def call(self, input, training=False):
weyl_base = weyl(input)
p, E = input
# selecting the right potential
p = p[:,:,self.pchoice[0]:self.pchoice[1]]
# reshape E to have the same shape as p
resized_E = (E[:,:,tf.newaxis]+0*p)
resized_E = resized_E[:,:,:1]
p = tf.concat([p, resized_E], axis=2)
#print(0, p.shape)
p = self.decider(p, training=training)
#print(1, p.shape)
out = [self.conv_down(model(p, training=training)[:,0,:]) for model in self.counters]
p = tf.concat(out, axis=1)
#print(2, p.shape)
#print(p.shape)
box_size = tf.minimum(1/tf.math.sqrt(E), 10000)
E = tf.concat([E, box_size], axis=1)
#print(3, E.shape)
E = self.weighters(E, training=training)
#print(4, E.shape)
#print(E.shape)
p = p*E
#print(5, p.shape)
p = self.out_layer(p)
return tf.nn.softplus(weyl_base + p)
def call(self, input, training=False):
weyl_base = weyl(input)
p, E = input
# selecting the right potential
p = p[:,:,self.pchoice[0]:self.pchoice[1]]
# reshape E to have the same shape as p
resized_E = (E[:,:,tf.newaxis]+0*p)
resized_E = resized_E[:,:,:1]
# pE = p and E
pE = tf.concat([p, resized_E], axis=2)
# compute features - this is the F(dW, W) in front of \int f_{FD}'
p = self.res(p)+p
p = self.emb(p, training=training)
# conv down potential
pE = self.conv1(pE)
# features as weight * conved down potential = F(dW, W) \int f_{FD}''
#print(p.shape, pE.shape)
p = pE*p
# weighted sum = integration
p = self.weighters(p, training=training)
pp = self.out_layer1(p)
p = self.out_layer2(pp)+p # residual layer
#print(p.shape)
return tf.nn.softplus(weyl_base + p)
def call(self, input):
weyl_base = weyl(input)
p, E = input
# selecting the right potential
p = p[:, :, self.pchoice[0]]
p = p - E
p = self.out_layer(p)
return tf.nn.softplus(weyl_base + p)
def call(self, input):
weyl_base = weyl(input)
p, E = input
# selecting the right potential
p = p[:, :, self.pchoice[0]:self.pchoice[1]]
E = E[:, :, tf.newaxis]
E = 0.0 * p + E
p = tf.concat([p, E], axis=2)
p = self.internal(p)
return tf.nn.softplus(weyl_base + p)
def call(self, input, training=False):
p, E = input
weyl_base = weyl((p, E))
# selecting the right potential
# p.shape = batch_size x 1 x domain_size x channel_size
p = p[:, tf.newaxis, :, self.pchoice[0]:self.pchoice[1]]
# reshape E to have the same shape as p
# resized_E.shape = batch_size x 1 x domain_sze x 1
resized_E = (E[:, tf.newaxis, :, tf.newaxis] + 0 * p)[:, :, :, :1]
#resized_E = resized_E[:,:,:1,:]
T = np.arange(self.momentum_size) / (self.momentum_size - 1)
T = T.reshape(1, self.momentum_size, 1, 1)
T = T**2
T = T * self.momentum_max
# phase space Hamiltonian
p = T + p - resized_E
p = self.decider(p, training=training)
#print(1, p.shape)
out = [
self.conv_down(model(p, training=training)[:, 0, 0, :])
for model in self.counters
]
p = tf.concat(out, axis=1)
#print(2, p.shape)
#print(p.shape)
box_size = tf.minimum(1 / tf.math.sqrt(E), 10000)
E = tf.concat([E, box_size], axis=1)
#print(3, E.shape)
E = self.weighters(E, training=training)
#print(4, E.shape)
#print(E.shape)
p = p * E
#print(5, p.shape)
p = self.out_layer(p)
res = tf.nn.softplus(weyl_base + p)
print(res.shape)
return res
def plot_model_pred(model, data_set_name, index, data_folder_path="data"):
p, E, target = load_data_set_(data_folder_path, data_set_name, [index])
d = {}
info_file = os.path.join(data_folder_path,
data_set_name + "_" + str(index) + "_info.txt")
with open(info_file) as f:
line = f.readline()
while line:
k, v = line.split(':', 1)
d[k] = v
line = f.readline()
#print(nve[:,0].shape, nve[:,1].shape)
fig = plt.figure(figsize=(18, 16), dpi=50, facecolor='w', edgecolor='k')
plt.scatter(E, target, color="blue", label="True Eigenvalue counting")
weyl_base = weyl((p, E)).numpy().astype(np.float32)
#print('weyl shape', weyl_base.shape)
plt.scatter(E, weyl_base, color="gray", label="Prediction by weyl's law")
pred = model((p, E))
#print('pred shape', pred.shape)
plt.scatter(E,
pred,
color="orange",
label="Prediction by {}".format(model.name))
#print(d)
subtitle = ""
for k in d:
if k != 'potential type':
subtitle += k + ":" + d[k].rstrip('\n') + '; '
plt.title(subtitle)
plt.suptitle(d['potential type'], fontsize=14, fontweight='bold')
plt.xlabel("Energy")
plt.ylabel("Eigenvalue Count")
plt.show()
def call(self, input):
weyl_base = weyl(input)
p, E = input
p = tf.concat([p, p - E[:, :, tf.newaxis]], axis=2)
#print(0, p.shape)
p = self.decider(p)
#print(1, p.shape)
out = [self.conv_down(model(p)[:, 0, :]) for model in self.counters]
p = tf.concat(out, axis=1)
#print(2, p.shape)
box_size = tf.minimum(1 / tf.math.sqrt(E), 10000)
E = tf.concat([E, box_size], axis=1)
#print(3, E.shape)
E = self.weighters(E)
#print(4, E.shape)
p = p * E
#print(5, p.shape)
p = self.out_layer(p)
#print(p.shape)
return tf.nn.relu(weyl_base + p)
