import numpy as np
import seaborn as sns
from matplotlib import pyplot as plt
import scipy.integrate as integrate

sns.set(style="white")
import os
from glob import glob
import pandas as pd
from sklearn.preprocessing import StandardScaler

# load Essen timeseries data and scrub resting state data
_fd_thr = 0.15
_perc_scrub_thr = 0.5
all_ts = []
root_dir = 'data/ex_timeseries/'
for f in glob(root_dir + 'bochum-*.tsv'):  # changed to bochum-*.tsv for replication with the Bochum dataset
    path, filename = os.path.split(f)
    ts = pd.read_csv(f, sep='\t').iloc[:, 1:].values
    ts = StandardScaler().fit_transform(ts)
    fd = pd.read_csv(path + '/FD_' + os.path.splitext(filename)[0] + '.txt', sep=',').values.flatten()
    fd = np.hstack(([0], fd))
    if np.sum(fd>_fd_thr)/len(fd) < _perc_scrub_thr:
        all_ts.append(ts[fd<_fd_thr])
    else:
        print('perc. scrubbed:', np.sum(fd>_fd_thr)/len(fd))
len(all_ts)

from connattractor import network
from nilearn.connectome import ConnectivityMeasure
# fcHNN solution
correlation_measure = ConnectivityMeasure(kind='partial correlation',  vectorize=False, discard_diagonal=False) # uses Ledoit-Wolf estimator
correlation_measure.fit_transform(all_ts)
parcor = 1 * correlation_measure.mean_
hopnet = network.Hopfield(parcor)
hopnet.W = (parcor-np.mean(parcor))/np.std(parcor) # overwrite the weights to put back the diagonal
attractors = {}
rng = np.random.default_rng(42)
for i in range(100):
    res = hopnet.update(np.tanh(rng.normal(0,1,122)), threshold=0, beta=0.071, num_iter=10000)
    if res[1]<10000: # converged
        state = tuple(np.round(res[0], 6))
        if state not in attractors.keys():
            attractors[state] = 1
        else:
            attractors[state] += 1

attractors_fchnn = sorted(attractors.items(), key=lambda item: item[1])[::-1]
len(attractors_fchnn)

attractor_states_fchnn = np.array([k for i, (k, v) in enumerate(attractors_fchnn)])
for i in range(attractor_states_fchnn.shape[0]):
    network.State(attractor_states_fchnn[i]).plot()

for i, (k, v) in enumerate(attractors_fchnn):
    print(i, v)

# comopute the negative precision matrix via nilearn
from nilearn.connectome import ConnectivityMeasure
correlation_measure = ConnectivityMeasure(kind='precision',  vectorize=False, discard_diagonal=False) # uses Ledoit-Wolf estimator
correlation_measure.fit_transform(all_ts)
J = -1 * correlation_measure.mean_

def compute_W_from_J(J, num_attractors=None):
    if num_attractors is None:
        num_attractors = J.shape[0]
    # Eigenvalue decomposition of J
    eigenvalues, eigenvectors = np.linalg.eigh(J)
    # Sort eigenvalues and corresponding eigenvectors in descending order
    idx = np.argsort(eigenvalues)[::-1]
    eigenvalues = eigenvalues[idx]
    eigenvectors = eigenvectors[:, idx]
    # Select the first `num_hidden_nodes` eigenvalues and eigenvectors
    selected_eigenvalues = eigenvalues[:num_attractors]
    selected_eigenvectors = eigenvectors[:, :num_attractors]
    # Compute the weight matrix W
    W = selected_eigenvectors * np.sqrt(-selected_eigenvalues)
    return W

from connattractor import network

num_attractors = 5

# analytical solution
W = compute_W_from_J(J, num_attractors = num_attractors)

# fcHNN solution already stored in attractors_fchnn
# we change the order to match the analytical solution
attractor_states_fchnn = np.array([k for i, (k, v) in enumerate(attractors_fchnn)])
indices = [None, 1, 0, 2, 3]

