# sta = (sta - np.min(sta)) / (np.max(sta) - np.min(sta))
# glm0.networks[0].layers[0].weights[:,0]=deepcopy((sta - np.min(sta)) / (np.max(sta) - np.min(sta)))
v2f0 = glm0.fit_variables(fit_biases=True)
# train initial model
_ = glm0.train(input_data=[Xstim],
output_data=Robs,
train_indxs=Ui,
test_indxs=Xi,
learning_alg='lbfgs',
opt_params=lbfgs_params,
fit_variables=v2f0)
# plot filters
DU.plot_3dfilters(glm0)
# Find best regularization
glmbest = glm0.copy_model()
[LLpath, glms] = NDNutils.reg_path(glmbest,
input_data=[Xstim],
output_data=Robs,
train_indxs=Ui,
test_indxs=Xi,
reg_type='glocal',
reg_vals=[1e-6, 1e-4, 1e-3, 1e-2, 0.1, 1],
layer_target=0,
ffnet_target=0,
learning_alg='lbfgs',
stas = (Xstim.T @ (Rvalid-np.mean(Rvalid, axis=0))) / np.sum(Rvalid, axis=0)
stas /= np.sum(stas,axis=0)
gqm0.networks[0].layers[0].weights[:] = deepcopy(stas[:])
gqm0.set_regularization('d2x', reg_val=bestreg, ffnet_target=0)
gqm0.set_regularization('d2x', reg_val=bestreg, ffnet_target=1)
gqm0.set_regularization('d2x', reg_val=bestreg, ffnet_target=2)
# train initial model
_ = gqm0.train(input_data=[Xstim], output_data=Rvalid,
train_indxs=Ui, test_indxs=Xi,
learning_alg='adam', opt_params=adam_params,
fit_variables=v2f0)
# f= plt.plot(np.asarray(LLxs))
#%% plot model
DU.plot_3dfilters(gqm0, ffnet=0)
DU.plot_3dfilters(gqm0, ffnet=1)
DU.plot_3dfilters(gqm0, ffnet=2)
#%% plot fit
LLx = gqm0.eval_models(input_data=Xstim, output_data=Rvalid, data_indxs=Xi, nulladjusted=True)
plt.figure()
plt.plot(LLx, '-o')
plt.axhline(0, color='k')
#%% Run eye correction
# Run it all once
eyeAtFrameCentered = (eyeAtFrame-(640, 380))
centers5, locs, LLspace1 = ne.get_corr_grid(gqm0, Stim, Robs, [NX,NY], cids,
# glm0.networks[0].layers[0].biases = np.mean(Rvalid,axis=0).astype('float32')
v2f0 = glm0.fit_variables(fit_biases=False)
v2f0[-1][-1]['biases'] = True
# train initial model
_ = glm0.train(input_data=[Xstim],
output_data=Rvalid,
train_indxs=Ui,
test_indxs=Xi,
learning_alg='lbfgs',
opt_params=lbfgs_params,
fit_variables=v2f0)
#%% plot filters
DU.plot_3dfilters(glm0)
LLx0 = glm0.eval_models(input_data=[Xstim],
output_data=Rvalid,
data_indxs=Xi,
nulladjusted=True)
plt.plot(LLx0, '-o')
plt.axhline(0)
# %% get crop indices
# Cxinds = ne.crop_indx(NX, range(1,30), range(1,30))
Cxinds = ne.crop_indx(NX, range(9, 24), range(9, 24))
# Cxinds = ne.crop_indx(NX, range(5,20), range(5,20))
# Cxinds = ne.crop_indx(NX, range(20,44), range(20,44))
NX2 = np.sqrt(len(Cxinds)).astype(int)
v2f = nim0.fit_variables(fit_biases=True)
# train
_ = nim0.train(input_data=[Xstim],
output_data=Robs,
train_indxs=Ui,
test_indxs=Xi,
learning_alg=optimizer,
opt_params=opt_params,
use_dropout=False,
fit_variables=v2f)
print("Done")
DU.plot_3dfilters(nim0)
# only include cells that are well fit
LLx = nim0.eval_models(input_data=[Xstim],
output_data=Robs,
data_indxs=Xi,
nulladjusted=True)
plt.plot(LLx, 'o')
cids = np.where(LLx > 0.05)[0]
NC = len(cids)
Robsv = deepcopy(Robs[:, cids])
num_subs = NC // 2
nim_par = NDNutils.ffnetwork_params(input_dims=[1, NX, NY, num_lags],
v2f = retV1.fit_variables(fit_biases=False)
v2f[0][0]['biases'] = True
v2f[-1][-1]['biases'] = True
#%% train
_ = retV1.train(input_data=[Xstim],
output_data=Robs,
train_indxs=Ui,
test_indxs=Xi,
silent=False,
learning_alg='adam',
opt_params=adam_params,
fit_variables=v2f)
# %% fit
DU.plot_3dfilters(retV1)
# ======================================================================
# ======================================================================
# ======================================================================
# ======================================================================
# STRAY CODE BELOW HERE
# ======================================================================
# ======================================================================
# ======================================================================
#%%
#%%
v2f = retV1.fit_variables(fit_biases=True)
v2f[0][0]['biases'] = True
v2f[-1][0]['biases'] = True
#%% train
_ = retV1b.train(input_data=[Xstim, Rvalid],
output_data=Rvalid,
train_indxs=Ui,
test_indxs=Xi,
silent=False,
learning_alg='adam',
opt_params=adam_params,
fit_variables=v2fb)
# %% fit
DU.plot_3dfilters(retV1b)
#%%
plt.figure(figsize=(10, 4))
plt.subplot(1, 2, 1)
plt.plot(retV1b.networks[0].layers[0].weights)
plt.title("Temporal Kernels")
plt.subplot(1, 2, 2)
plt.plot(retV1b.networks[1].layers[1].weights)
plt.title("Latent temporal kernel")
#%% get test likelihood
LLx1 = retV1b.eval_models(input_data=[Xstim, Rvalid],
output_data=Rvalid,
data_indxs=Xi,
nulladjusted=True,
use_gpu=False)