File size: 11,537 Bytes
229a3ba
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c864132
229a3ba
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
import torch
import numpy as np
import bitarray

from pytorch_transformers import GPT2LMHeadModel, GPT2Tokenizer

def decode(self, token_ids, **kwargs):
    filtered_tokens = self.convert_ids_to_tokens(token_ids)
    text = self.convert_tokens_to_string(filtered_tokens)
    return text
GPT2Tokenizer.decode = decode

def _convert_token_to_id(self, token):
    return self.encoder.get(token, 0)
GPT2Tokenizer._convert_token_to_id = _convert_token_to_id


def limit_past(past):
    past = list(past)
    for i in range(len(past)):
        past[i] = past[i][:, :, :, -1022:]
    return past

def kl(q, logq, logp):
    res = q*(logq-logp)/0.69315
    res[q==0] = 0
    return res.sum().item() # in bits

def entropy(q, logq):
    res = q*logq/0.69315
    res[q==0] = 0
    return -res.sum().item() # in bits

# e.g. [0, 1, 1, 1] looks like 1110=14
def bits2int(bits):
    res = 0
    for i, bit in enumerate(bits):
        res += bit*(2**i)
    return res

def int2bits(inp, num_bits):
    if num_bits == 0:
        return []
    strlist = ('{0:0%db}'%num_bits).format(inp)
    return [int(strval) for strval in reversed(strlist)]

def is_sent_finish(token_idx, enc):
    token = enc.decoder[token_idx]
    return '.' in token or '!' in token or '?' in token

def num_same_from_beg(bits1, bits2):
    assert len(bits1) == len(bits2)
    for i in range(len(bits1)):
        if bits1[i] != bits2[i]:
            break

    return i

def encode_context(raw_text, enc):
    context_tokens = [enc.encoder['<|endoftext|>']] + enc.encode(raw_text)
    return context_tokens

# Use gpt2-medium for 345M param model
# Use gpt2-large for 774M param model
def get_model(seed=1234, model_name='gpt2'):
    np.random.seed(seed)
    torch.random.manual_seed(seed)
    torch.cuda.manual_seed(seed)
    device = torch.device("cpu")

    enc = GPT2Tokenizer.from_pretrained(model_name)
    enc.unk_token = None
    enc.bos_token = None
    enc.eos_token = None
    
    model = GPT2LMHeadModel.from_pretrained(model_name)
    model.to(device)
    model.eval()
    #model.double()

    return enc, model

enc32_itoc = ['\0', 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', '.', ',', "'", '!', ' ']
enc32_ctoi = {k: v for v, k in enumerate(enc32_itoc)}
def enc32(text):
    bits = []
    for c in text:
        bits.extend(int2bits(enc32_ctoi[c], 5))
    return bits

def dec32(bits):
    text = ''
    for i in range(0, len(bits), 5):
        c = enc32_itoc[bits2int(bits[i:i+5])]
        if c == '\0':
            break
        text += c
    return text

# message should be bit string
# encoded should be text string
def expansion_ratio(message, encoded):
    message_bits = len(message)
    encoded_ba = bitarray.bitarray()
    encoded_ba.frombytes(encoded.encode('utf-8'))
    encoded_bits = len(encoded_ba.tolist())
    return encoded_bits/message_bits

#@title
import torch
import math
import random

def bin_sort(l, token_indices, total, entropy, device):
    #compute entropy for upper bound on the number of bins we need

    bucket_size = total
    num_bins = 2**int(entropy+1)
    bucket_size = total / num_bins

    bins = [torch.empty(0, dtype=torch.long, device=device)] * num_bins
    value_in_bins = [0] * num_bins
    space_left_after = [total - i*bucket_size for i in range(0,num_bins)]


    token_bins = [torch.empty(0, dtype=torch.long, device=device)] * num_bins

    # Figuring out what the search order should be
    step_size = num_bins/4
    search_order = []
    priorities = [0]*num_bins
    priority = 0
    search_order.append(int(num_bins/2))
    search_order.append(0)
    priorities[int(num_bins/2)] = 0
    priorities[0] = 0
    while(step_size>=1):
        priority += 1
        for x in range(num_bins-int(step_size), -1, -int(step_size*2)):
            search_order.append(x)
            priorities[x] = priority
        step_size = step_size/2

    # Adding the actual elements
    for (item, token_index) in zip(l.tolist(), token_indices.tolist()):
        found_single_bucket_fit = False
        single_bucket_index = -1
        single_bucket_value = bucket_size

        found_multi_bucket_bumpless_fit = False
        multi_bucket_bumpless_index = -1
        multi_bucket_bumpless_value = total

        found_multi_bucket_bumping_fit = False
        multi_bucket_bumping_index = -1
        multi_bucket_bumping_value = total

        for i in search_order:  # for index in search_order
            if(item > space_left_after[i]):
                continue
            if(value_in_bins[i] >= bucket_size):
                continue

            # Priority of choices
            #  1. Can i place this thing in an empty bucket all on its own?
            #  2. Can i plan this somewhere where is doesnt have to bump anything else around?
            #    2a. Minimize the wasted space.  Aka use the smallest space (of equal priority) that accomplishes this goal
            #  3. If not (1) and (2), then put it in the space the bumps stuff the least.

            if(value_in_bins[i] + item > bucket_size): #Would overflow. 

                space_before_next_block = bucket_size - value_in_bins[i]
                for j in range(i+1, len(bins)):
                    if(value_in_bins[j] > 0): # We have found a bucket with something in it.  This is how much space we have here.
                        space_before_next_block = space_before_next_block + (bucket_size - value_in_bins[i])
                        break
                    else: # This was a empty bucket
                        space_before_next_block = space_before_next_block + bucket_size

                if((not found_multi_bucket_bumpless_fit) or (found_multi_bucket_bumpless_fit and priorities[i] <= priorities[multi_bucket_bumpless_index])): #This could potentially be a match

