x-transformers 2.8.4__tar.gz → 2.9.1__tar.gz

{x_transformers-2.8.4 → x_transformers-2.9.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: x-transformers
-Version: 2.8.4
+Version: 2.9.1
 Summary: X-Transformers
 Project-URL: Homepage, https://pypi.org/project/x-transformers/
 Project-URL: Repository, https://github.com/lucidrains/x-transformers
@@ -40,6 +40,7 @@ Requires-Dist: loguru
 Requires-Dist: packaging>=21.0
 Requires-Dist: torch>=2.0
 Provides-Extra: examples
+Requires-Dist: adam-atan2-pytorch>=0.2.2; extra == 'examples'
 Requires-Dist: lion-pytorch; extra == 'examples'
 Requires-Dist: tqdm; extra == 'examples'
 Provides-Extra: test
@@ -2573,4 +2574,16 @@ ids_out, num_out, is_number_mask = model.generate(start_ids, start_nums, 17)
 }
 ```
 
+```bibtex
+@misc{yan2017hierarchicalmultiscaleattentionnetworks,
+    title   = {Hierarchical Multi-scale Attention Networks for Action Recognition}, 
+    author  = {Shiyang Yan and Jeremy S. Smith and Wenjin Lu and Bailing Zhang},
+    year    = {2017},
+    eprint  = {1708.07590},
+    archivePrefix = {arXiv},
+    primaryClass = {cs.CV},
+    url     = {https://arxiv.org/abs/1708.07590}, 
+}
+```
+
 *solve intelligence... then use that to solve everything else.* - Demis Hassabis

{x_transformers-2.8.4 → x_transformers-2.9.1}/README.md

@@ -2525,4 +2525,16 @@ ids_out, num_out, is_number_mask = model.generate(start_ids, start_nums, 17)
 }
 ```
 
+```bibtex
+@misc{yan2017hierarchicalmultiscaleattentionnetworks,
+    title   = {Hierarchical Multi-scale Attention Networks for Action Recognition}, 
+    author  = {Shiyang Yan and Jeremy S. Smith and Wenjin Lu and Bailing Zhang},
+    year    = {2017},
+    eprint  = {1708.07590},
+    archivePrefix = {arXiv},
+    primaryClass = {cs.CV},
+    url     = {https://arxiv.org/abs/1708.07590}, 
+}
+```
+
 *solve intelligence... then use that to solve everything else.* - Demis Hassabis

{x_transformers-2.8.4 → x_transformers-2.9.1}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "x-transformers"
-version = "2.8.4"
+version = "2.9.1"
 description = "X-Transformers"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }
@@ -35,6 +35,7 @@ Repository = "https://github.com/lucidrains/x-transformers"
 
 [project.optional-dependencies]
 examples = [
+    "adam-atan2-pytorch>=0.2.2",
     "lion-pytorch",
     "tqdm",
 ]

{x_transformers-2.8.4 → x_transformers-2.9.1}/tests/test_x_transformers.py

@@ -1373,3 +1373,11 @@ def test_muon_params():
 
     enc = Encoder(dim = 512, depth = 2)
     assert len(enc.muon_parameters()) > 0
+
+def test_stochastic_attn():
+    from x_transformers import Attention
+
+    attn = Attention(dim = 512, gumbel_softmax = True)
+    out = attn(torch.randn(1, 1024, 512))
+
+    assert out.shape == (1, 1024, 512)

