mirror of
https://github.com/wassname/Clover-Edition.git
synced 2026-09-11 11:51:53 +08:00
Manually merged in most of the changes from AccidentallyOnPurpose who changed the model to pytorch added suggested actions, a repetition penalty, and a number of other things.
This commit is contained in:
cloveranon
@@ -7,3 +7,5 @@ story*.json
|
||||
/prompts/*
|
||||
!/prompts/cloveranon-prompts/*
|
||||
/generator/gpt2/models/*
|
||||
/models/*
|
||||
!/models/.gitkeep
|
||||
|
||||
@@ -9,6 +9,8 @@
|
||||
# Ok coomers.
|
||||
temp = 0.4
|
||||
|
||||
repetition-penalty = 1.2
|
||||
|
||||
#The number of words the AI has to choose from.
|
||||
# It always chooses the "top k" most likely next words before randomly picking one according to temperature.
|
||||
# Low values reduce the randomness of the AI similar to temp.
|
||||
|
||||
@@ -0,0 +1,225 @@
|
||||
import os
|
||||
import torch
|
||||
import torch.nn.functional as F
|
||||
|
||||
from transformers import GPT2LMHeadModel, GPT2Tokenizer
|
||||
|
||||
from story.utils import cut_trailing_sentence, logger
|
||||
|
||||
# warnings.filterwarnings("ignore")
|
||||
MODEL_CLASSES = {
|
||||
"gpt2": (GPT2LMHeadModel, GPT2Tokenizer),
|
||||
}
|
||||
|
||||
def top_k_top_p_filtering(logits, top_k=0, top_p=0.0, filter_value=-float("Inf")):
|
||||
""" Filter a distribution of logits using top-k and/or nucleus (top-p) filtering
|
||||
Args:
|
||||
logits: logits distribution shape (batch size x vocabulary size)
|
||||
top_k > 0: keep only top k tokens with highest probability (top-k filtering).
|
||||
top_p > 0.0: keep the top tokens with cumulative probability >= top_p (nucleus filtering).
|
||||
Nucleus filtering is described in Holtzman et al. (http://arxiv.org/abs/1904.09751)
|
||||
From: https://gist.github.com/thomwolf/1a5a29f6962089e871b94cbd09daf317
|
||||
"""
|
||||
top_k = min(top_k, logits.size(-1)) # Safety check
|
||||
if top_k > 0:
|
||||
# Remove all tokens with a probability less than the last token of the top-k
|
||||
indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
|
||||
logits[indices_to_remove] = filter_value
|
||||
|
||||
if top_p > 0.0:
|
||||
sorted_logits, sorted_indices = torch.sort(logits, descending=True)
|
||||
cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
|
||||
|
||||
# Remove tokens with cumulative probability above the threshold
|
||||
sorted_indices_to_remove = cumulative_probs > top_p
|
||||
# Shift the indices to the right to keep also the first token above the threshold
|
||||
sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
|
||||
sorted_indices_to_remove[..., 0] = 0
|
||||
|
||||
# scatter sorted tensors to original indexing
|
||||
indices_to_remove = sorted_indices_to_remove.scatter(
|
||||
dim=1, index=sorted_indices, src=sorted_indices_to_remove
|
||||
)
|
||||
logits[indices_to_remove] = filter_value
|
||||
return logits
|
||||
|
||||
|
||||
def sample_sequence(
|
||||
model,
|
||||
length,
|
||||
context,
|
||||
num_samples=1,
|
||||
temperature=1,
|
||||
top_k=0,
|
||||
top_p=0.9,
|
||||
repetition_penalty=1.0,
|
||||
is_xlnet=False,
|
||||
is_xlm_mlm=False,
|
||||
xlm_mask_token=None,
|
||||
xlm_lang=None,
|
||||
device="cpu",
|
||||
):
|
||||
context = torch.tensor(context, dtype=torch.long, device=device)
|
||||
context = context.unsqueeze(0).repeat(num_samples, 1)
|
||||
generated = context
|
||||
with torch.no_grad():
|
||||
# for _ in tqdm(range(length), leave=False, desc='generating'):
|
||||
for _ in range(length):
|
||||
|
||||
inputs = {"input_ids": generated}
|
||||
|
||||
outputs = model(
|
||||
**inputs
|
||||
) # Note: we could also use 'past' with GPT-2/Transfo-XL/XLNet/CTRL (cached hidden-states)
|
||||
next_token_logits = outputs[0][:, -1, :] / (
|
||||
temperature if temperature > 0 else 1.0
|
||||
)
|