fig, axs = plt.subplots(ncols=2, nrows=num_attractors, figsize=(16, 2*num_attractors))
for i in range(num_attractors):
    print(i)
    if i == 0:
        network.State(np.zeros(122)).plot(plot_abs=False, figure=fig, axes=axs[i, 0], annotate=False, colorbar=False)
        state_analytic = network.State(-1*W[:,i]) # *-1 for visualization
        state_analytic.plot(plot_abs=False, figure=fig, axes=axs[i, 1], annotate=False, colorbar=False)
    else:
        network.State(attractor_states_fchnn[indices[i]]).plot(plot_abs=False, figure=fig, axes=axs[i, 0], annotate=False, colorbar=False)
        if i % 2 == 0:
            state_analytic = network.State(W[:,int((i+1)/2)]) # original
            state_analytic.plot(plot_abs=False, figure=fig, axes=axs[i, 1], annotate=False, colorbar=False)
        else:
            state_analytic = network.State(-1*W[:,int((i+1)/2)]) # inverse
            state_analytic.plot(plot_abs=False, figure=fig, axes=axs[i, 1], annotate=False, colorbar=False)
plt.show()

from connattractor import network

num_attractors = 5

from nilearn import surface, plotting, datasets
fsaverage = datasets.fetch_surf_fsaverage('fsaverage6')
surf = surface.load_surf_mesh('data/full_brain_left.surf.gii')
bg_map = np.hstack((surface.load_surf_data(fsaverage.sulc_left),
                    np.zeros(28935))) # magic number: number of cerebellum vertices (SUIT cerebellum template)

# analytical solution
W = compute_W_from_J(J, num_attractors = num_attractors)

# fcHNN solution already stored in attractors_fchnn
# we change the order to match the analytical solution
attractor_states_fchnn = np.array([k for i, (k, v) in enumerate(attractors_fchnn)])
indices = [None, 1, 0, 2, 3]

fig, axs = plt.subplots(ncols=2, nrows=num_attractors, figsize=(4, 2*num_attractors), subplot_kw={'projection': '3d'})
for i in range(num_attractors):
    print(i)
    if i == 0:
        network.State(np.ones(122)).plot(plot_abs=False, figure=fig, axes=axs[i, 0], annotate=False, colorbar=False)
        
        texture = np.zeros(surf.coordinates.shape[0])
        fig = plotting.plot_surf_stat_map(surf, texture, bg_map=bg_map, bg_on_data=True, hemi='left', colorbar=False, cmap='purple_green_r', threshold=0, figure=fig, axes=axs[i, 0])
        
        img = network.State(-1*W[:,i]).to_Nifti1Image()
        texture = surface.vol_to_surf(img, surf)
        fig = plotting.plot_surf_stat_map(surf, texture, bg_map=bg_map, bg_on_data=True, hemi='left', colorbar=False, cmap='purple_green_r', threshold=0, figure=fig, axes=axs[i, 1])
        
    else:
        img = network.State(attractor_states_fchnn[indices[i]]).to_Nifti1Image()
        texture = surface.vol_to_surf(img, surf)
        fig = plotting.plot_surf_stat_map(surf, texture, bg_map=bg_map, bg_on_data=True, hemi='left', colorbar=False, cmap='purple_green_r', threshold=0, figure=fig, axes=axs[i, 0])
        
        if i % 2 == 0:
            img = network.State(W[:,int((i+1)/2)]).to_Nifti1Image()
            texture = surface.vol_to_surf(img, surf)
            fig = plotting.plot_surf_stat_map(surf, texture, bg_map=bg_map, bg_on_data=True, hemi='left', colorbar=False, cmap='purple_green_r', threshold=0, figure=fig, axes=axs[i, 1])
        else:
            
