pytrainer 1.0.0

AI made easy. Based on pytorch.

Pytrainer

AI made Easy.

Tutorials

Train and Save a Model

from pytrainer.text import TextAI

# Create an instance of TextAI
ai = TextAI()

# Define some simple training data (input-output pairs)
training_data = {
    "Hello": "Hi there!",
    "How are you?": "I'm good, thank you!",
    "What's your name?": "I am TextAI."
}

# Train the model
ai.train(training_data, epochs=10, batch_size=1, learning_rate=0.001)

# Save the trained model
ai.save('trained_model.json')

Load and Prompt the Model

from pytrainer.text import TextAI

# Create a new instance of TextAI and load the trained model
ai_new = TextAI()
ai_new.load('trained_model.json')

# Generate a response to a new prompt
response = ai_new.prompt("What's your name?")
print(response)

Classes

TextAI

from pytrainer.text import TextAI

train(self, data, epochs=10, batch_size=1, learning_rate=0.001)

• Arguments:
  • data: A dictionary of input-output pairs for training (required).
  • epochs: The number of training epochs (default is 10).
  • batch_size: The batch size used during training (default is 1).
  • learning_rate: The learning rate for the optimizer (default is 0.001).

prompt(self, input_text, max_length=50, temperature=0.5)

• Arguments:
  • input_text: The text prompt to generate a response for (required).
  • max_length: The maximum length of the generated response (default is 50).
  • temperature: Controls the randomness of the generated text (default is 0.7).
  • nearestPunctuation: Should a word be rounded to the same word with different punctuation, if that word is not tokenized? (defualt is False)

save(self, path)

• Arguments:
  • path: The file path to save the model data (required).

load(self, path)

• Arguments:
  • path: The file path from which to load the model data (required).

Known Issues:

(Fixed errors will be removed next update, and moved to change log in github.)

• Exploding gradients during training of over 30 epochs, resulting in NaN errors when prompting. FIXED
• When prompting, using an un-tokenized word will fail. FIXED
  • Note: There is a new option to map un-tokenized words to the same word with different punctation, if one was tokenized. Otherwise, it will return nothing.

Planned Features:

• Individual classes for Tokenizers, allowing them to be customized.
• Option for built-in tokenizers, so the trained model is backward compatible with pytorch.