||||
|
||||
# repetition penalty from CTRL (https://arxiv.org/abs/1909.05858)
|
||||
for i in range(num_samples):
|
||||
for _ in set(generated[i].tolist()):
|
||||
next_token_logits[i, _] /= repetition_penalty
|
||||
|
||||
filtered_logits = top_k_top_p_filtering(
|
||||
next_token_logits, top_k=top_k, top_p=top_p
|
||||
).float()
|
||||
if temperature == 0: # greedy sampling:
|
||||
next_token = torch.argmax(filtered_logits, dim=-1).unsqueeze(-1)
|
||||
else:
|
||||
next_token = torch.multinomial(
|
||||
F.softmax(filtered_logits, dim=-1), num_samples=1
|
||||
)
|
||||
generated = torch.cat((generated, next_token), dim=1)
|
||||
return generated
|
||||
|
||||
|
||||
class GPT2Generator:
|
||||
def __init__(
|
||||
self, generate_num=60, temperature=0.4, top_k=40, top_p=0.9, censor=False, repetition_penalty=1,
|
||||
):
|
||||
self.generate_num = generate_num
|
||||
self.temp = temperature
|
||||
self.top_k = top_k
|
||||
self.top_p = top_p
|
||||
self.censor = censor
|
||||
self.samples = 1
|
||||
self.dtype = torch.half
|
||||
self.repetition_penalty = repetition_penalty
|
||||
self.batch_size = 1
|
||||
self.stop_token = None
|
||||
|
||||
self.model_name = "pytorch-gpt2-xl-aid2-v5"
|
||||
# self.model_name = "model_v5_pytorch_half"
|
||||
self.model_dir = "models"
|
||||
self.checkpoint_path = os.path.join(self.model_dir, self.model_name)
|
||||
# self.checkpoint_path = 'gpt2' # DEBUG quick test of a smaller untrained model
|
||||
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
||||
logger.info("Using device={}, checkpoint={}".format(self.device, self.checkpoint_path))
|
||||
|
||||
# Load tokenizer and model
|
||||
model_class, tokenizer_class = MODEL_CLASSES["gpt2"]
|
||||
self.tokenizer = tokenizer_class.from_pretrained(self.checkpoint_path)
|
||||
self.model = model_class.from_pretrained(self.checkpoint_path)
|
||||
self.model.to(self.device).to(self.dtype)
|
||||
self.model.eval()
|
||||
|
||||
# context_tokens = self.tokenizer.encode(' ', add_special_tokens=False)
|
||||
context_tokens = [
|
||||
self.tokenizer.pad_token_type_id,
|
||||
self.tokenizer.pad_token_type_id,
|
||||
]
|
||||
out = self.sample_sequence(context_tokens).tolist()
|
||||
# out = out[:, len(context_tokens):].tolist()
|
||||
for o in out:
|
||||
text = self.tokenizer.decode(o, clean_up_tokenization_spaces=True)
|
||||
if self.stop_token:
|
||||
index = text.find(self.stop_token)
|
||||
if index == -1:
|
||||
index = None
|
||||
text = text[:index]
|
||||
|
||||
def sample_sequence(self, context_tokens=None, generate_num=None, temperature=None):
|
||||
generate_num = generate_num if (generate_num is not None) else self.generate_num
|
||||
temperature = temperature if (temperature is not None) else self.temp
|
||||
out = sample_sequence(
|
||||
model=self.model,
|
||||
context=context_tokens,
|
||||
length=generate_num,
|
||||
# context=self.context,
|
||||
temperature=temperature,
|
||||
top_k=self.top_k,
|
||||
top_p=self.top_p,
|
||||
repetition_penalty=self.repetition_penalty,
|
||||
num_samples=self.samples,
|
||||
device=self.device
|
||||
# batch_size=self.batch_size,
|
||||
)
|
||||
return out
|
||||
|
||||
def prompt_replace(self, prompt):
|
||||
logger.debug("BEFORE PROMPT_REPLACE: `%s`", repr(prompt))
|
||||
if len(prompt) > 0 and prompt[-1] == " ":
|
||||
prompt = prompt[:-1]
|
||||
|
||||
# prompt = second_to_first_person(prompt)
|
||||
|
||||
logger.debug("AFTER PROMPT_REPLACE: `%s`", repr(prompt))
|
||||
return prompt
|
||||
|
||||
def result_replace(self, result):
|
||||
logger.debug("BEFORE RESULT_REPLACE: `%s`", repr(result))
|
||||
|
||||
result = cut_trailing_sentence(result)
|
||||
if len(result) == 0:
|
||||
return ""
|
||||
first_letter_capitalized = result[0].isupper()
|
||||
result = result.replace('."', '".')
