x-transformers 1.37.3__py3-none-any.whl → 1.37.5__py3-none-any.whl

x_transformers/attend.py

@@ -13,7 +13,7 @@ from functools import wraps
 from packaging import version
 from dataclasses import dataclass
 
-from einops import rearrange, repeat
+from einops import rearrange, repeat, pack, unpack
 
 # constants
 
@@ -39,9 +39,16 @@ def default(val, d):
 def compact(arr):
     return [*filter(exists, arr)]
 
-def softclamp(t, value):
+@torch.jit.script
+def softclamp(t: Tensor, value: float):
     return (t / value).tanh() * value
 
+def pack_one(t, pattern):
+    return pack([t], pattern)
+
+def unpack_one(t, ps, pattern):
+    return unpack(t, ps, pattern)[0]
+
 def once(fn):
     called = False
     @wraps(fn)
@@ -55,6 +62,18 @@ def once(fn):
 
 print_once = once(print)
 
+# alternative distance functions
+
+def qk_l2_distance(q, k):
+    if k.ndim == 3:
+        k = repeat(k, 'b j d -> b h j d', h = q.shape[1])
+
+    q, packed_shape = pack_one(q, '* i d')
+    k, _ = pack_one(k, '* j d')
+
+    distance = torch.cdist(q, k)
+    return unpack_one(distance, packed_shape, '* i j')
+
 # functions for creating causal mask
 # need a special one for onnx cpu (no support for .triu)
 
@@ -80,6 +99,7 @@ class Attend(Module):
         sparse_topk = None,
         scale = None,
         qk_norm = False,
+        l2_distance = False,
         flash = False,
         softclamp_logits = False,
         logit_softclamp_value = 50.,
@@ -123,6 +143,11 @@ class Attend(Module):
         assert not (flash and sigsoftmax), 'sigsoftmax not available for flash attention'
         self.sigsoftmax = sigsoftmax
 
+        # l2 distance attention
+
+        assert not (flash and l2_distance), 'l2 distance attention does not work with flash attention just yet'
+        self.l2_distance = l2_distance
+
         # add a key / value token composed of zeros
         # in case this helps controlling outliers, proposed by https://www.evanmiller.org/attention-is-off-by-one.html
 
@@ -325,7 +350,12 @@ class Attend(Module):
 
         kv_einsum_eq = 'b j d' if k.ndim == 3 else 'b h j d'
 
-        sim = einsum(f'b h i d, {kv_einsum_eq} -> b h i j', q, k) * scale
+        if not self.l2_distance:
+            sim = einsum(f'b h i d, {kv_einsum_eq} -> b h i j', q, k)
+        else:
+            sim = -qk_l2_distance(q, k)
+
+        sim = sim * scale
 
         if exists(prev_attn):
             sim = sim + prev_attn

x_transformers/autoregressive_wrapper.py

@@ -317,7 +317,7 @@ class AutoregressiveWrapper(Module):
             **kwargs
         )
 
-        loss_fn = F.cross_entropy if not self.net.is_log_prob else F.nll_loss
+        loss_fn = F.cross_entropy if not self.net.output_is_log_prob else F.nll_loss
 
         loss = loss_fn(
             rearrange(logits, 'b n c -> b c n'),

x_transformers/nonautoregressive_wrapper.py

@@ -309,9 +309,11 @@ class NonAutoregressiveWrapper(nn.Module):
         with context():
             logits = self.net(masked, **kwargs)
 
+        loss_fn = F.cross_entropy if not self.net.output_is_log_prob else F.nll_loss
+
         # cross entropy loss
 
-        loss = F.cross_entropy(
+        loss = loss_fn(
             logits[mask],
             orig_seq[mask]
         )

x_transformers/x_transformers.py

@@ -923,6 +923,7 @@ class Attention(Module):
         qk_norm_groups = 1,
         qk_norm_scale = 10,
         qk_norm_dim_scale = False,
+        l2_distance = False,
         one_kv_head = False,
         kv_heads = None,
         shared_kv = False,
@@ -1037,6 +1038,7 @@ class Attention(Module):
             sparse_topk = sparse_topk,
             qk_norm = qk_norm,
             scale = qk_norm_scale if qk_norm else self.scale,
+            l2_distance = l2_distance,
             add_zero_kv = add_zero_kv,
             flash = flash,
             softclamp_logits = softclamp_logits,
@@ -2078,7 +2080,7 @@ class TransformerWrapper(Module):
 
         # output type
 
-        self.is_log_prob = mixture_of_softmax
+        self.output_is_log_prob = mixture_of_softmax
 
         self.to_mixture = None
         self.combine_mixture = None

x_transformers/xl_autoregressive_wrapper.py

@@ -40,7 +40,7 @@ class XLAutoregressiveWrapper(nn.Module):
         eos_token = None,
         temperature = 1.,
         filter_logits_fn = top_k,
-        filter_thres = 0.9,
+        filter_kwargs: dict = dict(),
         mems = None,
         **kwargs
     ):
@@ -88,7 +88,7 @@ class XLAutoregressiveWrapper(nn.Module):
             mems = cache.mems
 
             logits = logits[:, -1]
-            filtered_logits = filter_logits_fn(logits, thres = filter_thres)
+            filtered_logits = filter_logits_fn(logits, **filter_kwargs)
             probs = F.softmax(filtered_logits / temperature, dim=-1)
 
             sample = torch.multinomial(probs, 1)
@@ -131,7 +131,9 @@ class XLAutoregressiveWrapper(nn.Module):
 
         split_x = x.split(max_seq_len, dim = -1)
         split_labels = labels.split(max_seq_len, dim = -1)
-        loss_weights = tuple(map(lambda t: t.shape[-1] / seq_len, split_x))
+        loss_weights = tuple((t.shape[-1] / seq_len) for t in split_x)
+
+        loss_fn = F.cross_entropy if not self.net.output_is_log_prob else F.nll_loss
 
