File size: 9,653 Bytes
de729bb
393ff84
de729bb
 
393ff84
de729bb
 
 
 
 
 
393ff84
de729bb
 
 
 
 
 
393ff84
 
 
de729bb
 
393ff84
 
de729bb
393ff84
 
de729bb
393ff84
 
 
de729bb
 
 
 
 
393ff84
 
 
 
 
de729bb
 
393ff84
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
de729bb
 
 
 
393ff84
de729bb
393ff84
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
de729bb
 
 
393ff84
de729bb
393ff84
 
 
 
 
 
 
 
 
 
 
de729bb
393ff84
 
 
de729bb
393ff84
 
 
 
de729bb
 
393ff84
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
de729bb
 
 
 
393ff84
 
 
 
de729bb
 
393ff84
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
de729bb
393ff84
 
de729bb
 
 
393ff84
 
 
 
de729bb
393ff84
de729bb
393ff84
de729bb
 
 
 
 
 
 
 
29318fb
de729bb
393ff84
 
 
 
 
 
de729bb
 
393ff84
 
de729bb
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
import logging
from typing import ClassVar, List, Optional

import numpy as np
import pywt
from tokenizers import ByteLevelBPETokenizer
from tokenizers.trainers import BpeTrainer
from transformers import PreTrainedTokenizerFast
from transformers.processing_utils import ProcessorMixin


class WaveletActionProcessor(ProcessorMixin):
    attributes: ClassVar[list[str]] = ["bpe_tokenizer"]
    bpe_tokenizer_class: str = "AutoTokenizer"

    def __init__(
        self,
        bpe_tokenizer: PreTrainedTokenizerFast,
        wavelet: str = "db1",
        level: int = 2,
        scale: float = 10.0,
        min_token: int = 0,
        *,
        action_dim: Optional[int] = None,
        time_horizon: Optional[int] = None,
    ):
        self.wavelet = wavelet
        self.level = level
        self.scale = scale
        self.min_token = int(min_token)

        # Used for decode (same logic as FAST)
        self.time_horizon = time_horizon
        self.action_dim = action_dim
        self.called_time_horizon = time_horizon
        self.called_action_dim = action_dim

        # Cache wavelet coefficient layout needed for decoding
        # We keep one slice-structure per dimension (they are typically identical for fixed T/wavelet/level)
        self._coeff_slices_per_dim = None  # list of slice dicts
        self._n_coeff = None               # number of wavelet coeffs per dim after coeffs_to_array

        super().__init__(bpe_tokenizer)

    def _ensure_coeff_layout(self, T: int, D: int):
        """Cache coeff slices and coeff vector length for given (T, wavelet, level)."""
        if (
            self._coeff_slices_per_dim is not None
            and self._n_coeff is not None
            and self.called_time_horizon == T
            and self.called_action_dim == D
        ):
            return

        dummy = np.zeros(T, dtype=np.float32)

        slices_per_dim = []
        n_coeff = None
        for _ in range(D):
            coeffs = pywt.wavedec(dummy, self.wavelet, level=self.level)
            arr, slc = pywt.coeffs_to_array(coeffs)
            slices_per_dim.append(slc)
            if n_coeff is None:
                n_coeff = int(arr.shape[0])

        self._coeff_slices_per_dim = slices_per_dim
        self._n_coeff = n_coeff

    def __call__(self, action_chunk: np.ndarray) -> List[List[int]]:
        """
        Encode actions to BPE tokens.

        action_chunk: (T,D) or (B,T,D)
        returns: List[List[int]] (batch of token id lists)
        """
        assert action_chunk.ndim <= 3, "Only 3 dimensions supported: [batch, timesteps, action_dim]"
        if action_chunk.ndim == 2:
            action_chunk = action_chunk[None, ...]

        B, T, D = action_chunk.shape

        # cache for decoding
        self.called_time_horizon, self.called_action_dim = T, D
        self._ensure_coeff_layout(T, D)

        batch_tokens: List[List[int]] = []
        for i in range(B):
            # wavelet per dim -> flattened coeffs of length (n_coeff * D)
            coeffs_by_dim = []
            for d in range(D):
                coeffs = pywt.wavedec(action_chunk[i, :, d], self.wavelet, level=self.level)
                flat, _ = pywt.coeffs_to_array(coeffs)  # shape (n_coeff,)
                coeffs_by_dim.append(flat)

            coeff_mat = np.stack(coeffs_by_dim, axis=1)      # (n_coeff, D)
            flat_all = coeff_mat.reshape(-1)                 # (n_coeff * D,)

            quant = np.around(flat_all * self.scale).astype(int)

            shifted = (quant - self.min_token).astype(int)

            # Optional safety check (unicode range). Keep it simple:
            if shifted.min() < 0:
                # This means min_token was not low enough for these coeffs.
                raise ValueError(
                    f"Shifted tokens became negative (min={shifted.min()}). "
                    f"Your min_token={self.min_token} is too high. Re-fit or lower min_token."
                )
            if shifted.max() > 0x10FFFF:
                raise ValueError(
                    f"Shifted tokens exceed Unicode max (max={shifted.max()}). "
                    f"Reduce scale or re-fit min/max range."
                )

            token_str = "".join(chr(int(x)) for x in shifted)
            batch_tokens.append(self.bpe_tokenizer(token_str)["input_ids"])

        return batch_tokens

    def decode(
        self,
        tokens: List[List[int]],
        *,
        time_horizon: Optional[int] = None,
        action_dim: Optional[int] = None,
    ) -> np.ndarray:
        """
        Decode BPE tokens back to actions.

