trainer = GraddescentMinibatch(
varin=model.varin, data=train_x,
truth=model.models_stack[-1].vartruth, truth_data=train_y,
supervised=True,
cost=model.models_stack[-1].cost(),
params=model.params,
batchsize=batchsize, learningrate=finetune_lr, momentum=momentum,
rng=npy_rng
)
init_lr = trainer.learningrate
prev_cost = numpy.inf
for epoch in xrange(finetune_epc):
cost = trainer.epoch()
if prev_cost <= cost:
if trainer.learningrate < (init_lr * 1e-7):
break
trainer.set_learningrate(trainer.learningrate*0.8)
prev_cost = cost
if epoch % 10 == 0:
print "*** error rate: train: %f, test: %f" % (train_error(), test_error())
try:
if epoch % 100 == 0:
save_params(model, 'svhn_CONV_flt2*5_pool2*5_nflt128*5_64*3-10.npy')
except:
pass
print "***FINAL error rate: train: %f, test: %f" % (train_error(), test_error())
print "Done."
pdb.set_trace()
#apply_mask[2]()
epc_cost += cost
if step % (50000 / batchsize) == 0 and step > 0:
# set stop rule
ind = (step / (50000 / batchsize)) % avg
hist_avg[ind] = crnt_avg[ind]
crnt_avg[ind] = epc_cost
if sum(hist_avg) < sum(crnt_avg):
break
# adjust learning rate
if prev_cost <= epc_cost:
patience += 1
if patience > 10:
trainer.set_learningrate(0.9 * trainer.learningrate)
patience = 0
prev_cost = epc_cost
# evaluate
print "***error rate: train: %f, test: %f" % (
train_error(), test_error())
epc_cost = 0.
print "Done."
print "***FINAL error rate, train: %f, test: %f" % (
train_error(), test_error()
)
save_params(model, __file__.split('.')[0] + '_params.npy')
pdb.set_trace()
params=layer_dropout.params_private,
supervised=False,
batchsize=batchsize,
learningrate=pretrain_lr,
momentum=momentum,
rng=npy_rng)
prev_cost = numpy.inf
patience = 0
for epoch in xrange(pretrain_epc):
cost = trainer.epoch()
if prev_cost <= cost:
patience += 1
if patience > 10:
patience = 0
trainer.set_learningrate(0.9 * trainer.learningrate)
if trainer.learningrate < 1e-10:
break
prev_cost = cost
save_params(
model,
'ZLIN_4000_1000_4000_1000_4000_1000_4000_normhid_nolinb_cae1_dropout.npy'
)
print "Done."
#########################
# BUILD FINE-TUNE MODEL #
#########################
print "\n\n... building fine-tune model -- contraction 1"
for imodel in model.models_stack:
