def val_reader():
traindataset = myreader.myreader_classify(val_datasetfile,
val_labelfile,
'val',
doshuffle=False)
for image, label in traindataset:
yield image, label
def debug_patchloss():
#debug loss NaN problem
from losses import SoftmaxLoss
from losses import ArcMarginLoss
traindataset = myreader.myreader_classify(train_datasetfile,
train_labelfile,
'train',
iscolor=0,
doshuffle=True,
preprocessfunc=preprocess)
embedding_size = 64
class_dim = 50000
batchsize = 128
model = l2net.L2Net()
#model = resnet18.ResNet18()
image = fluid.layers.data(name='image', shape=[1, 32, 32], dtype='uint8')
labelvar = fluid.layers.data(name='label', shape=[1], dtype='int64')
inputdata = trainmodule.preprocessimg(image)
out = model.net(input=inputdata, embedding_size=embedding_size)
metricloss = ArcMarginLoss(class_dim=class_dim,
margin=0.5,
scale=64,
easy_margin=False)
#metricloss = SoftmaxLoss(class_dim=class_dim)
cost, logit = metricloss.loss(out, labelvar)
avg_cost = fluid.layers.mean(x=cost)
imgs = []
labels = []
for img, label in traindataset:
if label < class_dim:
imgs.append(img.reshape(1, 1, 32, 32))
labels.append(label)
if len(imgs) == batchsize:
break
imgs = np.vstack(imgs)
labels = np.array(labels, np.int64).reshape((-1, 1))
optimizer = fluid.optimizer.SGD(learning_rate=0.1)
opts = optimizer.minimize(avg_cost)
place = fluid.CPUPlace()
#place = fluid.CUDAPlace(0)
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())
loss_value, output_value, label_value = exe.run(
fluid.default_main_program(),
feed={
'image': imgs,
'label': labels
},
fetch_list=[cost.name, out.name, labelvar.name])
for x, y, z in zip(loss_value, output_value, label_value):
print list(x.flat), np.sum(y.flat), list(z.flat)
def val_reader():
traindataset = myreader.myreader_classify(val_datasetfile,
val_labelfile,
'val',
doshuffle=False,
iscolor=0,
preprocessfunc=preprocess)
for image, label in traindataset:
yield image, label
def val_reader():
traindataset = myreader.myreader_classify(args.val_datasetfile,
args.val_labelfile, 'val')
for image, label in traindataset:
yield image, label
def debug_faceloss():
initrandom()
#debug loss NaN problem
from losses import SoftmaxLoss
from losses import ArcMarginLoss
train_datasetfile = 'dataset/face_ms1m_small/train.data'
train_labelfile = 'dataset/face_ms1m_small/train.label'
traindataset = myreader.myreader_classify(
train_datasetfile,
train_labelfile,
'train',
iscolor=1,
doshuffle=False, #disable shuffle for debug
)
embedding_size = 256
class_dim = 1000
batchsize = 128
imgshape = [3, 112, 112]
model = resnet18.ResNet18()
#model = SimpleNet()
image = fluid.layers.data(name='image', shape=imgshape, dtype='uint8')
labelvar = fluid.layers.data(name='label', shape=[1], dtype='int64')
inputdata = trainmodule.preprocessimg(image,
mean=[127.5, 127.5, 127.5],
std=[128.0, 128.0, 128.0])
out = model.net(input=inputdata, embedding_size=embedding_size)
metricloss = ArcMarginLoss(class_dim=class_dim,
margin=0.5,
scale=64,
easy_margin=False)
#metricloss = SoftmaxLoss(class_dim=class_dim)
cost, logit = metricloss.loss(out, labelvar)
avg_cost = fluid.layers.mean(x=cost)
if os.path.exists('debugbatch.npz'):
loaddata = np.load('debugbatch.npz')
imgs, labels = loaddata['imgs'], loaddata['labels']
print('img,label shape', imgs.shape, labels.shape, 'save mat sum',
np.sum(imgs), np.sum(labels))
else:
imgs = []
labels = []
for img, label in traindataset:
if label < class_dim:
imgs.append(img.reshape([1] + imgshape))
labels.append(label)
if len(imgs) == batchsize:
break
imgs = np.vstack(imgs)
labels = np.array(labels, np.int64).reshape((-1, 1))
np.savez('debugbatch.npz', imgs=imgs, labels=labels)
print('save mat sum', np.sum(imgs), np.sum(labels))
#place = fluid.CPUPlace()
place = fluid.CUDAPlace(0)
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())
if not os.path.exists('debugmodel'):
fluid.io.save_persistables(exe,
'debugmodel',
main_program=fluid.default_main_program())
else:
fluid.io.load_persistables(exe,
'debugmodel',
main_program=fluid.default_main_program())
fcvars = getvariable_value('embedded.w_0')
assert (len(fcvars) == 1)
fcname = fcvars[0].name
#optimizer = fluid.optimizer.SGD(learning_rate=0.1)
#opts = optimizer.minimize(avg_cost)
fetch_list = [cost.name, inputdata.name, logit.name]
for i in range(1):
outputlist = exe.run(fluid.default_main_program(),
feed={
'image': imgs,
'label': labels
},
fetch_list=fetch_list)
for row in range(len(outputlist[0])):
outsum = [(fetch_list[col], list(outputlist[col][row].flat[:3]))
for col in range(len(fetch_list))]
out = outputlist[2][row]
normout = np.sum(out * out)
print row, outsum, normout, out.shape