Files
crawl4ai/crawl4ai/train.py
2024-05-17 15:08:03 +08:00

146 lines
4.9 KiB
Python

import spacy
from spacy.training import Example
import random
import nltk
from nltk.corpus import reuters
import torch
def save_spacy_model_as_torch(nlp, model_dir="models/reuters"):
# Extract the TextCategorizer component
textcat = nlp.get_pipe("textcat_multilabel")
# Convert the weights to a PyTorch state dictionary
state_dict = {name: torch.tensor(param.data) for name, param in textcat.model.named_parameters()}
# Save the state dictionary
torch.save(state_dict, f"{model_dir}/model_weights.pth")
# Extract and save the vocabulary
vocab = extract_vocab(nlp)
with open(f"{model_dir}/vocab.txt", "w") as vocab_file:
for word, idx in vocab.items():
vocab_file.write(f"{word}\t{idx}\n")
print(f"Model weights and vocabulary saved to: {model_dir}")
def extract_vocab(nlp):
# Extract vocabulary from the SpaCy model
vocab = {word: i for i, word in enumerate(nlp.vocab.strings)}
return vocab
nlp = spacy.load("models/reuters")
save_spacy_model_as_torch(nlp, model_dir="models")
def train_and_save_reuters_model(model_dir="models/reuters"):
# Ensure the Reuters corpus is downloaded
nltk.download('reuters')
nltk.download('punkt')
if not reuters.fileids():
print("Reuters corpus not found.")
return
# Load a blank English spaCy model
nlp = spacy.blank("en")
# Create a TextCategorizer with the ensemble model for multi-label classification
textcat = nlp.add_pipe("textcat_multilabel")
# Add labels to text classifier
for label in reuters.categories():
textcat.add_label(label)
# Prepare training data
train_examples = []
for fileid in reuters.fileids():
categories = reuters.categories(fileid)
text = reuters.raw(fileid)
cats = {label: label in categories for label in reuters.categories()}
# Prepare spacy Example objects
doc = nlp.make_doc(text)
example = Example.from_dict(doc, {'cats': cats})
train_examples.append(example)
# Initialize the text categorizer with the example objects
nlp.initialize(lambda: train_examples)
# Train the model
random.seed(1)
spacy.util.fix_random_seed(1)
for i in range(5): # Adjust iterations for better accuracy
random.shuffle(train_examples)
losses = {}
# Create batches of data
batches = spacy.util.minibatch(train_examples, size=8)
for batch in batches:
nlp.update(batch, drop=0.2, losses=losses)
print(f"Losses at iteration {i}: {losses}")
# Save the trained model
nlp.to_disk(model_dir)
print(f"Model saved to: {model_dir}")
def train_model(model_dir, additional_epochs=0):
# Load the model if it exists, otherwise start with a blank model
try:
nlp = spacy.load(model_dir)
print("Model loaded from disk.")
except IOError:
print("No existing model found. Starting with a new model.")
nlp = spacy.blank("en")
textcat = nlp.add_pipe("textcat_multilabel")
for label in reuters.categories():
textcat.add_label(label)
# Prepare training data
train_examples = []
for fileid in reuters.fileids():
categories = reuters.categories(fileid)
text = reuters.raw(fileid)
cats = {label: label in categories for label in reuters.categories()}
doc = nlp.make_doc(text)
example = Example.from_dict(doc, {'cats': cats})
train_examples.append(example)
# Initialize the model if it was newly created
if 'textcat_multilabel' not in nlp.pipe_names:
nlp.initialize(lambda: train_examples)
else:
print("Continuing training with existing model.")
# Train the model
random.seed(1)
spacy.util.fix_random_seed(1)
num_epochs = 5 + additional_epochs
for i in range(num_epochs):
random.shuffle(train_examples)
losses = {}
batches = spacy.util.minibatch(train_examples, size=8)
for batch in batches:
nlp.update(batch, drop=0.2, losses=losses)
print(f"Losses at iteration {i}: {losses}")
# Save the trained model
nlp.to_disk(model_dir)
print(f"Model saved to: {model_dir}")
def load_model_and_predict(model_dir, text, tok_k = 3):
# Load the trained model from the specified directory
nlp = spacy.load(model_dir)
# Process the text with the loaded model
doc = nlp(text)
# gee top 3 categories
top_categories = sorted(doc.cats.items(), key=lambda x: x[1], reverse=True)[:tok_k]
print(f"Top {tok_k} categories:")
return top_categories
if __name__ == "__main__":
train_and_save_reuters_model()
train_model("models/reuters", additional_epochs=5)
model_directory = "reuters_model_10"
print(reuters.categories())
example_text = "Apple Inc. is reportedly buying a startup for $1 billion"
r =load_model_and_predict(model_directory, example_text)
print(r)