                    # If this is a valid space to put this without bumping and it is a better fit than previous spaces
                    if(space_before_next_block > item and space_before_next_block < multi_bucket_bumpless_value):
                        # set this to be the pointer!  we can fit stuff here
                        found_multi_bucket_bumpless_fit = True
                        multi_bucket_bumpless_index = i
                        multi_bucket_bumpless_value = space_before_next_block

                    # Find the overflow that will bump the least
                    if ( item - space_before_next_block < multi_bucket_bumping_value):
                        found_multi_bucket_bumping_fit = True
                        multi_bucket_bumping_index = i
                        multi_bucket_bumping_value = item - space_before_next_block

            if(value_in_bins[i] + item <= bucket_size): #Would fit
                if(single_bucket_value > value_in_bins[i]):
                    found_single_bucket_fit = True
                    single_bucket_value = value_in_bins[i]
                    single_bucket_index = i

        if (single_bucket_index == multi_bucket_bumpless_index == multi_bucket_bumping_index == -1):
            bins[0] = torch.cat( (torch.tensor([item], device=device), bins[0]), 0)
            token_bins[0] = torch.cat( (torch.tensor([token_index], device=device), token_bins[0]), 0)
            continue


        if found_single_bucket_fit:
            # We found somewhere we can actually fit!
            bins[single_bucket_index] = torch.cat( (bins[single_bucket_index], torch.tensor([item], device=device)), 0)  
            token_bins[single_bucket_index] = torch.cat( (token_bins[single_bucket_index], torch.tensor([token_index], device=device)), 0)  
            value_in_bins[single_bucket_index] += item
            for i in range(0, single_bucket_index+1):
                space_left_after[i] -= item

        elif found_multi_bucket_bumpless_fit:
            # Found somewhere we can put this without upsetting the force
            part_in_bucket = bucket_size - value_in_bins[multi_bucket_bumpless_index]
            part_overflow = item - part_in_bucket
            bins[multi_bucket_bumpless_index] = torch.cat( (bins[multi_bucket_bumpless_index], torch.tensor([item], device=device)), 0)
            token_bins[multi_bucket_bumpless_index] = torch.cat( (token_bins[multi_bucket_bumpless_index], torch.tensor([token_index], device=device)), 0)  
            value_in_bins[multi_bucket_bumpless_index] = bucket_size

            # Fill this bucket and continue overflowing
            j = multi_bucket_bumpless_index + 1
            for i in range(0, j):
                space_left_after[i] -= item

            while(part_overflow > 0):
                new_part_overflow = (value_in_bins[j] + part_overflow) - bucket_size
                value_in_bins[j] = min(bucket_size, part_overflow+value_in_bins[j]) # mark the bucket as filled
                space_left_after[j] -= part_overflow
                part_overflow = new_part_overflow
                j+=1

        else:
            part_in_bucket = bucket_size - value_in_bins[multi_bucket_bumping_index]
            part_overflow = item - part_in_bucket
            bins[multi_bucket_bumping_index] = torch.cat( (bins[multi_bucket_bumping_index], torch.tensor([item], device=device)), 0)
            token_bins[multi_bucket_bumping_index] = torch.cat( (token_bins[multi_bucket_bumping_index], torch.tensor([token_index], device=device)), 0)
            value_in_bins[multi_bucket_bumping_index] = bucket_size

            # Fill this bucket and continue overflowing
            j = multi_bucket_bumping_index + 1
            for i in range(0, j):
                space_left_after[i] -= item
            while(part_overflow > 0):
                new_part_overflow = (value_in_bins[j] + part_overflow) - bucket_size
                value_in_bins[j] = min(bucket_size, part_overflow+value_in_bins[j]) # mark the bucket as filled
                space_left_after[j] -= part_overflow
                part_overflow = new_part_overflow
                j+=1

    sorted_tensor = torch.cat(bins, 0)
    sorted_tokens = torch.cat(token_bins, 0)

    return sorted_tensor, sorted_tokens

def compute_ev(t, precision):
    expected_bits = []
    cum_probs = t.cumsum(0)

    for selection in range(0, len(cum_probs)):

        # Calculate new range as ints
        new_int_bottom = cum_probs[selection-1] if selection > 0 else 0
        new_int_top = cum_probs[selection]

        # Convert range to bits
        new_int_bottom_bits_inc = list(reversed(int2bits(new_int_bottom, precision)))
        new_int_top_bits_inc = list(reversed(int2bits(new_int_top-1, precision))) # -1 here because upper bound is exclusive

        # Consume most significant bits which are now fixed and update interval
        num_bits_encoded = num_same_from_beg(new_int_bottom_bits_inc, new_int_top_bits_inc)
        expected_bits.append(t[selection] * num_bits_encoded)

    return(float(sum(expected_bits).item())/(2**precision))

def visualize_bins(values_in_bins, bucket_size):
    out_str = "["
    for b in values_in_bins:
        out_str = out_str + "  " + str(round(100*b/bucket_size,2)) +  "  |"
    out_str = out_str + "]"
    print(out_str)

def visualize_distribution(l):
    total = sum(l)
    out_str = "["
    for b in l:
        out_str = out_str + "  " + str(round(100*b/total,2)) +  "  |"
    out_str = out_str + "]"
    print(out_str) 

def compute_entropy(lists):
    total = sum(lists)
    entropy = -1*sum([ (x/total) * math.log2(x/total) for x in lists])
    return entropy