def plot2d():
xs = [random_normal() for _ in range(1000)]
ys1 = [x + random_normal() / 2 for x in xs]
ys2 = [-x + random_normal() / 2 for x in xs]
print(correlation(xs, ys1))
print(correlation(xs, ys2))
plt.scatter(xs, ys1, marker='.', color='black', label='ys1')
plt.scatter(xs, ys2, marker='.', color='gray', label='ys2')
plt.xlabel('xs')
plt.ylabel('ys')
plt.legend(loc=9)
plt.title('Very Different Joint Distributions')
plt.show()
def main():
# compose data for xgboost
train, train_target, val, val_target = load_dataset() # np arrays [X 2048] [X 15] (large X)
# train = np.array([[1,2,3,4,5], [2,3,4,5,6], [4,5,6,7,8], [5,6,7,8,9]])
# train_target = np.array([3,4,5,6])
print(train.shape, train_target.shape, val[0].shape, val_target[0].shape)
regressor = xgb.XGBRegressor(tree_method='gpu_hist', predictor='gpu_predictor')
#
regressor.fit(train, train_target)
score = regressor.score(train, train_target)
print('Score:', score)
# print(regressor.predict(np.array([[6,7,8,9,10]])))
video_count = len(val)
corr_mean = 0.0
for i in range(video_count):
preds = regressor.predict(val[i])
preds = torch.from_numpy(preds).cuda()
preds = torch.cat([preds, preds[-1:]])
preds = preds.unsqueeze(1) # T -> T 1
preds = interpolate_output(preds, 1 , 6)
pred_len = val_target[i].shape[0]
preds = preds[:pred_len]
val_ti = torch.from_numpy(val_target[i]).cuda()
val_ti = val_ti.unsqueeze(1) # T -> T 1
corr, _ = correlation(preds, val_ti)
corr_mean += corr.item()
print('Correlation:', corr_mean/video_count)
def test_result(filt_emp):
global best_conf
# do a correlation evalutation
filt_emp = np.concatenate((filt_emp, filt_emp[-1:]))
output = interpolate_output(torch.from_numpy(filt_emp), 1, 6)
cor, _ = correlation(output[:vid_labels.shape[0]],
torch.from_numpy(vid_labels))
print('Correlation:', cor.item())
return cor.item()
def _generate_complete_graph(D: spn.Dataset, S: spn.Scope):
n = len(S)
G = graph_tool.generation.complete_graph(n,
self_loops=False,
directed=False)
W = G.new_edge_property('double')
G.edge_properties['weights'] = W
for e in G.edges():
s = int(e.source())
t = int(e.target())
if e not in W:
W[e] = utils.correlation(D, s, t)
return G
def _generate_sample_graph(D: spn.Dataset, S: spn.Scope, M: VertexMap,
sample_func):
G = graph_tool.Graph(directed=False)
W = G.new_edge_property('double')
G.edge_properties['weights'] = W
G.add_vertex(n=len(S))
for v in S:
N = sample_func(S, v)
for u in N:
s = M.g(v)
t = M.g(u)
if G.edge(s, t) is None:
e = G.add_edge(s, t)
W[e] = utils.correlation(D, s, t)
return G
def selected_data_using_corr(self):
