def train_gluon_ch7(trainer_name,
trainer_hyperparams,
features,
labels,
batch_size=10,
num_epochs=2):
"""Train a linear regression model with a given Gluon trainer."""
net = nn.Sequential()
net.add(nn.Dense(1))
net.initialize(init.Normal(sigma=0.01))
loss = gloss.L2Loss()
def eval_loss():
return loss(net(features), labels).mean().asscalar()
ls = [eval_loss()]
data_iter = gdata.DataLoader(gdata.ArrayDataset(features, labels),
batch_size,
shuffle=True)
trainer = gluon.Trainer(net.collect_params(), trainer_name,
trainer_hyperparams)
for _ in range(num_epochs):
start = time.time()
for batch_i, (X, y) in enumerate(data_iter):
with autograd.record():
l = loss(net(X), y)
l.backward()
trainer.step(batch_size)
if (batch_i + 1) * batch_size % 100 == 0:
ls.append(eval_loss())
print('loss: %f, %f sec per epoch' % (ls[-1], time.time() - start))
set_figsize()
plt.plot(np.linspace(0, num_epochs, len(ls)), ls)
plt.xlabel('epoch')
plt.ylabel('loss')
def train_ch3(net,
train_iter,
test_iter,
loss,
num_epochs,
batch_size,
params=None,
lr=None,
trainer=None):
"""Train and evaluate a model with CPU."""
for epoch in range(num_epochs):
train_l_sum, train_acc_sum, n = 0.0, 0.0, 0
for X, y in train_iter:
with autograd.record():
y_hat = net(X)
l = loss(y_hat, y).sum()
l.backward()
if trainer is None:
sgd(params, lr, batch_size)
else:
trainer.step(batch_size)
y = y.astype('float32')
train_l_sum += l.asscalar()
train_acc_sum += (y_hat.argmax(axis=1) == y).sum().asscalar()
n += y.size
test_acc = evaluate_accuracy(test_iter, net)
print('epoch %d, loss %.4f, train acc %.3f, test acc %.3f' %
(epoch + 1, train_l_sum / n, train_acc_sum / n, test_acc))
def train(train_iter, test_iter, net, loss, trainer, ctx, num_epochs):
"""Train and evaluate a model."""
print('training on', ctx)
if isinstance(ctx, mx.Context):
ctx = [ctx]
for epoch in range(num_epochs):
train_l_sum, train_acc_sum, n, m, start = 0.0, 0.0, 0, 0, time.time()
for i, batch in enumerate(train_iter):
Xs, ys, batch_size = _get_batch(batch, ctx)
ls = []
with autograd.record():
y_hats = [net(X) for X in Xs]
ls = [loss(y_hat, y) for y_hat, y in zip(y_hats, ys)]
for l in ls:
l.backward()
trainer.step(batch_size)
train_l_sum += sum([l.sum().asscalar() for l in ls])
n += sum([l.size for l in ls])
train_acc_sum += sum([(y_hat.argmax(axis=1) == y).sum().asscalar()
for y_hat, y in zip(y_hats, ys)])
m += sum([y.size for y in ys])
test_acc = evaluate_accuracy(test_iter, net, ctx)
print('epoch %d, loss %.4f, train acc %.3f, test acc %.3f, '
'time %.1f sec' % (epoch + 1, train_l_sum / n, train_acc_sum / m,
test_acc, time.time() - start))
def train(data_dir, pretrain_model, epoches=3, lr=0.001, wd=5e-4, momentum=0.9, batch_size=5, ctx=mx.cpu(), verbose_step=2, ckpt='ckpt'):
icdar_loader = ICDAR(data_dir=data_dir)
loader = DataLoader(icdar_loader, batch_size=batch_size, shuffle=True)
net = PSENet(num_kernels=7, ctx=ctx)
# initial params
net.collect_params().initialize(mx.init.Normal(sigma=0.01), ctx=ctx)
# net.initialize(ctx=ctx)
# net.load_parameters(pretrain_model, ctx=ctx, allow_missing=True, ignore_extra=True)
pse_loss = DiceLoss(lam=0.7)
cos_shc = ls.PolyScheduler(max_update=icdar_loader.length * epoches//batch_size, base_lr=lr)
trainer = Trainer(
