train model

alr_transformer/__init__.py

@@ -1,2 +1,4 @@
 from alr_transformer.model import ALRTransformer
 from alr_transformer.alr_block import ALRBlock
+
+__all__ = ['ALRTransformer', 'ALRBlock']

alr_transformer/at.py

@@ -0,0 +1,83 @@
+import torch
+import torch.nn.functional as F
+from einops import rearrange
+from torch import nn
+
+# helper function
+
+
+def exists(val):
+    return val is not None
+
+
+def eval_decorator(fn):
+    def inner(model, *args, **kwargs):
+        was_training = model.training
+        model.eval()
+        out = fn(model, *args, **kwargs)
+        model.train(was_training)
+        return out
+
+    return inner
+
+
+# top k filtering
+
+
+def top_k(logits, thres=0.9):
+    k = int((1 - thres) * logits.shape[-1])
+    val, ind = torch.topk(logits, k)
+    probs = torch.full_like(logits, float("-inf"))
+    probs.scatter_(1, ind, val)
+    return probs
+
+
+class AutoregressiveWrapper(nn.Module):
+    def __init__(self, net, max_seq_len=2048, pad_value=0):
+        super().__init__()
+        self.max_seq_len = max_seq_len
+        self.pad_value = pad_value
+        self.net = net
+
+    @torch.no_grad()
+    @eval_decorator
+    def generate(
+        self,
+        start_tokens,
+        seq_len,
+        eos_token=None,
+        temperature=1.0,
+        filter_thres=0.9,
+        **kwargs
+    ):
+        b, t, device = *start_tokens.shape, start_tokens.device
+
+        out = start_tokens
+
+        for _ in range(seq_len):
+            logits = self.net(out, **kwargs)[:, -1, :]
+
+            filtered_logits = top_k(logits, thres=filter_thres)
+            probs = F.softmax(filtered_logits / temperature, dim=-1)
+
+            sample = torch.multinomial(probs, 1)
+
+            out = torch.cat((out, sample), dim=-1)
+
+            if exists(eos_token):
+                is_eos_token = out == eos_token
+
+                if is_eos_token.any(dim=-1).all():
+                    # mask out everything after the eos tokens
+                    shifted_is_eos_tokens = F.pad(is_eos_token, (1, -1))
+                    mask = shifted_is_eos_tokens.float().cumsum(dim=-1) >= 1
+                    out = out.masked_fill(mask, self.pad_value)
+                    break
+
+        out = out[:, t:]
+        return out
+
+    def forward(self, x, **kwargs):
+        x_inp, x_labels = x[:, :-1], x[:, 1:]
+        logits = self.net(x_inp, **kwargs)
+        return F.cross_entropy(rearrange(logits, "b c n -> b n c"), x_labels)

train.py

@@ -0,0 +1,109 @@
+import gzip
+import random
+
+import numpy as np
+import torch
+import torch.optim as optim
+import tqdm
+from torch.nn import functional as F
+from torch.utils.data import DataLoader, Dataset
+
+from alr_transformer.at import AutoregressiveWrapper
+from alr_transformer.model import ALRTransformer
+
+# constants
+
+NUM_BATCHES = int(1e5)
+BATCH_SIZE = 4
+GRADIENT_ACCUMULATE_EVERY = 4
+LEARNING_RATE = 2e-4
+VALIDATE_EVERY = 100
+GENERATE_EVERY = 500
+GENERATE_LENGTH = 512
+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 = ALRTransformer(num_tokens=256, dim=512, depth=8)
+
+model = AutoregressiveWrapper(model, max_seq_len=SEQ_LEN)
+model.cuda()
+
+# prepare enwik8 data
+
+with gzip.open("./data/enwik8.gz") as file:
+    X = np.fromstring(file.read(int(95e6)), dtype=np.uint8)
+    trX, vaX = np.split(X, [int(90e6)])
+    data_train, data_val = torch.from_numpy(trX), torch.from_numpy(vaX)
+
+
+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,))
+        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))
+val_loader = cycle(DataLoader(val_dataset, batch_size=BATCH_SIZE))
+
+# optimizer
+
+optim = torch.optim.Adam(model.parameters(), lr=LEARNING_RATE)
+
+# training
+
+for i in tqdm.tqdm(range(NUM_BATCHES), mininterval=10.0, desc="training"):
+    model.train()
+
+    for __ in range(GRADIENT_ACCUMULATE_EVERY):
+        loss = model(next(train_loader))
+        loss.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 = model(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 = model.generate(inp[None, ...], GENERATE_LENGTH)
+        output_str = decode_tokens(sample[0])
+        print(output_str)