|
||||
result = result.replace("#", "")
|
||||
result = result.replace("*", "")
|
||||
result = result.replace("\n\n", "\n")
|
||||
# result = first_to_second_person(result)
|
||||
|
||||
if not first_letter_capitalized:
|
||||
result = result[0].lower() + result[1:]
|
||||
|
||||
logger.debug("nAFTER RESULT_REPLACE: `%s`", repr(result))
|
||||
|
||||
return result
|
||||
|
||||
def generate_raw(self, prompt, generate_num=None, temperature=None):
|
||||
context_tokens = self.tokenizer.encode(prompt, add_special_tokens=False)
|
||||
|
||||
generated = 0
|
||||
for _ in range(self.samples // self.batch_size):
|
||||
out = self.sample_sequence(
|
||||
context_tokens,
|
||||
generate_num=generate_num,
|
||||
temperature=temperature
|
||||
)
|
||||
out = out[:, len(context_tokens) :].tolist()
|
||||
for o in out:
|
||||
generated += 1
|
||||
text = self.tokenizer.decode(o, clean_up_tokenization_spaces=True)
|
||||
if self.stop_token:
|
||||
index = text.find(self.stop_token)
|
||||
if index == -1:
|
||||
index = None
|
||||
text = text[:index]
|
||||
return text
|
||||
|
||||
def generate(self, prompt, options=None, seed=1):
|
||||
|
||||
prompt = self.prompt_replace(prompt)
|
||||
|
||||
logger.debug("Prompt is: `%s`", repr(prompt))
|
||||
|
||||
text = self.generate_raw(prompt)
|
||||
|
||||
logger.debug("Generated result is: `%s`", repr(text))
|
||||
|
||||
result = text
|
||||
result = self.result_replace(result)
|
||||
if len(result) == 0:
|
||||
return self.generate(prompt)
|
||||
|
||||
return result
|
||||
@@ -0,0 +1,12 @@
|
||||
google-cloud-storage
|
||||
gsutil
|
||||
numpy
|
||||
profanityfilter
|
||||
pyyaml==5.2.0
|
||||
regex
|
||||
tracery
|
||||
transformers==2.3.0
|
||||
torch==1.2.0
|
||||
colorama
|
||||
textwrap
|
||||
pyjarowinkler
|
||||
@@ -0,0 +1,40 @@
|
||||
# Pytorch AI Dungeon2
|
||||
|
||||
A Fork of Nick Walton's [AI Dungeon2](https://github.com/AIDungeon/AIDungeon) and pytorch as a backend.
|
||||
|
||||
Uses prompts and play.py from the [Clover edition](https://github.com/cloveranon/Clover-Edition)
|
||||
|
||||
I did this because tensorflow is annoying to compile for my xeon processor an it was actually faster to port the generation code.
|
||||
|
||||
No colab yet.
|
||||
|
||||
If you want the converted model, just ask in the issues and I'll make it available.
|
||||
|
||||
Content warning: This model is trained on the internet which means there will be lots of toxic and offensive content along with the funny and wierd.
|
||||
|
||||
## Screenshot
|
||||
|
||||
|
||||
|
||||
## Changes
|
||||
|
||||
- user:
|
||||
- added suggested actions from the AI player
|
||||
- roll a d20 for speech or action to make it harder
|
||||
- d01: You fail to X
|
||||
- dX: You try to X
|
||||
- d20: You successfully X
|
||||
- use Clover edition ui, prompts, config
|
||||
- technical:
|
||||
- use half precision for smaller model, (but this might lead to lower quality, I need to test more)
|
||||
- better logging
|
||||
- use pytorch
|
||||
- set top_k to zero and just use top p
|
||||
- change clover config file to yaml
|
||||
|
||||
# Model
|
||||
|
||||
<a href="magnet:?xt=urn:btih:17dcfe3d12849db04a3f64070489e6ff5fc6f63f&dn=model_v5_pytorch&tr=udp%3a%2f%2ftracker.opentrackr.org%3a1337%2fannounce&tr=udp%3a%2f%2fopen.stealth.si%3a80%2fannounce&tr=udp%3a%2f%2fp4p.arenabg.com%3a1337%2fannounce&tr=udp%3a%2f%2ftracker.coppersurfer.tk%3a6969%2fannounce&tr=udp%3a%2f%2ftracker.cyberia.is%3a6969%2fannounce&tr=udp%3a%2f%2ftracker.moeking.me%3a6969%2fannounce&tr=udp%3a%2f%2f9.rarbg.me%3a2710%2fannounce&tr=udp%3a%2f%2ftracker3.itzmx.com%3a6961%2fannounce">magnet link to pytorch model torrent</a>
|
||||
|
||||
```magnet:?xt=urn:btih:17dcfe3d12849db04a3f64070489e6ff5fc6f63f&dn=model_v5_pytorch&tr=udp%3a%2f%2ftracker.opentrackr.org%3a1337%2fannounce&tr=udp%3a%2f%2fopen.stealth.si%3a80%2fannounce&tr=udp%3a%2f%2fp4p.arenabg.com%3a1337%2fannounce&tr=udp%3a%2f%2ftracker.coppersurfer.tk%3a6969%2fannounce&tr=udp%3a%2f%2ftracker.cyberia.is%3a6969%2fannounce&tr=udp%3a%2f%2ftracker.moeking.me%3a6969%2fannounce&tr=udp%3a%2f%2f9.rarbg.me%3a2710%2fannounce&tr=udp%3a%2f%2ftracker3.itzmx.com%3a6961%2fannounce```
