def train_network(self, data_dir, epochs, batch_size, weight_decay, lr,
output_dir):
self.cuda()
self.train(True)
data = zarr.open(data_dir, 'r')
train_data = data['train_data']
labels = data['labels']
n_batches = round(train_data.shape[0] / batch_size)
opt = optim.Adam(self.parameters(), weight_decay=weight_decay, lr=lr)
loss_f = nn.CrossEntropyLoss()
scheduler = optim.lr_scheduler.StepLR(opt, step_size=5, gamma=0.5)
losses = []
for epoch in range(epochs):
start = time.time()
scheduler.step()
avg_loss = numpy.zeros((1, ))
n_samples = 1
for idx in range(n_batches):
opt.zero_grad()
train_batch = train_data[idx * batch_size:idx * batch_size +
batch_size]
label_batch = labels[idx * batch_size:idx * batch_size +
batch_size]
train_batch = Variable(torch.LongTensor(train_batch)).cuda()
label_batch = Variable(torch.LongTensor(label_batch),
requires_grad=False).cuda()
output = self(train_batch)
loss = loss_f(output, label_batch)
loss.backward()
opt.step()
n_samples += 1
avg_loss += numpy.round(loss.cpu().data.numpy(), 3)
end = time.time()
avg_loss /= n_samples
losses.append(avg_loss)
logger.info(
'Epoch {0}, Average Loss {1}, Minutes spent {2}'.format(
epoch + 1, round(avg_loss.data[0], 4),
round((end - start) / 60., 2)))
Plotter.plot_training(epochs=epochs,
losses=losses,
embedding_dim=self.n_hidden,
output=output_dir)
def plot_behavior_payoff(self, results, imageFilePath):
"""Plots a graph that depicts the behavior and the
payoff of the players/agents from 'results' to one
or more image files, taking 'imageFilePath' as base name.
"""
plotter = Plotter(imageFilePath, self.image_file_types, "right")
plotter.beginDoublePlot("Period",
"Percentage in total subject population",
"Payoffs in MUs",
(0, self.chronicles.world.maxRounds),
(0,100), (30,60))
data = PC([results[FREE_RIDERS], results[HIGH_CONTRIBUTORS]])
plotter.barPlot([FREE_RIDERS, HIGH_CONTRIBUTORS], data, yaxis = "left")
plotter.linePlot([PAYOFF_HC], [results[PAYOFF_HC]], yaxis = "right")
plotter.linePlot([PAYOFF_FR], [results[PAYOFF_FR]], yaxis = "right")
plotter.endPlot()
def plot_agent_payoff(self, stats, imageFilePath, deviation = None):
"""Plots the agent statistics for either a single agent or for a group
of agents."""
plotter = Plotter(imageFilePath, self.image_file_types, "right")
plotter.beginDoublePlot("Period",
"Percentage total agent class",
"Payoffs in MUs",
(0, self.chronicles.world.maxRounds),
(0,100), None)
data = PC([stats[AG_SI], stats[AG_SFI]])
plotter.barPlot([AG_SI, AG_SFI], data, yaxis = "left")
if deviation:
error = deviation[AG_PAYOFF]
else:
error = None
plotter.linePlot([AG_PAYOFF], stats[AG_PAYOFF], error , "right")
plotter.endPlot()
def plot_agent_contrib(self, stats, imageFilePath, deviation = None):
"""Plots the contributions, sanctions and received punishements
or commendations of an agent."""
plotter = Plotter(imageFilePath, self.image_file_types, "right")
plotter.beginDoublePlot("Period",
"Coercions received",
"Contribs/Sanctions in percent of endowment",
(0, self.chronicles.world.maxRounds),
None, (0,100))
if deviation:
error = [deviation[AG_PUNISH], deviation[AG_COMMEND]]
else:
error = None
plotter.barPlot([AG_PUNISH, AG_COMMEND],
[stats[AG_PUNISH], stats[AG_COMMEND]],
errorBars = None, yaxis = "left") # graph gets too confusiung if errorBars are plotted here
if deviation:
error = PC([deviation[AG_CONTRIB], deviation[AG_SANCT]])
else:
error = None
plotter.linePlot([AG_CONTRIB, AG_SANCT],
PC([stats[AG_CONTRIB], stats[AG_SANCT]]),
errorBars = error, yaxis = "right")
plotter.endPlot()
def plot_payoff_punishment(self, results, imageFilePath):
"""Plots a graph that depicts the payoff and number of punishers
and non-punishers from the 'results' to one or more image files
at 'imageFilePath'.
