[arr[:shortest_len] for arr in all_running_mins])
rs_mean = np.mean(stacked_running_mins, axis=0)
rs_min = np.min(stacked_running_mins, axis=0)
rs_max = np.max(stacked_running_mins, axis=0)
# ==================================================================
# Load the vanilla ES results
# ==================================================================
base_dir = 'saves/many_hparams/es/val_sum_loss'
metric_key = 'perf/unroll_obj'
all_es_running_mins = []
all_es_iterations = []
for exp_dir in os.listdir(base_dir):
try:
log = plot_utils.load_log(os.path.join(base_dir, exp_dir),
'iteration.csv')
running_min = np.minimum.accumulate(log[metric_key])
all_es_running_mins.append(running_min)
all_es_iterations.append(log['perf/total_inner_iterations'])
except:
pass
shortest_len = min([len(arr) for arr in all_es_running_mins])
es_iterations = all_es_iterations[0][:shortest_len]
es_stacked_running_mins = np.stack(
[arr[:shortest_len] for arr in all_es_running_mins])
es_mean = np.mean(es_stacked_running_mins, axis=0)
es_min = np.min(es_stacked_running_mins, axis=0)
es_max = np.max(es_stacked_running_mins, axis=0)
# ==================================================================
plot_utils.plot_heatmap(
'saves/mnist_grid_sgdm/grid_mnist_mlp_lr:inverse-time-decay_sgdm_T_5000_N_40/seed_3/result.pkl',
key='unroll_obj',
xlabel='Log LR Decay',
ylabel='Log Init LR',
cmap=plt.cm.Purples_r,
levels=30,
sigma=1.0,
use_smoothing=True,
show_contours=True,
contour_alpha=0.2,
figsize=(8, 6),
)
es_K10 = plot_utils.load_log(
'saves/mnist_lr_decay/train_sum_loss/es-mnist-mlp-obj:train_sum_loss-tune:lr:inverse-time-decay-T:5000-K:10-nc:1-npc:1000-sigma:0.1-olr:0.01-ob1:0.9-ob2:0.999-ic:-1-oc:-1-seed:3',
fname='frequent.csv')
pes_K10 = plot_utils.load_log(
'saves/mnist_lr_decay/train_sum_loss/pes-mnist-mlp-obj:train_sum_loss-tune:lr:inverse-time-decay-T:5000-K:10-nc:1-npc:1000-sigma:0.1-olr:0.01-ob1:0.9-ob2:0.999-ic:-1-oc:-1-seed:3',
fname='frequent.csv')
es_K100 = plot_utils.load_log(
'saves/mnist_lr_decay/train_sum_loss/es-mnist-mlp-obj:train_sum_loss-tune:lr:inverse-time-decay-T:5000-K:100-nc:1-npc:1000-sigma:0.1-olr:0.01-ob1:0.9-ob2:0.999-ic:-1-oc:-1-seed:3',
fname='frequent.csv')
pes_K100 = plot_utils.load_log(
'saves/mnist_lr_decay/train_sum_loss/pes-mnist-mlp-obj:train_sum_loss-tune:lr:inverse-time-decay-T:5000-K:100-nc:1-npc:1000-sigma:0.1-olr:0.01-ob1:0.9-ob2:0.999-ic:-1-oc:-1-seed:3',
fname='frequent.csv')
plt.plot(np.log10(np.exp(pes_K10['lr_1']))[:2000],
np.log10(np.exp(pes_K10['lr_0']))[:2000],
color=colors[1],
plot_utils.plot_heatmap(
'saves/mnist_grid/grid_mnist_mlp_lr:inverse-time-decay_sgdm_T_5000_N_100/seed_3/result.pkl',
key='train_sum_loss',
xlabel='Decay',
ylabel='Initial LR',
cmap=plt.cm.Purples_r,
levels=10,
figsize=(8, 6),
)
plt.xticks([-2, -1, 0, 1, 2], fontsize=18)
plt.yticks([-3, -2, -1, 0, 1], fontsize=18)
es_K100 = plot_utils.load_log(
'saves/mnist_lr_decay/train_sum_loss/es-mnist-mlp-obj:train_sum_loss-tune:lr:itd-T:5000-K:100-N:100-sigma:0.1-olr:0.1-seed:3',
fname='frequent.csv')
es_K10 = plot_utils.load_log(
'saves/mnist_lr_decay/train_sum_loss/es-mnist-mlp-obj:train_sum_loss-tune:lr:itd-T:5000-K:10-N:100-sigma:0.1-olr:0.01-seed:3',
fname='frequent.csv')
es_K1 = plot_utils.load_log(
'saves/mnist_lr_decay/train_sum_loss/es-mnist-mlp-obj:train_sum_loss-tune:lr:itd-T:5000-K:1-N:100-sigma:0.1-olr:0.001-seed:3',
fname='frequent.csv')
pes_K100 = plot_utils.load_log(
'saves/mnist_lr_decay/train_sum_loss/pes-mnist-mlp-obj:train_sum_loss-tune:lr:itd-T:5000-K:100-N:100-sigma:0.1-olr:0.03-seed:3',
fname='frequent.csv')
pes_K10 = plot_utils.load_log(
'saves/mnist_lr_decay/train_sum_loss/pes-mnist-mlp-obj:train_sum_loss-tune:lr:itd-T:5000-K:10-N:100-sigma:0.1-olr:0.001-seed:3',
fname='frequent.csv')
pes_K1 = plot_utils.load_log(
),
('ES K=100',
'saves/control/es-K100/es-Swimmer-v2-lr:0.1-sigma:0.1-N:10-T:1000-K:100-c:0-d:0'
),
('PES K=100',
'saves/control/pes-K100/pes-Swimmer-v2-lr:0.1-sigma:0.3-N:10-T:1000-K:100-c:0-d:0'
),
]
for (label, exp_dir) in exps:
steps = []
rewards = []
min_num_rewards = 1e9
for seed_dir in os.listdir(exp_dir):
log = plot_utils.load_log(os.path.join(exp_dir, seed_dir),
fname='iteration.csv')
steps.append(log['total_steps'])
rewards.append(log['reward_mean'])
if len(log['reward_mean']) < min_num_rewards:
min_num_rewards = len(log['reward_mean'])
for list_idx in range(len(rewards)):
rewards[list_idx] = rewards[list_idx][:min_num_rewards]
means = np.array(rewards).mean(axis=0)
stds = np.array(rewards).std(axis=0)
steps = steps[0][:min_num_rewards]
plt.plot(steps, means, linewidth=2, label=label)
plt.fill_between(steps, means - stds, means + stds, linewidth=2, alpha=0.3)