alireza/cimai
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

first

Browse Source
This commit is contained in:
alireza
2025-02-19 13:45:07 +03:30
parent 729ae0b8c7
commit 8940a7d7f9
6 changed files with 1225 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

68
.gitignore vendored Normal file
View File

@@ -0,0 +1,68 @@
# Python
__pycache__/
*.py[cod]
*$py.class
*.so
.Python
build/
develop-eggs/
dist/
downloads/
eggs/
.eggs/
lib/
lib64/
parts/
sdist/
var/
wheels/
*.egg-info/
.installed.cfg
*.egg
# Virtual Environment
venv/
env/
ENV/
.env
.venv
# IDE
.idea/
.vscode/
*.swp
*.swo
.project
.pydevproject
.settings/
# Jupyter Notebook
.ipynb_checkpoints
*.ipynb
# Data and Models
*.pkl
*.h5
*.model
*.pt
*.pth
*.bin
encodings.pkl
# Logs and databases
*.log
*.sqlite
*.db
# OS generated files
.DS_Store
.DS_Store?
._*
.Spotlight-V100
.Trashes
ehthumbs.db
Thumbs.db
# Project specific
*.npy
*.npz venv/

337
load_tinystories.py Normal file
View File

@@ -0,0 +1,337 @@
from datasets import load_dataset
import numpy as np
from collections import Counter
import re
import pickle
import os
import matplotlib.pyplot as plt
from tqdm import tqdm
class BiDict:
"""
Bidirectional dictionary for word-to-vector and vector-to-word mappings
"""
def __init__(self):
self.word_to_vec = {}
self.vec_to_word = {}
def __setitem__(self, word, vector):
self.word_to_vec[word] = vector
self.vec_to_word[vector] = word
def __getitem__(self, key):
# Try word_to_vec first, then vec_to_word
return self.word_to_vec.get(key) or self.vec_to_word.get(key)
def __len__(self):
return len(self.word_to_vec)
def items(self):
return self.word_to_vec.items()
def values(self):
return self.word_to_vec.values()
def load_tinystories():
"""
Load the TinyStories dataset from Hugging Face.
Returns the dataset object containing train and validation splits.
"""
ds = load_dataset("roneneldan/TinyStories")
return ds
def tokenize_with_punctuation(text):
"""
Split text into words and punctuation marks as separate tokens.
Preserves spaces between words but treats punctuation as separate tokens.
"""
# Define pattern to split on word boundaries but keep punctuation as tokens
# Using raw string to properly escape special characters
pattern = r'([.,!?;:"\'()\[\]{}]|\s+|[a-zA-Z0-9]+)'
tokens = re.findall(pattern, text.lower())
# Filter out empty strings and pure whitespace, but keep punctuation
return [token for token in tokens if token.strip() or token in '.,!?;:"\'()[]{}']
def get_vocabulary(stories, N=12):
"""
Create vocabulary from the given stories.
Returns a bidirectional dictionary mapping words and vectors.
"""
# Get all unique tokens across all stories
all_tokens = set()
for story in stories:
tokens = tokenize_with_punctuation(story)
all_tokens.update(tokens)
# Sort tokens for consistent encoding
unique_tokens = sorted(all_tokens)
# Create unique N-bit vectors
num_tokens = len(unique_tokens)
if num_tokens > 2**N:
raise ValueError(f"Vocabulary size ({num_tokens}) exceeds {N}-bit capacity ({2**N})")
# Generate all possible N-bit numbers
all_possible = list(range(2**N))
np.random.shuffle(all_possible)
# Create unique random binary numbers for each token
token_to_vector = BiDict()
for i, token in enumerate(unique_tokens):
binary = format(all_possible[i], f'0{N}b')
token_to_vector[token] = binary
return token_to_vector
def save_encodings(vocab, encoded_stories, stories, filename='encodings.pkl'):
"""Save the encodings and vocabulary to a pickle file"""
data = {
'vocabulary': vocab,
'encoded_stories': encoded_stories,
'original_stories': stories
}
with open(filename, 'wb') as f:
pickle.dump(data, f)
def load_encodings(filename='encodings.pkl'):
"""Load encodings from pickle file if it exists"""
if os.path.exists(filename):
with open(filename, 'rb') as f:
data = pickle.load(f)
return data['vocabulary'], data['encoded_stories'], data['original_stories']
return None, None, None
def encode_stories(n_stories=30, force_encode=False, N=12):
"""
Encode the first n stories into N-bit vectors.
If encodings exist and force_encode is False, load from file.
Otherwise, create new encodings and save them.
"""
if not force_encode:
vocab, encoded_stories, stories = load_encodings()
if vocab is not None:
print("Loaded existing encodings from file")
return vocab, encoded_stories, stories
ds = load_tinystories()
stories = [ds['train'][i]['text'] for i in range(n_stories)]
print(stories)
# Get vocabulary mapping with specified N
vocab = get_vocabulary(stories, N=N)
# Encode stories
encoded_stories = []
for story in stories:
tokens = tokenize_with_punctuation(story)
encoded_tokens = [vocab[token] for token in tokens]
encoded_stories.append(encoded_tokens)
# Save the encodings
save_encodings(vocab, encoded_stories, stories)
print("Created and saved new encodings")
return vocab, encoded_stories, stories
def get_word_sequences(encoded_stories, M=100, N=12):
"""
Get sequences of M consecutive words from encoded stories.
Each word is N bits long.
"""
M_N_sequences = []
# Process each story with progress bar
for story in tqdm(encoded_stories, desc="Generating sequences"):
# Only process if story has enough words
if len(story) >= M:
# Get groups of M words, shifting by 1 word each time
for i in range(len(story) - M + 1):
word_group = story[i:i + M]
# Convert words to bit array
bits = []
for word in word_group:
bits.extend([int(bit) for bit in word])
vector = np.array(bits).reshape(M * N, 1)
M_N_sequences.append(vector)
return np.array(M_N_sequences)
def sequence_to_words(sequence, N=12):
"""
Convert a sequence vector back into a list of N-bit words
"""
# Convert sequence to flat list of bits
bits = [str(int(bit[0])) for bit in sequence]
# Split into N-bit chunks
words = [''.join(bits[i:i + N]) for i in range(0, len(bits), N)]
return words
def calculate_energy(sequences):
"""
Calculate the energy of each sequence.
