It's really challenging!!!

I'm just feeling that: When neural network goes deep into code, you have to go back to mathematics.

Implement a neural network framework from scratch, and train with 2 examples:

• MNIST Classifier
• Time Series Prediction

• Back Propagation Algorithm
• Model: Sequential(seq)
• Layers: dense(fc), conv2d, flatten, maxpooling2d, dropout
• Activations: relu, sigmoid, softmax
• Optimizor & Training
  • Initial learning rate 0.001, set to 5e-5 after loss < 1.0
  • SGD(with momentum)
  • Gradients Clipping
  • l2 Regularization

• Input:(None, 28, 28, 1)
• Conv2D(relu)
• Flatten
• Dropout(0.5)
• Full-Connect(softmax)

85% accuracy after 25w samples with a simple network above.

Actually I had achieved 90% accuracy, but somehow I lost it. It's obviously that, complex model with high accuracy and low loss is harder to train, especially in Python.

• Ubuntu 16.04.2 LTS
• Python 2.7.12(Python 3.6 Also works)

• train MNIST

python train_mnist.py

• test MNIST

python test_mnist.py --weight ./weights/weights-252160-0.797545-0.8575

• inference hand written digits

python inference_mnist.py --dir ./datasets/pics --weight ./weights/weights-252160-0.797545-0.8575

Training dataset goes like this:

     X             y
1 0 0 ... 0      1  0
0 0 0 ... 0      0  1
0 0 0 ... 0      0  1
1 0 0 ... 0      1  0
1 0 0 ... 0      1  0

Pattern is obviously, right?

• Input:(None, TIME_STEPS, INPUT_DIM)
• RNN(RNN_UNITS)
• Full-Connect(2, with softmax)

python train_rnn.py