"""
plotter = Plotter(imageFilePath, self.image_file_types, "right",
condense_factor = 5)
plotter.beginDoublePlot("Periods",
"Percentage of high contributers in SI",
"Payoffs in MUs",
(0, self.chronicles.world.maxRounds),
(0,100), (30,60))
data = PC([results[PUNISH_HC], results[NO_PUNISH_HC]])
plotter.barPlot([PUNISH_HC, NO_PUNISH_HC], data, yaxis = "left")
plotter.linePlot([PAYOFF_NOP_HC], [results[PAYOFF_NOP_HC]], yaxis = "right")
plotter.linePlot([PAYOFF_P_HC], [results[PAYOFF_P_HC]], yaxis = "right")
plotter.endPlot()
def plot_choice_contrib(self, results, imageFilePath):
"""Plots a graph that depcits the institution choice and the
contributions of the players/agents from 'results' to one
or more image files, taking 'imageFilePath' as base name.
"""
plotter = Plotter(imageFilePath, self.image_file_types, "right")
plotter.beginDoublePlot("Period",
"Percentage in total subject population",
"Contribution in percent of endowment",
(0, self.chronicles.world.maxRounds),
(0,100), (0,100))
data = PC([results[SI_MEMBERS], results[SFI_MEMBERS]])
plotter.barPlot([SI_MEMBERS, SFI_MEMBERS], data, yaxis = "left")
plotter.linePlot([AV_CONTRIB_SI], PC([results[AV_CONTRIB_SI]]), yaxis = "right")
plotter.linePlot([AV_CONTRIB_SFI], PC([results[AV_CONTRIB_SFI]]), yaxis = "right")
plotter.endPlot()
test_is_signal = test_dataframe.loc[:, "event_id"] == 0
new_masses = np.random.choice(masses, size=np.sum(~test_is_signal), replace=True)
# new_masses = np.array(test_dataframe.loc[test_is_signal, "true_mass"].sample(n=np.sum(~test_is_signal), replace=True))
test_dataframe.loc[~test_is_signal, "true_mass"] = new_masses
training_variables = ["ldgTrkPtFrac", "deltaPhiBjetMet", "deltaPhiTauMet", "deltaPhiTauBjet",
"MET", "tauPt", "bjetPt", "TransverseMass", "true_mass"]
preprocess_modes = ["MinMaxScale", "MinMaxScale", "MinMaxScale", "MinMaxScale",
"MinMaxScale", "MinMaxScale", "MinMaxScale", "MinMaxScale", "MinMaxScale"]
if only_make_comparisons:
dnns = ["model_for_even", "model_for_odds"]
labels = ["even", "odd"]
paths_to_dnn = ["./SavedModels/" + x for x in dnns]
plotter = Plotter(test_dataframe, None)
plotter.compareDnns(paths_to_dnn, training_variables, labels)
sys.exit(0)
# training_variables = ["ldgTrkPtFrac", "deltaPhiBjetMet", "deltaPhiTauMet", "deltaPhiTauBjet",
# "MET", "tauPt", "bjetPt", "TransverseMass"]
# preprocess_modes = ["MinMaxScale", "MinMaxScale", "MinMaxScale", "MinMaxScale",
# "MinMaxScale", "MinMaxScale", "MinMaxScale", "MinMaxScale"]
preprocessor = Preprocessor(training_variables, preprocess_modes)
dataframe = preprocessor._clean_inputs(dataframe, ["ldgTrkPtFrac", "TransverseMass"], [0.0, 0.0], [1.0, None])
test_dataframe = preprocessor._clean_inputs(test_dataframe, ["ldgTrkPtFrac", "TransverseMass"], [0.0, 0.0], [1.0, None])
is_signal = (dataframe.loc[:, "event_id"] == 0)
signal = dataframe.loc[is_signal, :]