# selected data using corr.
df = self._des2cat(self.df, self.category_cols)
X_data = df.drop(['price'], axis=1)
y_data = np.log(df['price'])
top_corr = correlation(df)
X_data = X_data.loc[:, top_corr[1:]]
# vis X_data
self.max_X_data = np.max(X_data.max())
vis_boxplot(X_data,
save_path='./selected_data_using_corr',
vmax=self.max_X_data)
y_data = pd.DataFrame(y_data)
return pd.concat((y_data, X_data), axis=1)
def extraFea(self, x):
output = []
input = x
for x in input:
result = []
x = abs(x[3].cpu().numpy())
result.append(utils.mean_abs_dev(x))
result.append(utils.average(x))
result.append(utils.fre_skewness(x))
result.append(utils.fre_kurtosis(x))
result.append(utils.energy(x))
result.append(utils.entropy(x))
#temp = utils.ar_coef(x)
#for i in temp:
# result.append(i)
x1 = x[0:int(len(x) / 2)]
x2 = x[int(len(x) / 2):]
result.append(utils.correlation(x1, x2))
result.append(utils.fswa(x))
output.append([result])
return output
def insight():
scope = [
'https://spreadsheets.google.com/feeds',
'https://www.googleapis.com/auth/drive'
]
creds = ServiceAccountCredentials.from_json_keyfile_name(
'SIOTfinal.json', scope)
client = gspread.authorize(creds)
sound_g = client.open('sound').sheet1
room_g = client.open('room').sheet1
outside_g = client.open('outside').sheet1
room_df = get_room_temp(room_g)
out_df = get_outside_temp(outside_g)
sound_df = get_sound(sound_g)
intervals = ["15min", "30min", "1h", "4h"]
dates = []
dates_obj = []
all_date = sound_df.resample('D').sum()
for i in range(len(all_date.index)):
dates.append(str(all_date.index[i].strftime('%Y-%m-%d')))
dates_obj.append(all_date.index[i])
dates = dates[1:]
dates_obj = dates_obj[1:]
variables = [
'room temperature', 'room humidity', 'local temperature',
'local humidity', 'atmospheric pressure', 'wind speed', 'cloud'
]
room_variables = ['room temperature', 'room humidity']
out_variables = [
'local temperature', 'local humidity', 'atmospheric pressure',
'wind speed', 'cloud'
]
current_date = request.args.get("dates")
current_interval = request.args.get("intervals")
current_variable = request.args.get("variables")
if current_interval == None:
current_interval = "15min"
if current_date == None:
current_date = str(datetime.today().strftime('%Y-%m-%d'))
if current_variable == None:
current_variable = 'room temperature'
date = current_date[8:10] + current_date[5:7] + current_date[0:4]
sound_date = time_select(sound_df, date)
room_date = time_select(room_df, date)
out_date = time_select(out_df, date)
if current_variable in room_variables:
col = room_date[current_variable]
elif current_variable in out_variables:
col = out_date[current_variable]
relation_plot = sound_line(sound_date['Trigger bool'], col,
current_interval)
correlation_plot = correlation(dates_obj, sound_df, room_df, out_df)
s1, div1 = components(relation_plot)
s2, div2 = components(correlation_plot)
return render_template('insight.html',
title="Insight",
s1=s1,
div1=div1,
s2=s2,
div2=div2,
intervals=intervals,
current_variable=current_variable,
variables=variables,
current_interval=current_interval,
dates=dates,
current_date=current_date)
def _generate_complete_network(D: spn.Dataset, S: spn.Scope):
G = nx.complete_graph(S)
for i, u in enumerate(S):
for j, v in enumerate(S[i + 1:]):
G.add_edge(u, v, weight=utils.correlation(D, i, j + i + 1))
return G
def future_policy_value(self,
x,
a,
trans,
seq_len,
seq_mask,
agent,
opt,
create_summary=False):
"""