"""
energies = []
hamiltonian = 0
for seq in sequences:
energy = -seq.dot(seq.T)/2
hamiltonian += energy
energies.append(energy)
plt.semilogy(-np.linalg.eigvals(hamiltonian), ".")
plt.show()
return energies, hamiltonian
def retrieve_sequences(sequences, partial_sequence, vocab, W, M=10, N=12, temperature=1.0):
"""
Retrieve the most likely next word using Ising Hamiltonian with temperature.
Uses associative memory to retrieve the last word of the sequence.
"""
# Convert partial sequence to vector
partial_vec = np.array([int(bit) for bit in partial_sequence]).reshape(-1, 1)
# Get all possible words from vocabulary
possible_words = list(vocab.values())
# Calculate weights matrix (Hebbian learning)
# Calculate energies for all possible words
word_energies = []
for word in possible_words:
# Create complete sequence with this word
complete_sequence = partial_sequence + word
if len(complete_sequence) == M*N: # Ensure correct length
complete_vec = np.array([int(bit) for bit in complete_sequence]).reshape(M * N, 1)
# Calculate energy using Ising Hamiltonian
energy_matrix = complete_vec.T.dot(W).dot(complete_vec)
energy = -0.5 * float(energy_matrix[0, 0])
word_energies.append((word, energy))
# Sort by energy
word_energies.sort(key=lambda x: x[1])
# Normalize energies to prevent overflow
energies = np.array([e[1] for e in word_energies])
energies = energies - np.min(energies) # Shift to make minimum energy 0
energies = energies / np.max(energies) if np.max(energies) > 0 else energies # Scale to [0,1]
# Calculate probabilities with normalized energies
probabilities = np.exp(-energies/temperature)
probabilities = probabilities / np.sum(probabilities)
# Check for valid probabilities
if np.any(np.isnan(probabilities)):
# Fallback to uniform distribution if numerical issues occur
probabilities = np.ones(len(word_energies)) / len(word_energies)
selected_idx = np.random.choice(len(word_energies), p=probabilities)
best_word, min_energy = word_energies[selected_idx]
# Find the word corresponding to the binary vector
for word, vector in vocab.items():
if vector == best_word:
return word, best_word, min_energy
def predict_sequence(initial_sequence, vocab, sequences, W, D=10, M=100, N=12, temperature=1.0):
"""
Predict D words iteratively by sliding the window.
"""
current_tokens = initial_sequence.copy()
predictions = []
energies = []
# Add progress bar for predictions
for _ in tqdm(range(D), desc="Predicting words"):
# Convert current tokens to binary sequence
partial_sequence = ""
for token in current_tokens:
partial_sequence += vocab[token]
# Predict next word
predicted_word, _, energy = retrieve_sequences(
sequences,
partial_sequence,
vocab,
W=W,
M=M,
N=N,
temperature=temperature
)
predictions.append(predicted_word)
energies.append(energy)
# Slide window: remove first token and add predicted word
current_tokens = current_tokens[1:] + [predicted_word]
return predictions, energies
if __name__ == "__main__":
N = 13 # Define N as a constant
M = 10 # Define M as a constant
D = 3 # Number of words to predict
temperature = 1.0 # Increased temperature for more diversity
print("Loading and encoding stories...")
# Force new encoding to ensure consistency
vocab, encoded_stories, original_stories = encode_stories(force_encode=True, N=N)
print("\nGenerating training sequences...")
# Get sequences for training
sequences = get_word_sequences(encoded_stories=encoded_stories, M=M, N=N)
print(f"Number of training sequences: {len(sequences)}")
print(f"Sequence shape: {sequences[0].shape if len(sequences) > 0 else 'No sequences found'}")
# Get initial sequence from first story
story_tokens = tokenize_with_punctuation(original_stories[0])
_, W = calculate_energy(sequences)
# Make sure we have enough tokens for M=100
if len(story_tokens) >= M-1:
initial_tokens = story_tokens[:M-1]
# Predict next D words
predicted_words, energies = predict_sequence(
initial_tokens,
vocab,
sequences,
W=W,
D=D,
M=M,
N=N,
temperature=temperature
)
# Print results
print("\nOriginal sequence:")
print(" ".join(initial_tokens[-10:])) # Last 10 tokens of initial sequence
print("\nPredicted sequence:")
print(" ".join(predicted_words))
print("\nEnergies:")
print(energies)
print("\nActual next words:")
print(" ".join(story_tokens[M-1:M-1+D])) # Next D actual words
else:
print(f"Story too short. Needs at least {M-1} tokens, but has {len(story_tokens)}")
# # Print example
# print(f"Total vocabulary size: {len(vocab)}")
# print("\nExample encoding for first story:")
# print("Original:", original_stories[0])
# print("First few tokens and their encodings:")
# tokens = tokenize_with_punctuation(original_stories[0])
# for token, encoding in zip(tokens[:10], encoded_stories[0][:10]):
# print(f"'{token}' -> {encoding}")
# # Get statistics about vector usage
# total_unique_in_vocab = len(vocab)
# total_unique_used = len(set([vec for story in encoded_stories for vec in story]))
# total_vectors = sum(len(story) for story in encoded_stories)
# print(f"\nTotal unique vectors in vocabulary: {total_unique_in_vocab}")
# print(f"Total unique vectors used in stories: {total_unique_used}")
# print(f"Total word occurrences: {total_vectors}")
# print(encoded_stories[0])
# print(sequences)
# plt.imshow(energies[0])
# plt.show()

61
predict_story.py Normal file
View File

@@ -0,0 +1,61 @@
from load_tinystories import (
tokenize_with_punctuation,
load_encodings,
predict_sequence,
get_word_sequences,
BiDict
)
def get_user_sentence():
"""Get a sentence from user and validate it has enough words"""
while True:
sentence = input("Enter a sentence (at least 99 words): ")
tokens = tokenize_with_punctuation(sentence)
if len(tokens) >= 99:
return tokens
print(f"Sentence too short. Got {len(tokens)} tokens, need 99.")
def main():
# Load saved encodings
vocab, encoded_stories, original_stories = load_encodings()
if vocab is None:
print("No saved encodings found. Please run load_tinystories.py first.")
return
# Constants
M = 100 # Sequence length
N = 13 # Bits per word
D = 10 # Number of words to predict
temperature = 1.0
print("Loading training sequences...")
