def train_epoch(rollouts):
actor.train()
critic.train()
loss = 0
total_len = 0
for states, actions, rewards, next_states in rollouts:
# 2nd step: use advantage function, estimated by critic
# bootstrap estimated next state values with rewards TD-1
values = critic(states)
last_state = next_states[-1].unsqueeze(0)
last_value = critic(last_state).item()
next_values = bootstrap(rewards, last_value, discounting)
advantage = normalize(next_values - values).flatten()
loss_critic = .5 * (advantage**2).sum()
# Get probabilities, shape (episode_length * num_actions)
# Then select only the probabilities corresponding to sampled actions
probabilities = actor(states)
probabilities = probabilities[range(states.shape[0]),
actions.flatten()]
loss_actor = (-torch.log(probabilities) * advantage).sum()
# Take the weighted average (helps convergence)
loss += loss_critic + loss_actor
total_len += states.shape[0]
loss = loss / total_len
optim_actor.zero_grad()
optim_critic.zero_grad()
loss.backward()
optim_actor.step()
optim_critic.step()
actor.eval()
critic.eval()
def random_forest_classifier(dataset, random_number=None, tree_number=10):
rf = split_tree.RandomForest()
end, feature = dataset.shape
feature = range(feature - 1)
if random_number == None:
dataset_number = end / 2
for i in range(tree_number):
sample_set = bootstrap(dataset, dataset_number)
sample_feature = random.sample(feature, len(feature) / 2)
tree = choose_feature(sample_set, 0, end / 2, sample_feature)
rf.Add_Tree(tree)
return rf
def main():
bootstrap()
('/skia-telemetry/skia_telemetry_info_page?',
skia_telemetry.TelemetryInfoPage),
('/skia-telemetry/skia_try', skia_telemetry.SkiaTryPage),
('/skia-telemetry/update_admin_tasks?',
skia_telemetry.UpdateAdminTasksPage),
('/skia-telemetry/update_chromium_build_tasks?',
skia_telemetry.UpdateChromiumBuildTasksPage),
('/skia-telemetry/update_chromium_try_tasks?',
skia_telemetry.UpdateChromiumTryTasksPage),
('/skia-telemetry/update_skia_try_tasks?',
skia_telemetry.UpdateSkiaTryTasksPage),
('/skia-telemetry/update_telemetry_tasks?',
skia_telemetry.UpdateTelemetryTasksPage),
('/skia-telemetry/update_lua_tasks?', skia_telemetry.UpdateLuaTasksPage),
('/skia-telemetry/update_telemetry_info?', skia_telemetry.UpdateInfoPage),
('/update_gpu_sheriffs_schedule', sheriff.UpdateGpuSheriffsSchedule),
('/update_robocops_schedule', sheriff.UpdateRobocopsSchedule),
('/update_sheriffs_schedule', sheriff.UpdateSheriffsSchedule),
('/update_troopers_schedule', sheriff.UpdateTroopersSchedule),
]
APPLICATION = webapp.WSGIApplication(URLS, debug=True)
# Do some one-time initializations.
base_page.bootstrap()
builder_status.bootstrap()
commit_queue.bootstrap()
status.bootstrap()
sheriff.bootstrap()
skia_telemetry.bootstrap()
utils.bootstrap()
('/skia-telemetry/skia_telemetry_info_page?',
skia_telemetry.TelemetryInfoPage),
('/skia-telemetry/skia_try', skia_telemetry.SkiaTryPage),
('/skia-telemetry/update_admin_tasks?', skia_telemetry.UpdateAdminTasksPage),
('/skia-telemetry/update_chromium_build_tasks?',
skia_telemetry.UpdateChromiumBuildTasksPage),
('/skia-telemetry/update_chromium_try_tasks?',
skia_telemetry.UpdateChromiumTryTasksPage),
('/skia-telemetry/update_skia_try_tasks?',
skia_telemetry.UpdateSkiaTryTasksPage),
('/skia-telemetry/update_telemetry_tasks?',
skia_telemetry.UpdateTelemetryTasksPage),
('/skia-telemetry/update_lua_tasks?', skia_telemetry.UpdateLuaTasksPage),
('/skia-telemetry/update_telemetry_info?', skia_telemetry.UpdateInfoPage),
('/update_gpu_sheriffs_schedule', sheriff.UpdateGpuSheriffsSchedule),
('/update_robocops_schedule', sheriff.UpdateRobocopsSchedule),
('/update_sheriffs_schedule', sheriff.UpdateSheriffsSchedule),
('/update_troopers_schedule', sheriff.UpdateTroopersSchedule),
]
APPLICATION = webapp.WSGIApplication(URLS, debug=True)
