introvoyz041's picture
Migrated from GitHub
e11c92f verified
import torch
import pickle
import os
import numpy as np
from sklearn.decomposition import PCA
import matplotlib
import matplotlib.pyplot as plt
import matplotlib.ticker as ticker
from matplotlib.patches import Ellipse
import seaborn as sns
from . import clustering
from . import modules
from . import dms
def load_network(f):
"""Load RNN network."""
noise_std = 5e-2
alpha = 0.2
hidden_size = 500
load = torch.load(f, map_location="cpu")
if len(load) == 2:
z, state = load
else:
state = load
z = None
net = modules.LowRankRNN(2, hidden_size, 1, noise_std, alpha, rank=2)
net.load_state_dict(state)
net.svd_reparametrization()
return z, net
def sample_network(net, f, seed=0):
if os.path.exists(f):
print('Network found with same name. Loading...')
return torch.load(open(f, "rb"))
n_pops = 2
z, _ = clustering.gmm_fit(net, n_pops, algo="bayes", random_state=seed)
net_sampled = clustering.to_support_net(net, z)
if os.path.exists(f):
z, state = torch.load(f)
net_sampled.load_state_dict(state)
else:
x_train, y_train, mask_train, x_val, y_val, mask_val = dms.generate_dms_data(1000)
modules.train(
net_sampled,
x_train,
y_train,
mask_train,
20,
lr=1e-6,
resample=True,
keep_best=True,
clip_gradient=1,
)
torch.save([z, net_sampled], f)
return z, net_sampled
def generate_trajectories(
net, input=None, epochs=None, n_traj=None, fname="./data/RNN_trajectories.pkl"
):
if fname is not None:
if os.path.exists(fname):
print('Trajectory file found with same name. Loading...')
return pickle.load(open(fname, "rb"))
traj = []
for i in range(len(input)):
conds = []
for _ in range(n_traj):
_, traj_ = net(input[i].unsqueeze(0))
traj_ = traj_.squeeze().detach().numpy()
traj_epoch = [traj_[e : epochs[j + 1]] for j, e in enumerate(epochs[:-1])]
conds.append(traj_epoch)
traj.append(conds)
pickle.dump(traj, open(fname, "wb"))
return traj
def load_trajectories(fname):
return pickle.load(open(fname, "rb"))
def plot_ellipse(ax, w, color="silver", std_factor=1):
X = np.array([w[:, 0], w[:, 1]]).T
cov = X.T @ X / X.shape[0]
eigvals, eigvecs = np.linalg.eig(cov)
v1 = eigvecs[:, 0]
angle = np.arctan(v1[1] / v1[0])
angle = angle * 180 / np.pi
ax.add_artist(
Ellipse(
xy=[0, 0],
angle=angle,
width=np.sqrt(eigvals[0]) * 2 * std_factor,
height=np.sqrt(eigvals[1]) * 2 * std_factor,
fill=True,
fc=color,
ec=color,
lw=1,
alpha=0.4,
)
)
return ax
def plot_coefficients(net, z=None):
if z is None:
n_pops = 2
z, _ = clustering.gmm_fit(net, n_pops, algo="bayes", random_state=0)
m1 = net.m[:, 0].detach().numpy()
n1 = net.n[:, 0].detach().numpy()
m2 = net.m[:, 1].detach().numpy()
n2 = net.n[:, 1].detach().numpy()
wi1 = net.wi[0].detach().numpy()
wi2 = net.wi[1].detach().numpy()
fig, ax = plt.subplots(1, 4, figsize=(12, 2))
colors = ["#364285", "#E5BA52"]
n_pops = 2
clustering.pop_scatter_linreg(wi1, wi2, z, n_pops, colors=colors, ax=ax[0])
plot_ellipse(
ax[0], np.array([wi1[z.astype(bool)], wi2[z.astype(bool)]]).T, std_factor=3, color=colors[1]
)
plot_ellipse(
ax[0],
np.array([wi1[~z.astype(bool)], wi2[~z.astype(bool)]]).T,
std_factor=3,
color=colors[0],
)
clustering.pop_scatter_linreg(m1, m2, z, n_pops, colors=colors, ax=ax[1])
clustering.pop_scatter_linreg(n1, n2, z, n_pops, colors=colors, ax=ax[2])
