def train_comparison(args, theta, dataset):
hyper_params = {k: args[k] for k in ['b_size']}
hyper_params['to_fix'] = [
] # a selection of parameters can be fixed, e.g. the word embeddings
# initialize optimizer with learning rate (other hyperparams: default values)
opt = args['optimizer']
if opt == 'adagrad':
optimizer = Optimizer.Adagrad(theta,
lr=args['learningrate'],
lambda_l2=args['lambda'])
elif opt == 'adam':
optimizer = Optimizer.Adam(theta,
lr=args['learningrate'],
lambda_l2=args['lambda'])
elif opt == 'sgd':
optimizer = Optimizer.SGD(theta,
lr=args['learningrate'],
lambda_l2=args['lambda'])
else:
raise RuntimeError("No valid optimizer chosen")
# train model
evals = plain_train(optimizer,
dataset,
hyper_params,
n_epochs=args['n_epochs'],
outdir=args['out_dir'])
# store learned model
store_theta(
theta, os.path.join(args['out_dir'], 'comparisonFinalModel.theta.pik'))
# run final evaluation
for name, tb in dataset.iteritems():
print('Evaluation on ' + name + ' data (' + str(len(tb.examples)) +
' examples)')
tb.evaluate(optimizer.theta, verbose=1)
# create convergence plot
for name, eval in evals.items():
toplot = [e[key] for e in eval for key in e if 'loss' in key]
plt.plot(xrange(len(toplot)), toplot, label=name)
plt.legend()
plt.title([key for key in eval[0].keys() if 'loss' in key][0])
plt.savefig(os.path.join(args['out_dir'], 'comparisonConvergence.png'))
m = 2000
case = "quadratic"
nclass = 3
X, Y = example_classification.example(nfeature, m, case, nclass)
# (2) Define model
model = NeuralNetwork.NeuralNetwork(nfeature)
model.add_layer(11, "tanh")
model.add_layer(9, "tanh")
model.add_layer(6, "tanh")
model.add_layer(3, "tanh")
model.add_layer(nclass, "softmax")
# (3) Compile model
#optimizer = Optimizer.GradientDescent(0.3)
#optimizer = Optimizer.Momentum(0.3,0.9)
#optimizer = Optimizer.RmsProp(0.02,0.9,1e-8)
#optimizer = Optimizer.Adam(0.02,0.9,0.999,1e-8)
optimizer = Optimizer.Adagrad(0.1, 1e-8)
model.compile("crossentropy", optimizer)
# (4) Train model
epochs = 100
time_start = time.time()
history = model.fit(X, Y, epochs, batch_size=1000)
time_end = time.time()
print("Train time: {}".format(time_end - time_start))
# (5) Results
# plot loss and accuracy
plot_results.plot_results_history(history, ["loss"])
plot_results.plot_results_history(history, ["accuracy"])
# plot heatmap in x0-x1 plane
plot_results.plot_results_classification((X, Y), model, nclass)
plt.show()
