testing_candidates = []
testing_jobs = []
testing_job_ranks = []
for i, year in enumerate(year_range):
if year in hold_out:
testing_candidates.append(candidate_pools[i])
testing_jobs.append(job_pools[i])
testing_job_ranks.append(job_ranks[i])
else:
training_candidates.append(candidate_pools[i])
training_jobs.append(job_pools[i])
training_job_ranks.append(job_ranks[i])
# Find a decent starting place (using the training set)
simulator = SimulationEngine(training_candidates, training_jobs, training_job_ranks, inst, model, power=args.power, reg=args.reg, iters=20)
w0 = None
best_error = np.inf
for i in xrange(args.num_steps):
wtemp = np.random.randn(model.num_weights())
error = simulator.simulate(weights=wtemp)
if error < best_error:
w0 = wtemp.copy()
best_error = error
# Optimize from there (for the training set)
simulator = SimulationEngine(training_candidates, training_jobs, training_job_ranks, inst, model, power=args.power, reg=args.reg, iters=args.num_iters)
opt = {'maxiter':args.num_steps}
res = minimize(simulator.simulate, w0, method='Nelder-Mead', options=opt)
final_weights = res.x
print 'FINAL_WEIGHTS:', final_weights
school_info=inst,
ranking='pi_rescaled',
year_start=1970,
year_stop=2012,
year_step=1)
if args.validation: # if specified years are to be evaluated
hold_out = [int(year) for year in args.validation.split(',')]
testing_candidates, testing_jobs, testing_job_ranks = [], [], []
for i, year in enumerate(year_range):
if year in hold_out:
testing_candidates.append(candidate_pools[i])
testing_jobs.append(job_pools[i])
testing_job_ranks.append(job_ranks[i])
# Overwrite originals
candidate_pools, job_pools, job_ranks = testing_candidates, testing_jobs, testing_job_ranks
model = SigmoidModel(prob_function=args.prob_function)
simulator = SimulationEngine(candidate_pools, job_pools, job_ranks, inst, model, iters=1, reg=0)
if model.num_weights() > 0:
w = np.array([float(x) for x in args.weights.split(',')])
if len(w) != model.num_weights():
print len(w), model.num_weights()
raise ValueError('Invalid number of weights/model parameters!')
else:
w = None
for i in xrange(args.num_iters):
simulator.simulate(weights=w, ranking=args.ranking)
for i, year in enumerate(year_range):
if year in hold_out:
testing_candidates.append(candidate_pools[i])
testing_jobs.append(job_pools[i])
testing_job_ranks.append(job_ranks[i])
else:
training_candidates.append(candidate_pools[i])
training_jobs.append(job_pools[i])
training_job_ranks.append(job_ranks[i])
# Find a decent starting place (using the training set)
simulator = SimulationEngine(training_candidates,
training_jobs,
training_job_ranks,
inst,
model,
power=args.power,
reg=args.reg,
iters=20)
w0 = None
best_error = np.inf
for i in xrange(args.num_steps):
wtemp = np.random.randn(model.num_weights())
error = simulator.simulate(weights=wtemp)
if error < best_error:
w0 = wtemp.copy()
best_error = error
# Optimize from there (for the training set)
simulator = SimulationEngine(training_candidates,
training_jobs,
for i, year in enumerate(year_range):
if year not in hold_out:
training_candidates.append(candidate_pools[i])
training_jobs.append(job_pools[i])
training_job_ranks.append(job_ranks[i])
training_orders.append(hiring_orders[i])
training_probs.append(hiring_probs[i])
# Overwrite originals:
candidate_pools, job_pools, job_ranks = training_candidates, training_jobs, training_job_ranks