        tokens: List[List[int]] (batch)
        returns: (B, T, D)
        """
        T = time_horizon or self.time_horizon or self.called_time_horizon
        D = action_dim or self.action_dim or self.called_action_dim

        assert T is not None and D is not None, (
            "Tokenizer not initialized: call encode() once or pass time_horizon and action_dim."
        )

        # cache for next call + ensure layout
        self.time_horizon, self.action_dim = T, D
        self.called_time_horizon, self.called_action_dim = T, D
        self._ensure_coeff_layout(T, D)

        decoded_actions = []
        for tok_list in tokens:
            # decode to string of chars
            s = self.bpe_tokenizer.decode(tok_list, clean_up_tokenization_spaces=False)

            ints = np.array([ord(c) for c in s], dtype=np.int64)

            # unshift + dequantize
            quant = ints + self.min_token
            flat_coeffs = quant.astype(np.float32) / self.scale  # (n_coeff * D,)

            # reshape to (n_coeff, D)
            expected = self._n_coeff * D
            if flat_coeffs.shape[0] != expected:
                raise ValueError(
                    f"Decoded coeff length mismatch: got {flat_coeffs.shape[0]}, expected {expected}. "
                    f"(T={T}, D={D}, n_coeff={self._n_coeff}). "
                    "This usually means you decoded with different T/D than encoding."
                )

            coeff_mat = flat_coeffs.reshape(self._n_coeff, D)

            # inverse wavelet per dimension
            recon = np.zeros((T, D), dtype=np.float32)
            for d in range(D):
                arr = coeff_mat[:, d]
                coeff_list = pywt.array_to_coeffs(
                    arr,
                    self._coeff_slices_per_dim[d],
                    output_format="wavedec",
                )
                sig = pywt.waverec(coeff_list, self.wavelet)
                recon[:, d] = sig[:T]  # waverec can return a bit longer due to padding

            decoded_actions.append(recon)

        return np.stack(decoded_actions, axis=0)

    @classmethod
    def fit(
        cls,
        action_data: List[np.ndarray],  # each (T,D)
        wavelet: str = "db1",
        level: int = 2,
        scale: float = 10.0,
        vocab_size: int = 1024,
        *,
        time_horizon: Optional[int] = None,
        action_dim: Optional[int] = None,
    ) -> "WaveletActionProcessor":
        """
        Fit BPE tokenizer on wavelet-quantized coefficient streams.
        """

        # Compute quantized coefficient streams to estimate min/max token range
        all_streams = []
        for a in action_data:
            assert a.ndim == 2, "Each item must be (T,D)"
            T, D = a.shape
            # wavelet per dim -> flatten (n_coeff * D)
            coeffs_by_dim = []
            for d in range(D):
                coeffs = pywt.wavedec(a[:, d], wavelet, level=level)
                flat, _ = pywt.coeffs_to_array(coeffs)
                coeffs_by_dim.append(flat)
            coeff_mat = np.stack(coeffs_by_dim, axis=1)
            stream = np.around(coeff_mat.reshape(-1) * scale).astype(int)
            all_streams.append(stream)

        all_vals = np.concatenate(all_streams)
        min_token = int(all_vals.min())
        max_token = int(all_vals.max())

        token_range = max_token - min_token + 1
        if token_range > vocab_size:
            raise ValueError(
                f"Vocab size {vocab_size} too small for token range {token_range}. "
                "Increase vocab_size or reduce scale."
            )
        if token_range + 100 > vocab_size:
            logging.warning(
                f"Initial alphabet size {token_range} is close to vocab_size {vocab_size}. "
                "Consider increasing vocab_size for better BPE merges."
            )

        def _token_iter():
            for stream in all_streams:
                shifted = (stream - min_token).astype(int)
                # no clamp; must be >=0
                yield "".join(chr(int(x)) for x in shifted)

        # Train BPE
        bpe = ByteLevelBPETokenizer()
        alphabet = [chr(i) for i in range(token_range)]
        trainer = BpeTrainer(
            vocab_size=vocab_size,
            min_frequency=2,
            show_progress=True,
            special_tokens=[],
            initial_alphabet=alphabet,
            max_token_length=10000,
        )
        bpe._tokenizer.train_from_iterator(_token_iter(), trainer=trainer)

        # infer T/D defaults if not provided
        if time_horizon is None:
            time_horizon = int(action_data[0].shape[0])
        if action_dim is None:
            action_dim = int(action_data[0].shape[1])

        return cls(
            PreTrainedTokenizerFast(tokenizer_object=bpe, clean_up_tokenization_spaces=False),
            wavelet=wavelet,
            level=level,
            scale=scale,
            min_token=min_token,
            time_horizon=time_horizon,
            action_dim=action_dim,
        )