|
||||
|
||||
@@ -1,16 +1,31 @@
|
||||
import os
|
||||
import configparser
|
||||
import gc
|
||||
import textwrap
|
||||
import logging
|
||||
from pathlib import Path
|
||||
from random import shuffle
|
||||
from shutil import get_terminal_size
|
||||
|
||||
from generator.gpt2.gpt2_generator import *
|
||||
from story.story_manager import *
|
||||
from story.utils import *
|
||||
import textwrap
|
||||
from gpt2generator import GPT2Generator
|
||||
|
||||
#silence transformers outputs when loading model
|
||||
logging.getLogger("transformers.tokenization_utils").setLevel(logging.WARN)
|
||||
logging.getLogger("transformers.modeling_utils").setLevel(logging.WARN)
|
||||
logging.getLogger("transformers.configuration_utils").setLevel(logging.WARN)
|
||||
|
||||
logging.basicConfig(
|
||||
format='%(asctime)s - %(levelnames)s - %(messages)s',
|
||||
datefmt='%m/%d/%Y %H:%M%S',
|
||||
level=logging.INFO
|
||||
)
|
||||
logger.setLevel(10)#settings['log-level'])
|
||||
|
||||
#TODO: Move all these utilty functions to seperate utily and config file
|
||||
|
||||
#add color for windows users that install colorama
|
||||
# It is not necessary to install colorama on most systems, but it does come with pip
|
||||
try:
|
||||
import colorama
|
||||
colorama.init()
|
||||
@@ -20,13 +35,11 @@ except ModuleNotFoundError:
|
||||
with open(Path('interface', 'clover'), 'r', encoding='utf-8') as file:
|
||||
print(file.read())
|
||||
|
||||
#perhaps all the following should be put in a seperate utils file like original
|
||||
config = configparser.ConfigParser()
|
||||
config.read('config.ini')
|
||||
settings=config["Settings"]
|
||||
colors=config["Colors"]
|
||||
|
||||
os.environ["TF_CPP_MIN_LOG_LEVEL"] = settings["log-level"]
|
||||
|
||||
|
||||
#ECMA-48 set graphics codes for the curious. Check out "man console_codes"
|
||||
@@ -44,7 +57,7 @@ def getNumberInput(n):
|
||||
val=colInput("Enter a number from above (default 0):", colors["selection-prompt"], colors["selection-value"])
|
||||
if val=='':
|
||||
return 0
|
||||
elif 0>int(val) or int(val)>n:
|
||||
elif not re.match('^\d+$', val) or 0>int(val) or int(val)>n:
|
||||
colPrint("Invalid choice.", colors["error"])
|
||||
return getNumberInput(n)
|
||||
else:
|
||||
@@ -55,7 +68,9 @@ def selectFile(p=Path('prompts')):
|
||||
files=[x for x in p.iterdir()]
|
||||
shuffle(files)
|
||||
for n in range(len(files)):
|
||||
colPrint('{}: {}'.format(n, re.sub(r'\.txt$', '', files[n].name)), colors["menu"])
|
||||
colPrint(
|
||||
'{}: {}'.format(n, re.sub(r'\.txt$', '', files[n].name)),
|
||||
colors["menu"])
|
||||
return selectFile(files[getNumberInput(len(files)-1)])
|
||||
else:
|
||||
with p.open('r', encoding='utf-8') as file:
|
||||
@@ -63,38 +78,71 @@ def selectFile(p=Path('prompts')):
|
||||
rest=file.read()
|
||||
return (line1, rest)
|
||||
|
||||
#print files done several times and probably deserves own function
|
||||
def instructions():
|
||||
with open('interface/instructions.txt', 'r', encoding='utf-8') as file:
|
||||
colPrint(file.read(), colors["instructions"], False)
|
||||
|
||||
def getGenerator():
|
||||
colPrint("\nInitializing AI Engine! (This might take a few minutes)\n", colors["loading-message"])
|
||||
colPrint(
|
||||
"\nInitializing AI Engine! (This might take a few minutes)\n",
|
||||
colors["loading-message"])
|
||||
return GPT2Generator(
|
||||
generate_num=settings.getint('generate-num'),
|
||||
temperature=settings.getfloat("temp"),
|
||||
top_k=settings.getint("top-keks"),
|
||||
top_p=settings.getfloat("top-p"))
|
||||
temperature=settings.getfloat('temp'),
|
||||
top_k=settings.getint('top-keks'),
|
||||
top_p=settings.getfloat('top-p'),
|
||||
repetition_penalty=settings.getfloat('repetition-penalty')
|
||||
)
|
||||
|
||||
if not Path('prompts', 'Anime').exists():
|
||||
try:
|
||||
import pastebin
|
||||
except:
|
||||
logger.warning('Failed to scrape pastebin: %e', e)
|
||||
