def load_df(exp_id, min_shots = None):
dirname = variational_model_PATH + "/trained_models/{0}_good/".format(exp_id)
data_dict_list = []
for filename in os.listdir(dirname):
if "info" not in filename:
continue
try:
info_dict = pickle.load(open(dirname + filename, "rb"))
except:
print("Cannot load {0}".format(filename))
data_record = info_dict["data_record"]
data_dict = parse_dict(filename)
if data_dict["exp_mode"] == "meta":
model = load_model_dict(info_dict["model_dict"][-1])
data_dict["mse_few_shot"] = plot_few_shot_loss(model, tasks_test, min_shots = min_shots)
elif data_dict["exp_mode"] == "baseline":
print(filename)
model = load_model_dict_net(info_dict["model_dict"][-1])
data_dict["mse_few_shot"] = [None]
print("\n{0}".format(filename))
plot_data_record(info_dict["data_record"], is_VAE = is_VAE)
data_dict["mean_last"] = np.mean(data_record['mse_mean_test'][-4:-1])
data_dict["median_last"] = np.mean(data_record['mse_median_test'][-4:-1])
data_dict["mean_min"] = np.min(data_record['mse_mean_test'])
data_dict["median_min"] = np.min(data_record['mse_median_test'])
print("mean_last: {0:6f}\tmean_min: {1:.6f}\tmedian_last: {2:.6f}\tmedian_min: {3:.6f}".format(data_dict["mean_last"], data_dict["mean_min"], data_dict["median_last"], data_dict["median_min"]))
if data_dict["exp_mode"] == "meta":
statistics_list_test, z_list_test = plot_task_ensembles(tasks_test, model.statistics_Net, model.generative_Net, is_VAE = is_VAE, title = "y_pred_test vs. y_test")
print("test statistics vs. z:")
data_dict["corr_info"] = plot_statistics_vs_z(z_list_test, statistics_list_test)
if "bounce" in task_id_list[0]:
if data_dict["exp_mode"] == "meta":
plot_individual_tasks_bounce(tasks_test, num_examples_show = 40, num_tasks_show = 9, master_model = model, num_shots = 200)
elif data_dict["exp_mode"] == "baseline":
plot_individual_tasks_bounce(tasks_test, num_examples_show = 40, num_tasks_show = 9, model = model, num_shots = 200)
else:
_ = plot_individual_tasks(tasks_test, master_model.statistics_Net, master_model.generative_Net, is_VAE = is_VAE, xlim = task_settings["xlim"])
data_dict_list.append(data_dict)
df = pd.DataFrame(data_dict_list)
df = df.sort_values(by = "median_last")
return df
def load_file(exp_id, filename):
dirname = variational_model_PATH + "/trained_models/{0}_good/".format(exp_id)
info_dict = pickle.load(open(dirname + filename, "rb"))
data_record = info_dict["data_record"]
data_dict = parse_dict(filename)
master_model = load_model_dict(info_dict["model_dict"][-1])
print("\n{0}".format(filename))
data_dict["mse_few_shot"] = plot_few_shot_loss(master_model, tasks_test)
plot_data_record(info_dict["data_record"], is_VAE = is_VAE)
data_dict["mean_last"] = np.mean(data_record['mse_mean_test'][-6:-1])
data_dict["median_last"] = np.mean(data_record['mse_median_test'][-6:-1])
data_dict["mean_min"] = np.min(data_record['mse_mean_test'])
data_dict["median_min"] = np.min(data_record['mse_median_test'])
print("mean_last: {0:6f}\tmean_min: {1:.6f}\tmedian_last: {2:.6f}\tmedian_min: {3:.6f}".format(data_dict["mean_last"], data_dict["mean_min"], data_dict["median_last"], data_dict["median_min"]))
statistics_list_test, z_list_test = plot_task_ensembles(tasks_test, master_model.statistics_Net, master_model.generative_Net, is_VAE = is_VAE, title = "y_pred_test vs. y_test")
print("test statistics vs. z:")
data_dict["corr_info"] = plot_statistics_vs_z(z_list_test, statistics_list_test)
if "bounce" in task_id_list[0]:
plot_individual_tasks_bounce(tasks_test, num_examples_show = 40, num_tasks_show = 9, master_model = master_model, num_shots = 200)
else:
_ = plot_individual_tasks(tasks_test, master_model.statistics_Net, master_model.generative_Net, is_VAE = is_VAE, xlim = task_settings["xlim"])
return master_model, info_dict, data_dict
statistics_list_train, z_list_train = plot_task_ensembles(tasks_train, statistics_Net, generative_Net, is_VAE = is_VAE, is_regulated_net = is_regulated_net, title = "y_pred_train vs. y_train", isplot = isplot)
statistics_list_test, z_list_test = plot_task_ensembles(tasks_test, statistics_Net, generative_Net, is_VAE = is_VAE, is_regulated_net = is_regulated_net, title = "y_pred_test vs. y_test", isplot = isplot)
record_data(data_record, [np.array(z_list_train), np.array(z_list_test), np.array(statistics_list_train), np.array(statistics_list_test)],
["z_list_train_list", "z_list_test_list", "statistics_list_train_list", "statistics_list_test_list"])
if isplot:
print("train statistics vs. z:")
plot_statistics_vs_z(z_list_train, statistics_list_train)
print("test statistics vs. z:")
plot_statistics_vs_z(z_list_test, statistics_list_test)
# Plotting individual test data:
if "bounce" in task_id_list[0]:
plot_individual_tasks_bounce(tasks_test, num_examples_show = 40, num_tasks_show = 6, master_model = master_model, num_shots = 200)
else:
print("train tasks:")
plot_individual_tasks(tasks_train, statistics_Net, generative_Net, generative_Net_logstd = generative_Net_logstd, is_VAE = is_VAE, is_regulated_net = is_regulated_net, xlim = task_settings["xlim"])
print("test tasks:")
plot_individual_tasks(tasks_test, statistics_Net, generative_Net, generative_Net_logstd = generative_Net_logstd, is_VAE = is_VAE, is_regulated_net = is_regulated_net, xlim = task_settings["xlim"])
print("=" * 50 + "\n\n")
try:
sys.stdout.flush()
except:
pass
if i % save_interval == 0 or to_stop:
record_data(info_dict, [master_model.model_dict, i], ["model_dict", "iter"])
pickle.dump(info_dict, open(filename + "info.p", "wb"))
if to_stop:
print("The training loss stops decreasing for {0} steps. Early stopping at {1}.".format(patience, i))
break
# In[ ]:
# plot_types = ["standard"]
plot_types = ["gradient"]
# plot_types = ["slider"]
for plot_type in plot_types:
if plot_type == "standard":
plot_data_record(data_record, is_VAE = is_VAE)
statistics_list_test, z_list_test = plot_task_ensembles(tasks_test, statistics_Net, generative_Net, is_VAE = is_VAE, title = "y_pred_test vs. y_test")
print("test statistics vs. z:")
plot_statistics_vs_z(z_list_test, statistics_list_test)
_ = plot_individual_tasks(tasks_test, statistics_Net, generative_Net, is_VAE = is_VAE, xlim = task_settings["xlim"])
elif plot_type == "gradient":
batch_size = 256
sample_task_id = task_id_list[0] + "_{0}".format(np.random.randint(num_test_tasks))
print("sample_task_id: {0}".format(sample_task_id))
((X_train, y_train), (X_test, y_test)), _ = tasks_test[sample_task_id]
epochs_statistics = 50
lr_statistics = 1e-3
optim_type_statistics = "adam"
# epochs = 50
# lr = 1e-3
# optimizer = "adam"
epochs = 50
statistics_list_train, z_list_train = plot_task_ensembles(tasks_train, master_model = master_model, model = model, is_time_series = is_time_series, is_VAE = is_VAE, is_uncertainty_net = is_uncertainty_net, is_regulated_net = is_regulated_net, title = "y_pred_train vs. y_train", isplot = isplot)
statistics_list_test, z_list_test = plot_task_ensembles(tasks_test, master_model = master_model, model = model, is_time_series = is_time_series, is_VAE = is_VAE, is_uncertainty_net = is_uncertainty_net, is_regulated_net = is_regulated_net, title = "y_pred_test vs. y_test", isplot = isplot)
record_data(data_record, [np.array(z_list_train), np.array(z_list_test), np.array(statistics_list_train), np.array(statistics_list_test)],
["z_list_train_list", "z_list_test_list", "statistics_list_train_list", "statistics_list_test_list"])
if isplot:
print("train statistics vs. z:")
plot_statistics_vs_z(z_list_train, statistics_list_train)
print("test statistics vs. z:")
plot_statistics_vs_z(z_list_test, statistics_list_test)
# Plotting individual test data:
if "bounce" in task_id_list[0]:
plot_individual_tasks_bounce(tasks_test, num_examples_show = 40, num_tasks_show = 6, master_model = master_model, model = model, num_shots = 200, valid_input_dims = input_size - z_size, target_forward_steps = len(forward_steps), eval_forward_steps = len(forward_steps))
else:
print("train tasks:")
plot_individual_tasks(tasks_train, master_model = master_model, model = model, is_time_series = is_time_series, is_VAE = is_VAE, is_uncertainty_net = is_uncertainty_net, is_regulated_net = is_regulated_net, is_oracle = is_oracle, xlim = task_settings["xlim"])
print("test tasks:")
plot_individual_tasks(tasks_test, master_model = master_model, model = model, is_time_series = is_time_series, is_VAE = is_VAE, is_uncertainty_net = is_uncertainty_net, is_regulated_net = is_regulated_net, is_oracle = is_oracle, xlim = task_settings["xlim"])
print("=" * 50 + "\n\n")
try:
sys.stdout.flush()
except:
pass
if i % save_interval == 0 or to_stop:
if master_model is not None:
record_data(info_dict, [master_model.model_dict], ["model_dict"])
else:
record_data(info_dict, [model.model_dict], ["model_dict"])
pickle.dump(info_dict, open(filename + "info.p", "wb"))
if to_stop:
print("The training loss stops decreasing for {0} steps. Early stopping at {1}.".format(patience, i))
