def train_model_and_plot_stats(model,
error,
learning_rule,
train_data,
valid_data,
num_epochs,
stats_interval,
notebook=True):
# As well as monitoring the error over training also monitor classification
# accuracy i.e. proportion of most-probable predicted classes being equal to targets
data_monitors = {'acc': lambda y, t: (y.argmax(-1) == t.argmax(-1)).mean()}
# Use the created objects to initialise a new Optimiser instance.
optimiser = Optimiser(model,
error,
learning_rule,
train_data,
valid_data,
data_monitors,
notebook=notebook)
# Run the optimiser for 5 epochs (full passes through the training set)
# printing statistics every epoch.
stats, keys, run_time = optimiser.train(num_epochs=num_epochs,
stats_interval=stats_interval)
return stats, keys
def train_model_and_plot_stats(
model, error, learning_rule, train_data, valid_data, num_epochs, stats_interval, notebook=True):
# As well as monitoring the error over training also monitor classification
# accuracy i.e. proportion of most-probable predicted classes being equal to targets
data_monitors={'acc': lambda y, t: (y.argmax(-1) == t.argmax(-1)).mean()}
# Use the created objects to initialise a new Optimiser instance.
optimiser = Optimiser(
model, error, learning_rule, train_data, valid_data, data_monitors, notebook=notebook)
# Run the optimiser for 5 epochs (full passes through the training set)
# printing statistics every epoch.
stats, keys, run_time = optimiser.train(patience=10, max_num_epochs=num_epochs, stats_interval=stats_interval)
# # Plot the change in the validation and training set error over training.
# fig_1 = plt.figure(figsize=(8, 4))
# ax_1 = fig_1.add_subplot(111)
# for k in ['error(train)', 'error(valid)']:
# ax_1.plot(np.arange(1, stats.shape[0]) * stats_interval,
# stats[1:, keys[k]], label=k)
# ax_1.legend(loc=0)
# ax_1.set_xlabel('Epoch number')
# # Plot the change in the validation and training set accuracy over training.
# fig_2 = plt.figure(figsize=(8, 4))
# ax_2 = fig_2.add_subplot(111)
# for k in ['acc(train)', 'acc(valid)']:
# ax_2.plot(np.arange(1, stats.shape[0]) * stats_interval,
# stats[1:, keys[k]], label=k)
# ax_2.legend(loc=0)
# ax_2.set_xlabel('Epoch number')
return stats, keys, #run_time, fig_1, ax_1, fig_2, ax_2
# to ensure reproducibility of results
rng.seed(seed)
train_data.reset()
valid_data.reset()
print('Regularisation: {0}'.format(weights_penalty))
model = MultipleLayerModel([
AffineLayer(input_dim, hidden_dim, weights_init, biases_init,
weights_penalty),
ReluLayer(),
AffineLayer(hidden_dim, hidden_dim, weights_init, biases_init,
weights_penalty),
ReluLayer(),
AffineLayer(hidden_dim, output_dim, weights_init, biases_init,
weights_penalty)
])
optimiser = Optimiser(model, error, learning_rule, train_data, valid_data,
data_monitors)
run_info[weights_penalty] = optimiser.train(num_epochs, stats_interval)
models[weights_penalty] = model
# In[5]:
import matplotlib.pyplot as plt
get_ipython().magic(u'matplotlib inline')
plt.style.use('ggplot')
fig = plt.figure(figsize=(12, 6))
ax1 = fig.add_subplot(1, 2, 1)
ax2 = fig.add_subplot(1, 2, 2)
for weight_penalty, run in run_info.items():
stats, keys, run_time = run
from mlp.optimisers import SGDOptimiser
from support import train_dp_flat, valid_dp_flat, create_one_hid_model, rng
from mlp.costs import CECost
from mlp.layers import MLP, Softmax, Sigmoid
from mlp.schedulers import LearningRateFixed
from noise_tries import show_mnist_images
import numpy as np
from mlp.dataset import MNISTDataProvider
from mlp.optimisers import Optimiser
tsk3_2_optimiser = Optimiser()
tsk3_2_model = create_one_hid_model()
def tsk_3_2_draw_handler(cur_layer_id, cur_model, get_inputs):
if cur_layer_id != 0:
return
mnist_dp = MNISTDataProvider(dset='valid',
batch_size=4,
max_num_examples=4,
randomize=False)
for batch in mnist_dp:
features, targets = batch
inputs, pure = get_inputs(features)
output_dc = cur_model.fprop(inputs)
num_imgs = features.shape[0]
imgs = features.reshape(num_imgs, 28, 28)
#images noisy
imgs_ns = inputs.reshape(num_imgs, 28, 28)
#images decoded