x_transformers-2.9.1/train_with_muon.py

@@ -0,0 +1,132 @@
+# /// script
+# dependencies = [
+#   "x-transformers",
+#   "adam-atan2-pytorch>=0.2.2",
+# ]
+# ///
+
+from x_transformers import TransformerWrapper, Decoder
+from x_transformers.autoregressive_wrapper import AutoregressiveWrapper
+
+import random
+import tqdm
+import gzip
+import numpy as np
+import torch
+import torch.optim as optim
+from torch.nn import functional as F
+from torch.utils.data import DataLoader, Dataset
+
+from adam_atan2_pytorch import MuonAdamAtan2
+
+# constants
+
+NUM_BATCHES = int(1e5)
+BATCH_SIZE = 4
+GRADIENT_ACCUMULATE_EVERY = 4
+LEARNING_RATE = 1e-4
+MUON_LEARNING_RATE = 1e-3
+VALIDATE_EVERY  = 100
+GENERATE_EVERY  = 500
+GENERATE_LENGTH = 1024
+SEQ_LEN = 1024
+
+# helpers
+
+def cycle(loader):
+    while True:
+        for data in loader:
+            yield data
+
+def decode_token(token):
+    return str(chr(max(32, token)))
+
+def decode_tokens(tokens):
+    return ''.join(list(map(decode_token, tokens)))
+
+# instantiate GPT-like decoder model
+
+model = TransformerWrapper(
+    num_tokens = 256,
+    max_seq_len = SEQ_LEN,
+    attn_layers = Decoder(
+        dim = 512,
+        depth = 6,
+        heads = 8,
+        rotary_pos_emb = True
+    )
+)
+
+ar_wrapper = AutoregressiveWrapper(model)
+model.cuda()
+
+# prepare enwik8 data
+
+with gzip.open('./data/enwik8.gz') as file:
+    data = np.frombuffer(file.read(int(95e6)), dtype=np.uint8).copy()
+    train_x, valid_x = np.split(data, [int(90e6)])
+    data_train, data_val = torch.from_numpy(train_x), torch.from_numpy(valid_x)
+
+class TextSamplerDataset(Dataset):
+    def __init__(self, data, seq_len):
+        super().__init__()
+        self.data = data
+        self.seq_len = seq_len
+
+    def __getitem__(self, index):
+        rand_start = torch.randint(0, self.data.size(0) - self.seq_len - 1, (1,))
+        full_seq = self.data[rand_start: rand_start + self.seq_len + 1].long()
+        return full_seq.cuda()
+
+    def __len__(self):
+        return self.data.size(0) // self.seq_len
+
+train_dataset = TextSamplerDataset(data_train, SEQ_LEN)
+val_dataset   = TextSamplerDataset(data_val, SEQ_LEN)
+train_loader  = cycle(DataLoader(train_dataset, batch_size = BATCH_SIZE, drop_last = True))
+val_loader    = cycle(DataLoader(val_dataset, batch_size = BATCH_SIZE, drop_last = True))
+
+# optimizer
+
+optim = MuonAdamAtan2(
+    muon_params = model.muon_parameters(),
+    params = model.parameters(),
+    remove_muon_params_from_params = True,
+    lr = LEARNING_RATE,
+    muon_lr = MUON_LEARNING_RATE,
+)
+
+# training
+
+for i in tqdm.tqdm(range(NUM_BATCHES), mininterval=10., desc='training'):
+    model.train()
+
+    for __ in range(GRADIENT_ACCUMULATE_EVERY):
+        loss = ar_wrapper(next(train_loader))
+        (loss / GRADIENT_ACCUMULATE_EVERY).backward()
+
+    print(f'training loss: {loss.item()}')
+    torch.nn.utils.clip_grad_norm_(model.parameters(), 0.5)
+    optim.step()
+    optim.zero_grad()
+
+    if i % VALIDATE_EVERY == 0:
+        model.eval()
+        with torch.no_grad():
+            loss = ar_wrapper(next(val_loader))
+            print(f'validation loss: {loss.item()}')
+
+    if i % GENERATE_EVERY == 0:
+        model.eval()
+        inp = random.choice(val_dataset)[:-1]
+        prime = decode_tokens(inp)
+        print(f'%s \n\n %s', (prime, '*' * 100))
+
+        sample = ar_wrapper.generate(
+            prompts = inp,
+            seq_len = GENERATE_LENGTH,
+            cache_kv = True
+        )
+
+        output_str = decode_tokens(sample)
+        print(output_str)

{x_transformers-2.8.4 → x_transformers-2.9.1}/x_transformers/attend.py

@@ -171,6 +171,9 @@ class Attend(Module):
         qk_norm = False,
         l2_distance = False,
         sigmoid = False,
+        gumbel_softmax = False,
+        gumbel_softmax_temp = 1.,
+        gumbel_softmax_hard = True,
         custom_attn_fn: Callable | None = None,
         flash = False,
         softclamp_logits = False,
@@ -203,7 +206,7 @@ class Attend(Module):
         assert not (flash and hard), 'hard attention not available for flash'
         assert not (flash and is_sparse_topk_attn), 'topk attention not available for flash'
 
-        assert at_most_one_of(sigmoid, hard, l2_distance, is_sparse_topk_attn)
+        assert at_most_one_of(sigmoid, hard, l2_distance, gumbel_softmax, is_sparse_topk_attn)
 
         if exists(custom_attn_fn):
             self.attn_fn = custom_attn_fn
@@ -213,6 +216,8 @@ class Attend(Module):
             self.attn_fn = one_hot_straight_through
         elif is_sparse_topk_attn:
             self.attn_fn = partial(sparse_topk_attn, sparse_topk = sparse_topk, straight_through = sparse_topk_straight_through)
+        elif gumbel_softmax:
+            self.attn_fn = partial(F.gumbel_softmax, dim = -1, tau = gumbel_softmax_temp, hard = gumbel_softmax_hard)
         else:
             softmax_fn = partial(F.softmax, dim = -1)
             self.attn_fn = partial(softmax_fn, dtype = torch.float32) if not qk_norm else softmax_fn

{x_transformers-2.8.4 → x_transformers-2.9.1}/x_transformers/x_transformers.py

@@ -1336,6 +1336,9 @@ class Attention(Module):
         value_rmsnorm = False,      # used in alphagenome and bytedance's GR3 for further stability
         l2_distance = False,
         sigmoid = False,
+        gumbel_softmax = False,
+        gumbel_softmax_temp = 1.,
+        gumbel_softmax_hard = True,
         selective = False,
         custom_attn_fn: Callable | None = None,
         hybrid_module: Module | None = None,
@@ -1541,6 +1544,9 @@ class Attention(Module):
             scale = qk_norm_scale if qk_norm else self.scale,
             l2_distance = l2_distance,
             sigmoid = sigmoid,
+            gumbel_softmax = gumbel_softmax,
+            gumbel_softmax_temp = gumbel_softmax_temp,
+            gumbel_softmax_hard = gumbel_softmax_hard,
             selective = selective,
             custom_attn_fn = custom_attn_fn,
             add_zero_kv = add_zero_kv,