sequences = get_word_sequences(encoded_stories=encoded_stories, M=M, N=N)
print(f"Loaded {len(sequences)} training sequences")
# Get sentence from user
print("\nI'll help you continue your story!")
initial_tokens = get_user_sentence()
# Predict next words
print("\nPredicting next words...")
predicted_words, energies = predict_sequence(
initial_tokens[:M-1], # Use first M-1 tokens
vocab,
sequences,
D=D,
M=M,
N=N,
temperature=temperature
)
# Print results
print("\nYour input ended with:")
print(" ".join(initial_tokens[-10:])) # Last 10 tokens
print("\nPredicted continuation:")
print(" ".join(predicted_words))
print("\nEnergies of predictions:")
for i, (word, energy) in enumerate(zip(predicted_words, energies)):
print(f"Word {i+1} ('{word}'): {energy:.4f}")
if __name__ == "__main__":
main()

332
sentences.txt Normal file
View File

@@ -0,0 +1,332 @@
“You can’t, all right? You can’t help me. No one can help me. My wife is dead, and the police think
I killed her.” His voice is rising, spots of colour appear on his cheeks. “They think I killed her.”
“But . . . Kamal Abdic . . .”
The chair crashes against the kitchen wall with such force that one of the legs splinters away. I
jump back in fright, but Scott has barely moved. His hands are back at his sides, balled into fists. I can
see the veins under his skin.
“Kamal Abdic,” he says, teeth gritted, “is no longer a suspect.” His tone is even, but he is
struggling to restrain himself. I can feel the anger vibrating off him. I want to get to the front door,
but he is in my way, blocking my path, blocking out what little light there was in the room.
“Do you know what he’s been saying?” he asks, turning away from me to pick up the chair. Of
course I don’t, I think, but I realize once again that he’s not really talking to me. “Kamal’s got all sorts
of stories. Kamal says that Megan was unhappy, that I was a jealous, controlling husband, a—what
was the word?—an emotional abuser.” He spits the words out in disgust. “Kamal says Megan was
afraid of me.”
“But he’s—”
“He isn’t the only one. That friend of hers, Tara—she says that Megan asked her to cover for her
sometimes, that Megan wanted her to lie to me about where she was, what she was doing.”
He places the chair back at the table and it falls over. I take a step towards the hallway, and he looks
at me then. “I am a guilty man,” he says, his face a twist of anguish. “I am as good as convicted.”
He kicks the broken chair aside and sits down on one of the three remaining good ones. I hover,
unsure. Stick or twist? He starts to talk again, his voice so soft I can barely hear him. “Her phone was
in her pocket,” he says. I take a step closer to him. “There was a message on it from me. The last thing
I ever said to her, the last words she ever read, were Go to hell you lying bitch.”
His chin on his chest, his shoulders start to shake. I am close enough to touch him. I raise my hand
and, trembling, put my fingers lightly on the back of his neck. He doesn’t shrug me away.
“I’m sorry,” I say, and I mean it, because although I’m shocked to hear the words, to imagine that
he could speak to her like that, I know what it is to love someone and to say the most terrible things to
them, in anger or anguish. “A text message,” I say. “It’s not enough. If that’s all they have . . .”
“It’s not, though, is it?” He straightens up then, shrugging my hand away from him. I walk back
around the table and sit down opposite him. He doesn’t look up at me. “I have a motive. I didn’t
behave . . . I didn’t react the right way when she walked out. I didn’t panic soon enough. I didn’t call
her soon enough.” He gives a bitter laugh. “And there is a pattern of abusive behaviour, according to
Kamal Abdic.” It’s then that he looks up at me, that he sees me, that a light comes on. Hope. “You . . .
you can talk to the police. You can tell them that it’s a lie, that he’s lying. You can at least give another
side of the story, tell them that I loved her, that we were happy.”
I can feel panic rising in my chest. He thinks I can help him. He is pinning his hopes on me and all I
have for him is a lie, a bloody lie.
“They won’t believe me,” I say weakly. “They don’t believe me. I’m an unreliable witness.”
The silence between us swells and fills the room; a fly buzzes angrily against the French doors.
Scott picks at the dried blood on his cheek, I can hear his nails scraping against his skin. I push my
chair back, the legs scraping on the tiles, and he looks up.
“You were here,” he says, as though the piece of information I gave him fifteen minutes ago is
only now sinking in. “You were in Witney the night Megan went missing?”
I can barely hear him above the blood thudding in my ears. I nod.
“Why didn’t you tell the police that?” he asks. I can see the muscle tic in his jaw.
“I did. I did tell them that. But I didn’t have . . . I didn’t see anything. I don’t remember anything.”
He gets to his feet, walks over to the French doors and pulls back the curtain. The sunshine is
momentarily blinding. Scott stands with his back to me, his arms folded.
“You were drunk,” he says matter-of-factly. “But you must remember something. You must—that’s
why you keep coming back here, isn’t it?” He turns around to face me. “That’s it, isn’t it? Why you
keep contacting me. You know something.” He’s saying this as though it’s fact: not a question, not an
accusation, not a theory. “Did you see his car?” he asks. “Think. Blue Vauxhall Corsa. Did you see it?”
I shake my head and he throws his arms up in frustration. “Don’t just dismiss it. Really think. What did
you see? You saw Anna Watson, but that doesn’t mean anything. You saw—come on! Who did you
see?”
Blinking into the sunlight, I try desperately to piece together what I saw, but nothing comes.
Nothing real, nothing helpful. Nothing I could say out loud. I was in an argument. Or perhaps I
witnessed an argument. I stumbled on the station steps, a man with red hair helped me up—I think that
he was kind to me, although now he makes me feel afraid. I know that I had a cut on my head, another
on my lip, bruises on my arms. I think I remember being in the underpass. It was dark. I was
frightened, confused. I heard voices. I heard someone call Megan’s name. No, that was a dream. That
wasn’t real. I remember blood. Blood on my head, blood on my hands. I remember Anna. I don’t
remember Tom. I don’t remember Kamal or Scott or Megan.
He is watching me, waiting for me to say something, to offer him some crumb of comfort, but I
have none.
