def mlp_from_cascade(layers, loss):
"""From a list of layers, build up an mlp with the given loss."""
# Determine transfer functions.
transfers = [l.feature_transfer for l in layers[:-1]]
out_transfer = layers[-1].out_transfer
# Determine sizes.
sizes = [l.n_feature for l in layers[:-1]]
n_inpt = layers[0].n_inpt
n_output = layers[-1].n_output
# Build up mlp.
mlp = MultiLayerPerceptron(n_inpt, sizes, n_output, transfers, out_transfer, loss)
mlp.parameters.data[:] = np.zeros(mlp.parameters.data.shape)
# Transfer weights and biases.
weights, biases = [], []
for l in layers:
w, b = get_affine_parameters(l)
weights.append(w)
biases.append(b)
weight_names = ["in_to_hidden"] + ["hidden_to_hidden_%i" % i for i in range(len(weights) - 2)] + ["hidden_to_out"]
bias_names = ["hidden_bias_%i" % i for i in range(len(biases) - 1)] + ["out_bias"]
for wname, bname, w, b in zip(weight_names, bias_names, weights, biases):
mlp.parameters[wname][:] = w
mlp.parameters[bname][:] = b
return mlp
def test_mlp2():
l = MultiLayerPerceptron(2, [10, 12], 3, ['tanh', 'sigmoid'], 'softabs', 'squared')
f = l.function(['inpt', 'target'], 'loss', mode='FAST_COMPILE')
grad = T.grad(l.exprs['loss'], l.parameters.flat)
fprime = l.function(['inpt', 'target'], grad, mode='FAST_COMPILE')
f(np.random.random((10, 2)), np.random.random((10, 3)))
fprime(np.random.random((10, 2)), np.random.random((10, 3)))
def test_mlp1():
l = MultiLayerPerceptron(2, [10], 3, ['tanh'], 'softabs', 'squared')
f = l.function(['inpt', 'target'], 'loss', mode='FAST_COMPILE')
grad = T.grad(l.exprs['loss'], l.parameters.flat)
fprime = l.function(['inpt', 'target'], grad, mode='FAST_COMPILE')
f(np.random.random((10, 2)), np.random.random((10, 3)))
fprime(np.random.random((10, 2)), np.random.random((10, 3)))
def mlp_from_cascade(layers, loss):
"""From a list of layers, build up an mlp with the given loss."""
# Determine transfer functions.
transfers = [l.feature_transfer for l in layers[:-1]]
out_transfer = layers[-1].out_transfer
# Determine sizes.
sizes = [l.n_feature for l in layers[:-1]]
n_inpt = layers[0].n_inpt
n_output = layers[-1].n_output
# Build up mlp.
mlp = MultiLayerPerceptron(n_inpt, sizes, n_output, transfers,
out_transfer, loss)
mlp.parameters.data[:] = np.zeros(mlp.parameters.data.shape)
# Transfer weights and biases.
weights, biases = [], []
for l in layers:
w, b = get_affine_parameters(l)
weights.append(w)
biases.append(b)
weight_names = (
['in_to_hidden'] +
['hidden_to_hidden_%i' % i
for i in range(len(weights) - 2)] + ['hidden_to_out'])
bias_names = ['hidden_bias_%i' % i
for i in range(len(biases) - 1)] + ['out_bias']
for wname, bname, w, b in zip(weight_names, bias_names, weights, biases):
mlp.parameters[wname][:] = w
mlp.parameters[bname][:] = b
return mlp