         # go through each chunk and derive weighted losses
 
@@ -146,7 +148,7 @@ class XLAutoregressiveWrapper(nn.Module):
                 **kwargs
             )
 
-            loss = F.cross_entropy(
+            loss = loss_fn(
                 rearrange(logits, 'b n c -> b c n'),
                 chunk_labels,
                 ignore_index = ignore_index

{x_transformers-1.37.3.dist-info → x_transformers-1.37.5.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: x-transformers
-Version: 1.37.3
+Version: 1.37.5
 Summary: X-Transformers - Pytorch
 Home-page: https://github.com/lucidrains/x-transformers
 Author: Phil Wang

x_transformers-1.37.5.dist-info/RECORD

@@ -0,0 +1,15 @@
+x_transformers/__init__.py,sha256=-MkQrSc37cTVDX7AOykxunYnqVtFlQ7lb0Cse5dsGWU,793
+x_transformers/attend.py,sha256=4RnX1yhWZIf8holucqnYXTIP7U1m40UpP58RZNT_2sM,13128
+x_transformers/autoregressive_wrapper.py,sha256=DOJJCMMDOqDYKWy_IaG5IyKsXD3AW6amzfUgdAADOLY,10500
+x_transformers/continuous.py,sha256=cIVEdhfei258__ziV7kQBrJMxCel54bExBTDrO9rfCI,6450
+x_transformers/dpo.py,sha256=LjvWgCkqTl-UuehrzQ8nkX5guLr4whYwsmm7SKSwdls,3450
+x_transformers/multi_input.py,sha256=tCh-fTJDj2ib4SMGtsa-AM8MxKzJAQSwqAXOu3HU2mg,9252
+x_transformers/nonautoregressive_wrapper.py,sha256=2NU58hYMgn-4Jzg3mie-mXb0XH_dCN7fjlzd3K1rLUY,10510
+x_transformers/x_transformers.py,sha256=-2fj6QcDSfMI5lJA_fzOW2mdzdS1C1LD6jMBtGQY48E,83752
+x_transformers/xl_autoregressive_wrapper.py,sha256=CvZMJ6A6PA-Y_bQAhnORwjJBSl6Vjq2IdW5KTdk8NI8,4195
+x_transformers/xval.py,sha256=QE1ltYZTR_eGgIHPP2BrMWVWVLqMW-OpDZh87BSmQEg,8563
+x_transformers-1.37.5.dist-info/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
+x_transformers-1.37.5.dist-info/METADATA,sha256=zHUhvP1bQjFbMtxnVO9iDESgXpGOQxuBCsm4b6K1w44,661
+x_transformers-1.37.5.dist-info/WHEEL,sha256=GV9aMThwP_4oNCtvEC2ec3qUYutgWeAzklro_0m4WJQ,91
+x_transformers-1.37.5.dist-info/top_level.txt,sha256=hO6KGpFuGucRNEtRfme4A_rGcM53AKwGP7RVlRIxS5Q,15
+x_transformers-1.37.5.dist-info/RECORD,,

x_transformers-1.37.3.dist-info/RECORD

@@ -1,15 +0,0 @@
-x_transformers/__init__.py,sha256=-MkQrSc37cTVDX7AOykxunYnqVtFlQ7lb0Cse5dsGWU,793
-x_transformers/attend.py,sha256=mV7duZ7ON2puS3-k4ctBifb2rq-jTJqrMbof7tI5jR4,12326
-x_transformers/autoregressive_wrapper.py,sha256=2FN4ZobFcdDGDGWEnUof_geb16dRGSJycZGwG899Pa4,10493
-x_transformers/continuous.py,sha256=cIVEdhfei258__ziV7kQBrJMxCel54bExBTDrO9rfCI,6450
-x_transformers/dpo.py,sha256=LjvWgCkqTl-UuehrzQ8nkX5guLr4whYwsmm7SKSwdls,3450
-x_transformers/multi_input.py,sha256=tCh-fTJDj2ib4SMGtsa-AM8MxKzJAQSwqAXOu3HU2mg,9252
-x_transformers/nonautoregressive_wrapper.py,sha256=ys_p8obc7lTeeodCqvkRKxOXQ1C9T3j5Jwr-JbVgnXk,10432
-x_transformers/x_transformers.py,sha256=gOJBZzOJMu5RkIsxw9TZtde4Sx--D18yX8LjrYIsPbE,83677
-x_transformers/xl_autoregressive_wrapper.py,sha256=DCx4n0_c1tFai4nOqaWVnqx2p9eutsZsDMiMP1ckxNU,4117
-x_transformers/xval.py,sha256=QE1ltYZTR_eGgIHPP2BrMWVWVLqMW-OpDZh87BSmQEg,8563
-x_transformers-1.37.3.dist-info/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
-x_transformers-1.37.3.dist-info/METADATA,sha256=SIGTCQMrLkyq_aksJAst0iXw9VfFT6QWlGvtUElbTMg,661
-x_transformers-1.37.3.dist-info/WHEEL,sha256=GV9aMThwP_4oNCtvEC2ec3qUYutgWeAzklro_0m4WJQ,91
-x_transformers-1.37.3.dist-info/top_level.txt,sha256=hO6KGpFuGucRNEtRfme4A_rGcM53AKwGP7RVlRIxS5Q,15
-x_transformers-1.37.3.dist-info/RECORD,,