Fix number of quantisation buckets (#182)

Fixes https://github.com/amazon-science/chronos-forecasting/issues/181.

Chronos' tokenizer has a vocabulary size of `n_tokens`. Among these,
there are `n_special_tokens` reserved for EOS, PAD, etc. and `n_tokens -
n_special_tokens` allocated to numerical values. However, the provided
`MeanScaleUniformBins` tokenizer creates` n_tokens - n_special_tokens +
1` different buckets, resulting in a total of `n_tokens + 1` possible
tokens. This causes training and inference errors when one of the data
points gets allocated to the largest bucket, as the model requires 0 <=
token_id < n_tokens.

This PR modifies the `MeanScaleUniformBins` tokenizer, so that it
creates one less bucket for numerical values.

---

By submitting this pull request, I confirm that you can use, modify,
copy, and redistribute this contribution, under the terms of your
choice.

---------

Co-authored-by: Lorenzo Stella <lorenzostella@gmail.com>

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "chronos"
-version = "1.2.0"
+version = "1.2.1"
 requires-python = ">=3.8"
 license = { file = "LICENSE" }
 dependencies = [

src/chronos/chronos.py

@@ -5,7 +5,6 @@ import warnings
 from dataclasses import dataclass
 from typing import Any, Dict, List, Literal, Optional, Tuple, Union
 
-import chronos
 import torch
 import torch.nn as nn
 from transformers import (
@@ -16,6 +15,8 @@ from transformers import (
     PreTrainedModel,
 )
 
+import chronos
+
 
 @dataclass
 class ChronosConfig:
@@ -187,6 +188,9 @@ class MeanScaleUniformBins(ChronosTokenizer):
             )
             + self.config.n_special_tokens
         )
+
+        token_ids.clamp_(0, self.config.n_tokens - 1)
+
         token_ids[~attention_mask] = self.config.pad_token_id
 
         return token_ids, attention_mask, scale

test/test_chronos.py

@@ -4,8 +4,8 @@
 from pathlib import Path
 from typing import Tuple
 
-import torch
 import pytest
+import torch
 
 from chronos import ChronosConfig, ChronosPipeline, MeanScaleUniformBins
 
@@ -244,3 +244,59 @@ def test_pipeline_embed(torch_dtype: str):
     embedding, scale = pipeline.embed(context[0, ...])
     validate_tensor(embedding, (1, expected_embed_length, d_model))
     validate_tensor(scale, (1,))
+
+
+@pytest.mark.parametrize("n_tokens", [10, 1000, 10000])
+def test_tokenizer_number_of_buckets(n_tokens):
+    config = ChronosConfig(
+        tokenizer_class="MeanScaleUniformBins",
+        tokenizer_kwargs=dict(low_limit=-1.0, high_limit=1.0),
+        n_tokens=n_tokens,
+        n_special_tokens=2,
+        pad_token_id=0,
+        eos_token_id=1,
+        use_eos_token=True,
+        model_type="seq2seq",
+        context_length=512,
+        prediction_length=64,
+        num_samples=20,
+        temperature=1.0,
+        top_k=50,
+        top_p=1.0,
+    )
+    tokenizer = config.create_tokenizer()
+
+    n_numerical_tokens = config.n_tokens - config.n_special_tokens
+
+    # The tokenizer has one bucket too many as a result of an early bug. In order to
+    # keep consistent with the original trained models, this is kept as it is. However,
+    # token ids are clipped to a maximum of `n_tokens - 1` to avoid out-of-bounds errors.
+    assert len(tokenizer.centers) == (n_numerical_tokens - 1)
+    assert len(tokenizer.boundaries) == n_numerical_tokens
+
+
+@pytest.mark.parametrize("n_tokens", [10, 1000, 10000])
+def test_token_clipping(n_tokens):
+    config = ChronosConfig(
+        tokenizer_class="MeanScaleUniformBins",
+        tokenizer_kwargs={"low_limit": -15, "high_limit": 15},
+        n_tokens=n_tokens,
+        n_special_tokens=2,
+        pad_token_id=0,
+        eos_token_id=1,
+        use_eos_token=True,
+        model_type="seq2seq",
+        context_length=512,
+        prediction_length=64,
+        num_samples=20,
+        temperature=1.0,
+        top_k=50,
+        top_p=1.0,
+    )
+    tokenizer = config.create_tokenizer()
+
+    huge_value = 1e22  # this large value is assigned to the largest bucket
+    token_ids, _, _ = tokenizer._input_transform(
+        context=torch.tensor([[huge_value]]), scale=torch.tensor(([1]))
+    )
+    assert token_ids[0, 0] == config.n_tokens - 1  # and it's clipped to n_tokens - 1