colPrint("Failed to scrape pastebin, possible connection issue.\nTry again later. Continuing without downloading prompts...", colors['error'])
|
||||
|
||||
class AIPlayer:
|
||||
def __init__(self, generator):
|
||||
self.generator = generator
|
||||
|
||||
def get_action(self, prompt):
|
||||
result_raw = self.generator.generate_raw(
|
||||
prompt, generate_num=settings.getint('action-generate-num'), temperature=settings.getint('temp'))
|
||||
return clean_suggested_action(result_raw, min_length=settings.getint('action-min-length'))
|
||||
|
||||
def get_actions(self, prompt):
|
||||
suggested_actions = [
|
||||
self.get_action(prompt)
|
||||
for _ in range(settings.getint('action-alternatives'))
|
||||
]
|
||||
logger.debug("Suggested actions before filter and dedup %s", suggested_actions)
|
||||
#remove short ones
|
||||
suggested_actions = [
|
||||
s
|
||||
for s in suggested_actions
|
||||
if len(s) > settings.getint('action-min-length')
|
||||
]
|
||||
#remove dups
|
||||
suggested_actions = list(set(suggested_actions))
|
||||
return suggested_actions
|
||||
|
||||
def play():
|
||||
story_manager = UnconstrainedStoryManager(getGenerator())
|
||||
generator = getGenerator()
|
||||
story_manager = UnconstrainedStoryManager(generator)
|
||||
ai_player = AIPlayer(generator)
|
||||
print("\n")
|
||||
|
||||
with open("interface/mainTitle.txt", "r", encoding="utf-8") as file:
|
||||
colPrint(file.read(), colors["title"])
|
||||
with open(Path('interface', 'mainTitle.txt'), 'r', encoding='utf-8') as file:
|
||||
colPrint(file.read(), colors['title'])
|
||||
|
||||
with open('interface/subTitle.txt', 'r', encoding="utf-8") as file:
|
||||
with open(Path('interface', 'subTitle.txt'), 'r', encoding='utf-8') as file:
|
||||
cols=get_terminal_size()[0]
|
||||
for line in file:
|
||||
line=re.sub(r'\n', '', line)
|
||||
line=line[:cols]
|
||||
#fills in the graphic using reverse video mode substituted into the areas between |'s
|
||||
colPrint(re.sub(r'\|[ _]*\|', lambda x: '\x1B[7m'+x.group(0)+'\x1B[27m', line), colors["subtitle"], False)
|
||||
|
||||
|
||||
while True:
|
||||
if story_manager.story != None:
|
||||
@@ -107,32 +155,53 @@ def play():
|
||||
if getNumberInput(1) == 1:
|
||||
with open(Path('interface', 'prompt-instructions.txt'), 'r', encoding='utf-8') as file:
|
||||
colPrint(file.read(), colors['instructions'], False)
|
||||
context=colInput('Context>', colors['main-prompt'], colors['user-text'])
|
||||
prompt=colInput('Prompt>', colors['main-prompt'], colors['user-text'])
|
||||
context = colInput('Context>', colors['main-prompt'], colors['user-text'])
|
||||
prompt = colInput('Prompt>', colors['main-prompt'], colors['user-text'])
|
||||
filename=colInput('Name to save prompt as? (Leave blank for no save): ', colors['query'], colors['user-text'])
|
||||
filename=re.sub('-$','',re.sub('^-', '', re.sub('[^a-zA-Z0-9_-]+', '-', filename)))
|
||||
if filename != '':
|
||||
with open(Path('prompts', filename+'.txt'), 'w', encoding='utf-8') as f:
|
||||
#this saves unix style line endings which might be an issue
|
||||
#don't know how to do this properly
|
||||
f.write(context+'\n'+prompt+'\n')
|
||||
f.write(context+'\n'+prompt)
|
||||
else:
|
||||
context, prompt = selectFile()
|
||||
|
||||
instructions()
|
||||
|
||||
colPrint("\nGenerating story...", colors["loading-message"])
|
||||
print()
|
||||
colPrint("Generating story...", colors['loading-message'])
|
||||
|
||||
story_manager.start_new_story(
|
||||
prompt, context=context
|
||||
)
|
||||
#TODO:seperate out AI generated part of story and print with different color
|
||||
story_manager.start_new_story(prompt, context=context)
|
||||
print("\n")
|
||||
colPrint(str(story_manager.story), colors["ai-text"])
|
||||
|
||||
while True:
|
||||
#Generate suggested actions
|
||||
if settings.getint('action-alternatives') > 0:
|
||||
|
||||
#TODO change this to two messages for different colors
|
||||
action_prompt = (
|
||||
story_manager.story.results[-1]
|
||||
if story_manager.story.results
|
||||
else "\nWhat do you do now?"
