x-transformers 2.1.9__py3-none-any.whl → 2.1.11__py3-none-any.whl

x_transformers/belief_state_wrapper.py

@@ -12,6 +12,11 @@ from torch.nn import Module, ModuleList
 from torch import nn, cat, stack, tensor, arange, cartesian_prod
 import torch.nn.functional as F
 
+from x_transformers.autoregressive_wrapper import (
+    eval_decorator,
+    min_p,
+)
+
 from x_transformers.x_transformers import (
     Decoder,
     TransformerWrapper
@@ -28,6 +33,15 @@ def exists(v):
 def default(v, d):
     return v if exists(v) else d
 
+def eval_decorator(fn):
+    def inner(self, *args, **kwargs):
+        was_training = self.training
+        self.eval()
+        out = fn(self, *args, **kwargs)
+        self.train(was_training)
+        return out
+    return inner
+
 # wrappers
 
 class BeliefStateWrapper(Module):
@@ -38,12 +52,14 @@ class BeliefStateWrapper(Module):
     def __init__(
         self,
         forward_decoder: TransformerWrapper,
-        backward_decoder: TransformerWrapper,
+        backward_decoder: TransformerWrapper | None = None,
         train_frac_forward_backward_pairs: float = 1.,
         text_head: Module | None = None,
         backward_ar_loss_weight: float = 1. # can weigh the training of the backwards decoder differently, perhaps fwd/bwd have a shared backbone etc etc
     ):
         super().__init__()
+        backward_decoder = default(backward_decoder, forward_decoder) # if backward decoder not set, use the same transformer, assume it knows how to switch gears based on suffix token
+
         assert forward_decoder.emb_dim == backward_decoder.emb_dim, 'forward and backwards model must have the same embedding dimension'
         assert forward_decoder.num_tokens == backward_decoder.num_tokens, 'forward and backwards model must have the same number of tokens'
 
@@ -85,6 +101,88 @@ class BeliefStateWrapper(Module):
 
         self.register_buffer('loss_weights', tensor([1., self.backward_ar_loss_weight]))
 
+        # sampling
+
+        self.max_seq_len = self.forward_decoder.max_seq_len
+
+    @torch.no_grad()
+    @eval_decorator
+    def generate_with_suffix_token_only(
+        self,
+        prompts,
+        seq_len,
+        temperature = 1.25,
+        cache_kv = True,
+        suffix: Tensor | None = None, # the goal conditioning
+        filter_logits_fn = min_p,
+        filter_kwargs = dict(
+            min_p = 0.1
+        ),
+        **kwargs
+    ):
+        max_seq_len, greedy, device = self.max_seq_len, temperature == 0., prompts.device
+
+        batch, orig_seq_len = prompts.shape
+
+        out = prompts
+
+        # kv caches
+
+        cache = None
+
+        # get the encoded suffix token once
+
+        if not exists(suffix):
+            suffix = out[:, 0:0]
+
+        if suffix.ndim == 1:
+            suffix = repeat(suffix, 'n -> b n', b = batch)
+
+        suffix_sos_tokens = rearrange(self.suffix_token, 'd -> 1 1 d')
+
+        suffix_sos_tokens = repeat(suffix_sos_tokens, '1 1 d -> b 1 d', b = batch)
+
+        suffix_embed = self.backward_decoder(
+            suffix,
+            prepend_embeds = suffix_sos_tokens,
+            return_embeddings = True
+        )
+
+        # sampling up to seq_len
+
+        for _ in range(seq_len):
+
+            embeds, new_cache = self.forward_decoder(
+                out,
+                return_intermediates = True,
+                return_embeddings = True,
+                cache = cache,
+                **kwargs
+            )
+
+            last_embeds = embeds[:, -1:]
+            embeds = cat((last_embeds, suffix_embed), dim = -1)
+
+            if cache_kv and self.forward_decoder.can_cache_kv:
+                cache = new_cache
+
+            logits, _ = self.text_head(embeds).chunk(2, dim = -1)
+
+            logits = logits[:, -1]
+
+            if greedy:
+                sample = logits.argmax(dim = -1, keepdim = True)
+            else:
+                filtered_logits = filter_logits_fn(logits, **filter_kwargs)
+                probs = F.softmax(filtered_logits / temperature, dim = -1)
+                sample = torch.multinomial(probs, 1)
+
+            # concat sample
+
+            out = torch.cat((out, sample), dim=-1)
+
+        return out[:, orig_seq_len:]
+
     def forward(
         self,
         seq,

{x_transformers-2.1.9.dist-info → x_transformers-2.1.11.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: x-transformers
-Version: 2.1.9
+Version: 2.1.11
 Summary: X-Transformers
 Project-URL: Homepage, https://pypi.org/project/x-transformers/
 Project-URL: Repository, https://github.com/lucidrains/x-transformers

{x_transformers-2.1.9.dist-info → x_transformers-2.1.11.dist-info}/RECORD

@@ -1,7 +1,7 @@
 x_transformers/__init__.py,sha256=NDoiBivau559WQ0FvXG4ssU3Il9aoHmTIUFN_1juz0s,911
 x_transformers/attend.py,sha256=-5BWWhFsp7tvZTdN91Ay5SqOjyj9uOs-122vFvoO6b4,17253
 x_transformers/autoregressive_wrapper.py,sha256=reLCno9Z9pchVU79tBF8OMo21LwSZ67KAeB83jqkyAc,10505
-x_transformers/belief_state_wrapper.py,sha256=bA-H9BqfyVRI5Q7GIcGbzdLjmon0CKKGHb08-BnpJOs,5990
+x_transformers/belief_state_wrapper.py,sha256=mpfTNZb8gadbtlpG2TxyfIMWkMVM4vigFDqCJ_mjxSU,8711
 x_transformers/continuous.py,sha256=p0sCAiH1na236ygwgL1Yyhu36eZBf9cZvoW1JyP_fFE,7073
 x_transformers/dpo.py,sha256=xt4OuOWhU8pN3OKN2LZAaC2NC8iiEnchqqcrPWVqf0o,3521
 x_transformers/multi_input.py,sha256=tCh-fTJDj2ib4SMGtsa-AM8MxKzJAQSwqAXOu3HU2mg,9252
@@ -10,7 +10,7 @@ x_transformers/nonautoregressive_wrapper.py,sha256=2NU58hYMgn-4Jzg3mie-mXb0XH_dC
 x_transformers/x_transformers.py,sha256=-80N4sqUr3sR51Ms4wCfc4jhxnPwf0ApNR4xfIsasfQ,110142
 x_transformers/xl_autoregressive_wrapper.py,sha256=CvZMJ6A6PA-Y_bQAhnORwjJBSl6Vjq2IdW5KTdk8NI8,4195
 x_transformers/xval.py,sha256=7S00kCuab4tWQa-vf-z-XfzADjVj48MoFIr7VSIvttg,8575
-x_transformers-2.1.9.dist-info/METADATA,sha256=YXRYQPw90873on2JyVr56N8Tz4ua2ibPawTUDvVE35g,87570
-x_transformers-2.1.9.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-x_transformers-2.1.9.dist-info/licenses/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
-x_transformers-2.1.9.dist-info/RECORD,,
+x_transformers-2.1.11.dist-info/METADATA,sha256=miwxNJVS0ZNJlw3qeJaGaxunTRxLgmz9WITWi_jnXcc,87571
+x_transformers-2.1.11.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+x_transformers-2.1.11.dist-info/licenses/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
+x_transformers-2.1.11.dist-info/RECORD,,