            img = network.State(-1*W[:,int((i+1)/2)]).to_Nifti1Image()
            texture = surface.vol_to_surf(img, surf)
            fig = plotting.plot_surf_stat_map(surf, texture, bg_map=bg_map, bg_on_data=True, hemi='left', colorbar=False, cmap='purple_green_r', threshold=0, figure=fig, axes=axs[i, 1])
plt.show()

from connattractor import network

num_attractors = 14

from nilearn import surface, plotting, datasets
fsaverage = datasets.fetch_surf_fsaverage('fsaverage6')
surf = surface.load_surf_mesh('data/full_brain_left.surf.gii')
bg_map = np.hstack((surface.load_surf_data(fsaverage.sulc_left),
                    np.zeros(28935))) # magic number: number of cerebellum vertices (SUIT cerebellum template)

# analytical solution
W = compute_W_from_J(J, num_attractors = num_attractors)

# fcHNN solution already stored in attractors_fchnn
# we change the order to match the analytical solution
attractor_states_fchnn = np.array([k for i, (k, v) in enumerate(attractors_fchnn)])

fig, axs = plt.subplots(ncols=2, nrows=num_attractors, figsize=(4, 2*num_attractors), subplot_kw={'projection': '3d'})

for i in range(0,num_attractors):
    print(i)
        
    # cmap: PiYG, cyan_orange, cold_hot, blue_red
        
    if i % 2 == 0:
        img = network.State(W[:,int((i)/2)]).to_Nifti1Image()
        img.to_filename(f'data/attractor_{int((i)/2)}_pos.nii.gz')
        texture = surface.vol_to_surf(img, surf)
        #texture[texture<0] = 0
        fig = plotting.plot_surf_stat_map(surf, texture, bg_map=bg_map, bg_on_data=False, hemi='left', colorbar=False, cmap='PiYG', threshold=0, figure=fig, axes=axs[i, 0], symmetric_cbar=True, vmin=-0.2, vmax=0.2, darkness=1)
        fig = plotting.plot_surf_stat_map(surf, texture, bg_map=bg_map, bg_on_data=False, hemi='right', colorbar=False, cmap='PiYG', threshold=0, figure=fig, axes=axs[i, 1], symmetric_cbar=True, vmin=-0.2, vmax=0.2, darkness=1)
    else:
        img = network.State(-1*W[:,int((i)/2)]).to_Nifti1Image()
        img.to_filename(f'data/attractor_{int((i)/2)}_neg.nii.gz')
        texture = surface.vol_to_surf(img, surf)
        #texture[texture<0] = 0
        fig = plotting.plot_surf_stat_map(surf, texture, bg_map=bg_map, bg_on_data=False, hemi='left', colorbar=False, cmap='PiYG', threshold=0, figure=fig, axes=axs[i, 0], symmetric_cbar=True, vmin=-0.25, vmax=0.25, darkness=1)
        fig = plotting.plot_surf_stat_map(surf, texture, bg_map=bg_map, bg_on_data=False, hemi='right', colorbar=False, cmap='PiYG', threshold=0, figure=fig, axes=axs[i, 1], symmetric_cbar=True, vmin=-0.25, vmax=0.25, darkness=1)
plt.show()

    from nilearn import surface, plotting, datasets
    fsaverage = datasets.fetch_surf_fsaverage('fsaverage6')
    surf = surface.load_surf_mesh('data/full_brain_left.surf.gii')
    bg_map = np.hstack((surface.load_surf_data(fsaverage.sulc_left),
                        np.zeros(28935))) # magic number: number of cerebellum vertices (SUIT cerebellum template)
    # plot control signal
    signal = np.zeros(122)
    img = network.State(signal).to_Nifti1Image()
    # due to numerical instabilities, the sign may flips, we flip ti back by np.sign(W[0,i])
    texture = surface.vol_to_surf(img, surf)
    fig = plotting.plot_surf_stat_map(surf, texture, bg_map=bg_map, bg_on_data=True, hemi='right', colorbar=False, cmap='coolwarm', threshold=0.2)

mist_regions = pd.read_csv('data/MIST122_relabeled.tsv', sep='\t')
mist_regions.columns = ['label', 'MIST_index', 'region', 'module']
mist_mapper=mist_regions[['label', 'region']]
mist_mapper.set_index('region', inplace=True)
mist_mapper.loc['CER6_p', 'label']
mist_regions