|
||||
) + "\n>"
|
||||
#can this just be a for loop?
|
||||
suggested_actions = ai_player.get_actions(action_prompt)
|
||||
if len(suggested_actions):
|
||||
suggested_actions_enum = [
|
||||
"{}> {}\n".format(i, a) for i, a in enumerate(suggested_actions)
|
||||
]
|
||||
suggested_action = "".join(suggested_actions_enum)
|
||||
#TODO: check color
|
||||
colPrint('Suggested actions\n' + suggested_action, colors['selection-value'])
|
||||
print()
|
||||
|
||||
if settings.getboolean('console-bell'):
|
||||
print('\x07', end='')
|
||||
action = colInput("> ", colors["main-prompt"], colors["user-text"])
|
||||
|
||||
#TODO:Clear suggestions and user input
|
||||
|
||||
setRegex = re.search('^set ([^ ]+) ([^ ]+)$', action)
|
||||
if setRegex:
|
||||
if setRegex.group(1) in settings:
|
||||
@@ -144,9 +213,6 @@ def play():
|
||||
if colInput('y/n? >', colors['selection-prompt'], colors['selection-value']) == 'y':
|
||||
with open('config.ini', 'w', encoding='utf-8') as file:
|
||||
config.write(file)
|
||||
del story_manager.generator
|
||||
gc.collect()
|
||||
story_manager.generator = getGenerator()
|
||||
else:
|
||||
colPrint('Invalid Setting', colors['error'])
|
||||
instructions()
|
||||
@@ -158,11 +224,11 @@ def play():
|
||||
instructions()
|
||||
elif action == "print":
|
||||
print("\nPRINTING\n")
|
||||
colPrint(str(story_manager.story), colors["print-story"])
|
||||
colPrint(str(story_manager.story), colors['print-story'])
|
||||
elif action == "revert":
|
||||
|
||||
if len(story_manager.story.actions) is 0:
|
||||
colPrint("You can't go back any farther. ", colors["error"])
|
||||
if len(story_manager.story.actions) == 0:
|
||||
colPrint("You can't go back any farther. ", colors['error'])
|
||||
continue
|
||||
|
||||
story_manager.story.actions = story_manager.story.actions[:-1]
|
||||
@@ -225,5 +291,5 @@ def play():
|
||||
else:
|
||||
colPrint(result, colors["ai-text"])
|
||||
|
||||
|
||||
#TODO: there's no reason for this to be enclosed in a function
|
||||
play()
|
||||
|
||||
@@ -0,0 +1,35 @@
|
||||
{
|
||||
"attn_pdrop": 0.1,
|
||||
"embd_pdrop": 0.1,
|
||||
"finetuning_task": null,
|
||||
"id2label": {
|
||||
"0": "LABEL_0",
|
||||
"1": "LABEL_1"
|
||||
},
|
||||
"initializer_range": 0.02,
|
||||
"is_decoder": false,
|
||||
"label2id": {
|
||||
"LABEL_0": 0,
|
||||
"LABEL_1": 1
|
||||
},
|
||||
"layer_norm_epsilon": 1e-05,
|
||||
"n_ctx": 1024,
|
||||
"n_embd": 1600,
|
||||
"n_head": 25,
|
||||
"n_layer": 48,
|
||||
"n_positions": 1024,
|
||||
"num_labels": 2,
|
||||
"output_attentions": false,
|
||||
"output_hidden_states": false,
|
||||
"output_past": true,
|
||||
"pruned_heads": {},
|
||||
"resid_pdrop": 0.1,
|
||||
"summary_activation": null,
|
||||
"summary_first_dropout": 0.1,
|
||||
"summary_proj_to_labels": true,
|
||||
"summary_type": "cls_index",
|
||||
"summary_use_proj": true,
|
||||
"torchscript": false,
|
||||
"use_bfloat16": false,
|
||||
"vocab_size": 50257
|
||||
}
|
||||
@@ -0,0 +1,110 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2018 The HuggingFace Inc. team.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
"""Convert OpenAI GPT checkpoint."""
