x-transformers 2.2.5__py3-none-any.whl → 2.2.7__py3-none-any.whl

x_transformers/entropy_based_tokenizer.py

@@ -6,12 +6,6 @@ import torch.nn.functional as F
 from torch.nn import Module
 from torch.nn.utils.rnn import pad_sequence
 
-from x_transformers.x_transformers import (
-    Decoder,
-    TransformerWrapper,
-    calc_entropy
-)
-
 import einx
 from einops import repeat, rearrange, pack, unpack
 
@@ -23,6 +17,13 @@ def exists(v):
 def default(v, d):
     return v if exists(v) else d
 
+def log(t, eps = 1e-20):
+    return t.clamp(min = eps).log()
+
+def calc_entropy_from_logits(logits):
+    prob = logits.softmax(dim = -1)
+    return -(prob * log(prob)).sum(dim = -1)
+
 # entropy based tokenizer applied in byte-latent transformer paper
 # they use a simple entropy threshold for segmenting a string into variable sized tokens
 
@@ -43,7 +44,8 @@ class EntropyBasedTokenizer(Module):
         self,
         seq,            # Float['b n'] | Float['n']
         lens = None,    # Int['b']
-        return_segmented_seq = False
+        return_segmented_seq = False,
+        decoder_forward_kwargs: dict = dict()
     ):
         no_batch_dim = seq.ndim == 1
         seq, maybe_batch_ps = pack((seq,), '* n')
@@ -57,9 +59,9 @@ class EntropyBasedTokenizer(Module):
 
         # forward through a small trained decoder and get the entropies of the logits
 
-        logits = self.decoder(seq)
+        logits = self.decoder(seq, **decoder_forward_kwargs)
 
-        entropies = calc_entropy(logits)
+        entropies = calc_entropy_from_logits(logits)
 
         # get length mask for boundaries
 

{x_transformers-2.2.5.dist-info → x_transformers-2.2.7.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: x-transformers
-Version: 2.2.5
+Version: 2.2.7
 Summary: X-Transformers
 Project-URL: Homepage, https://pypi.org/project/x-transformers/
 Project-URL: Repository, https://github.com/lucidrains/x-transformers

{x_transformers-2.2.5.dist-info → x_transformers-2.2.7.dist-info}/RECORD

@@ -4,14 +4,14 @@ x_transformers/autoregressive_wrapper.py,sha256=reLCno9Z9pchVU79tBF8OMo21LwSZ67K
 x_transformers/belief_state_wrapper.py,sha256=YLUMk6t2MhFBEw5lHDDHJHcoCxTIkHvxTNY__GGZEKU,13374
 x_transformers/continuous.py,sha256=p0sCAiH1na236ygwgL1Yyhu36eZBf9cZvoW1JyP_fFE,7073
 x_transformers/dpo.py,sha256=xt4OuOWhU8pN3OKN2LZAaC2NC8iiEnchqqcrPWVqf0o,3521
-x_transformers/entropy_based_tokenizer.py,sha256=Dr7Hvb0WbdhCeD3_61jAQeONr6IWFFSCRrqBfn0d46Y,3772
+x_transformers/entropy_based_tokenizer.py,sha256=c9dcERmoK454mD3JKlxeU3cNwYVPi8I8bI91xrvVvOA,3938
 x_transformers/multi_input.py,sha256=tCh-fTJDj2ib4SMGtsa-AM8MxKzJAQSwqAXOu3HU2mg,9252
 x_transformers/neo_mlp.py,sha256=XCNnnop9WLarcxap1kGuYc1x8GHvwkZiDRnXOxSl3Po,3452
 x_transformers/nonautoregressive_wrapper.py,sha256=2NU58hYMgn-4Jzg3mie-mXb0XH_dCN7fjlzd3K1rLUY,10510
 x_transformers/x_transformers.py,sha256=Fl2CuAKTxJDOQvqwQo2FK8eO2s1iNLO-P1PP2Yw64rQ,111826
 x_transformers/xl_autoregressive_wrapper.py,sha256=CvZMJ6A6PA-Y_bQAhnORwjJBSl6Vjq2IdW5KTdk8NI8,4195
 x_transformers/xval.py,sha256=7S00kCuab4tWQa-vf-z-XfzADjVj48MoFIr7VSIvttg,8575
-x_transformers-2.2.5.dist-info/METADATA,sha256=5hJR7TPpRznsPZrYAHsbk3Eq9E9XKrAuwybmc3r6B1s,88686
-x_transformers-2.2.5.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-x_transformers-2.2.5.dist-info/licenses/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
-x_transformers-2.2.5.dist-info/RECORD,,
+x_transformers-2.2.7.dist-info/METADATA,sha256=vkh3IO9_ibFfNcvg0vYiwtqAHM51QNu3aMOwJnyhaNQ,88686
+x_transformers-2.2.7.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+x_transformers-2.2.7.dist-info/licenses/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
+x_transformers-2.2.7.dist-info/RECORD,,