(),
"beam_lengths": counts.tolist(),
"length": counts[0].item()
}
return sampling_result
