def test_gradient_finite_differences(net, input_data):
# ######## calculate deltas ##########
net.provide_external_data({"default": input_data})
net.forward_pass(training_pass=True)
net.backward_pass()
gradient_calc = net.buffer.gradients
# ######## estimate deltas ##########
def f(x):
net.buffer.parameters[:] = x
net.forward_pass()
return net.get_loss_values()["total_loss"]
initial_weigths = net.buffer.parameters.copy()
gradient_approx = approx_fprime(initial_weigths, f, 1e-6)
# ######## compare them #############
nr_sequences = input_data.shape[1]
diff = gradient_approx - gradient_calc
mse = np.sum(diff ** 2) / nr_sequences
if mse > 1e-4:
# Hijack the network gradient buffer for the view
net.buffer.gradients[:] = diff
for layer_name in net.layers:
if not net.buffer[layer_name]:
continue
print("============= Layer: {} =============".format(layer_name))
for view_name in net.buffer[layer_name].gradients.keys():
print("------------- {} -------------".format(view_name))
print(net.buffer[layer_name].gradients[view_name])
# print(">> Checking Gradient = %0.4f" % mse)
assert mse < 1e-4
def test_deltas_finite_differences(net, input_data):
# ######## calculate deltas ##########
net.provide_external_data({"default": input_data})
net.forward_pass(training_pass=True)
net.backward_pass()
delta_calc = net.buffer.Input.output_deltas.default.flatten()
# ######## estimate deltas ##########
def f(x):
net.provide_external_data({"default": x.reshape(input_data.shape)})
net.forward_pass()
return net.get_loss_values()["total_loss"]
delta_approx = approx_fprime(input_data.copy().flatten(), f, 1e-5)
# ######## compare them #############
nr_sequences = input_data.shape[1]
mse = np.sum((delta_approx - delta_calc) ** 2) / nr_sequences
if mse > 1e-4:
diff = (delta_approx - delta_calc).reshape(input_data.shape)
for t in range(diff.shape[0]):
print("======== t=%d =========" % t)
print(diff[t])
# print("Checking Deltas = %0.4f" % mse)
assert mse < 1e-4
def test_gradient_finite_differences(net, input_data):
# ######## calculate deltas ##########
net.provide_external_data({'default': input_data})
net.forward_pass(training_pass=True)
net.backward_pass()
gradient_calc = net.buffer.gradients
# ######## estimate deltas ##########
def f(x):
net.buffer.parameters[:] = x
net.forward_pass()
return net.get_loss_values()['total_loss']
initial_weigths = net.buffer.parameters.copy()
gradient_approx = approx_fprime(initial_weigths, f, 1e-6)
# ######## compare them #############
nr_sequences = input_data.shape[1]
diff = gradient_approx - gradient_calc
mse = np.sum(diff ** 2) / nr_sequences
if mse > 1e-4:
# Hijack the network gradient buffer for the view
net.buffer.gradients[:] = diff
for layer_name in net.layers:
if not net.buffer[layer_name]:
continue
print("============= Layer: {} =============".format(layer_name))
for view_name in net.buffer[layer_name].gradients.keys():
print("------------- {} -------------".format(view_name))
print(net.buffer[layer_name].gradients[view_name])
# print(">> Checking Gradient = %0.4f" % mse)
assert mse < 1e-4
def test_deltas_finite_differences(net, input_data):
# ######## calculate deltas ##########
net.provide_external_data({'default': input_data})
net.forward_pass(training_pass=True)
net.backward_pass()
delta_calc = net.buffer.Input.output_deltas.default.flatten()
# ######## estimate deltas ##########
def f(x):
net.provide_external_data({'default': x.reshape(input_data.shape)})
net.forward_pass()
return net.get_loss_values()['total_loss']
delta_approx = approx_fprime(input_data.copy().flatten(), f, 1e-5)
# ######## compare them #############
nr_sequences = input_data.shape[1]
mse = np.sum((delta_approx - delta_calc) ** 2) / nr_sequences
if mse > 1e-4:
diff = (delta_approx - delta_calc).reshape(input_data.shape)
for t in range(diff.shape[0]):
print("======== t=%d =========" % t)
print(diff[t])
# print("Checking Deltas = %0.4f" % mse)
assert mse < 1e-4
def test_elementwise_act_func_gradients():
pairs_to_test = [
(HANDLER.sigmoid, HANDLER.sigmoid_deriv),
(HANDLER.tanh, HANDLER.tanh_deriv),
(HANDLER.rel, HANDLER.rel_deriv),
]
test_shape = (3, 2, 4)
for fwd, bwd in pairs_to_test:
inputs = HANDLER.create_from_numpy(np.random.randn(*test_shape))
outputs = HANDLER.zeros(test_shape)
doutputs = HANDLER.ones(test_shape)
dinputs = HANDLER.zeros(test_shape)
fwd(inputs, outputs)
bwd(inputs, outputs, doutputs, dinputs)
grad_calc = HANDLER.get_numpy_copy(dinputs)
size = inputs.size
x0 = HANDLER.get_numpy_copy(inputs).reshape((size,))
def f(x):
flat_inputs = inputs.reshape((size,))
HANDLER.set_from_numpy(flat_inputs, x)
HANDLER.fill(outputs, 0.0)
fwd(inputs, outputs)
return HANDLER.get_numpy_copy(outputs).sum()
grad_approx = approx_fprime(x0, f, 1e-5).reshape(grad_calc.shape)
close = np.allclose(grad_approx, grad_calc, rtol=1e-4, atol=1e-4)
if not close:
print("-----------------------------")
print("Testing", fwd.__name__)
print("-- Approximated Gradient ----")
print(grad_approx)
print("---- Calculated Gradient ----")
print(grad_calc)
print("------------- Difference ----")
print(grad_approx - grad_calc)
assert close
def test_elementwise_act_func_gradients():
pairs_to_test = [(HANDLER.sigmoid, HANDLER.sigmoid_deriv),
(HANDLER.tanh, HANDLER.tanh_deriv),
(HANDLER.rel, HANDLER.rel_deriv)]
test_shape = (3, 2, 4)
for fwd, bwd in pairs_to_test:
inputs = HANDLER.create_from_numpy(np.random.randn(*test_shape))
outputs = HANDLER.zeros(test_shape)
doutputs = HANDLER.ones(test_shape)
dinputs = HANDLER.zeros(test_shape)
fwd(inputs, outputs)
bwd(inputs, outputs, doutputs, dinputs)
grad_calc = HANDLER.get_numpy_copy(dinputs)
size = inputs.size
x0 = HANDLER.get_numpy_copy(inputs).reshape((size, ))
def f(x):
flat_inputs = inputs.reshape((size, ))
HANDLER.set_from_numpy(flat_inputs, x)
HANDLER.fill(outputs, 0.)
fwd(inputs, outputs)
return HANDLER.get_numpy_copy(outputs).sum()
grad_approx = approx_fprime(x0, f, 1e-5).reshape(grad_calc.shape)
close = np.allclose(grad_approx, grad_calc, rtol=1e-4, atol=1e-4)
if not close:
print("-----------------------------")
print("Testing", fwd.__name__)
print('-- Approximated Gradient ----')
print(grad_approx)
print('---- Calculated Gradient ----')
print(grad_calc)
print('------------- Difference ----')
print(grad_approx - grad_calc)
assert close