hiring_orders, hiring_probs = training_orders, training_probs
# Which model?
model = SigmoidModel(prob_function=args.prob_function)
# Find a starting place
simulator = SimulationEngine(candidate_pools, job_pools, job_ranks, inst, model, power=1, reg=args.reg,
hiring_orders=hiring_orders, hiring_probs=hiring_probs)
w0 = None
best_neg_likelihood = np.inf
for i in xrange(10): #args.num_steps):
#wtemp = 100*np.random.randn(model.num_weights()) # ~[-100, 100]
wtemp = 100*(np.random.random(model.num_weights()) - 0.5) # ~[-100, 100]
temp = simulator.calculate_neg_log_likelihood(weights=wtemp)
if temp < best_neg_likelihood:
w0 = wtemp.copy()
best_neg_likelihood = temp
# Optimize from there
simulator = SimulationEngine(candidate_pools, job_pools, job_ranks, inst, model, power=1, reg=args.reg,
hiring_orders=hiring_orders, hiring_probs=hiring_probs)
opt = {'maxiter':args.num_steps}
training_candidates, training_jobs, training_job_ranks = [], [], []
for i, year in enumerate(year_range):
if year not in hold_out:
training_candidates.append(candidate_pools[i])
training_jobs.append(job_pools[i])
training_job_ranks.append(job_ranks[i])
# Overwrite originals
candidate_pools, job_pools, job_ranks = training_candidates, training_jobs, training_job_ranks
model = SigmoidModel(prob_function=args.prob_function)
# Find a decent starting place
simulator = SimulationEngine(candidate_pools,
job_pools,
job_ranks,
inst,
model,
power=1,
reg=args.reg,
iters=20)
w0 = None
best_error = np.inf
for i in xrange(args.num_steps):
wtemp = 200 * (np.random.random(model.num_weights()) - 0.5
) # ~[-100, 100]
error = simulator.simulate(weights=wtemp)
if error < best_error:
w0 = wtemp.copy()
best_error = error
# Optimize from there
simulator = SimulationEngine(candidate_pools,
for person, rank in pool:
actual_hires.append((person, person.first_asst_job_location))
compute_network_stats(actual_hires, inst, output)
output.close()
exit()
# Otherwise, set up the model + simulator
if args.prob_function == 'flat':
# 'flat' could very easily be a prob_function in the sigmoid model -- this is faster
model = ConfigurationModel()
else:
model = SigmoidModel(prob_function=args.prob_function)
simulator = SimulationEngine(candidate_pools,
job_pools,
job_ranks,
inst,
model,
iters=1,
reg=0)
if model.num_weights() > 0:
w = np.array([float(x) for x in args.weights.split(',')])
if len(w) != model.num_weights():
print len(w), model.num_weights()
raise ValueError('Invalid number of weights/model parameters!')
else:
w = None
# Run however many iterations and compute stats
for i in xrange(args.num_iters):
hires = simulator.generate_network(weights=w, one_list=True)
compute_network_stats(hires, inst, output)
if args.validation: # if specified years are to be evaluated
hold_out = [int(year) for year in args.validation.split(',')]
testing_candidates, testing_jobs, testing_job_ranks = [], [], []
for i, year in enumerate(year_range):
if year in hold_out:
testing_candidates.append(candidate_pools[i])
testing_jobs.append(job_pools[i])
testing_job_ranks.append(job_ranks[i])
# Overwrite originals
candidate_pools, job_pools, job_ranks = testing_candidates, testing_jobs, testing_job_ranks
model = SigmoidModel(prob_function=args.prob_function)
simulator = SimulationEngine(candidate_pools,
job_pools,
job_ranks,
inst,
model,
iters=1,
reg=0)
if model.num_weights() > 0:
w = np.array([float(x) for x in args.weights.split(',')])
if len(w) != model.num_weights():
print len(w), model.num_weights()
raise ValueError('Invalid number of weights/model parameters!')
else:
w = None
for i in xrange(args.num_iters):
simulator.simulate(weights=w, ranking=args.ranking)
if args.actual:
actual_hires = []
for pool in candidate_pools:
for person, rank in pool:
actual_hires.append((person, person.first_asst_job_location))
compute_network_stats(actual_hires, inst, output)
output.close()
exit()
# Otherwise, set up the model + simulator
if args.prob_function == 'flat':
# 'flat' could very easily be a prob_function in the sigmoid model -- this is faster
model = ConfigurationModel()
else:
model = SigmoidModel(prob_function=args.prob_function)
simulator = SimulationEngine(candidate_pools, job_pools, job_ranks, inst, model, iters=1, reg=0)
if model.num_weights() > 0:
w = np.array([float(x) for x in args.weights.split(',')])
if len(w) != model.num_weights():
print len(w), model.num_weights()
raise ValueError('Invalid number of weights/model parameters!')
else:
w = None
# Run however many iterations and compute stats
for i in xrange(args.num_iters):
hires = simulator.generate_network(weights=w, one_list=True)
compute_network_stats(hires, inst, output)
output.close()
if args.validation: # if specified years are to be left out
hold_out = [int(year) for year in args.validation.split(',')]
training_candidates, training_jobs, training_job_ranks = [], [], []
for i, year in enumerate(year_range):
if year not in hold_out:
training_candidates.append(candidate_pools[i])
training_jobs.append(job_pools[i])
training_job_ranks.append(job_ranks[i])
# Overwrite originals
candidate_pools, job_pools, job_ranks = training_candidates, training_jobs, training_job_ranks
model = SigmoidModel(prob_function=args.prob_function)
# Find a decent starting place
simulator = SimulationEngine(candidate_pools, job_pools, job_ranks, inst, model, power=1, reg=args.reg, iters=20)
w0 = None
best_error = np.inf
for i in xrange(args.num_steps):
wtemp = 200*(np.random.random(model.num_weights()) - 0.5) # ~[-100, 100]
error = simulator.simulate(weights=wtemp)
if error < best_error:
w0 = wtemp.copy()
best_error = error
# Optimize from there
simulator = SimulationEngine(candidate_pools, job_pools, job_ranks, inst, model, power=1, reg=args.reg, iters=args.num_iters)
opt = {'maxiter':args.num_steps}
res = minimize(simulator.simulate, w0, method='Nelder-Mead', options=opt)
print res