net.collect_params(),
'sgd',
{
'learning_rate': lr,
'wd': wd,
'momentum': momentum,
'lr_scheduler':cos_shc
})
summary_writer = SummaryWriter(ckpt)
for e in range(epoches):
cumulative_loss = 0
for i, item in enumerate(loader):
im, score_maps, kernels, training_masks, ori_img = item
im = im.as_in_context(ctx)
score_maps = score_maps[:, ::4, ::4].as_in_context(ctx)
kernels = kernels[:, ::4, ::4, :].as_in_context(ctx)
training_masks = training_masks[:, ::4, ::4].as_in_context(ctx)
with autograd.record():
kernels_pred = net(im)
loss = pse_loss(score_maps, kernels, kernels_pred, training_masks)
loss.backward()
trainer.step(batch_size)
if i%verbose_step==0:
global_steps = icdar_loader.length * e + i * batch_size
summary_writer.add_image('score_map', score_maps[0:1, :, :], global_steps)
summary_writer.add_image('score_map_pred', kernels_pred[0:1, -1, :, :], global_steps)
summary_writer.add_image('kernel_map', kernels[0:1, :, :, 0], global_steps)
summary_writer.add_image('kernel_map_pred', kernels_pred[0:1, 0, :, :], global_steps)
summary_writer.add_scalar('loss', mx.nd.mean(loss).asscalar(), global_steps)
summary_writer.add_scalar('c_loss', mx.nd.mean(pse_loss.C_loss).asscalar(), global_steps)
summary_writer.add_scalar('kernel_loss', mx.nd.mean(pse_loss.kernel_loss).asscalar(), global_steps)
summary_writer.add_scalar('pixel_accuracy', pse_loss.pixel_acc, global_steps)
print("step: {}, loss: {}, score_loss: {}, kernel_loss: {}, pixel_acc: {}".format(i * batch_size, mx.nd.mean(loss).asscalar(), \
mx.nd.mean(pse_loss.C_loss).asscalar(), mx.nd.mean(pse_loss.kernel_loss).asscalar(), \
pse_loss.pixel_acc))
cumulative_loss += mx.nd.mean(loss).asscalar()
print("Epoch {}, loss: {}".format(e, cumulative_loss))
net.save_parameters(os.path.join(ckpt, 'model_{}.param'.format(e)))
summary_writer.close()
def train_and_predict_rnn_gluon(model, num_hiddens, vocab_size, ctx,
corpus_indices, idx_to_char, char_to_idx,
num_epochs, num_steps, lr, clipping_theta,
batch_size, pred_period, pred_len, prefixes):
"""Train an Gluon RNN model and predict the next item in the sequence."""
loss = gloss.SoftmaxCrossEntropyLoss()
model.initialize(ctx=ctx, force_reinit=True, init=init.Normal(0.01))
trainer = gluon.Trainer(model.collect_params(), 'sgd', {
'learning_rate': lr,
'momentum': 0,
'wd': 0
})
for epoch in range(num_epochs):
l_sum, n, start = 0.0, 0, time.time()
data_iter = data_iter_consecutive(corpus_indices, batch_size,
num_steps, ctx)
state = model.begin_state(batch_size=batch_size, ctx=ctx)
for X, Y in data_iter:
for s in state:
s.detach()
with autograd.record():
(output, state) = model(X, state)
y = Y.T.reshape((-1, ))
l = loss(output, y).mean()
l.backward()
params = [p.data() for p in model.collect_params().values()]
grad_clipping(params, clipping_theta, ctx)
trainer.step(1)
l_sum += l.asscalar() * y.size
n += y.size
if (epoch + 1) % pred_period == 0:
print('epoch %d, perplexity %f, time %.2f sec' %
(epoch + 1, math.exp(l_sum / n), time.time() - start))
for prefix in prefixes:
print(
' -',
predict_rnn_gluon(prefix, pred_len, model, vocab_size, ctx,
idx_to_char, char_to_idx))
def train_ch5(net, train_iter, test_iter, batch_size, trainer, ctx,
num_epochs):
"""Train and evaluate a model with CPU or GPU."""