# top 5 positive and negative weights for all attractors
num_attractors = 6
W = compute_W_from_J(J, num_attractors=num_attractors)
for i in range(num_attractors):
    print(f'Attractor {i}')
    print('Positive weights:')
    positive_indices = np.argsort(W[:, i])[::-1][:10]
    for idx in positive_indices:
        region = mist_mapper.loc[mist_mapper['label'] == idx+1].index[0]
        weight = W[idx, i]
        print(f'{region}: {weight}')
    
    print('Negative weights:')
    negative_indices = np.argsort(W[:, i])[:10]
    for idx in negative_indices:
        region = mist_mapper.loc[mist_mapper['label'] == idx+1].index[0]
        weight = W[idx, i]
        print(f'{region}: {weight}')

W = compute_W_from_J(J, num_attractors=12)
# do tsne on the weights
from sklearn.manifold import TSNE
tsne = TSNE(n_components=2, perplexity=5, random_state=42)
W_tsne = tsne.fit_transform(W)
#plot it
plt.figure(figsize=(6,6))
plt.scatter(W_tsne[:,0], W_tsne[:,1], alpha=0.5)
# add labels
for i, txt in enumerate(range(1, len(W_tsne))):
    plt.annotate(mist_regions.iloc[i-1, 2], (W_tsne[i,0], W_tsne[i,1]), fontsize=6)
plt.show()

attractor_states_fchnn.shape

def hn_energy(state, threshold):
    return  -0.5 * state @ J @ state + state @ threshold

def plot_manifold(num_attractors = 6, perplexity=None, smooth_sigma=0.8, filename=None):

    from openTSNE import TSNE
    W = compute_W_from_J(J, num_attractors=num_attractors)

    if perplexity is None:
        perplexity = (num_attractors+1)/3

    W_paired = []
    for i in range(0, num_attractors):
        W_paired.append(W[:,i])
        W_paired.append(-W[:,i])
    W_paired = np.array(W_paired).T
    
    #W_paired = attractor_states_fchnn.T
    
    # do tsne on the weights
    tsne = TSNE(n_components=2, perplexity=perplexity, initialization='pca', metric="euclidean", random_state=42, n_jobs=16)
    tsne_embedding = tsne.fit(W_paired.transpose())
    W_tsne = tsne_embedding[:num_attractors*2,:]
    W_paired = W_paired[:,:num_attractors*2]
    #transform sample
    sample = np.vstack(all_ts)
    sample_tsne = tsne_embedding.transform(sample, perplexity=perplexity)

    from scipy.stats import binned_statistic_2d
    energies = np.array([hn_energy(sample[i], np.zeros(J.shape[0])) for i in range(sample.shape[0])])
    mean_energy = binned_statistic_2d(sample_tsne[:,0], sample_tsne[:,1], values=energies, statistic='mean', bins=20)
    # smooth ean_energy.statistic
    from scipy.ndimage import gaussian_filter
    mean_energy_smoothed = np.copy(mean_energy.statistic)
    grand_mean = np.mean(mean_energy_smoothed[~ np.isnan(mean_energy_smoothed)])
    mean_energy_smoothed[np.isnan(mean_energy_smoothed)] = grand_mean  # mean inputting
    mean_energy_smoothed = gaussian_filter(mean_energy_smoothed, sigma=smooth_sigma)
    mean_energy_smoothed_nan = mean_energy_smoothed.copy()
    mean_energy_smoothed_nan[np.isnan(mean_energy.statistic)] = np.nan  # write back nan values
    # plot
    fig, axes = plt.subplots(1, 4, figsize=(16, 4))
    
    attractor_energies = [hn_energy(W_paired[:,i], np.zeros(J.shape[0])) for i in range(W_paired.shape[1])]
    
