def get_mode_connections(p1, t1, p2, t2, eval_task, config):
w1 = flatten_params(load_task_model_by_policy(t1, p1, config['exp_dir']))
w2 = flatten_params(load_task_model_by_policy(t2, p2, config['exp_dir']))
loss, acc, ts = calculate_mode_connectivity(w1, w2, loaders['sequential'][eval_task]['val'], config)
save_path = '{}/mc_{}_{}_to_{}_{}_on_{}'.format(config['exp_dir'],p1, t1, p2, t2, eval_task)
res = {'loss': loss, 'acc': acc, 'ts': ts}
save_np_arrays(res, path=save_path)
return res
def calculate_l2_distance(p1, t1, p2, t2, config):
m1 = load_task_model_by_policy(t1, p1, config['exp_dir'])
m2 = load_task_model_by_policy(t2, p2, config['exp_dir'])
key = 'l2_{}_{}_to_{}_{}'.format(p1, t1, p2, t2)
l2_dist = get_norm_distance(m1, m2)
log_comet_metric(experiment, key, l2_dist, 1)
return l2_dist
def plot_cka(p1, t1, p2, t2, eval_task, config):
m1 = load_task_model_by_policy(t1, p1, config['exp_dir'])
m2 = load_task_model_by_policy(t2, p2, config['exp_dir'])
save_path = '{}/cka_on_{}_{}_vs_{}_{}'.format(config['exp_dir'], p1, t1,
p2, t2)
scores, keys = calculate_CKA(m1,
m2,
loaders['sequential'][eval_task]['val'],
num_batches=50)
res = {'scores': scores, 'keys': keys}
save_np_arrays(res, path=save_path)
ylabel = r'$w^*_{}$'.format(
t1) if p1 == 'mtl' else r'$\hat{{w}}_{}$'.format(t1)
xlabel = r'$w^*_{}$'.format(
t2) if p2 == 'mtl' else r'$\hat{{w}}_{}$'.format(t2)
plot_heat_map(scores, keys, save_path, xlabel, ylabel)
return res
def plot_graphs(config):
# load models
models = {'seq': {}, 'mtl': {}}
for t in range(1, config['num_tasks'] + 1):
models['seq'][t] = flatten_params(
load_task_model_by_policy(t, 'seq', config['exp_dir']))
if t >= 2:
models['mtl'][t] = flatten_params(
load_task_model_by_policy(t, 'mtl', config['exp_dir']))
# plot_l2_distances(config)
# acc_fig_path = "{}/accs".format(config['exp_dir'])
# plot_accs(config['num_tasks'], seq_meter.data, mtl_meter.data, acc_fig_path)
# plot_cka_scores(config)
# --- task 1 ---
plot_mode_connections_for_minima('seq', 1, config)
plot_mode_connections_for_minima('seq', 2, config)
# get_custom_mode_connections_for_minima('seq', 1, config)
path = '{}/surface_{}_{}_{}_{}_{}_{}_on_{}'.format(config['exp_dir'],
'seq', 1, 'mtl', 2,
'seq', 2, 1)
labels = [r"$\hat{w}_1$", r"$w^*_{2}$", r"$\hat{w}_{2}$"]
plot_loss_plane([models['seq'][1], models['mtl'][2], models['seq'][2]],
loaders['sequential'][1]['val'], path, labels, config)
path = '{}/surface_{}_{}_{}_{}_{}_{}_on_{}'.format(config['exp_dir'],
'seq', 1, 'mtl', 2,
'seq', 2, 2)
labels = [r"$\hat{w}_1$", r"$w^*_{2}$", r"$\hat{w}_{2}$"]
plot_loss_plane([models['seq'][1], models['mtl'][2], models['seq'][2]],
loaders['sequential'][2]['val'], path, labels, config)
def plot_loss_plane(w, eval_loader, path, w_labels, config):
u = w[2] - w[0]
dx = np.linalg.norm(u)
u /= dx
v = w[1] - w[0]
v -= np.dot(u, v) * u
dy = np.linalg.norm(v)
v /= dy
m = load_task_model_by_policy(0, 'init', config['exp_dir'])
m.eval()
coords = np.stack(get_xy(p, w[0], u, v) for p in w)
# print("coords", coords)
G = 15
margin = 0.2
alphas = np.linspace(0.0 - margin, 1.0 + margin, G)
betas = np.linspace(0.0 - margin, 1.0 + margin, G)
tr_loss = np.zeros((G, G))
grid = np.zeros((G, G, 2))
for i, alpha in enumerate(alphas):
for j, beta in enumerate(betas):
p = w[0] + alpha * dx * u + beta * dy * v
m = assign_weights(m, p).to(DEVICE)
err = eval_single_epoch(m, eval_loader)['loss']
c = get_xy(p, w[0], u, v)
#print(c)
grid[i, j] = [alpha * dx, beta * dy]
tr_loss[i, j] = err
contour = {'grid': grid, 'values': tr_loss, 'coords': coords}
save_np_arrays(contour, path=path)
plot_contour(grid,
tr_loss,
coords,
log_alpha=-5.0,