“That night,” he says, “that’s the key time.” He sits back down at the table, closer to me now, his
back to the window. There is a sheen of sweat on his forehead and his upper lip, and he shivers as
though with fever. “That’s when it happened. They think that’s when it happened. They can’t be
sure . . .” He tails off. “They can’t be sure. Because of the condition . . . of the body.” He takes a deep
breath. “But they think it was that night. Or soon after.” He’s back on autopilot, speaking to the room,
not to me. I listen in silence as he tells the room that the cause of death was head trauma, her skull was
fractured in several places. No sexual assault, or at least none that they could confirm, because of her
condition. Her condition, which was ruined.
When he comes back to himself, back to me, there is fear in his eyes, desperation.
“If you remember anything,” he says, “you have to help me. Please, try to remember, Rachel.” The
sound of my name on his lips makes my stomach flip, and I feel wretched.
On the train, on the way home, I think about what he said, and I wonder if it’s true. Is the reason that
I can’t let go of this trapped inside my head? Is there some knowledge I’m desperate to impart? I
know that I feel something for him, something I can’t name and shouldn’t feel. But is it more than
that? If there’s something in my head, then maybe someone can help me get it out. Someone like a
psychiatrist. A therapist. Someone like Kamal Abdic.
T UESDAY, AUGUST 6, 2013
MORNING
I’ve barely slept. All night, I lay awake thinking about it, turning it over and over in my mind. Is this
stupid, reckless, pointless? Is it dangerous? I don’t know what I’m doing. I made an appointment
yesterday morning to see Dr. Kamal Abdic. I rang his surgery and spoke to a receptionist, asked for
him by name. I might have been imagining it, but I thought she sounded surprised. She said he could
see me today at four thirty. So soon? My heart battering my ribs, my mouth dry, I said that would be
fine. The session costs £75. That £300 from my mother is not going to last very long.
Ever since I made the appointment, I haven’t been able to think of anything else. I’m afraid, but I’m
excited, too. I can’t deny that there’s a part of me that finds the idea of meeting Kamal thrilling.
Because all this started with him: a glimpse of him and my life changed course, veered off the tracks.
The moment I saw him kiss Megan, everything changed.
And I need to see him. I need to do something, because the police are only interested in Scott. They
had him in for questioning again yesterday. They won’t confirm it, of course, but there’s footage on
the Internet: Scott, walking into the police station, his mother at his side. His tie was too tight, he
looked strangled.
Everyone speculates. The newspapers say that the police are being more circumspect, that they
cannot afford to make another hasty arrest. There is talk of a botched investigation, suggestions that a
change in personnel may be required. On the Internet, the talk about Scott is horrible, the theories
wild, disgusting. There are screen grabs of him giving his first tearful appeal for Megan’s return, and
next to them are pictures of killers who had also appeared on television, sobbing, seemingly
distraught at the fate of their loved ones. It’s horrific, inhuman. I can only pray that he never looks at
this stuff. It would break his heart.
So, stupid and reckless I may be, but I am going to see Kamal Abdic, because unlike all the
speculators, I have seen Scott. I’ve been close enough to touch him, I know what he is, and he isn’t a
murderer.
EVENING
My legs are still trembling as I climb the steps to Corly station. I’ve been shaking like this for hours,
it must be the adrenaline, my heart just won’t slow down. The train is packed—no chance of a seat
here, it’s not like getting on at Euston, so I have to stand, midway through a carriage. It’s like a
sweatbox. I’m trying to breathe slowly, my eyes cast down to my feet. I’m just trying to get a handle
on what I’m feeling.
Exultation, fear, confusion and guilt. Mostly guilt.
It wasn’t what I expected.
By the time I got to the practice, I’d worked myself up into a state of complete and utter terror: I
was convinced that he was going to look at me and somehow know that I knew, that he was going to
view me as a threat. I was afraid that I would say the wrong thing, that somehow I wouldn’t be able to
stop myself from saying Megan’s name. Then I walked into a doctor ’s waiting room, boring and
bland, and spoke to a middle-aged receptionist, who took my details without really looking at me. I
sat down and picked up a copy of Vogue and flicked through it with trembling fingers, trying to focus
my mind on the task ahead while at the same time attempting to look unremarkably bored, just like
any other patient.
There were two others in there: a twentysomething man reading something on his phone and an
older woman who stared glumly at her feet, not once looking up, even when her name was called by
the receptionist. She just got up and shuffled off, she knew where she was going. I waited there for
five minutes, ten. I could feel my breathing getting shallow. The waiting room was warm and airless,
and I felt as though I couldn’t get enough oxygen into my lungs. I worried that I might faint.
Then a door flew open and a man came out, and before I’d even had time to see him properly, I
knew that it was him. I knew the way I knew that he wasn’t Scott the first time I saw him, when he was
nothing but a shadow moving towards her—just an impression of tallness, of loose, languid
movement. He held out his hand to me.
“Ms. Watson?”
I raised my eyes to meet his and felt a jolt of electricity all the way down my spine. I put my hand
into his. It was warm and dry and huge, enveloping the whole of mine.
“Please,” he said, indicating for me to follow him into his office, and I did, feeling sick, dizzy all
the way. I was walking in her footsteps. She did all this. She sat opposite him in the chair he told me to
sit in, he probably folded his hands just below his chin the way he did this afternoon, he probably
nodded at her in the same way, saying, “OK, what would you like to talk to me about today?”
Everything about him was warm: his hand, when I shook it; his eyes; the tone of his voice. I
searched his face for clues, for signs of the vicious brute who smashed Megan’s head open, for a
glimpse of the traumatized refugee who had lost his family. I couldn’t see any. And for a while, I
forgot myself. I forgot to be afraid of him. I was sitting there and I wasn’t panicking any longer. I
swallowed hard and tried to remember what I had to say, and I said it. I told him that for four years I’d
had problems with alcohol, that my drinking had cost me my marriage and my job, it was costing me
my health, obviously, and I feared it might cost me my sanity, too.
“I don’t remember things,” I said. “I black out and I can’t remember where I’ve been or what I’ve
done. Sometimes I wonder if I’ve done or said terrible things, and I can’t remember. And if . . . if
someone tells me something I’ve done, it doesn’t even feel like me. It doesn’t feel like it was me who
was doing that thing. And it’s so hard to feel responsible for something you don’t remember. So I
never feel bad enough. I feel bad, but the thing that I’ve done—it’s removed from me. It’s like it
doesn’t belong to me.”
