Deep Learning Architectures β€” pesos (versoes reduzidas)

Pesos treinados do projeto portfolio/projects/deep-learning-architectures, em versoes reduzidas para rodar em CPU.

Arquivo Arquitetura Dataset Resultado
cnn_cifar10.keras CNN CIFAR-10 val_accuracy 0.666 (5 epocas; original usa 15)
lstm_spam.keras + lstm_spam_tokenizer.pkl LSTM SMS Spam Collection val_accuracy 0.9641
gan_generator.keras / gan_discriminator.keras GAN (DCGAN) MNIST 3 epocas (original usa 100)
gan_samples.png β€” β€” amostras geradas pelo generator

Como carregar

import tensorflow as tf
modelo = tf.keras.models.load_model("cnn_cifar10.keras")

LSTM (o tokenizador e manual, pois o Keras 3 removeu tf.keras.preprocessing):

import pickle, tensorflow as tf
tok = pickle.load(open("lstm_spam_tokenizer.pkl", "rb"))
word2idx, maxlen = tok["word2idx"], tok["maxlen"]
modelo = tf.keras.models.load_model("lstm_spam.keras")

Notas

  • Receitas reduzidas de proposito (menos epocas/camadas) para treinar em CPU; os notebooks originais estao em portfolio/projects/deep-learning-architectures/notebooks.
  • O Transformer (notebook 04, PTβ†’EN) exige tensorflow-datasets, tensorflow_text e dataset web-archive: use o notebook original (Colab). Nao ha pesos aqui.
  • Datasets: CIFAR-10 e MNIST (Keras), SMS Spam Collection via mirror TSV do pycon-2016-tutorial.
Downloads last month
124
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support