clustering.pop_scatter_linreg(m1, n1, z, n_pops, colors=colors, ax=ax[3])
def setup_matplotlib():
plt.rcParams["axes.titlesize"] = 24
plt.rcParams["axes.labelsize"] = 19
plt.rcParams["xtick.labelsize"] = 16
plt.rcParams["ytick.labelsize"] = 16
plt.rcParams["figure.figsize"] = (6, 4)
plt.rcParams["axes.titlepad"] = 24
plt.rcParams["axes.labelpad"] = 10
plt.rcParams["axes.spines.top"] = False
plt.rcParams["axes.spines.right"] = False
plt.rcParams["font.size"] = 14
def get_lower_tri_heatmap(
ov, bounds=None, figsize=None, cbar=False, cbar_shrink=0.9, cbar_pad=0.3, ax=None
):
mask = np.zeros_like(ov, dtype=np.bool)
mask[np.triu_indices_from(mask)] = True
ov[np.diag_indices_from(ov)] = 0
# mask[np.diag_indices_from(mask)] = False
mask = mask.T
print(mask)
if figsize is None:
figsize = matplotlib.rcParams["figure.figsize"]
if bounds is None:
bound = np.max((np.abs(np.min(ov)), np.abs(np.max(ov))))
bounds = [-bound, bound]
# Set up the matplotlib figure
if ax is None:
f, ax = plt.subplots(figsize=figsize)
# Generate a custom diverging colormap
cmap = sns.diverging_palette(220, 10, sep=10, as_cmap=True)
print(ov)
# Draw the heatmap with the mask and correct aspect ratio
if not cbar:
mesh = sns.heatmap(
ov[:-1, 1:],
mask=mask[:-1, 1:],
cmap=cmap,
center=0,
square=True,
linewidths=0.5,
cbar=False,
vmin=bounds[0],
vmax=bounds[1],
ax=ax,
)
else:
mesh = sns.heatmap(
ov[:-1, 1:],
mask=mask[:-1, 1:],
cmap=cmap,
center=0,
square=True,
linewidths=0.5,
cbar=True,
vmin=bounds[0],
vmax=bounds[1],
ax=ax,
cbar_kws={"shrink": cbar_shrink, "ticks": ticker.MaxNLocator(3), "pad": cbar_pad},
)
ax.xaxis.tick_top()
ax.yaxis.tick_right()
return ax, mesh
def set_size(size, ax=None):
"""to force the size of the plot, not of the overall figure, from
https://stackoverflow.com/questions/44970010/axes-class-set-explicitly-size-width-height-of-axes-in-given-units"""
if not ax:
ax = plt.gca()
l = ax.figure.subplotpars.left
r = ax.figure.subplotpars.right
t = ax.figure.subplotpars.top
b = ax.figure.subplotpars.bottom
w, h = size
figw = float(w) / (r - l)
figh = float(h) / (t - b)
ax.figure.set_size_inches(figw, figh)
def center_limits(ax):
xmin, xmax = ax.get_xlim()
xbound = max(-xmin, xmax)
ax.set_xlim(-xbound, xbound)
ymin, ymax = ax.get_ylim()
ybound = max(-ymin, ymax)
ax.set_ylim(-ybound, ybound)
def plot_all_scatters(vectors):
fig, ax = plt.subplots(len(vectors), len(vectors), figsize=(6, 8))
for i in range(len(vectors)):
for j in range(len(vectors)):
if i is not j:
ax[i, j].scatter(vectors[i], vectors[j])
return ax
def plot_rates_single_neurons(rates, offset=1, colors=None, deltaT=1.0, figsize=(6, 8), ax=None):
if ax is None:
fig, ax = plt.subplots(figsize=figsize)
cur_max = 0.0
for i in range(rates.shape[1]):
color = colors[i] if colors is not None else "red"
ax.plot(
np.arange(rates.shape[0]) * deltaT / 1000,
rates[:, i] + cur_max + offset - np.min(rates[:, i]),
color=color,
)
cur_max = cur_max + offset - np.min(rates[:, i]) + np.max(rates[:, i])
return ax
def bar_plots_vectors(n, wi, wi_ctx1, wi_ctx2, title, xticks):
fig, ax = plt.subplots()
x = np.arange(3)
ctx_derivative1 = phi_prime(wi_ctx1)
ctx_derivative2 = phi_prime(wi_ctx2)
win = wi
Nfull = n.shape[0]
neff_ctx1 = n.reshape(Nfull, 1) * ctx_derivative1.reshape(Nfull, 1)