All this came out, all this truth, I just spilled it in front of him in the first few minutes of being in
his presence. I was so ready to say it, I’d been waiting to say it to someone. But it shouldn’t have been
him. He listened, his clear amber eyes on mine, his hands folded, motionless. He didn’t look around
the room or make notes. He listened. And eventually he nodded slightly and said, “You want to take
responsibility for what you have done, and you find it difficult to do that, to feel fully accountable if
you cannot remember it?”
“Yes, that’s it, that’s exactly it.”
“So, how do we take responsibility? You can apologize—and even if you cannot remember
committing your transgression, that doesn’t mean that your apology, and the sentiment behind your
apology, is not sincere.”
“But I want to feel it. I want to feel . . . worse.”
It’s an odd thing to say, but I think this all the time. I don’t feel bad enough. I know what I’m
responsible for, I know all the terrible things I’ve done, even if I don’t remember the details—but I
feel distanced from those actions. I feel them at one remove.
“You think that you should feel worse than you do? That you don’t feel bad enough for your
mistakes?”
“Yes.”
Kamal shook his head. “Rachel, you have told me that you lost your marriage, you lost your job—
do you not think this is punishment enough?”
I shook my head.
He leaned back a little in his chair. “I think perhaps you are being rather hard on yourself.”
“I’m not.”
“All right. OK. Can we go back a bit? To when the problem started. You said it was . . . four years
ago? Can you tell me about that time?”
I resisted. I wasn’t completely lulled by the warmth of his voice, by the softness of his eyes. I
wasn’t completely hopeless. I wasn’t going to start telling him the whole truth. I wasn’t going to tell
him how I longed for a baby. I told him that my marriage broke down, that I was depressed, and that
I’d always been a drinker, but that things just got out of hand.
“Your marriage broke down, so . . . you left your husband, or he left you, or . . . you left each
other?”
“He had an affair,” I said. “He met another woman and fell in love with her.” He nodded, waiting
for me to go on. “It wasn’t his fault, though. It was my fault.”
“Why do you say that?”
“Well, the drinking started before . . .”
“So your husband’s affair was not the trigger?”
“No, I’d already started, my drinking drove him away, it was why he stopped . . .”
Kamal waited, he didn’t prompt me to go on, he just let me sit there, waiting for me to say the
words out loud.
“Why he stopped loving me,” I said.
I hate myself for crying in front of him. I don’t understand why I couldn’t keep my guard up. I
shouldn’t have talked about real things, I should have gone in there with some totally made-up
problems, some imaginary persona. I should have been better prepared.
I hate myself for looking at him and believing, for a moment, that he felt for me. Because he
looked at me as though he did, not as though he pitied me, but as though he understood me, as though
I was someone he wanted to help.
“So then, Rachel, the drinking started before the breakdown of your marriage. Do you think you
can point to an underlying cause? I mean, not everyone can. For some people, there is just a general
slide into a depressive or an addicted state. Was there something specific for you? A bereavement,
some other loss?”
I shook my head, shrugged. I wasn’t going to tell him that. I will not tell him that.
He waited for a few moments and then glanced quickly at the clock on his desk.
“We will pick up next time, perhaps?” he said, and then he smiled and I went cold.
Everything about him is warm—his hands, his eyes, his voice—everything but the smile. You can
see the killer in him when he shows his teeth. My stomach a hard ball, my pulse skyrocketing again, I
left his office without shaking his outstretched hand. I couldn’t stand to touch him.
I understand, I do. I can see what Megan saw in him, and it’s not just that he’s arrestingly handsome.
He’s also calm and reassuring, he exudes a patient kindness. Someone innocent or trusting or simply
troubled might not see through all that, might not see that under all that calm he’s a wolf. I understand
that. For almost an hour, I was drawn in. I let myself open up to him. I forgot who he was. I betrayed
Scott, and I betrayed Megan, and I feel guilty about that.
But most of all, I feel guilty because I want to go back.
WEDNESDAY, AUGUST 7, 2013
MORNING
I had it again, the dream where I’ve done something wrong, where everyone is against me, sides with
Tom. Where I can’t explain, or even apologize, because I don’t know what the thing is. In the space
between dreaming and wakefulness, I think of a real argument, long ago—four years ago—after our
first and only round of IVF failed, when I wanted to try again. Tom told me we didn’t have the money,
and I didn’t question that. I knew we didn’t—we’d taken on a big mortgage, he had some debts left
over from a bad business deal his father had coaxed him into pursuing—I just had to deal with it. I just
had to hope that one day we would have the money, and in the meantime I had to bite back the tears
that came, hot and fast, every time I saw a stranger with a bump, every time I heard someone else’s
happy news.
It was a couple of months after we’d found out that the IVF had failed that he told me about the trip.
Vegas, for four nights, to watch the big fight and let off some steam. Just him and a couple of his
mates from the old days, people I had never met. It cost a fortune, I know, because I saw the booking
receipt for the flight and the room in his email inbox. I’ve no idea what the boxing tickets cost, but I
can’t imagine they were cheap. It wasn’t enough to pay for a round of IVF, but it would have been a
start. We had a horrible fight about it. I don’t remember the details because I’d been drinking all
afternoon, working myself up to confront him about it, so when I did it was in the worst possible way.
I remember his coldness the next day, his refusal to speak about it. I remember him telling me, in flat
disappointed tones, what I’d done and said, how I’d smashed our framed wedding photograph, how
I’d screamed at him for being so selfish, how I’d called him a useless husband, a failure. I remember
how much I hated myself that day.
I was wrong, of course I was, to say those things to him, but what comes to me now is that I wasn’t
unreasonable to be angry. I had every right to be angry, didn’t I? We were trying to have a baby—
shouldn’t we have been prepared to make sacrifices? I would have cut off a limb if it meant I could
have had a child. Couldn’t he have forgone a weekend in Vegas?
I lie in bed for a bit, thinking about that, and then I get up and decide to go for a walk, because if I
don’t do something I’m going to want to go round to the corner shop. I haven’t had a drink since
Sunday and I can feel the fight going on within me, the longing for a little buzz, the urge to get out of
my head, smashing up against the vague feeling that something has been accomplished and that it
would be a shame to throw it away now.
Ashbury isn’t really a good place to walk, it’s just shops and suburbs, there isn’t even a decent
park. I head off through the middle of town, which isn’t so bad when there’s no one else around. The
trick is to fool yourself into thinking that you’re headed somewhere: just pick a spot and set off
towards it. I chose the church at the top of Pleasance Road, which is about two miles from Cathy’s flat.
