def train_model_task4(path_data, path_label):
data, label = read_input.read_entire_dataset(path_data, path_label)
module_desc = [(nn_modules.Linear, {
'n_in': 28 * 28,
'n_out': 200
}), (nn_modules.Tanh, {}), (nn_modules.Linear, {
'n_in': 200,
'n_out': 10
}), (nn_modules.LogSoftmax, {})]
my_model = model.NNModel(module_desc)
loss = nn_modules.LogCrossEntropyLoss()
train.train(my_model, loss, data, label, 100, 600, 0.1)
return my_model
def test_model(my_model, path_data, path_label):
data, label = read_input.read_entire_dataset(path_data, path_label)
batch_size = 100
accurate_pred = 0
for batch_num in range(0, int(np.ceil(data.shape[0] / float(batch_size)) + 0.1)):
logging.info('Batch %d' % batch_num)
batch_indices = np.arange(batch_num * batch_size, min((batch_num + 1) * batch_size, data.shape[0]))
batch_data = data[batch_indices, :]
batch_label = label[batch_indices]
model_out = my_model.fprop(batch_data)
predictions = np.argmax(model_out, 1)
accurate_batch_pred = np.isclose(predictions, batch_label)
accurate_pred += np.sum(accurate_batch_pred)
print("Errors:", data.shape[0] - accurate_pred)
print("Test accuracy %f" % ((float(accurate_pred) / data.shape[0])))
def train_model_task2(path_data, path_label):
data, label = read_input.read_entire_dataset(path_data, path_label)
# Module description
module_desc = [(nn_modules.Linear, {
'n_in': 28 * 28,
'n_out': 10
}), (nn_modules.Softmax, {})]
# Construct model with given components/modules
my_model = model.NNModel(module_desc)
# Get loss
loss = nn_modules.CrossEntropyLoss()
train.train(my_model,
loss,
data,
label,
nbr_epochs=100,
batch_size=600,
lr=0.1)
return my_model