|
||||
|
||||
from __future__ import absolute_import, division, print_function
|
||||
|
||||
import argparse
|
||||
from io import open
|
||||
|
||||
import torch
|
||||
|
||||
from transformers import (
|
||||
CONFIG_NAME,
|
||||
WEIGHTS_NAME,
|
||||
GPT2Config,
|
||||
GPT2Model,
|
||||
load_tf_weights_in_gpt2,
|
||||
)
|
||||
|
||||
import logging
|
||||
|
||||
logging.basicConfig(level=logging.INFO)
|
||||
|
||||
|
||||
def convert_gpt2_checkpoint_to_pytorch(
|
||||
gpt2_checkpoint_path, gpt2_config_file, pytorch_dump_folder_path
|
||||
):
|
||||
# Construct model
|
||||
if gpt2_config_file == "":
|
||||
config = GPT2Config()
|
||||
else:
|
||||
config = GPT2Config.from_json_file(gpt2_config_file)
|
||||
model = GPT2Model(config)
|
||||
|
||||
# Load weights from numpy
|
||||
load_tf_weights_in_gpt2(model, config, gpt2_checkpoint_path)
|
||||
|
||||
# Save pytorch-model
|
||||
pytorch_weights_dump_path = pytorch_dump_folder_path + "/" + WEIGHTS_NAME
|
||||
pytorch_config_dump_path = pytorch_dump_folder_path + "/" + CONFIG_NAME
|
||||
print("Save PyTorch model to {}".format(pytorch_weights_dump_path))
|
||||
torch.save(model.state_dict(), pytorch_weights_dump_path)
|
||||
|
||||
print("Save configuration file to {}".format(pytorch_config_dump_path))
|
||||
with open(pytorch_config_dump_path, "w", encoding="utf-8") as f:
|
||||
f.write(config.to_json_string())
|
||||
|
||||
# Also save as half precision to save transfer, loading, and inference time and memory
|
||||
pytorch_weights_dump_path += 'half'
|
||||
model.half()
|
||||
torch.save(model.state_dict(), pytorch_weights_dump_path)
|
||||
with open(pytorch_config_dump_path "w", encoding="utf-8") as f:
|
||||
f.write(config.to_json_string())
|
||||
print("Save configuration file to {}".format(pytorch_config_dump_path))
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
parser = argparse.ArgumentParser()
|
||||
## Required parameters
|
||||
parser.add_argument(
|
||||
"--gpt2_checkpoint_path",
|
||||
default=None,
|
||||
type=str,
|
||||
required=True,
|
||||
help="Path to the TensorFlow checkpoint path.",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--pytorch_dump_folder_path",
|
||||
default=None,
|
||||
type=str,
|
||||
required=True,
|
||||
help="Path to the output PyTorch model.",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--gpt2_config_file",
|
||||
default="",
|
||||
type=str,
|
||||
help="An optional config json file corresponding to the pre-trained OpenAI model. \n"
|
||||
"This specifies the model architecture.",
|
||||
)
|
||||
args = parser.parse_args()
|
||||
convert_gpt2_checkpoint_to_pytorch(
|
||||
args.gpt2_checkpoint_path, args.gpt2_config_file, args.pytorch_dump_folder_path
|
||||
)
|
||||
|
||||
|
||||
"""
|
||||
download aidungeon2 v5 model from this torrent or elsewhere
|
||||
- magnet:?xt=urn:btih:b343b83b35bff774dab13e0281ce13b3daf37d3e&dn=model_v5&tr=udp%3a%2f%2ftracker.coppersurfer.tk%3a6969%2fannounce&tr=udp%3a%2f%2ftracker.leechers-paradise.org%3a6969%2fannounce
|
||||
export OPENAI_GPT2_CHECKPOINT_PATH=../generator/gpt2/models/model_v5
|
||||
export PYTORCH_DUMP_OUTPUT=../generator/gpt2/models/model_v5_pytorch
|
||||
python convert_gpt2_model.py \
|
||||
--gpt2_checkpoint_path $OPENAI_GPT2_CHECKPOINT_PATH \
|
||||
--pytorch_dump_folder_path $PYTORCH_DUMP_OUTPUT \
|
||||
--gpt2_config_file ./aidungeonv2_model_v5_config.json
|
||||
wget https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-xl-merges.txt -o $PYTORCH_DUMP_OUTPUT/merges.txt
|
||||
wget https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-xl-vocab.json -o $PYTORCH_DUMP_OUTPUT/vocab.json
|
||||
"""
|
||||
+72
-7
@@ -1,6 +1,13 @@
|
||||
# coding: utf-8
|
||||
import re
|
||||
from difflib import SequenceMatcher
|
||||
import logging
|
||||
from pyjarowinkler import distance
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
def console_print(text, width=75):
|
||||
@@ -18,8 +25,26 @@ def console_print(text, width=75):