I’ve been to an AA meeting there. I didn’t go to the local one because I didn’t want to bump into
anyone I might see on the street, in the supermarket, on the train.
When I get to the church, I turn around and walk back, striding purposefully towards home, a
woman with things to do, somewhere to go. Normal. I watch the people I pass—the two men running,
backpacks on, training for the marathon, the young woman in a black skirt and white trainers, heels in
her bag, on her way to work—and I wonder what they’re hiding. Are they moving to stop drinking,
running to stand still? Are they thinking about the killer they met yesterday, the one they’re planning
to see again?
I’m not normal.
I’m almost home when I see it. I’ve been lost in thought, thinking about what these sessions with
Kamal are actually supposed to achieve: am I really planning to rifle through his desk drawers if he
happens to leave the room? To try to trap him into saying something revealing, to lead him into
dangerous territory? Chances are he’s a lot cleverer than I am; chances are he’ll see me coming. After
all, he knows his name has been in the papers—he must be alert to the possibility of people trying to
get stories on him or information from him.
This is what I’m thinking about, head down, eyes on the pavement, as I pass the little Londis shop
on the right and try not to look at it because it raises possibilities, but out of the corner of my eye I see
her name. I look up and it’s there, in huge letters on the front of a tabloid newspaper: WAS MEGAN A
CHILD KILLER?
ANNA
• • •
WEDNESDAY, AUGUST 7, 2013
MORNING
I was with the National Childbirth Trust girls at Starbucks when it happened. We were sitting in our
usual spot by the window, the kids were spreading Lego all over the floor, Beth was trying (yet again)
to persuade me to join her book club, and then Diane showed up. She had this look on her face, the
self-importance of someone who is about to deliver a piece of particularly juicy gossip. She could
barely contain herself as she struggled to get her double buggy through the door.
“Anna,” she said, her face grave, “have you seen this?” She held up a newspaper with the headline
WAS MEGAN A CHILD KILLER? I was speechless. I just stared at it and, ridiculously, burst into tears. Evie
was horrified. She howled. It was awful.
I went to the loos to clean myself (and Evie) up, and when I got back they were all speaking in
hushed tones. Diane glanced slyly up at me and asked, “Are you all right, sweetie?” She was enjoying
it, I could tell.
I had to leave then, I couldn’t stay. They were all being terribly concerned, saying how awful it
must be for me, but I could see it on their faces: thinly disguised disapproval. How could you entrust
your child to that monster? You must be the worst mother in the world.
I tried to call Tom on the way home, but his phone just went straight to voice mail. I left him a
message to ring me back as soon as possible—I tried to keep my voice light and even, but I was
trembling and my legs felt shaky, unsteady.
I didn’t buy the paper, but I couldn’t resist reading the story online. It all sounds rather vague.
“Sources close to the Hipwell investigation” claim an allegation has been made that Megan “may have
been involved in the unlawful killing of her own child” ten years ago. The “sources” also speculate
that this could be a motive for her murder. The detective in charge of the whole investigation—
Gaskill, the one who came to speak to us after she went missing—made no comment.
Tom rang me back—he was in between meetings, he couldn’t come home. He tried to placate me,
he made all the right noises, he told me it was probably a load of rubbish anyway. “You know you
can’t believe half the stuff they print in the newspapers.” I didn’t make too much of a fuss, because he
was the one who suggested she come and help out with Evie in the first place. He must be feeling
horrible.
And he’s right. It may not even be true. But who would come up with a story like that? Why would
you make up a thing like that? And I can’t help thinking, I knew. I always knew there was something
off about that woman. At first I just thought she was a bit immature, but it was more than that, she was
sort of absent. Self-involved. I’m not going to lie—I’m glad she’s gone. Good riddance.
EVENING
I’m upstairs, in the bedroom. Tom’s watching TV with Evie. We’re not talking. It’s my fault. He
walked in the door and I just went for him.
I was building up to it all day. I couldn’t help it, couldn’t hide from it, she was everywhere I looked.
Here, in my house, holding my child, feeding her, changing her, playing with her while I was taking a
nap. I kept thinking of all the times I left Evie alone with her, and it made me sick.
And then the paranoia came, that feeling I’ve had almost all the time I’ve lived in this house, of
being watched. At first, I used to put it down to the trains. All those faceless bodies staring out of the
windows, staring right across at us, it gave me the creeps. It was one of the many reasons why I didn’t
want to move in here in the first place, but Tom wouldn’t leave. He said we’d lose money on the sale.
At first the trains, and then Rachel. Rachel watching us, turning up on the street, calling us up all the
time. And then even Megan, when she was here with Evie: I always felt she had half an eye on me, as
though she were assessing me, assessing my parenting, judging me for not being able to cope on my
own. Ridiculous, I know. Then I think about that day when Rachel came to the house and took Evie,
and my whole body goes cold and I think, I’m not being ridiculous at all.
So by the time Tom came home, I was spoiling for a fight. I issued an ultimatum: we have to leave,
there’s no way I can stay in this house, on this road, knowing everything that has gone on here.
Everywhere I look now I have to see not only Rachel, but Megan, too. I have to think about everything
she touched. It’s too much. I said I didn’t care whether we got a good price for the house or not.
“You will care when we’re forced to live in a much worse place, when we can’t make our
mortgage payments,” he said, perfectly reasonably. I asked whether he couldn’t ask his parents to help
out—they have plenty of money—but he said he wouldn’t ask them, that he’d never ask them for
anything again, and he got angry then, said he didn’t want to talk about it anymore. It’s because of
how his parents treated him when he left Rachel for me. I shouldn’t even have mentioned them, it
always pisses him off.
But I can’t help it. I feel desperate, because now every time I close my eyes I see her, sitting there at
the kitchen table with Evie on her lap. She’d be playing with her and smiling and chattering, but it
never seemed real, it never seemed as if she really wanted to be there. She always seemed so happy to
be handing Evie back to me when it was time for her to go. It was almost as though she didn’t like the
feel of a child in her arms.