    ##########
    sctplt = axes[0].scatter(sample_tsne[:,0], sample_tsne[:,1], alpha=0.5, c=[hn_energy(sample[i], np.zeros(J.shape[0])) for i in range(sample.shape[0])],cmap='coolwarm', s=6, vmin=100, vmax=350, edgecolor='none')
    plt.colorbar(sctplt,ax=axes[0], location='bottom', shrink=0.6, ticks=[100, 350])
    axes[0].scatter(W_tsne[:,0], W_tsne[:,1], alpha=0.5, c=attractor_energies,cmap='coolwarm', s=30, edgecolors= "black", vmin=100, vmax=350)
    ##########
    
    sns.kdeplot(x=sample_tsne[:,0], y=sample_tsne[:,1], ax=axes[1], cmap='Greens', cbar=True, fill=True, levels=8, thresh=0.01, alpha=1.0,antialiased=True, cbar_kws={"location":"bottom", "shrink":0.6, "ticks":[0.01, 0.1]})
    axes[1].scatter(W_tsne[:,0], W_tsne[:,1], alpha=0.5, c=attractor_energies,
    cmap='coolwarm', edgecolors= "black", vmin=200, vmax=280, color=None)
    
    ###########
    implt = axes[2].imshow(mean_energy_smoothed_nan.transpose(), origin='lower', extent=(sample_tsne[:,0].min(), sample_tsne[:,0].max(), sample_tsne[:,1].min(), sample_tsne[:,1].max()), aspect='auto', cmap='coolwarm', vmin=200, vmax=280,
               #interpolation='bicubic'
               )
    plt.colorbar(implt,ax=axes[2], location='bottom', shrink=0.6, ticks=[200, 280])
    axes[2].scatter(W_tsne[:,0], W_tsne[:,1], alpha=0.5, c=attractor_energies,
    cmap='coolwarm', edgecolors= "black", vmin=200, vmax=280, s=30)
    # set limits
    axes[2].set_xlim(sample_tsne[:,0].min()*1.1, sample_tsne[:,0].max()*1.1)
    axes[2].set_ylim(sample_tsne[:,1].min()*1.1, sample_tsne[:,1].max()*1.1)
    
    
    for i, txt in enumerate(range(0, W_tsne.shape[0])):
        axes[0].annotate(txt, (W_tsne[i,0], W_tsne[i,1]))
        axes[1].annotate(txt, (W_tsne[i,0], W_tsne[i,1]))
        axes[2].annotate(txt, (W_tsne[i,0], W_tsne[i,1]))
    
    sns.despine() #left=True, bottom=True)
    
    ## Surface plot:
    # swap 4th subplot to 3d
    axes[3].remove()
    axes[3] = fig.add_subplot(1, 4, 4, projection='3d')

    # Extracting the necessary information
    mean_values = mean_energy_smoothed  # The mean energy in each bin
    x_edges = mean_energy.x_edge         # The edges of the bins in the x direction
    y_edges = mean_energy.y_edge         # The edges of the bins in the y direction
    
    # Calculate bin centers
    x_centers = (x_edges[:-1] + x_edges[1:]) / 2
    y_centers = (y_edges[:-1] + y_edges[1:]) / 2
    
    # Create a meshgrid for the surface plot
    X, Y = np.meshgrid(x_centers, y_centers)
    
    # Plotting the 3D surface    
    # Surface plot
    for i in range(0, W_tsne.shape[0]):
        # search closes point from the meshgrid
        x_idx = np.argmin(np.abs(x_centers - W_tsne[i,0]))
        y_idx = np.argmin(np.abs(y_centers - W_tsne[i,1]))
        #ax.scatter(x_centers[x_idx], y_centers[y_idx], mean_values[x_idx, y_idx], color='black', s=30)
        axes[3].plot([x_centers[x_idx], x_centers[x_idx]],
                [y_centers[y_idx],y_centers[y_idx]],
                zs=[mean_values[x_idx, y_idx]+2, mean_values[x_idx, y_idx]+20], color='black')
    
