on_unused_input='warn')
funcs['acc_loss'] = theano.function(
[inp.input_var, patch_op.input_var, target], [acc, cost],
on_unused_input='warn')
funcs['predict'] = theano.function([inp.input_var, patch_op.input_var], [pred],
on_unused_input='warn')
''' Training (a bit simplified) '''
n_epochs = 50
eval_freq = 1
start_time = time.time()
best_trn = 1e5
best_tst = 1e5
kvs = snapshotter.Snapshotter('/home/mabbasloo/ShapeNet/training.snap')
for it_count in range(n_epochs):
tic = time.time()
b_l, b_c, b_s, b_r, b_g, b_a = [], [], [], [], [], []
for x_ in ds.train_iter():
tmp = funcs['train'](*x_)
# do some book keeping (store stuff for training curves etc)
b_l.append(tmp[0])
b_c.append(tmp[1])
b_r.append(tmp[2])
b_g.append(tmp[3])
b_a.append(tmp[4])
epoch_cost = np.asarray(
[np.mean(b_l),
funcs['train'] = theano.function([inp.input_var, patch_op.input_var, target],
[cost, cla, l2_weight * regL2, grads_norm, acc], updates=updates,
on_unused_input='warn')
funcs['acc_loss'] = theano.function([inp.input_var, patch_op.input_var, target],
[acc, cost], on_unused_input='warn')
funcs['predict'] = theano.function([inp.input_var, patch_op.input_var],
[pred], on_unused_input='warn')
n_epochs = 50
eval_freq = 1
start_time = time.time()
best_trn = 1e5
best_tst = 1e5
kvs = snapshotter.Snapshotter('demo_training.snap')
n = 0;
for it_count in xrange(n_epochs):
tic = time.time()
b_l, b_c, b_s, b_r, b_g, b_a = [], [], [], [], [], []
for x_ in ds.train_iter():
tmp = funcs['train'](*x_)
b_l.append(tmp[0])
b_c.append(tmp[1])
b_r.append(tmp[2])
b_g.append(tmp[3])
b_a.append(tmp[4])
epoch_cost = np.asarray([np.mean(b_l), np.mean(b_c), np.mean(b_r), np.mean(b_g), np.mean(b_a)])
print(('[Epoch %03i][trn] cost %9.6f (cla %6.4f, reg %6.4f), |grad| = %.06f, acc = %7.5f %% (%.2fsec)') %
(it_count, epoch_cost[0], epoch_cost[1], epoch_cost[2], epoch_cost[3], epoch_cost[4] * 100,