x-transformers 1.42.9__py3-none-any.whl → 1.42.11__py3-none-any.whl

x_transformers/autoregressive_wrapper.py

@@ -48,7 +48,7 @@ def align_right(t, lens, pad_id = 0):
     batch_arange = torch.arange(batch, device = device, dtype = torch.long)[..., None]
     prompt_len_arange = torch.arange(seq_len, device = device, dtype = torch.long)
 
-    t = F.pad(t, (max_pad_len, 0), value = 0)
+    t = F.pad(t, (max_pad_len, 0), value = pad_id)
     offset = max_pad_len - pad_lens
 
     aligned = t[batch_arange, prompt_len_arange + offset[..., None]]

x_transformers/continuous.py

@@ -2,7 +2,8 @@ import torch
 from torch import nn
 import torch.nn.functional as F
 
-from einops import pack, repeat, unpack
+import einx
+from einops import reduce, pack, repeat, unpack
 
 from x_transformers.x_transformers import (
     AttentionLayers,
@@ -24,6 +25,15 @@ def default(val, d):
         return val
     return d() if callable(d) else d
 
+def masked_mean(t, mask):
+    t = einx.where('b n, b n d, -> b n d', mask, t, 0.)
+
+    num = reduce(t, 'b n d -> b', 'sum')
+    den = mask.sum(dim = -1)
+
+    masked_average = num / den.clamp(min = 1.)
+    return masked_average
+
 # main classes
 
 class ContinuousTransformerWrapper(nn.Module):
@@ -169,12 +179,15 @@ class ContinuousAutoregressiveWrapper(nn.Module):
         net: ContinuousTransformerWrapper,
         ignore_index = -100,
         pad_value = 0,
-        loss_fn = nn.MSELoss(reduction = 'none')
+        loss_fn = nn.MSELoss(reduction = 'none'),
+        equal_loss_weight_batch = False  # setting this to True, if the mask is passed in and sequences are variable in length, each sequence will be weighted the same (as opposed to each token)
     ):
         super().__init__()
         self.net = net
         self.max_seq_len = net.max_seq_len
+
         self.loss_fn = loss_fn
+        self.equal_loss_weight_batch = equal_loss_weight_batch
 
     @torch.no_grad()
     def generate(self, start_tokens, seq_len, **kwargs):
@@ -222,6 +235,10 @@ class ContinuousAutoregressiveWrapper(nn.Module):
 
         if exists(mask):
             assert loss.ndim > 1, 'loss should not be reduced if mask is passed in'
-            loss = loss[mask]
+
+            if self.equal_loss_weight_batch:
+                loss = masked_mean(loss, mask)
+            else:
+                loss = loss[mask]
 
         return loss.mean()

{x_transformers-1.42.9.dist-info → x_transformers-1.42.11.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: x-transformers
-Version: 1.42.9
+Version: 1.42.11
 Summary: X-Transformers - Pytorch
 Home-page: https://github.com/lucidrains/x-transformers
 Author: Phil Wang

{x_transformers-1.42.9.dist-info → x_transformers-1.42.11.dist-info}/RECORD

@@ -1,7 +1,7 @@
 x_transformers/__init__.py,sha256=l0dom8ZYkRzFvnDdgzDboXqrI1tKav3beVE7TN2nHko,844
 x_transformers/attend.py,sha256=SdWlV8Vp5DtpsOzAd0LRhm4VGrJf0lJCGiV2_j_CtoA,17284
-x_transformers/autoregressive_wrapper.py,sha256=DOJJCMMDOqDYKWy_IaG5IyKsXD3AW6amzfUgdAADOLY,10500
-x_transformers/continuous.py,sha256=cIVEdhfei258__ziV7kQBrJMxCel54bExBTDrO9rfCI,6450
+x_transformers/autoregressive_wrapper.py,sha256=reLCno9Z9pchVU79tBF8OMo21LwSZ67KAeB83jqkyAc,10505
+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
 x_transformers/neo_mlp.py,sha256=XCNnnop9WLarcxap1kGuYc1x8GHvwkZiDRnXOxSl3Po,3452
@@ -9,8 +9,8 @@ x_transformers/nonautoregressive_wrapper.py,sha256=2NU58hYMgn-4Jzg3mie-mXb0XH_dC
 x_transformers/x_transformers.py,sha256=VxdA44EYQhVH1Rp7wreJ83I2e0Ea7VN_bFRE-iDXOI8,93833
 x_transformers/xl_autoregressive_wrapper.py,sha256=CvZMJ6A6PA-Y_bQAhnORwjJBSl6Vjq2IdW5KTdk8NI8,4195
 x_transformers/xval.py,sha256=7S00kCuab4tWQa-vf-z-XfzADjVj48MoFIr7VSIvttg,8575
-x_transformers-1.42.9.dist-info/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
-x_transformers-1.42.9.dist-info/METADATA,sha256=k9r-D0b0xnf8gwE-SwwgybnfQpoRwiY0wthOn66xc6Y,689
-x_transformers-1.42.9.dist-info/WHEEL,sha256=P9jw-gEje8ByB7_hXoICnHtVCrEwMQh-630tKvQWehc,91
-x_transformers-1.42.9.dist-info/top_level.txt,sha256=hO6KGpFuGucRNEtRfme4A_rGcM53AKwGP7RVlRIxS5Q,15
-x_transformers-1.42.9.dist-info/RECORD,,
+x_transformers-1.42.11.dist-info/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
+x_transformers-1.42.11.dist-info/METADATA,sha256=0dlrRj5RehRfEhgK7M4ESmaNHuthe912XQiC7Hsim_8,690
+x_transformers-1.42.11.dist-info/WHEEL,sha256=P9jw-gEje8ByB7_hXoICnHtVCrEwMQh-630tKvQWehc,91
+x_transformers-1.42.11.dist-info/top_level.txt,sha256=hO6KGpFuGucRNEtRfme4A_rGcM53AKwGP7RVlRIxS5Q,15
+x_transformers-1.42.11.dist-info/RECORD,,