Computes the value of a policy according to the critic when updated using the objective function
:param x: observations
:param a: actions
:param trans: entire tuple of transition (s_t, a_t, r_t, d_t, s_{t+1})
:param seq_len: Length of trajectories
:param seq_mask: Binary mask of trajectories
:param agent: agent to compute value for
:param opt: optimizer to use for the policy update
:param create_summary: whether to create summary ops
:return: tensor of batched future policy value
"""
with tf.variable_scope('future_policy_value'):
policy = agent.main.policy
policy_vars = policy.trainable_variables
# The replace manager can replace the policy variables with updated variables
replace_manager = policy.variable_scope.custom_getter
use_adam = self.dconfig.obj_func_second_order_adam
step_size = self.dconfig.obj_func_second_order_stepsize
step_count = self.dconfig.obj_func_second_order_steps + 1
batch_size = self.dconfig.buffer_sample_size
# Split tensors according to number of inner gradient descent steps
x_s = tf.split(x, step_count, axis=0)
a_s = tf.split(a, step_count, axis=0)
if seq_len is not None:
seq_len_s = tf.split(seq_len, step_count, axis=0)
seq_mask_s = tf.split(seq_mask, step_count, axis=0)
else:
seq_len_s = utils.ConstArray()
seq_mask_s = utils.ConstArray(seq_mask)
trans_s = list(
zip(*(tf.split(e, step_count, axis=0) for e in trans)))
objective_val = None
policy_grads = None
opt_args_dict = {}
current_vars = policy_vars
var_names = [var.op.name for var in policy_vars]
for i in range(step_count - 1):
# Run policy
policy_result = policy(x_s[i], seq_len=seq_len_s[i])
# Run objective
objective_val = self.objective(x_s[i], a_s[i], trans_s[i],
seq_len_s[i], seq_mask_s[i],
agent, policy_result,
create_summary)
# Compute policy gradients
policy_grads = tf.gradients(objective_val * seq_mask_s[i],
current_vars)
if use_adam:
def grad_transform(grad, var, var_name):
if var_name in opt_args_dict:
opt_args = opt_args_dict[var_name]
else:
opt_args = []
new_grad, *opt_args = opt.adapt_gradients(grad,
var,
*opt_args,
lr=step_size)
opt_args_dict[var_name] = opt_args
return new_grad
else:
def grad_transform(grad, *args):
return step_size * grad
# Use adam or vanilla SGD for inner gradient step
transformed_grads = [
grad_transform(grad, var,
var_name) for grad, var, var_name in zip(
policy_grads, current_vars, var_names)
]
one_step_updated_policy_vars = [
var - grad
for var, grad in zip(current_vars, transformed_grads)
]
one_step_updated_policy_vars_dict = OrderedDict(
zip(var_names, one_step_updated_policy_vars))
# # Updates replace manager to run policy with updated variables in the next loop iteration
replace_manager.replace_dict = one_step_updated_policy_vars_dict
current_vars = one_step_updated_policy_vars
# Run policy with final parameters
future_policy = policy(x, seq_len=seq_len)
replace_manager.replace_dict = None
# Estimate the final policy value
future_policy_value = agent.main.critic(
x, future_policy.action) * seq_mask
if create_summary:
orig_policy = policy(x_s[-1], seq_len=seq_len_s[-1])
partial_future_policy_value = future_policy_value[-batch_size:]
tf.summary.histogram('objective_value', objective_val)
tf.summary.histogram('policy_grads', utils.flat(policy_grads))
tf.summary.histogram('policy_value', orig_policy.value)
tf.summary.histogram('future_policy_value',
partial_future_policy_value)
tf.summary.histogram(
'policy_value_gain',
partial_future_policy_value - orig_policy.value)
sample_axis = [
0, 1