neff_ctx2 = n.reshape(Nfull, 1) * ctx_derivative2.reshape(Nfull, 1)
ov1 = np.sum(n.reshape(Nfull, 1) * win.reshape(Nfull, 1))
ov2 = np.sum(neff_ctx1.reshape(Nfull, 1) * win.reshape(Nfull, 1))
ov3 = np.sum(neff_ctx2.reshape(Nfull, 1) * win.reshape(Nfull, 1))
y = [ov1, ov2, ov3]
ax.bar(x, y)
plt.ylabel(title, fontsize=30)
plt.xticks(x, xticks, fontsize=25)
return ax
def radial_distribution_plot(x, N=80, bottom=0.1, cmap_scale=0.05, points=True):
"""
Plot a radial histogram of angles
:param x: if points=True, an array of shape nx2 of points in 2d space. if points=False a series of angles
:param N: num bins
:param bottom: radius of base circle
:param cmap_scale: to adjust the colormap
:param points: see x
:return:
"""
if points:
assert len(x.shape) == 2 and x.shape[1] == 2
x_cplx = np.array(x[:, 0], dtype=np.complex64)
x_cplx.imag = x[:, 1]
angles = np.angle(x_cplx)
else:
angles = x
angles = angles % (2 * np.pi)
theta = np.linspace(0.0, 2 * np.pi, N, endpoint=False)
radii = [
np.mean(np.logical_and(angles > theta[i], angles < theta[i + 1]))
for i in range(len(theta) - 1)
]
radii.append(np.mean(angles > theta[-1]))
width = (2 * np.pi) / N
offset = np.pi / N
ax = plt.subplot(111, polar=True)
bars = ax.bar(theta + offset, radii, width=width, bottom=bottom)
for r, bar in zip(radii, bars):
bar.set_facecolor(plt.cm.jet(r / cmap_scale))
bar.set_alpha(0.8)
plt.yticks([])
def dimensionality_plot(trajectories, vecs, labels, figsize=None):
"""
plot cumulative percentage of variance explained by vectors vecs, while ordering with most explicative vectors first
:param trajectories: numpy array of shape #time_points x #neurons (with trials already flattened)
:param vecs: list of numpy arrays of shape #neurons
:param labels: labels associated with each vector
:param figsize:
:return: axes
"""
total_var = np.sum(np.var(trajectories, axis=0))
print(total_var)
vars_nonorth = []
for v in vecs:
proj = trajectories @ v / np.linalg.norm(v)
vars_nonorth.append(np.var(proj))
indices = np.argsort(vars_nonorth)
vecs_ordered = [vecs[i] for i in indices[::-1]]
labels_ordered = [labels[i] for i in indices[::-1]]
vecs_orth = gram_schmidt(vecs_ordered)
variances = []
for v in vecs_orth:
proj = trajectories @ v / np.linalg.norm(v)
variances.append(np.var(proj))
print(variances)
cumvar = np.cumsum(variances).tolist()
print(cumvar)
fig, ax = plt.subplots(figsize=figsize)
ax.bar(
range(1, len(variances) + 1), cumvar / total_var * 100, color="lightslategray", alpha=0.5
)
ax.axhline(100, c="r")
ax.spines["top"].set_visible(False)
ax.spines["right"].set_visible(False)
ax.set(xticks=range(1, len(variances) + 1), xticklabels=labels_ordered, yticks=[0, 100])
return ax
def phi_prime(x):
return 1 - np.tanh(x) ** 2
def overlap_matrix(vectors):
hidden_size = len(vectors[0])
ov = np.zeros((len(vectors), len(vectors)))
for i in range(len(vectors)):
for j in range(i, len(vectors)):
ov[i, j] = 1 / hidden_size * np.sum(vectors[i] * vectors[j])
return ov
def boxplot_accuracies(accs, figsize=None, labels=None):
fig, ax = plt.subplots(1, 1, figsize=figsize)
for i, acc in enumerate(accs):
if not isinstance(acc, list):
plt.scatter(i, acc, marker="*", s=90, c="k")
else:
bp = ax.boxplot(acc, positions=[i], widths=0.5)
ax.scatter([i] * len(acc), acc, c="gray", alpha=0.5, s=5)