N=7,
path=path,
w_labels=w_labels,
dataset='mnist') #config['dataset'])
return contour
def main():
print('Started the trial >>', TRIAL_ID, 'for experiment 1')
# init and save9
setup_experiment(experiment, config)
# convention: init => initialization
# convention: t_i_seq => task i (sequential)
# convention: t_i_mtl => task 1 ... i (multitask)
# convention: t_i_lcm => task 1 ... i (Linear Mode Connectivity)
eigen_spectrum = {1: {}, 2: {}}
for task in range(1, config['num_tasks'] + 1):
print('---- Task {} (seq) ----'.format(task))
seq_model = train_task_sequentially(
task, loaders['sequential'][task]['train'], config)
eigenvals, eigenvecs = get_model_eigenspectrum(
seq_model, loaders['sequential'][task]['val'])
eigen_spectrum[task]['eigenvals'], eigen_spectrum[task][
'eigenvecs'] = eigenvals, eigenvecs
save_task_model_by_policy(seq_model, task, 'seq', config['exp_dir'])
for prev_task in range(1, task + 1):
metrics = eval_single_epoch(
seq_model, loaders['sequential'][prev_task]['val'])
seq_meter.update(task, prev_task, metrics['accuracy'])
print(prev_task, metrics)
log_comet_metric(experiment, 't_{}_seq_acc'.format(prev_task),
metrics['accuracy'], task)
log_comet_metric(experiment, 't_{}_seq_loss'.format(prev_task),
round(metrics['loss'], 5), task)
if task == 1:
log_comet_metric(experiment, 'avg_acc', metrics['accuracy'],
task)
log_comet_metric(experiment, 'avg_loss', metrics['loss'], task)
if task > 1:
accs_mtl, losses_mtl = [], []
print('---- Task {} (mtl) ----'.format(task))
mtl_model = train_task_MTL(
task, loaders['full-multitask'][task]['train'], config,
loaders['sequential'][1]['val'])
#grads_t1 = get_model_grads(mtl_model, loaders['sequential'][1]['val'])
# grads_t2 = get_model_grads(mtl_model, loaders['sequential'][2]['val'])
grads_t1 = get_model_grads(
load_task_model_by_policy(1, 'seq',
config['exp_dir']).to(DEVICE),
loaders['full-multitask'][2]['train'])
grads_t3 = get_model_grads(
load_task_model_by_policy(1, 'seq',
config['exp_dir']).to(DEVICE),
loaders['sequential'][2]['train'])
seq_1 = flatten_params(
load_task_model_by_policy(1, 'seq', config['exp_dir']),
False).cpu()
seq_2 = flatten_params(
load_task_model_by_policy(2, 'seq', config['exp_dir']),
False).cpu()
cosines_t1 = compute_direction_cosines(
grads_t1, eigen_spectrum[1]['eigenvecs'])
# cosines_t2 = compute_direction_cosines(grads_t2, eigen_spectrum[2]['eigenvecs'])
cosines_t3 = compute_direction_cosines(
grads_t3, eigen_spectrum[1]['eigenvecs'])
cosine_d1 = compute_direction_cosines(
(flatten_params(mtl_model, False).cpu() - seq_1),
eigen_spectrum[1]['eigenvecs'])
cosine_d2 = compute_direction_cosines(
seq_2 - seq_1, eigen_spectrum[1]['eigenvecs'])
print("cos 1 >> ", cosines_t1)
# print("cos 2 >> ", cosines_t2)
print("cos 3 >> ", cosines_t3)
print("dir 1 >>", cosine_d1)
print("dir 2 >>", cosine_d2)
save_task_model_by_policy(mtl_model, task, 'mtl',
config['exp_dir'])
for prev_task in range(1, task + 1):
metrics_mtl = eval_single_epoch(
mtl_model, loaders['sequential'][prev_task]['val'])
accs_mtl.append(metrics_mtl['accuracy'])
losses_mtl.append(metrics_mtl['loss'])
mtl_meter.update(task, prev_task, metrics['accuracy'])
print('MTL >> ', prev_task, metrics_mtl)
log_comet_metric(experiment, 't_{}_mtl_acc'.format(prev_task),
metrics_mtl['accuracy'], task)
log_comet_metric(experiment, 't_{}_mtl_loss'.format(prev_task),
round(metrics_mtl['loss'], 5), task)
log_comet_metric(experiment, 'avg_acc_mtl', np.mean(accs_mtl),
task)
log_comet_metric(experiment, 'avg_loss_mtl', np.mean(losses_mtl),
task)
print()
seq_meter.save(config)
mtl_meter.save(config)
experiment.log_asset_folder(config['exp_dir'])
experiment.end()