# Do some one-time initializations.
base_page.bootstrap()
builder_status.bootstrap()
commit_queue.bootstrap()
status.bootstrap()
sheriff.bootstrap()
skia_telemetry.bootstrap()
utils.bootstrap()
def main():
bootstrap()
def cli():
bootstrap()
def test_bootstrap():
with mock.patch("numpy.random.randint", return_value=np.arange(len(x))):
X_subset, z_subset = bootstrap(X, z, 1)
assert (np.allclose(X_subset, X) and np.allclose(z, z_subset))
def singleDriverTrainer2(file_to_classify,
training_files,
threshold=0.2,
in_model=RandomForestClassifier()):
"""
Takes in the file path of the driver file we want to classify (the target),
the paths of the files we will use as our 'noise' files,
and the input model
First, trains the input model on all of the files, with file_to_classify
as class 1 and training_files as class 0
Then, uses the model to make probabilistic predictions on file_to_classify
Changes:
1. Upsamples target data to balance classes for model training
2. Uses probabilistic predictions relabels 1s to 0s based on threshhold
"""
# first, grab the target data
x_target, y_target, id_target = extractCSV(file_to_classify, target=1)
# remove na's
x_target = np.nan_to_num(x_target)
y_target = np.nan_to_num(y_target)
# copy target data
x_target_upsampled = copy.copy(x_target)
y_target_upsampled = copy.copy(y_target)
#upsample target to balance classes
if len(training_files) > 1:
num_samples = len(x_target_upsampled) * len(training_files)
x_target_upsampled, y_target_upsampled = bootstrap(
x_target, y_target, num_samples)
x_trains = None
y_trains = None
# loop through all of our training/noise files, keep separate from target
for filepath in training_files:
# open the file
x_current, y_current, ids = extractCSV(filepath, target=0)
# and add the contents to our training data
if x_trains is None or y_trains is None:
x_trains = x_current
y_trains = y_current
else:
x_trains = np.concatenate((x_trains, x_current))
y_trains = np.concatenate((y_trains, y_current))
# repeat for every filepath in our training files list
# remove NAs from train data
x_trains = np.nan_to_num(x_trains)
y_trains = np.nan_to_num(y_trains)
# now combine with target data
x_all = np.concatenate((x_target_upsampled, x_trains))
y_all = np.concatenate((y_target_upsampled, y_trains))
# with all of our data, now we can train our model
in_model.fit(x_all, y_all)
# now we are ready to provide class probabilities for our predictions
predictions = in_model.predict_proba(x_target)
# note that we must extract the index of the class 1 probability
prob_idx = np.where(in_model.classes_ == 1)[0][0]
class_probs = [pred[prob_idx] for pred in predictions]
#get new data labels by comparing threshold to class probs
new_labels = np.array([1 if p > threshold else 0 for p in class_probs])
#redo upsampling
if len(training_files) > 1:
num_samples = len(x_target) * len(training_files)
x_target_relabeled, y_target_relabeled = bootstrap(
x_target, new_labels, num_samples)
else:
x_target_relabeled = copy.copy(x_target)
y_target_relabeled = copy.copy(new_labels)
#combine with non-target data from before
x_all_new = np.concatenate((x_target_relabeled, x_trains))
y_all_new = np.concatenate((y_target_relabeled, y_trains))
# refit model
in_model.fit(x_all_new, y_all_new)
# provide class probabilities for our predictions
predictions = in_model.predict_proba(x_target)
# extract the index of the class 1 probability
prob_idx = np.where(in_model.classes_ == 1)[0][0]
class_probs = [pred[prob_idx] for pred in predictions]
# and return a matrix of the id's and the corresponding probabilities
return_mat = [[id_target[idx], class_probs[idx]] \
for idx in xrange(len(class_probs))]
# report
print 'completed driver %s' % file_to_classify
return np.asarray(return_mat)
def random_forest(N, M, F, table, attr_indexes, attr_domains, class_index,
strat_index):
random.shuffle(table)
test, remainder = test_remainder_stratified(table, strat_index)
boot_samples = []
attr_subsets = []
trees = []
accuracies = []
trees = []
#setup boot straps
for _ in range(N):
attr_subsets.append(utils.rand_attributes(attr_indexes, F))
boot = utils.bootstrap(remainder)
valid = []
#build validator set
for item in remainder:
if item not in boot:
valid.append(item)
boot_samples.append([boot, valid])
#build trees
for i in range(N):
#returns predictions, tree
pred, tree = train_test_tree(boot_samples[i][0], boot_samples[i][1],
attr_subsets[i], attr_domains,
class_index)
correct = 0
for j in range(len(boot_samples[i][1])):
if boot_samples[i][1][j][class_index] == pred[j]:
correct += 1
trees.append([tree, utils.div(correct, len(boot_samples[i][1]))])