] if self.dconfig.recurrent_time_steps > 1 else 0
cor = utils.correlation(-orig_policy.value, objective_val,
sample_axis)
tf.summary.scalar('objective_critic_correlation',
tf.squeeze(cor))
grad, = tf.gradients(objective_val, policy_result.value)
if grad is not None:
tf.summary.histogram('objective_critic_grads', grad)
return future_policy_value
def recusive_application_performance(self,
net,
dataset,
split_point,
samples=20):
print('===> Evaluating performance of recursive application')
path = os.path.join(config['output_dir'], self.output_name,
'recursive')
mkdir(path)
if split_point - samples / 2 < 0:
start_index = 0
end_index = int(split_point + samples)
else:
start_index = int(split_point - samples / 2)
end_index = int(split_point + samples / 2)
print('-- Start index:', start_index)
print('-- End index:', end_index)
mse = []
cor = []
psnr = []
ssim = []
diff_avrg = []
diff_max = []
diff_x = []
diff_y = []
change_psnr_x = []
change_psnr_y = []
change_diff_x = []
change_diff_y = []
input_img = dataset[start_index][0].expand(1, -1, -1,
-1).to(self.device)
if self.parameterized:
params = dataset[start_index][2].expand(1, -1, -1,
-1).to(self.device)
for index in range(start_index, end_index):
pred_input = self._prepare_tensor_img(input_img[0].clone(),
is_input=True)
if self.parameterized:
predicted = net((input_img, params))
else:
predicted = net(input_img)
target = dataset[index][1].expand(1, -1, -1, -1).to(self.device)
if self.args.mask:
for i, j in itertools.product(range(predicted.shape[0]),
range(predicted.shape[1])):
predicted[i][j] = self.MASK * predicted[i][j]
input_img = torch.cat((torch.tensor(
predicted.clone().detach()[0][0:3]).expand(
1, -1, -1, -1), self.MASK.expand(1, -1, -1, -1)),
1)
else:
input_img = predicted.clone().detach()
cur_mse = self.criterionMSE(predicted, target).item()
if not self.args.use_pressure:
predicted_x, predicted_y = self._prepare_tensor_img(
predicted[0])
target_x, target_y = self._prepare_tensor_img(
dataset[index][1])
else:
predicted_x, predicted_y, predicted_p = self._prepare_tensor_img(
predicted[0])
target_x, target_y, target_p = self._prepare_tensor_img(
dataset[index][1])
merge_and_save(
target_x, predicted_x, 'Real', 'Predicted',
os.path.join(path,
'x_recursive_{}.png'.format(index - start_index)))
merge_and_save(
target_y, predicted_y, 'Real', 'Predicted',
os.path.join(path,
'y_recursive_{}.png'.format(index - start_index)))
predicted_img = self.denormalize_output(
predicted).detach().cpu().numpy()
target_img = self.denormalize_input(target).detach().cpu().numpy()
mse += [cur_mse]
psnr += [10 * math.log10(1 / cur_mse)]
cor += [
np.average(
np.array([
correlation(predicted_img[i], target_img[i])
for i in range(predicted_img.shape[0])
]))
]
ssim += [
np.average(
np.array([
ssim_metr(predicted_img[i].T,
target_img[i].T,
multichannel=True)
for i in range(predicted_img.shape[0])
]))
]
diff_avrg_, _, diff_max_ = imgs_perc_diff(target_img,
predicted_img)
diff_avrg.append(diff_avrg_)
diff_max.append(diff_max_)
diff_x.append(
imgs_perc_diff(target_img[0][0], predicted_img[0][0])[0])
diff_y.append(
imgs_perc_diff(target_img[0][1], predicted_img[0][1])[0])
real_input = self._prepare_tensor_img(dataset[index][0], True)
change_x_real = np.abs(target_x - real_input[0])
change_x_predicted = np.abs(pred_input[0] - predicted_x)
change_y_real = np.abs(target_y - real_input[1])
change_y_predicted = np.abs(pred_input[1] - predicted_y)