242
spin_glass.py Normal file
View File

@@ -0,0 +1,242 @@
import numpy as np
from itertools import product
import re
class SpinGlassHamiltonian:
def __init__(self, n_words=4, word_bits=8, text_file='sentences.txt', seed=None):
"""
Initialize a fully connected spin glass Hamiltonian with J_ij couplings
constructed from text patterns
Args:
n_words (int): Number of words per pattern
word_bits (int): Number of bits to encode each word
text_file (str): Path to file containing training sentences
seed (int): Random seed for reproducibility
"""
if seed is not None:
np.random.seed(seed)
self.n_words = n_words
self.word_bits = word_bits
self.n_spins = n_words * word_bits
# Load text from file
try:
with open(text_file, 'r') as f:
text = f.read()
except FileNotFoundError:
print(f"Warning: {text_file} not found. Using default text.")
text = """
The quantum spin glass model shows fascinating behavior
in statistical physics and complex systems research
many body interactions lead to frustration effects
ground state properties reveal emergent phenomena
"""
# Generate patterns from text
self.patterns, self.words = self.text_to_patterns(text, n_words, word_bits)
# Initialize J matrix for 2-point interactions
self.J = np.zeros((self.n_spins, self.n_spins))
# Construct J using patterns
for pattern in self.patterns:
flat_pattern = pattern.flatten()
# 2-point interactions
self.J += np.outer(flat_pattern, flat_pattern)
# Normalize by number of patterns
self.J = self.J / len(self.patterns)
np.fill_diagonal(self.J, 0) # No self-interactions
print("Words used to construct patterns:")
for i, words in enumerate(self.words):
print(f"\nPattern {i} words: {words}")
print(f"Pattern {i} configuration ({self.n_words}x{self.word_bits}):")
print(self.patterns[i])
def text_to_patterns(self, text, n_words, word_bits):
"""Convert text to patterns where each row encodes a full word"""
words = re.findall(r'\b\w+\b', text.lower())
patterns = []
pattern_words = []
# Modified to shift one word at a time
for i in range(len(words) - n_words + 1):
word_group = words[i:i+n_words]
pattern_words.append(word_group)
# Create pattern (n_words x word_bits)
pattern = np.zeros((n_words, word_bits))
for row, word in enumerate(word_group):
# Hash the word to a unique pattern
word_hash = sum(ord(c) for c in word)
# Generate word_bits number of bits
for col in range(word_bits):
bit_val = (word_hash >> col) & 1
pattern[row, col] = 1 if bit_val else -1
patterns.append(pattern)
if len(patterns) >= 5: # Limit to 5 patterns
break
return np.array(patterns), pattern_words
def calculate_energy(self, state):
"""
Calculate energy using only 2-point interactions
Args:
state (numpy.array): Array of +1/-1 spins
Returns:
float: Energy of the configuration
"""
# 2-point interaction energy
energy = -0.5 * np.sum(self.J * np.outer(state, state))
return energy
def state_to_2d(self, state):
"""Convert 1D state array to 2D grid (n_words x word_bits)"""
return state.reshape(self.n_words, self.word_bits)
def state_to_words(self, state_2d):
"""Analyze a 2D state pattern to find closest matching words from patterns"""
closest_words = []
for row_idx, row in enumerate(state_2d):
# Find the pattern row that has the highest overlap with this state row
max_overlap = -1
best_word = None
for pattern_idx, pattern in enumerate(self.patterns):
for word_idx, pattern_row in enumerate(pattern):
overlap = abs(np.sum(row * pattern_row) / self.word_bits)
if overlap > max_overlap:
max_overlap = overlap
best_word = self.words[pattern_idx][word_idx]
closest_words.append(f"{best_word} (overlap: {max_overlap:.2f})")
return closest_words
def generate_states(n_spins):
for state in product([-1, 1], repeat=n_spins):
yield np.array(state)
def main():
# Create a 4x5 spin glass system (4 words per pattern, 5 bits per word)
n_words = 24
word_bits = 16
sg = SpinGlassHamiltonian(n_words=n_words, word_bits=word_bits, seed=42)
# Initialize the best state with random configuration
current_state = np.random.choice([-1, 1], size=n_words * word_bits)
current_energy = sg.calculate_energy(current_state)
# Learn one word at a time
for word_idx in range(n_words):
print(f"\nOptimizing word {word_idx + 1}...")
# Try all possibilities for current word while keeping others fixed
best_energy = current_energy
best_state = current_state.copy()
# Generate all possibilities for one word (2^word_bits combinations)
for word_state in product([-1, 1], repeat=word_bits):
# Create test state by replacing only the current word's bits
test_state = current_state.copy()
start_idx = word_idx * word_bits
end_idx = start_idx + word_bits
test_state[start_idx:end_idx] = word_state
# Calculate energy
energy = sg.calculate_energy(test_state)
# Update if better
if energy < best_energy:
best_energy = energy
best_state = test_state.copy()
# Update current state with best found for this word
current_state = best_state
current_energy = best_energy
# Show intermediate result
state_2d = sg.state_to_2d(current_state)
print(f"Current energy: {current_energy:.4f}")
print("Current state:")
print(state_2d)
words = sg.state_to_words(state_2d)
print("Current words:")
for i, word_info in enumerate(words):
print(f"Word {i+1}: {word_info}")
# Store final result
state_energies = [(current_state, current_energy)]
print(f"\nSpin Glass System with {n_words}x{word_bits} lattice")
print("\nOptimized state:")
# Get the single optimized state
state, energy = state_energies[0]
state_2d = sg.state_to_2d(state)
# Calculate absolute overlap with each pattern
overlaps = []
for p, pattern in enumerate(sg.patterns):
overlap = abs(np.sum(state_2d * pattern) / (n_words*word_bits))
overlaps.append(f"P{p}: {overlap:.2f}")
print(f"\nEnergy: {energy:.4f}")
print(f"State configuration:\n{state_2d}")
print("\nGenerated words:")
closest_words = sg.state_to_words(state_2d)
for row_idx, word_info in enumerate(closest_words):
print(f"Word {row_idx + 1}: {word_info}")
print(f"Absolute overlaps with patterns: {', '.join(overlaps)}")
# Show which original sentence this state is most similar to
best_pattern_idx = np.argmax([float(o.split(': ')[1]) for o in overlaps])
print(f"Most similar to sentence: {' '.join(sg.words[best_pattern_idx])}")
# Generate a 30-word sentence
print("\nGenerating 30-word sentence:")
# Start with the first n_words-1 words from the best matching pattern
sentence = list(sg.words[best_pattern_idx][:-1]) # Take all but last word as initial context
# Generate remaining words until we reach 30
while len(sentence) < 30:
# Create state from current context
context_state = np.zeros((n_words, word_bits))
for i, word in enumerate(sentence[-n_words+1:]):
word_hash = sum(ord(c) for c in word)
for col in range(word_bits):
bit_val = (word_hash >> col) & 1
context_state[i, col] = 1 if bit_val else -1
# Optimize the next word (last row of state)
best_energy = float('inf')
best_word_state = None
# Try all possibilities for the next word
for word_state in product([-1, 1], repeat=word_bits):
test_state = context_state.copy()
test_state[-1] = word_state
# Calculate energy
energy = sg.calculate_energy(test_state.flatten())
if energy < best_energy:
best_energy = energy
best_word_state = test_state.copy()
# Get the predicted word
words = sg.state_to_words(best_word_state)
next_word = words[-1].split(" (")[0] # Get just the word, not the overlap info
sentence.append(next_word)
print(" ".join(sentence))
if __name__ == "__main__":
main()

185
spin_glass_gpu.py Normal file
View File

@@ -0,0 +1,185 @@
import jax
import jax.numpy as jnp
from functools import partial
from load_tinystories import (
tokenize_with_punctuation,
load_encodings,
BiDict
)
from tqdm import tqdm
@partial(jax.jit, static_argnums=(3, 4))
def calculate_weights(sequences, M, N):
"""
Calculate weight matrix using JAX for GPU acceleration.
