def activation(x, w_conv, stride, act_conv, w_dense, b_dense, act_dense):
# convolution
tmp = conv.conv_1d(x, w_conv, stride)
tmp = act.dict_activations[act_conv](tmp)
# dense layer
tmp = np.dot(tmp, w_dense) + bias_dense
tmp = act.dict_activations[act_dense](tmp)
return np.sum(tmp)
def activation(x, w_conv_1, stride_1, act_conv_1, w_conv_2, stride_2,
act_conv_2, w_dense_1, b_dense_1, act_dense_1, w_dense_2,
b_dense_2, act_dense_2):
# convolutions
tmp = conv.conv_1d(x, w_conv_1, stride_1)
tmp = act.dict_activations[act_conv_1](tmp)
tmp = conv.conv_1d(tmp, w_conv_2, stride_2)
tmp = act.dict_activations[act_conv_2](tmp)
# dense layers
tmp = np.dot(tmp, w_dense_1) + b_dense_1
tmp = act.dict_activations[act_dense_1](tmp)
tmp = np.dot(tmp, w_dense_2) + b_dense_2
tmp = act.dict_activations[act_dense_2](tmp)
return np.sum(tmp)
def activation(self, input_, accumulate=False):
if accumulate is True:
self.input_ = cp.copy(input_)
output = conv.conv_1d(input_, self.weights, self.stride)
output = act.dict_activations[self.act](output)
self.output = cp.copy(output)
return output
def activation(input_, w_conv, stride, act_conv, w_dense, b_dense, act_dense):
tmp = input_
# convolutions
for i in range(len(w_conv)):
tmp = conv.conv_1d(tmp, w_conv[i], stride[i])
tmp = act.dict_activations[act_conv[i]](tmp)
# dense layers
for i in range(len(w_dense)):
tmp = np.dot(tmp, w_dense[i]) + b_dense[i]
tmp = act.dict_activations[act_dense[i]](tmp)
return tmp
def exp(series, weights, stride):
res = conv.conv_1d(series, weights, stride)
return np.exp(res)
def linear(series, weights, stride):
res = conv.conv_1d(series, weights, stride)
return res
weights = np.random.rand(1, np.random.randint(1, 10))
stride = np.random.randint(1, 10)
weights_c = np.random.rand(1, np.random.randint(1, 10))
stride_c = np.random.randint(1, 10)
### linear activation test ###
# derivative by implementation (chain rule)
derivative_by_implementation = stm.series_to_matrix(
inp, weights.shape[1], stride).T
tmp = np.zeros(shape=(
weights.shape[1],
linear(linear(inp, weights, stride), weights_c, stride_c).shape[1]))
for i in range(derivative_by_implementation.shape[0]):
tmp[i] = conv.conv_1d(derivative_by_implementation[np.newaxis, i],
weights_c, stride_c)
derivative_by_implementation = np.sum(tmp, axis=1)
# derivative by definition
derivative_by_def = list()
epsilon = 1e-5
for i in range(weights.shape[1]):
weights[:, i] += epsilon
f_plus = linear(linear(inp, weights, stride), weights_c, stride_c)
weights[:, i] -= 2 * epsilon
f_minus = linear(linear(inp, weights, stride), weights_c, stride_c)
weights[:, i] += epsilon
derivative_by_def.append(np.sum((f_plus - f_minus) / (2 * epsilon)))
# create the net
net = nn.NN(net_blocks)
# initialize the parameters to random values between [-1, 1]
net.init_parameters(['uniform', -1., 1.])
input_ = np.random.rand(1, net.n_inputs)
tmp = cp.copy(input_)
for i in range(len(net.layers)):
if isinstance(net.layers[i], layer_conv.Conv):
weights = net.layers[i].weights
stride = net.layers[i].stride
tmp = conv.conv_1d(tmp, weights, stride)
tmp = act.dict_activations[net.layers[i].act](tmp)
elif isinstance(net.layers[i], layer_dense.Dense):
weights = net.layers[i].weights
bias = net.layers[i].bias
tmp = np.dot(tmp, weights) + bias
tmp = act.dict_activations[net.layers[i].act](tmp)
output_by_definition = tmp
# calculate net's output
net.activation(input_)
output_by_calculation = cp.copy(net.output)
import convolution as conv
import series_to_matrix as stm
if __name__ == '__main__':
np.random.seed(43) # the answer to everything, plus 1
series = np.random.rand(1, 100)
for k in range(1, 10):
for s in range(1, 10):
kernel = np.random.rand(1, k)
# calculate convolution with different couples of (kernels, stridings)
res = conv.conv_1d(series, kernel, striding=s)
print("Series shape ", series.shape, "\nKernel shape ",
kernel.shape, "\nStride shape ", s, "\nResult shape ",
res.shape)
print("#######\n")
# case where striding is 1 and we have a numpy counterpart function
if s == 1:
# test convolution, should rise exception if convolution is wrong
res_numpy = np.correlate(series[0, :], kernel[0, :])
np.testing.assert_almost_equal(res[0, :], res_numpy)
# test SIMD convolution, should rise exception if convolution is
# wrong