These notebooks are used in Udacity\Google course :
- 1_notmnist.ipynb
- 2_fullyconnected.ipynb
- 3_regularization.ipynb
- 4_convolutions.ipynb
- 5_word2vec.ipynb
- 6_lstm.ipynb
###Architecture & scores summary :
1_notmnist :
- Logistic regression (this one via scikit-learn) => 87% score on test dataset
2_fullyconnected :
- Logistic regression => 83%
- Logistic regression with SGD => 86%
- Fully connected with 1 hidden layer and ReLU => 88%
3_regularization :
- Adding L2 regularization to Logistic Regression => 88%
- Adding L2 regularization to NN => 92%
- Replacing L2 regul by dropout on NN => 87%
4_convolutions :
- Conv > ReLU > Conv > ReLU > FullyConnected > ReLU > FullyConnected => 90%
- Conv > MaxPool > ReLU > Conv > MaxPool > ReLU > FullyConnected > ReLU > FullyConnected => 89%
- Conv > ReLU > MaxPool > Conv > ReLU > MaxPool > FullyConnected > ReLU > FullyConnected => 91%
- Conv > ReLU > MaxPool > Conv > ReLU > MaxPool > DropOut > FullyConnected > ReLU > DropOut > FullyConnected => 94.2%
Karpathy_Simple_RNN.ipynb is a minimal RNN using the input.txt dataset.
Simple_TF_1.ipynb is a trivial tensorflow classification example.