trees.sort(key=lambda x: x[1])
mtrees = trees[len(trees) - M:]
#predict and determine accuracy
print(" grouping test set")
minutes, groups = utils.groupBy(test, 1)
print(" running classifier")
accuracies = []
overall_correct = 0
total_instance = len(test)
for count in range(len(minutes)):
correct = 0
for item in groups[count]:
votes = []
for tree in mtrees:
votes.append(classify_tdidt(tree[0], item))
vote = utils.majority_vote(votes)
if item[class_index] == vote:
correct += 1
overall_correct += 1
accuracies.append([
minutes[count], correct / len(groups[count]), correct,
len(groups[count])
])
print("Sorting accuracies")
accuracies.sort(key=lambda x: x[0])
count = 0
for item in accuracies:
print('Minute: ', item[0])
print(' Accuracy: ', item[1])
print(' Correct: ', item[2])
print(' Instances: ', item[3])
print()
count += 1
print("Overll Accurracy: ", overall_correct / total_instance)
print("Instances: ", total_instance)
print("Correct: ", overall_correct)
return accuracies
def resample(models, lmd, X, z, nboots, split_size=0.2):
""" Dictionaires to keep track of the results """
z_test = {"ridge": [], "lasso": [], "ols": []}
z_pred_test = {"ridge": [], "lasso": [], "ols": []}
bias = {"ridge": [], "lasso": [], "ols": []}
var = {"ridge": [], "lasso": [], "ols": []}
beta = {"ridge": [], "lasso": [], "ols": []}
mse_test = {"ridge": [], "lasso": [], "ols": []}
#r2_test = {"ridge": [], "lasso": [], "ols": []}
" ----------------------"
mse_train = {"ridge": [], "lasso": [], "ols": []}
#r2_train = {"ridge": [], "lasso": [], "ols": []}
np.random.seed(2018)
# Spilt the data in tran and split
X_train, X_test, z_train, z_test_ = train_test_split(X,
z,
test_size=split_size)
# # extract data from design matrix
# x = X[:, 1]
# y = X[:, 2]
# x_test = X_test[:, 1]
# y_test = X_test[:, 2]
for name, model in models.items():
# creating a model with the previosly known best lmd
estimator = model(lmd[name])
# Train a model for this pair of lambda and random state
""" Keeping information for test """
estimator.fit(X_train, z_train)
z_pred_test_ = np.empty((z_test_.shape[0], nboots))
z_pred_train_ = np.empty((z_train.shape[0], nboots))
beta_ = np.empty((X.shape[1], nboots))
for i in range(nboots):
X_, z_ = bootstrap(
X_train, z_train,
i) # i is now also the random state for the bootstrap
estimator.fit(X_, z_)
# Evaluate the new model on the same test data each time.
z_pred_test_[:, i] = np.squeeze(estimator.predict(X_test))
z_pred_train_[:, i] = np.squeeze(estimator.predict(X_train))
beta_[:, i] = np.squeeze(estimator.coef_)
beta[name] = beta_
z_pred_test[name] = z_pred_test_
z_test_ = z_test_.reshape((z_test_.shape[0], 1))
z_test[name] = z_test_
mse_test[name] = (np.mean(
np.mean((z_test_ - z_pred_test_)**2, axis=1, keepdims=True)))
bias[name] = np.mean(
(z_test_ - np.mean(z_pred_test_, axis=1, keepdims=True))**2)
var[name] = np.mean(np.var(z_pred_test_, axis=1, keepdims=True))
z_train = z_train.reshape((z_train.shape[0], 1))
mse_train[name] = np.mean(
np.mean((z_train - z_pred_train_)**2, axis=1, keepdims=True))
# print('Error:', mse_test)
# print('Bias^2:', bias)
# print('Var:', var)
# print('{} >= {} + {} = {}'.format(mse_test, bias, variance, bias + variance))
# plt.figure(1, figsize=(11, 7))
# plt.subplot(121)
# plt.plot(x, z, label='f')
# plt.scatter(x_test, z_test, label='Data points')
# plt.scatter(x_test, np.mean(z_pred, axis=1), label='Pred')
# plt.legend()
# plt.xlabel('x')
# plt.ylabel('z')
#
# plt.subplot(122)
# plt.plot(y, z, label='f')
# plt.scatter(y_test, z_test, label='Data points')
# plt.scatter(y_test, np.mean(z_pred, axis=1), label='Pred')
# plt.legend()
# plt.xlabel('y')
# plt.ylabel('z')
# plt.show()
# Confidence intervals
ci_beta = np.empty((2, beta_.shape[0]))
poly = []
for p in range(beta_.shape[0]):
ci_beta[:, p] = np.array(ci(beta_[p, :])).T
poly.append(p)
# plt.plot(poly, ci_beta[0, :], label='Upper CI (95%)') # --> Vise i tabell
# plt.plot(poly, np.mean(beta, axis=1), label='Beta')
# plt.plot(poly, ci_beta[1, :], label='Lower CI (95%)')
# plt.legend()
# plt.show()
return z_test, z_pred_test, bias, var, beta, mse_test, mse_train, ci_beta
import pickle
from functools import partial
from utils import breakout_strategy, bootstrap, m_proc
import sys
sys.path.append('/home/tomek/ib_tools/')
from datastore_pytables import Store # noqa
store = Store()
contract = store.read('/cont/min/NQ_20191220_GLOBEX_USD').sort_index()
table = bootstrap(contract, start='20180701', end='20181231', paths=100)
func = partial(
breakout_strategy,
time_int=30,
periods=[
5,
10,
20,
40,
80,
160,
],
ema_fast=10,
ema_slow=120,
atr_periods=80,
sl_atr=1,
)
results = m_proc(table, func)