def get_data(simple=False, look_back=5):
if simple:
n_traj = 20
train_data = gen_timeline(40, n_traj)
x = np.linspace(0, n_traj, n_traj)
train_labels = np.zeros(
train_data.shape) #np.vstack([np.sin(k*x) for i in range(n_data)])
n_data = 40
k = 0.05
for dim in range(3):
new_train_labels = np.vstack(
[np.sin(k * (1 + dim) * x) for i in range(n_data)])
train_labels[:, :, dim] = new_train_labels
#plot_forces(train_labels[0, :,:])
#x1_train = train_data
trainX, trainY = create_dataset(train_data, look_back=look_back)
return trainX, trainY
#y1_train = train_labels
#return y1_train, y1_train
else:
good_trajs, bad_trajs = get_good_bad_traj()
good_trajs = [smooth_traj(traj, n=5) for traj in good_trajs]
bad_trajs = [smooth_traj(traj, n=5) for traj in bad_trajs]
good_trajs = [subsample_traj(traj, n=5) for traj in good_trajs]
bad_trajs = [subsample_traj(traj, n=5) for traj in bad_trajs]
n_traj = good_trajs[0].shape[0]
print("traj shape", np.array(good_trajs).shape)
if PLOT:
plot_forces(good_trajs[2])
num_train_good = int(0.75 * len(good_trajs))
num_train_bad = int(0 * len(bad_trajs))
#data, _ = make_good_and_bad_dataset(num_train_good, num_train_bad, good_trajs, bad_trajs, test=False)
#data = np.zeros((num_train_good+num_train_bad, n_traj, 3)) #HACK
#for i in range(num_train_good+num_train_bad):
# data[i, :, :] = good_trajs[0]
#timeline = gen_timeline(data.shape[0], n_traj)
#data = data[:,:,:,0]
#data shifted back by repeating the first one lag+1 times
trainX, trainY = create_dataset(good_trajs[:num_train_good],
look_back=look_back)
return trainX, trainY
train_data = np.zeros(data.shape)
first_row = data[:, 0, :]
lag = 1
for i in range(lag):
train_data[:, i, :] = data[:, i, :] #first_row
#and the rest
train_data[lag:, :, :] = data[:lag, :, :]
return train_data, data
def predict(model, curr_set=np.zeros((1, 3, 5)), numsteps=20, upsample=1):
full_prediction = []
full_prediction = None
for i in range(int(numsteps / upsample)):
curr_set[:, :, :-1] = curr_set[:, :, 1:]
curr_set[:, :, -1] = model.predict(curr_set)[0].T
for i in range(upsample):
if full_prediction is None:
full_prediction = curr_set[:, :, -1]
else:
full_prediction = np.vstack(
[full_prediction, curr_set[:, :, -1]])
plot_forces(full_prediction)
return full_prediction
forces = predict(model, curr_set=curr_forces, numsteps=100, upsample=5)
#print("Min", np.min(forces.flatten()))
#print("Max", np.max(forces.flatten()))
ipdb.set_trace()
cortical_response, classification_vector = encoder(
forces) #we want the second term to be 1
dist = distance(cortical_response, good_responses)
prediction = classification_vector[0][1]
force_list.append(forces)
class_list.append(prediction)
force_to_dist[i] = dist
dist_list.append(dist)
#best_i = min(force_to_dist.keys(), key=lambda x: force_to_dist[x])
best_i = np.argmax(class_list)
lowest_dist = force_to_dist[best_i]
print("lowest dist", lowest_dist)
print("highest class", class_list[best_i])
if PLOT:
plt.scatter(dist_list, class_list)
plt.xlabel("Distance")
plt.ylabel("Probability successful | theta")
plt.show()
return force_list[best_i]
N = 30
#plot_forces(gen_parameterized_forces(gen_weights(N=N), 100, N=N))
plot_forces(find_best_encoding(N=N))