102 lines
3.2 KiB
Python
102 lines
3.2 KiB
Python
#%%
|
|
import pandas as pd
|
|
import torch
|
|
import numpy as np
|
|
|
|
#%%
|
|
df = pd.read_csv('../temp/2body_data.csv').sort_values(by='c')
|
|
df.loc[df['re_E'] < 0, 'im_E'] = 0 # set im_E = 0 for bound states (to avoid square root issues)
|
|
df['E'] = df['re_E'] + 1j * df['im_E']
|
|
df['k'] = np.sqrt(df['E'])
|
|
|
|
train_data = df[df['re_E'] < 0]
|
|
target_data = df[df['re_E'] > 0]
|
|
|
|
train_cs = train_data['c'].to_numpy()
|
|
train_ks = torch.tensor(train_data['k'].to_numpy(), dtype=torch.complex128)
|
|
|
|
#%%
|
|
# hyperparameters
|
|
N = 5
|
|
|
|
# initialize random Hamiltonians
|
|
H0 = torch.randn(N, N, dtype=torch.complex128)
|
|
H0 = (H0 + torch.transpose(H0, 0, 1)).requires_grad_() # symmetric
|
|
H1 = torch.randn(N, N, dtype=torch.complex128)
|
|
H1 = (H1 + torch.transpose(H1, 0, 1)).requires_grad_() # symmetric
|
|
|
|
#%%
|
|
# training
|
|
|
|
# generate a set of c values to follow by subdividing the training cs
|
|
subdivisions = 2
|
|
c_steps = np.concatenate([np.linspace(start, stop, subdivisions, endpoint=False) for (start, stop) in zip(train_cs, train_cs[1:])])
|
|
c_steps = np.append(c_steps, train_cs[-1])
|
|
|
|
lr = 0.01
|
|
epochs = 100000
|
|
for epoch in range(epochs):
|
|
ks = torch.empty(len(train_data), dtype=torch.complex128)
|
|
current_k = 0.0 # start at the threshold
|
|
for c in c_steps:
|
|
H = H0 + c * H1
|
|
evals = torch.linalg.eigvals(H)
|
|
current_k = evals[torch.argmin(torch.abs(evals - current_k))]
|
|
if np.any(c == train_cs):
|
|
index = np.where(c == train_cs)[0][0]
|
|
ks[index] = current_k
|
|
|
|
loss = ((ks - train_ks).abs() ** 2).sum()
|
|
|
|
if epoch % 1000 == 0:
|
|
print(f"Training {(epoch+1)/epochs:.1%} \t Loss: {loss}")
|
|
|
|
if H0.grad is not None:
|
|
H0.grad.zero_()
|
|
if H1.grad is not None:
|
|
H1.grad.zero_()
|
|
loss.backward()
|
|
|
|
with torch.no_grad():
|
|
H0 -= lr * H0.grad
|
|
H1 -= lr * H1.grad
|
|
|
|
# %%
|
|
# evaluate for all points
|
|
all_c = torch.tensor(df['c'].values, dtype=torch.float64)
|
|
exact_k = torch.tensor(df['k'].values, dtype=torch.complex128)
|
|
pred_ks = np.empty(len(df), dtype=np.complex128)
|
|
with torch.no_grad():
|
|
for (index, (c, k)) in enumerate(zip(all_c, exact_k)):
|
|
H = H0 + c * H1
|
|
evals = torch.linalg.eigvals(H)
|
|
i = torch.argmin(torch.abs(evals - k)) # TODO: more robust way to identify the eigenvector
|
|
pred_ks[index]= evals[i]
|
|
|
|
pred_Es = pred_ks ** 2
|
|
|
|
# %%
|
|
# plot the results
|
|
import matplotlib.pyplot as plt
|
|
|
|
fig, axs = plt.subplots(2, 1, figsize=(8, 12)) # Create a figure with two vertical panels
|
|
|
|
# First panel: k values
|
|
axs[0].scatter(np.real(train_data['k']), np.imag(train_data['k']), label='training')
|
|
axs[0].scatter(np.real(target_data['k']), np.imag(target_data['k']), label='target')
|
|
axs[0].scatter(np.real(pred_ks), np.imag(pred_ks), marker='x', label='predicted')
|
|
axs[0].set_xlabel('Re(k)')
|
|
axs[0].set_ylabel('Im(k)')
|
|
axs[0].legend()
|
|
|
|
# Second panel: E values
|
|
axs[1].scatter(np.real(train_data['E']), np.imag(train_data['E']), label='training')
|
|
axs[1].scatter(np.real(target_data['E']), np.imag(target_data['E']), label='target')
|
|
axs[1].scatter(np.real(pred_Es), np.imag(pred_Es), marker='x', label='predicted')
|
|
axs[1].set_xlabel('Re(E)')
|
|
axs[1].set_ylabel('Im(E)')
|
|
axs[1].legend()
|
|
|
|
plt.tight_layout() # Adjust spacing between panels
|
|
plt.show()
|