change_mse_x = (np.square(change_x_real -
change_x_predicted)).mean(axis=None)
change_mse_y = (np.square(change_y_real -
change_y_predicted)).mean(axis=None)
change_psnr_x += [10.0 * np.log10(255.0 / np.sqrt(change_mse_x))]
change_psnr_y += [10.0 * np.log10(255.0 / np.sqrt(change_mse_y))]
change_diff_x += [
imgs_perc_diff(change_x_real, change_x_predicted)[0]
]
change_diff_y += [
imgs_perc_diff(change_y_real, change_y_predicted)[0]
]
merge_and_save(
change_x_real, change_x_predicted, 'Real', 'Predicted',
os.path.join(path,
'x_diff_{}.png'.format(index - start_index)))
merge_and_save(
change_y_real, change_y_predicted, 'Real', 'Predicted',
os.path.join(path,
'y_diff_{}.png'.format(index - start_index)))
print('> Recursive application {} completed'.format(index -
start_index))
with open(
os.path.join(self.root_dir, self.output_name,
'recursive_application.txt'), 'w') as list_hand:
list_hand.write('Split index: {}\n'.format(str(samples / 2)))
list_hand.write('{} {}\n'.format('mse: ',
','.join(str(i) for i in mse)))
list_hand.write('{} {}\n'.format('cor: ',
','.join(str(i) for i in cor)))
list_hand.write('{} {}\n'.format('psnr: ',
','.join(str(i) for i in psnr)))
list_hand.write('{} {}\n'.format('ssim: ',
','.join(str(i) for i in ssim)))
list_hand.write('{} {}\n'.format(
'diff_avrg: ', ','.join(str(i) for i in diff_avrg)))
list_hand.write('{} {}\n'.format(
'diff_max: ', ','.join(str(i) for i in diff_max)))
list_hand.write('{} {}\n'.format('x_diff_avrg: ',
','.join(str(i) for i in diff_x)))
list_hand.write('{} {}\n'.format('y_diff_max: ',
','.join(str(i) for i in diff_y)))
list_hand.write('{} {}\n'.format(
'change_psnr_x: ', ','.join(str(i) for i in change_psnr_x)))
list_hand.write('{} {}\n'.format(
'change_psnr_y: ', ','.join(str(i) for i in change_psnr_y)))
list_hand.write('{} {}\n'.format(
'change_diff_x: ', ','.join(str(i) for i in change_diff_x)))
list_hand.write('{} {}\n'.format(
'change_diff_y: ', ','.join(str(i) for i in change_diff_y)))
def individual_images_performance(self, net, test_dataloader):
print('===> Evaluating performance on individual images')
mse = []
cor = []
psnr = []
ssim = []
diff_avrg = []
diff_max = []
diff_x = []
diff_y = []
change_mse_x = []
change_mse_y = []
change_psnr_x = []
change_psnr_y = []
change_diff_x = []
change_diff_y = []
change_psnr = []
for iteration, batch in enumerate(test_dataloader, 1):
real_a, real_b = batch[0].to(self.device), batch[1].to(self.device)
if self.parameterized:
params = batch[2].to(self.device)
predicted = net((real_a, params))
else:
predicted = net(real_a)
if self.args.mask:
for i, j in itertools.product(range(predicted.shape[0]),
range(predicted.shape[1])):
predicted[i][j] = self.MASK * predicted[i][j]
cur_mse = self.criterionMSE(predicted, real_b).item()
predicted = self.denormalize_output(
predicted).detach().cpu().numpy()
real_a = self.denormalize_output(real_a).detach().cpu().numpy()
real_b = self.denormalize_output(real_b).detach().cpu().numpy()
mse += [cur_mse]
psnr += [10 * math.log10(1 / cur_mse)]
cor += [
np.average(
np.array([
correlation(predicted[i], real_b[i])
for i in range(predicted.shape[0])
]))
]
ssim += [
np.average(
np.array([
ssim_metr(predicted[i].T,
real_b[i].T,
multichannel=True)
for i in range(predicted.shape[0])
]))
]
diff_avrg_, _, diff_max_ = imgs_perc_diff(real_b, predicted)
diff_avrg.append(diff_avrg_)
diff_max.append(diff_max_)
for i in range(predicted.shape[0]):
diff_x.append(imgs_perc_diff(real_b[i][0], predicted[i][0])[0])