|
||||
print(text)
|
||||
|
||||
|
||||
#TODO: get rid if pyjarowinker dependency
|
||||
def get_similarity(a, b):
|
||||
return SequenceMatcher(None, a, b).ratio()
|
||||
return distance.get_jaro_distance(
|
||||
a, b, winkler=True, scaling = 0.1
|
||||
)
|
||||
|
||||
|
||||
def get_num_options(num):
|
||||
|
||||
while True:
|
||||
choice = input("Enter the number of your choice: ")
|
||||
try:
|
||||
result = int(choice)
|
||||
if result >= 0 and result < num:
|
||||
return result
|
||||
else:
|
||||
print("Error invalid choice. ")
|
||||
except ValueError:
|
||||
print("Error invalid choice. ")
|
||||
|
||||
|
||||
def player_died(text):
|
||||
"""
|
||||
@@ -33,6 +58,7 @@ def player_died(text):
|
||||
"you (die|pass away|perish|suffocate|drown|bleed out)",
|
||||
"you('ve| have) (died|perished|suffocated|drowned|been (killed|slain))",
|
||||
"you (\w* )?(yourself )?to death",
|
||||
"you (\w* )*(collapse|bleed out|chok(e|ed|ing)|drown|dissolve) (\w* )*and (die(|d)|pass away|cease to exist|(\w* )+killed)",
|
||||
]
|
||||
return any(re.search(regexp, lower_text) for regexp in you_dead_regexps)
|
||||
|
||||
@@ -40,14 +66,20 @@ def player_died(text):
|
||||
def player_won(text):
|
||||
lower_text = text.lower()
|
||||
won_phrases = [
|
||||
"you live happily ever after",
|
||||
"you live (forever|eternally|for eternity)",
|
||||
"you (are|become|turn into) (a)? (deity|god)",
|
||||
"you ((go|get) (in)?to|arrive (at|in)) (heaven|paradise)",
|
||||
"you ((\w* )*and |)live happily ever after",
|
||||
"you ((\w* )*and |)live (forever|eternally|for eternity)",
|
||||
"you ((\w* )*and |)(are|become|turn into) ((a|now) )?(deity|god|immortal)",
|
||||
"you ((\w* )*and |)((go|get) (in)?to|arrive (at|in)) (heaven|paradise)",
|
||||
"you ((\w* )*and |)celebrate your (victory|triumph)",
|
||||
"you ((\w* )*and |)retire",
|
||||
]
|
||||
return any(re.search(regexp, lower_text) for regexp in won_phrases)
|
||||
|
||||
|
||||
def remove_profanity(text):
|
||||
return pf.censor(text)
|
||||
|
||||
|
||||
def cut_trailing_quotes(text):
|
||||
num_quotes = text.count('"')
|
||||
if num_quotes % 2 is 0:
|
||||
@@ -79,11 +111,44 @@ def cut_trailing_action(text):
|
||||
or "You ask" in last_line
|
||||
or "you say" in last_line
|
||||
or "You say" in last_line
|
||||
):
|
||||
) and len(lines) > 1:
|
||||
text = "\n".join(lines[0:-1])
|
||||
return text
|
||||
|
||||
|
||||
def clean_suggested_action(result_raw, min_length=4):
|
||||
result_raw = standardize_punctuation(result_raw)
|
||||
|
||||
# The generations actions carry on into the next prompt, so lets remove the prompt
|
||||
results = result_raw.split("\n")
|
||||
results = [s.strip() for s in results]
|
||||
results = [s for s in results if len(s) > min_length]
|
||||
# Sometimes actions are generated with leading > ! . or ?. Likely the model trying to finish the prompt or start an action.
|
||||
result = results[0].strip().lstrip(" >!.?")
|
||||
result = cut_trailing_quotes(result)
|
||||
logger.debug(
|
||||
"full suggested action '%s'. Cropped: '%s'. Split '%s'",
|
||||
result_raw,
|
||||
result,
|
||||
results,
|
||||
)
|
||||
|
||||
# Often actions are cropped with sentance fragment, lets remove. Or we could just turn up config_act["generate-number"]
|
||||
last_punc = max(result.rfind("."), result.rfind("!"), result.rfind("?"))
|
||||
if (last_punc / (len(result) + 1)) > 0.7:
|
||||
result = result[:last_punc]
|
||||
elif last_punc == len(result):
|
||||
pass
|
||||
else:
|
||||
result += "..."
|
||||
|
||||
# Remove you from start
|
||||
result = first_to_second_person(result)
|
||||
result = re.sub('^ ?[Yy]ou ?', '', result)
|
||||
logger.debug("suggested action after cleaning %s", result)
|
||||
return result
|
||||
|
||||
|
||||
def cut_trailing_sentence(text):
|
||||
text = standardize_punctuation(text)
|
||||
last_punc = max(text.rfind("."), text.rfind("!"), text.rfind("?"))
|
||||
|
||||
Reference in New Issue
Block a user