See this post for more details.
Train a convolutional neural network. Network topology may be specified in a dict passed to the train function.
Below is an example model configuration.
FP_PARAMS = 'params_aug_misclass1.pkl'
BATCH_SIZE = 50
IMAGE_DIM_SIZE = 28
NUM_CLASSES = 10
NUM_FC_UNITS = 1024
MAX_TRAIN_STEPS = 20000
conv_layers = [
{'filter': [5, 5, 1, 32], 'strides': [1] * 4, 'padding': 'SAME', 'act_fn': tf.nn.relu},
{'filter': [5, 5, 32, 64], 'strides': [1] * 4, 'padding': 'SAME', 'act_fn': tf.nn.relu},
]
pool_layers = [
{'filter': [1, 2, 2, 1], 'strides': [1, 2, 2, 1], 'padding': 'SAME'},
{'filter': [1, 2, 2, 1], 'strides': [1, 2, 2, 1], 'padding': 'SAME'},
]
model_options = {'image_dim_size': IMAGE_DIM_SIZE,
'num_features': IMAGE_DIM_SIZE ** 2,
'fp_params': FP_PARAMS,
'conv_layers': conv_layers,
'pool_layers': pool_layers,
'num_fc_units': NUM_FC_UNITS,
'num_classes': NUM_CLASSES,
'batch_size': BATCH_SIZE,
'max_train_steps': MAX_TRAIN_STEPS,
}And we can being training:
mnist = input_data.read_data_sets('MNIST_data', one_hot=True)
train(mnist, model_options)Backwork out spurious yet strongly classified inputs.
Below is an example of additional model information that must be provided.
num_layers = len(model_options['conv_layers']) + 3
cost_factors = np.logspace(-num_layers, -1, num=num_layers)
cost_factors[-1] *= 100
model_options_update = {
'cost_factors': cost_factors,
'save_freq': 1,
'prob_cap_min': 0.999,
'num_examples': 8,
'prev_coeff': 0.5,
'sim_coeff': 0.5,
'num_mono_dec_saves': 5,
}
model_options.update(model_options_update)And to run the backworking model while saving pickled results to file:
mnist = input_data.read_data_sets('MNIST_data', one_hot=True)
for label in range(10):
model_options['label'] = label
results = get_faulty_input_layer(mnist, model_options)
with open('results_{}.pkl'.format(label), 'wb') as f:
pickle.dump(results, f)Augment the training data.
To augment with backworked data saved as results_i.pkl (for label i):
mnist = input_data.read_data_sets('MNIST_data', one_hot=True)
augment_data(mnist)