diff_y.append(imgs_perc_diff(real_b[i][1], predicted[i][1])[0])
# error images
batch_change_mse = []
batch_change_mse_x = []
batch_change_mse_y = []
batch_change_psnr_x = []
batch_change_psnr_y = []
batch_change_diff_x = []
batch_change_diff_y = []
for ind in range(real_a.shape[0]):
if self.args.use_pressure:
real_change_img = np.abs(real_a[ind][0:3] - real_b[ind])
else:
real_change_img = np.abs(real_a[ind][0:2] - real_b[ind])
predicted_change_img = np.abs(predicted[ind] - real_b[ind])
cur_mse = (np.square(real_change_img -
predicted_change_img)).mean(axis=None)
cur_psnr = 10 * np.log10(255.0 / np.sqrt(cur_mse))
batch_change_mse.append(cur_psnr)
real_change_img_x = np.abs(real_a[ind][0] - real_b[ind][0])
predicted_change_img_x = np.abs(predicted[ind][0] -
real_b[ind][0])
real_change_img_y = np.abs(real_a[ind][1] - real_b[ind][1])
predicted_change_img_y = np.abs(predicted[ind][1] -
real_b[ind][1])
x_cur_mse = (np.square(real_change_img_x -
predicted_change_img_x)).mean(axis=None)
y_cur_mse = (np.square(real_change_img_y -
predicted_change_img_y)).mean(axis=None)
x_cur_psnr = 10 * np.log10(255.0 / np.sqrt(x_cur_mse))
y_cur_psnr = 10 * np.log10(255.0 / np.sqrt(y_cur_mse))
cur_diff_x, _, _ = imgs_perc_diff(real_change_img_x,
predicted_change_img_x)
cur_diff_y, _, __ = imgs_perc_diff(real_change_img_y,
predicted_change_img_y)
batch_change_mse_x.append(x_cur_mse)
batch_change_mse_y.append(y_cur_mse)
batch_change_psnr_x.append(x_cur_psnr)
batch_change_psnr_y.append(y_cur_psnr)
batch_change_diff_x.append(cur_diff_x)
batch_change_diff_y.append(cur_diff_y)
change_psnr.append(np.array(batch_change_mse).mean())
change_mse_x.append(np.array(batch_change_mse_x).mean())
change_mse_y.append(np.array(batch_change_mse_y).mean())
change_psnr_x.append(np.array(batch_change_psnr_x).mean())
change_psnr_y.append(np.array(batch_change_psnr_y).mean())
change_diff_x.append(np.array(batch_change_diff_x).mean())
change_diff_y.append(np.array(batch_change_diff_y).mean())
if iteration % 10 == 0:
print('> Evaluation {} completed'.format(iteration))
mse = np.array(mse)
cor = np.array(cor)
psnr = np.array(psnr)
ssim = np.array(ssim)
diff_avrg = np.array(diff_avrg)
diff_max = np.array(diff_max)
change_mse_x = np.array(change_mse_x)
change_mse_y = np.array(change_mse_y)
change_psnr_x = np.array(change_psnr_x)
change_psnr_y = np.array(change_psnr_y)
change_diff_x = np.array(change_diff_x)
change_diff_y = np.array(change_diff_y)
with open(
os.path.join(self.root_dir, self.output_name,
'metrics_avrg.txt'), 'w') as avrg_hand:
avrg_hand.write('{} {}\n'.format('Avrg mse: ', np.average(mse)))
avrg_hand.write('{} {}\n'.format('Avrg cor: ', np.average(cor)))
avrg_hand.write('{} {}\n'.format('Avrg psnr: ', np.average(psnr)))
avrg_hand.write('{} {}\n'.format('Avrg ssim: ', np.average(ssim)))
avrg_hand.write('{} {}\n'.format('Avrg avrg_diff_perc: ',
np.average(diff_avrg)))
avrg_hand.write('{} {}\n'.format('Avrg max_diff_perc: ',
np.average(diff_max)))
avrg_hand.write('{} {}\n'.format('Avrg avrt_diff_x: ',
np.average(diff_x)))
avrg_hand.write('{} {}\n'.format('Avrg avrt_diff_y: ',
np.average(diff_y)))
avrg_hand.write('{} {}\n'.format('Var mse: ', np.var(mse)))
avrg_hand.write('{} {}\n'.format('Var cor: ', np.var(cor)))
avrg_hand.write('{} {}\n'.format('Var psnr: ', np.var(psnr)))
avrg_hand.write('{} {}\n'.format('Var ssim: ', np.var(ssim)))
avrg_hand.write('{} {}\n'.format('Var avrg_diff_perc: ',
np.var(diff_avrg)))