    surf = axes[3].plot_surface(X, Y, mean_values.T, cmap='coolwarm', linewidth=0.1)
    
    axes[3].set_zlim(200, 280)
    axes[3].view_init(elev=50, azim=-70)
    # zoom in a bit more
    #axes[3].dist = 3
    
    zoom = 0.8
    axes[3].set_xlim3d(sample_tsne[:,0].min()*zoom, sample_tsne[:,0].max()*zoom)
    axes[3].set_ylim3d(sample_tsne[:,0].min()*zoom, sample_tsne[:,0].max()*zoom)
    axes[3].set_zlim3d(100, 350)     
    # remove axes and everything
    axes[3].set_axis_off()
    
    if filename is not None:
        plt.savefig(filename)
    
    plt.show()
    

plot_manifold(num_attractors = 4)

plot_manifold(num_attractors = 5)

plot_manifold(num_attractors = 6, filename='fig/attractor_manifold.pdf')

plot_manifold(num_attractors = 8)

plot_manifold(num_attractors = 10)

plot_manifold(num_attractors = 16)

from connattractor import network
from nilearn import surface, plotting, datasets

fsaverage = datasets.fetch_surf_fsaverage('fsaverage6')
surf = surface.load_surf_mesh('data/full_brain_left.surf.gii')
bg_map = np.hstack((surface.load_surf_data(fsaverage.sulc_left),
                        np.zeros(28935))) # magic number: number of cerebellum vertices (SUIT cerebellum template)

W = compute_W_from_J(J, num_attractors=num_attractors)

num_attractors = 6

W_paired = []
for i in range(0, num_attractors):
     W_paired.append(W[:,i])
     W_paired.append(-W[:,i])
W_paired = np.array(W_paired).T

fig, axes = plt.subplots(nrows=num_attractors*2, ncols=2, figsize=(4, 3*num_attractors), subplot_kw={'projection': '3d'})

for i in range(num_attractors*2):
    img = network.State(W_paired[:,i]).to_Nifti1Image()
    texture = surface.vol_to_surf(img, surf)
    fig = plotting.plot_surf_stat_map(surf, texture, bg_map=bg_map, bg_on_data=True, hemi='left', colorbar=False, cmap='coolwarm', threshold=0, figure=fig, axes=axes[i, 0], title='Attractor {}'.format(i))
    
    fig = plotting.plot_surf_stat_map(surf, texture, bg_map=bg_map, bg_on_data=True, hemi='right', colorbar=False, cmap='coolwarm', threshold=0, figure=fig, axes=axes[i, 1])
    
#plt.savefig('fig/attractor_states.pdf')
    
plt.show()

from connattractor import network
from nilearn import surface, plotting, datasets

fsaverage = datasets.fetch_surf_fsaverage('fsaverage6')
surf = surface.load_surf_mesh('data/full_brain_left.surf.gii')
bg_map = np.hstack((surface.load_surf_data(fsaverage.sulc_left),
                        np.zeros(28935))) # magic number: number of cerebellum vertices (SUIT cerebellum template)

W = compute_W_from_J(J, num_attractors=num_attractors)

num_attractors = 6

W_paired = []
for i in range(0, num_attractors):
     W_paired.append(W[:,i])
     W_paired.append(-W[:,i])
W_paired = np.array(W_paired).T

fig, axes = plt.subplots(nrows=num_attractors*2, ncols=1, figsize=(3, 3*num_attractors))

for i in range(num_attractors*2):
    network.State(W_paired[:,i]).plot(display_mode='x', plot_abs=False, figure=fig, axes=axes[i],
                                      title=str(i), annotate=False, colorbar=False)
#plt.savefig('fig/attractor_states.pdf')
    
plt.show()