[l.set_linewidth(2) for l in bp["medians"]]
ax.set_xlim(-0.5, len(accs) - 0.5)
ax.set_ylim(0, 1.1)
ax.set_xticks(list(range(len(accs))))
ax.spines["top"].set_visible(False)
ax.spines["right"].set_visible(False)
ax.set_xticklabels(labels, rotation=45)
ax.set_yticks([0.0, 0.25, 0.5, 0.75, 1.0])
ax.set_ylabel("accuracy")
ax.axhline(1, c="k", zorder=-10, lw=1, ls="--")
return ax
def gram_schmidt(vecs):
vecs_orth = []
vecs_orth.append(vecs[0] / np.linalg.norm(vecs[0]))
for i in range(1, len(vecs)):
v = vecs[i]
for j in range(i):
v = v - (v @ vecs_orth[j]) * vecs_orth[j]
v = v / np.linalg.norm(v)
vecs_orth.append(v)
return vecs_orth
def gram_factorization(G):
"""
The rows of the returned matrix are the basis vectors whose Gramian matrix is G
:param G: ndarray representing a symmetric semidefinite positive matrix
:return: ndarray
"""
w, v = np.linalg.eigh(G)
x = v * np.sqrt(w)
return x
def angle(v, w, deg=True):
res = np.arccos((v @ w) / (np.linalg.norm(v) * np.linalg.norm(w)))
if not deg:
return res
else:
return res * 180 / np.pi
def plot_experiment(net, input, traj, epochs, rect=(-8, 8, -6, 6), traj_to_show=1):
fig, ax = plt.subplots(4, 5, figsize=(25, 20))
idx = np.floor(np.linspace(0, len(input) - 1, 4))
for i in range(4):
for j, e in enumerate(epochs[:-1]):
dms._plot_field(net, input[int(idx[i]), e], ax[i][j], rect=rect, sizes=1.3)
epoch = [c[j] for c in traj[i]]
dms.plot_trajectories(net, epoch, ax[i][j], c="#C30021", n_traj=traj_to_show)
if j > 0:
epoch = [c[j - 1] for c in traj[i]]
dms.plot_trajectories(
net, epoch, ax[i][j], c="#C30021", style="--", n_traj=traj_to_show
)
ax[0][0].set_title("Fix")
ax[0][1].set_title("Stim 1")
ax[0][2].set_title("Delay")
ax[0][3].set_title("Stim 2")
ax[0][4].set_title("Decision")
for i in range(4):
ax[i][0].set_ylabel(r"$\kappa_2$")
for i in range(5):
ax[3][i].set_xlabel(r"$\kappa_1$")
fig.subplots_adjust(hspace=0.1, wspace=0.1)
def aggregate_data(traj, epochs, transient=10, only_stim=False, pca=True, n_pca=3):
n_conds = len(traj)
n_epochs = len(epochs) - 1
n_traj = len(traj[0])
# fit PCA to all data
pos = []
for i in range(n_conds): # conditions
for k in range(n_epochs):
for j in range(n_traj): # trajectories
pos.append(traj[i][j][k][transient:])
if pca:
pca = PCA(n_components=n_pca)
pca.fit(np.vstack(pos))
print("Explained variance: ", pca.explained_variance_ratio_)
# aggregate data under baseline condition (no input)
pos, vel = [], []
if not only_stim:
for i in range(n_conds): # conditions
pos_, vel_ = [], []
for k in [0, 2, 4]:
for j in range(n_traj): # trajectories
pos_proj = traj[i][j][k][transient:]
if pca:
pos_proj = pca.transform(pos_proj)
pos_.append(pos_proj[:-1]) # stack trajectories
vel_.append(np.diff(pos_proj, axis=0)) # compute differences
pos_, vel_ = np.vstack(pos_), np.vstack(vel_) # stack trajectories
pos.append(pos_)
vel.append(vel_)
# aggregate data under stimulated condition
for i in range(n_conds): # conditions
pos_, vel_ = [], []
for k in [1, 3]:
for j in range(n_traj): # trajectories
pos_proj = traj[i][j][k][transient:]
if pca:
pos_proj = pca.transform(pos_proj)
pos_.append(pos_proj[:-1])
vel_.append(np.diff(pos_proj, axis=0))
pos_, vel_ = np.vstack(pos_), np.vstack(vel_) # stack trajectories
pos.append(pos_)
vel.append(vel_)
return pos, vel