avrg_hand.write('{} {}\n'.format('Var max_diff_perc: ',
np.var(diff_max)))
avrg_hand.write('{} {}\n'.format('Var avrt_diff_x: ',
np.var(diff_x)))
avrg_hand.write('{} {}\n'.format('Var avrt_diff_y: ',
np.var(diff_y)))
avrg_hand.write('{} {}\n'.format('avrg_Change_mse_x: ',
np.mean(change_psnr)))
avrg_hand.write('{} {}\n'.format('avrg_Change_mse_x: ',
np.mean(change_mse_x)))
avrg_hand.write('{} {}\n'.format('avrg_Change_mse_y: ',
np.mean(change_mse_y)))
avrg_hand.write('{} {}\n'.format('avrg_Change_psnr_x: ',
np.mean(change_psnr_x)))
avrg_hand.write('{} {}\n'.format('avrg_Change_psnr_y: ',
np.mean(change_psnr_y)))
avrg_hand.write('{} {}\n'.format('avrg_Change_diff_x: ',
np.mean(change_diff_x)))
avrg_hand.write('{} {}\n'.format('avrg_Change_diff_y: ',
np.mean(change_diff_y)))
with open(
os.path.join(self.root_dir, self.output_name,
'metrics_list.txt'), 'w') as list_hand:
list_hand.write('{} {}\n'.format('mse: ',
','.join(str(i) for i in mse)))
list_hand.write('{} {}\n'.format('cor: ',
','.join(str(i) for i in cor)))
list_hand.write('{} {}\n'.format('psnr: ',
','.join(str(i) for i in psnr)))
list_hand.write('{} {}\n'.format('ssim: ',
','.join(str(i) for i in ssim)))
list_hand.write('{} {}\n'.format(
'diff_avrg: ', ','.join(str(i) for i in diff_avrg)))
list_hand.write('{} {}\n'.format(
'diff_max: ', ','.join(str(i) for i in diff_max)))
import json
data = json.loads(open("titles.json").read())
counts = [x["count"] for x in data]
lengths = [len(x["title"]) for x in data]
words = [x["title"].count(" ") for x in data]
word_lengths = [len(x["title"])/x["title"].count(" ") for x in data]
colons = [1 if ":" in x["title"] else 0 for x in data]
with_colon = 0
with_num = 0
wo_colon = 0
wo_num = 0
for entry in data:
if ":" in entry["title"]:
with_colon += entry["count"]
with_num += 1
else:
wo_colon += entry["count"]
wo_num += 1
print "Correlation between length and count:", correlation(lengths, counts)
print "Correlation between num words and count:", correlation(words, counts)
print "Correlation between avg word length and count:", correlation(word_lengths, counts)
print "Correlation between colon and count:", correlation(colons, counts)
print "Avg count with colon:", with_colon / with_num
print "Avg count without colon:", wo_colon / wo_num
def matrix_entry(i, j):
return correlation(data[:,i], data[:,j])
pert_idose = ft['pert_idose']
else:
pert_idose = None
predict = model(drug, data.gene, mask, pert_type, cell_id,
pert_idose)
loss = model.loss(lb, predict)
epoch_loss += loss.item()
lb_np = np.concatenate((lb_np, lb.cpu().numpy()), axis=0)
predict_np = np.concatenate((predict_np, predict.cpu().numpy()),
axis=0)
print('Dev loss:')
print(epoch_loss / (i + 1))
rmse_score = rmse(lb_np, predict_np)
rmse_list_dev.append(rmse_score)
print('RMSE: %.4f' % rmse_score)
pearson, _ = correlation(lb_np, predict_np, 'pearson')
pearson_list_dev.append(pearson)
print('Pearson\'s correlation: %.4f' % pearson)
spearman, _ = correlation(lb_np, predict_np, 'spearman')
spearman_list_dev.append(spearman)
print('Spearman\'s correlation: %.4f' % spearman)
precision = []
for k in precision_degree:
precision_neg, precision_pos = precision_k(lb_np, predict_np, k)
print("Precision@%d Positive: %.4f" % (k, precision_pos))
print("Precision@%d Negative: %.4f" % (k, precision_neg))
precision.append([precision_pos, precision_neg])
precisionk_list_dev.append(precision)
if best_dev_pearson < pearson:
best_dev_pearson = pearson