from connattractor import network
# fcHNN solution
correlation_measure = ConnectivityMeasure(kind='partial correlation',  vectorize=False, discard_diagonal=False) # uses Ledoit-Wolf estimator
correlation_measure.fit_transform(all_ts)
parcor = 1 * correlation_measure.mean_
hopnet = network.Hopfield(parcor)
hopnet.W = (parcor-np.mean(parcor))/np.std(parcor) # overwrite the weights to put back the diagonal
attractors = {}
rng = np.random.default_rng(42)
for i in range(1000):
    res = hopnet.update(np.tanh(rng.normal(0,1,122)), threshold=0, beta=0.055, num_iter=10000)
    if res[1]<10000: # converged
        state = tuple(np.round(res[0], 6))
        if state not in attractors.keys():
            attractors[state] = 1
        else:
            attractors[state] += 1

attractors_fchnn = sorted(attractors.items(), key=lambda item: item[1])[::-1]
len(attractors_fchnn)

def_control_signals = [1, 2, 1000] #np.logspace(-0.5,0.5,5)

def control_signal_analysis(conrol_sigmal_index, control_signals = def_control_signals, num_attractors = 5):
    
    W_null = compute_W_from_J(J, num_attractors = num_attractors)
    
    from nilearn import surface, plotting, datasets
    fsaverage = datasets.fetch_surf_fsaverage('fsaverage6')
    surf = surface.load_surf_mesh('data/full_brain_left.surf.gii')
    bg_map = np.hstack((surface.load_surf_data(fsaverage.sulc_left),
                        np.zeros(28935))) # magic number: number of cerebellum vertices (SUIT cerebellum template)
    # plot control signal
    signal = np.zeros(122)
    print(conrol_sigmal_index)
    for idx in conrol_sigmal_index:
        signal[idx-1] = 1
    img = network.State(signal).to_Nifti1Image()
    # due to numerical instabilities, the sign may flips, we flip ti back by np.sign(W[0,i])
    texture = surface.vol_to_surf(img, surf)
    fig = plotting.plot_surf_stat_map(surf, texture, bg_map=bg_map, bg_on_data=True, hemi='left', colorbar=False, cmap='coolwarm', threshold=0.2)
    
    fig, axs = plt.subplots(ncols=len(control_signals), nrows=num_attractors, figsize=(len(control_signals)*2, 2*num_attractors), subplot_kw={'projection': '3d'})
    for i in range(num_attractors):
        
        for j, control_signal in enumerate(control_signals):
            
            J_stim = np.copy(J)
            for idx in conrol_sigmal_index:
                J_stim[idx,idx] *= control_signal 
            W = compute_W_from_J(J_stim, num_attractors = num_attractors)
            
            if i == 0: # eigencentrality
                signflip = -1*np.sign(np.corrcoef(W_null[:,i], W[:,i])[0,1])
                img = network.State(signflip*W[:,i]).to_Nifti1Image()
                # due to numerical instabilities, the sign may flips, we flip ti back by np.sign(W[0,i])
                texture = surface.vol_to_surf(img, surf)
                fig = plotting.plot_surf_stat_map(surf, texture, bg_map=bg_map, bg_on_data=True, hemi='left', colorbar=False, cmap='coolwarm', threshold=0, figure=fig, axes=axs[i, j])
                
            else:  
                signflip = np.sign(np.corrcoef(W_null[:,int((i+1)/2)], W[:,int((i+1)/2)])[0,1])
                if i % 2 == 0:
                    img = network.State(signflip*W[:,int((i+1)/2)]).to_Nifti1Image()
                    texture = surface.vol_to_surf(img, surf)
                    fig = plotting.plot_surf_stat_map(surf, texture, bg_map=bg_map, bg_on_data=True, hemi='left', colorbar=False, cmap='coolwarm', threshold=0, figure=fig, axes=axs[i, j])
                else:
                    img = network.State(-1*signflip*W[:,int((i+1)/2)]).to_Nifti1Image()
                    texture = surface.vol_to_surf(img, surf)
                    fig = plotting.plot_surf_stat_map(surf, texture, bg_map=bg_map, bg_on_data=True, hemi='left', colorbar=False, cmap='coolwarm', threshold=0, figure=fig, axes=axs[i, j])
    plt.show()