print('training on', ctx)
loss = gloss.SoftmaxCrossEntropyLoss()
for epoch in range(num_epochs):
train_l_sum, train_acc_sum, n, start = 0.0, 0.0, 0, time.time()
for X, y in train_iter:
X, y = X.as_in_context(ctx), y.as_in_context(ctx)
with autograd.record():
y_hat = net(X)
l = loss(y_hat, y).sum()
l.backward()
trainer.step(batch_size)
y = y.astype('float32')
train_l_sum += l.asscalar()
train_acc_sum += (y_hat.argmax(axis=1) == y).sum().asscalar()
n += y.size
test_acc = evaluate_accuracy(test_iter, net, ctx)
print('epoch %d, loss %.4f, train acc %.3f, test acc %.3f, '
'time %.1f sec' % (epoch + 1, train_l_sum / n, train_acc_sum / n,
test_acc, time.time() - start))
def test(ctx=mx.cpu()):
net = MySSD(1, num_anchors)
net.initialize(init="Xavier", ctx=ctx)
net.hybridize()
# print(net)
# x = nd.random.normal(0,1,(100,3,256,256), ctx=ctx)
# net(x)
batch_size, edge_size = args.batch_size, args.input_size
train_iter, _ = predata.load_data_uav(args.data_path, batch_size,
edge_size)
batch = train_iter.next()
batch.data[0].shape, batch.label[0].shape
if batch_size >= 25: # show f*****g pikachuus in grid
imgs = (batch.data[0][0:25].transpose((0, 2, 3, 1))) / 255
axes = utils.show_images(imgs, 5, 5).flatten()
for ax, label in zip(axes, batch.label[0][0:25]):
utils.show_bboxes(ax, [label[0][1:5] * edge_size], colors=['w'])
plt.show()
# net.initialize(init=init.Xavier(), ctx=ctx)
trainer = mx.gluon.Trainer(net.collect_params(), 'sgd', {
'learning_rate': 0.2,
'wd': 5e-4
})
cls_loss = gloss.SoftmaxCrossEntropyLoss()
bbox_loss = gloss.L1Loss()
def calc_loss(cls_preds, cls_labels, bbox_preds, bbox_labels, bbox_masks):
cls = cls_loss(cls_preds, cls_labels)
bbox = bbox_loss(bbox_preds * bbox_masks, bbox_labels * bbox_masks)
return cls + bbox
def cls_eval(cls_preds, cls_labels):
# the result from class prediction is at the last dim
# argmax() should be assigned with the last dim of cls_preds
return (cls_preds.argmax(axis=-1) == cls_labels).sum().asscalar()
def bbox_eval(bbox_preds, bbox_labels, bbox_masks):
return ((bbox_labels - bbox_preds) * bbox_masks).abs().sum().asscalar()
IF_LOAD_MODEL = args.load
if IF_LOAD_MODEL:
net.load_parameters(args.model_path)
else:
for epoch in range(args.num_epoches):
acc_sum, mae_sum, n, m = 0.0, 0.0, 0, 0
train_iter.reset(
) # reset data iterator to read-in images from beginning
start = time.time()
for batch in train_iter:
X = batch.data[0].as_in_context(ctx)
Y = batch.label[0].as_in_context(ctx)
with autograd.record():
# generate anchors and generate bboxes
im, anchors, cls_preds, bbox_preds = net(X)
# assign classes and bboxes for each anchor
bbox_labels, bbox_masks, cls_labels = nd.contrib.MultiBoxTarget(
anchors, Y, cls_preds.transpose((0, 2, 1)))
# calc loss
l = calc_loss(cls_preds, cls_labels, bbox_preds,
bbox_labels, bbox_masks)
l.backward()
trainer.step(batch_size)
acc_sum += cls_eval(cls_preds, cls_labels)
n += cls_labels.size
mae_sum += bbox_eval(bbox_preds, bbox_labels, bbox_masks)
m += bbox_labels.size
if (epoch + 1) % 1 == 0:
print(
'epoch %2d, class err %.2e, bbox mae %.2e, time %.1f sec' %
(epoch + 1, 1 - acc_sum / n, mae_sum / m,
time.time() - start))
net.save_parameters("myssd.params")
def predict(X):
im, anchors, cls_preds, bbox_preds = net(X.as_in_context(ctx))
# im = im.transpose((2, 3, 1, 0)).asnumpy()
# imgs = [im[1:-2,1:-2, k, 0] for k in range(256)] # why are there boundary effect?
# utils.show_images_np(imgs, 16, 16)
# plt.show()
# plt.savefig("./activdation/figbase%s"%nd.random.randint(0,100,1).asscalar())
# plt.imshow(nd.sum(nd.array(im[1:-2, 1:-2, :, :]), axis=2).asnumpy()[:, :, 0], cmap='gray')
# plt.savefig("./suming_act")
cls_probs = cls_preds.softmax().transpose((0, 2, 1))
output = contrib.nd.MultiBoxDetection(cls_probs, bbox_preds, anchors)
idx = [i for i, row in enumerate(output[0]) if row[0].asscalar() != -1]
if idx == []:
raise ValueError("NO TARGET. Seq Terminated.")
return output[0, idx]
def display(img, output, threshold):
lscore = []
for row in output:
lscore.append(row[1].asscalar())
for row in output:
score = row[1].asscalar()
if score < min(max(lscore), threshold):
continue
h, w = img.shape[0:2]
bbox = [row[2:6] * nd.array((w, h, w, h), ctx=row.context)]
cv.rectangle(img, (bbox[0][0].asscalar(), bbox[0][1].asscalar()),
(bbox[0][2].asscalar(), bbox[0][3].asscalar()),
(0, 255, 0), 3)
cv.imshow("res", img)
cv.waitKey(60)
cap = cv.VideoCapture(args.test_path)
rd = 0
while True:
ret, frame = cap.read()
img = nd.array(frame)
feature = image.imresize(img, 512, 512).astype('float32')
X = feature.transpose((2, 0, 1)).expand_dims(axis=0)
countt = time.time()
output = predict(X)
if rd == 0: net.export('ssd')
countt = time.time() - countt
print("SPF: %3.2f" % countt)
utils.set_figsize((5, 5))
display(frame / 255, output, threshold=0.8)
plt.show()
rd += 1