"""
# Convert sequences to JAX array
sequences_jax = jnp.array(sequences)
# Calculate weights using matrix multiplication
W = jnp.sum(jnp.matmul(sequences_jax, sequences_jax.transpose(0, 2, 1)), axis=0)
# Normalize weights
W = W / len(sequences)
return W
@partial(jax.jit, static_argnums=(3, 4, 5))
def calculate_energies(complete_sequences, W, temperature):
"""
Calculate energies for all possible completions using JAX.
"""
# Calculate energies using matrix operations
energies = -0.5 * jnp.diagonal(
jnp.matmul(
jnp.matmul(complete_sequences, W),
complete_sequences.transpose(0, 2, 1)
)
).reshape(-1)
# Normalize energies
energies = energies - jnp.min(energies)
energies = energies / (jnp.max(energies) + 1e-10)
# Calculate probabilities
probabilities = jnp.exp(-energies/temperature)
probabilities = probabilities / jnp.sum(probabilities)
return energies, probabilities
def get_word_sequences_gpu(encoded_stories, M=100, N=12):
"""
Get sequences of M consecutive words, optimized for GPU.
"""
sequences = []
for story in tqdm(encoded_stories, desc="Generating sequences"):
if len(story) >= M:
for i in range(len(story) - M + 1):
word_group = story[i:i + M]
bits = []
for word in word_group:
bits.extend([int(bit) for bit in word])
vector = jnp.array(bits).reshape(M * N, 1)
sequences.append(vector)
return jnp.array(sequences)
def retrieve_sequences_gpu(sequences, partial_sequence, vocab, M=100, N=12, temperature=1.0):
"""
GPU-accelerated version of sequence retrieval using JAX.
"""
# Convert partial sequence to vector
partial_vec = jnp.array([int(bit) for bit in partial_sequence]).reshape(-1, 1)
# Get all possible words
possible_words = list(vocab.values())
# Calculate weight matrix
W = calculate_weights(sequences, M, N)
# Create complete sequences for all possible words
complete_sequences = []
for word in possible_words:
complete_sequence = partial_sequence + word
if len(complete_sequence) == M*N:
complete_vec = jnp.array([int(bit) for bit in complete_sequence]).reshape(M * N, 1)
complete_sequences.append(complete_vec)
complete_sequences = jnp.array(complete_sequences)
# Calculate energies and probabilities
energies, probabilities = calculate_energies(complete_sequences, W, temperature)
# Select word based on probabilities
selected_idx = jax.random.choice(
jax.random.PRNGKey(0),
len(possible_words),
p=probabilities
)
best_word = possible_words[selected_idx]
min_energy = float(energies[selected_idx])
# Find corresponding word
for word, vector in vocab.items():
if vector == best_word:
return word, best_word, min_energy
def predict_sequence_gpu(initial_sequence, vocab, sequences, D=10, M=100, N=12, temperature=1.0):
"""
GPU-accelerated version of sequence prediction.
"""
current_tokens = initial_sequence.copy()
predictions = []
energies = []
for _ in tqdm(range(D), desc="Predicting words"):
partial_sequence = ""
for token in current_tokens:
partial_sequence += vocab[token]
predicted_word, _, energy = retrieve_sequences_gpu(
sequences,
partial_sequence,
vocab,
M=M,
N=N,
temperature=temperature
)
predictions.append(predicted_word)
energies.append(energy)
current_tokens = current_tokens[1:] + [predicted_word]
return predictions, energies
def main():
# Load saved encodings
vocab, encoded_stories, original_stories = load_encodings()
if vocab is None:
print("No saved encodings found. Please run load_tinystories.py first.")
return
# Constants
M = 100
N = 13
D = 10
temperature = 1.0
print("Loading training sequences...")
sequences = get_word_sequences_gpu(encoded_stories=encoded_stories, M=M, N=N)
print(f"Loaded {len(sequences)} training sequences")
# Get sentence from user
print("\nEnter your story:")
sentence = input("Enter a sentence (at least 99 words): ")
initial_tokens = tokenize_with_punctuation(sentence)
if len(initial_tokens) < M-1:
print(f"Sentence too short. Got {len(initial_tokens)} tokens, need {M-1}.")
return
# Predict next words
print("\nPredicting next words...")
predicted_words, energies = predict_sequence_gpu(
initial_tokens[:M-1],
vocab,
sequences,
D=D,
M=M,
N=N,
temperature=temperature
)
# Print results
print("\nYour input ended with:")
print(" ".join(initial_tokens[-10:]))
print("\nPredicted continuation:")
print(" ".join(predicted_words))
print("\nEnergies of predictions:")
for i, (word, energy) in enumerate(zip(predicted_words, energies)):
print(f"Word {i+1} ('{word}'): {energy:.4f}")
if __name__ == "__main__":
main()