mist_mapper.index.values
# see also: https://simexp.github.io/multiscale_dashboard/index.html

control_signal_analysis(mist_mapper.loc[['AMY'], 'label'].values)

control_signal_analysis(mist_mapper.loc[['L_VLPFcor'], 'label'].values)

control_signal_analysis(mist_mapper.loc[['PCsul_d', 'l_PCsul', 'R_PCsul'], 'label'].values)

control_signal_analysis(mist_mapper.loc[['MOTnet_vl'], 'label'].values)

control_signal_analysis(mist_mapper.loc[['L_ANGgyr'], 'label'].values)

control_signal_analysis(mist_mapper.loc[['PVISnet_vm'], 'label'].values)

control_signal_analysis(mist_mapper.loc[['PINS_d'], 'label'].values)

correlation_measure = ConnectivityMeasure(kind='precision', vectorize=False, discard_diagonal=False) # uses Ledoit-Wolf estimator
M = correlation_measure.fit_transform(all_ts)
J = -1* correlation_measure.mean_
hopnet = network.Hopfield(J, scale=True)
res = hopnet.update(np.tanh(np.random.normal(0,1,122)), threshold=0, beta=0.06, num_iter=100000)
print(res[1])
network.State(res[0]).plot(plot_abs=False)

from nilearn import surface, plotting, datasets
fsaverage = datasets.fetch_surf_fsaverage('fsaverage6')
surf = surface.load_surf_mesh('data/full_brain_left.surf.gii')
bg_map = np.hstack((surface.load_surf_data(fsaverage.sulc_left),
                    np.zeros(28935)))

img = network.State(res[0]).to_Nifti1Image()
texture = surface.vol_to_surf(img, surf)
fig = plotting.plot_surf_stat_map(surf, texture, bg_map=bg_map, bg_on_data=True, hemi='left', title='lateral', colorbar=False, cmap='coolwarm', threshold=0)
fig = plotting.plot_surf_stat_map(surf, texture, bg_map=bg_map, bg_on_data=True, hemi='right', title='medial', colorbar=False, cmap='coolwarm', threshold=0)

bg_map.shape, surf.coordinates.shape

plotting.plot_surf_stat_map(
        surf, texture, hemi='left',
        title='title', colorbar=True, cmap='coolwarm',
        threshold=0)#, bg_map=fsaverage.sulc_left)

from nilearn.image import smooth_img
img = network.State(res[0]).to_Nifti1Image()
# plot surface
from nilearn import plotting
fig, axes = plotting.plot_img_on_surf(img, surf_mesh=s, threshold=0, inflate=False, 
                           cmap='coolwarm', bg_on_data=True, alpha=1)

network.Hopfield(J).plot_weights()
plt.show()
network.Hopfield(J2).plot_weights()

def compute_W_from_J(J, num_attractors=None):
    if num_attractors is None:
        num_attractors = J.shape[0]
    # Eigenvalue decomposition of J
    eigenvalues, eigenvectors = np.linalg.eigh(J)
    # Sort eigenvalues and corresponding eigenvectors in descending order
    idx = np.argsort(eigenvalues)[::-1]
    eigenvalues = eigenvalues[idx]
    eigenvectors = eigenvectors[:, idx]
    # Select the first `num_hidden_nodes` eigenvalues and eigenvectors
    selected_eigenvalues = eigenvalues[:num_attractors]
    selected_eigenvectors = eigenvectors[:, :num_attractors]
    # Compute the weight matrix W
    W = selected_eigenvectors * np.sqrt(-selected_eigenvalues)

    print(np.sqrt(-selected_eigenvalues))
    return W

x =compute_W_from_J(J, 5)