def test_softmax0(self):
# is this a correct test? Should the gradients always be this small?
my_graph = ad.Softmax(self.my_w4)
tf_graph = tf.nn.softmax(self.tf_w4)
wrt_vars = [self.my_w4]
tf_vars = [self.tf_w4]
utils.custom_test(self, my_graph, wrt_vars, tf_graph, tf_vars)
def sample_char(
logits): # sample a character from the multinomial distribution
next_char_onehot = np.random.multinomial(
n=1, pvals=ad.Softmax(logits / temperature)()[0])
next_char = text_loader.ind_to_char[np.argmax(next_char_onehot)]
next_char_onehot = np.expand_dims(next_char_onehot, axis=0)
return next_char_onehot, next_char
def test_softmax2(self):
my_graph = ad.Softmax(self.my_w6)
tf_graph = tf.nn.softmax(self.tf_w6)
wrt_vars = [self.my_w6]
tf_vars = [self.tf_w6]
utils.custom_test(self, my_graph, wrt_vars, tf_graph, tf_vars)
def network_output(x):
probs = ad.Softmax(nn(x))
return probs()