Experimental result

[1292765944]

Cool code! I wonder whether you have reimplemented the exact results of the FPN paper?

Best!
guangxing

[Yc174]

Have you trained this ?Can you tell me the results? @1292765944

[xmyqsh]

I'm starting training...

[chl916185]

When I was loading the trained model，I met this error：
Traceback (most recent call last):
File "./faster_rcnn/test_net.py", line 44, in
network = get_network(network_name)
File "./faster_rcnn/../lib/networks/factory.py", line 19, in get_network
return FPN_test()
File "./faster_rcnn/../lib/networks/FPN_test.py", line 26, in init
self.setup()
File "./faster_rcnn/../lib/networks/FPN_test.py", line 356, in setup
.fc(n_classes, relu=False, name='cls_score'))
File "./faster_rcnn/../lib/networks/network.py", line 34, in layer_decorated
layer_output = op(self, layer_input, *args, **kwargs)
File "./faster_rcnn/../lib/networks/network.py", line 407, in fc
regularizer=self.l2_regularizer(cfg.TRAIN.WEIGHT_DECAY))
File "./faster_rcnn/../lib/networks/network.py", line 97, in make_var
return tf.get_variable(name, shape, initializer=initializer, trainable=trainable, regularizer=regularizer)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/variable_scope.py", line 988, in get_variable
custom_getter=custom_getter)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/variable_scope.py", line 890, in get_variable
custom_getter=custom_getter)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/variable_scope.py", line 348, in get_variable
validate_shape=validate_shape)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/variable_scope.py", line 333, in _true_getter
caching_device=caching_device, validate_shape=validate_shape)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/variable_scope.py", line 639, in _get_single_variable
name, "".join(traceback.format_list(tb))))
ValueError: Variable cls_score/weights already exists, disallowed. Did you mean to set reuse=True in VarScope? Originally defined at:

File "./faster_rcnn/../lib/networks/network.py", line 97, in make_var
return tf.get_variable(name, shape, initializer=initializer, trainable=trainable, regularizer=regularizer)
File "./faster_rcnn/../lib/networks/network.py", line 407, in fc
regularizer=self.l2_regularizer(cfg.TRAIN.WEIGHT_DECAY))
File "./faster_rcnn/../lib/networks/network.py", line 34, in layer_decorated
layer_output = op(self, layer_input, *args, **kwargs)

The reason is that：
#========= RCNN ============
.fc(n_classes, relu=False, name='cls_score'))
in lib/networks/FPN_test.py.
@xmyqsh

[xmyqsh]

@1292765944 @Yc174 @chl916185

train result, rpn_loss is on the fly...
What is the reason that may cause rpn_loss not converge, do you think ?

speed: 0.412s / iter
image: 2008_000307.jpg iter: 199870 / 200000, total loss: 21.4851, rpn_loss_cls: 1.2544, rpn_loss_box: 19.9337, loss_cls: 0.2095, loss_box: 0.0876, lr: 0.0000
00
speed: 0.401s / iter
image: 2008_005319.jpg iter: 199880 / 200000, total loss: 11.8511, rpn_loss_cls: 1.3227, rpn_loss_box: 10.2611, loss_cls: 0.2030, loss_box: 0.0643, lr: 0.0000
00
speed: 0.406s / iter
image: 006483.jpg iter: 199890 / 200000, total loss: 25.3140, rpn_loss_cls: 1.1014, rpn_loss_box: 24.0750, loss_cls: 0.1176, loss_box: 0.0201, lr: 0.000000
speed: 0.395s / iter
image: 002967.jpg iter: 199900 / 200000, total loss: 10.4668, rpn_loss_cls: 1.5063, rpn_loss_box: 8.6171, loss_cls: 0.2443, loss_box: 0.0990, lr: 0.000000
speed: 0.397s / iter
image: 2011_001198.jpg iter: 199910 / 200000, total loss: 14.9206, rpn_loss_cls: 1.4058, rpn_loss_box: 13.2595, loss_cls: 0.2015, loss_box: 0.0537, lr: 0.000000
speed: 0.380s / iter
image: 2008_003645.jpg iter: 199920 / 200000, total loss: 19.0740, rpn_loss_cls: 1.2057, rpn_loss_box: 17.1077, loss_cls: 0.5278, loss_box: 0.2328, lr: 0.000000
speed: 0.426s / iter
image: 005811.jpg iter: 199930 / 200000, total loss: 10.8934, rpn_loss_cls: 1.4540, rpn_loss_box: 9.0603, loss_cls: 0.2846, loss_box: 0.0945, lr: 0.000000
speed: 0.439s / iter
image: 2009_003860.jpg iter: 199940 / 200000, total loss: 14.3258, rpn_loss_cls: 1.6082, rpn_loss_box: 12.5498, loss_cls: 0.1635, loss_box: 0.0044, lr: 0.000000
speed: 0.388s / iter
image: 2011_000556.jpg iter: 199950 / 200000, total loss: 22.7038, rpn_loss_cls: 1.2282, rpn_loss_box: 20.8446, loss_cls: 0.3576, loss_box: 0.2734, lr: 0.000000
speed: 0.397s / iter
image: 2008_004585.jpg iter: 199960 / 200000, total loss: 11.1179, rpn_loss_cls: 1.3557, rpn_loss_box: 9.2402, loss_cls: 0.3512, loss_box: 0.1708, lr: 0.000000
speed: 0.346s / iter
image: 2008_000393.jpg iter: 199970 / 200000, total loss: 11.2609, rpn_loss_cls: 1.7995, rpn_loss_box: 8.7590, loss_cls: 0.4851, loss_box: 0.2173, lr: 0.000000
speed: 0.371s / iter
image: 2011_003081.jpg iter: 199980 / 200000, total loss: 10.4372, rpn_loss_cls: 1.4303, rpn_loss_box: 8.8616, loss_cls: 0.1440, loss_box: 0.0013, lr: 0.000000
speed: 0.342s / iter
image: 2008_007069.jpg iter: 199990 / 200000, total loss: 19.8492, rpn_loss_cls: 1.3173, rpn_loss_box: 18.3676, loss_cls: 0.1637, loss_box: 0.0005, lr: 0.000000
speed: 0.356s / iter

[chl916185]

the lr is 0.000000,why？
@xmyqsh

[1292765944]

@xmyqsh I met similar problem. It is not easy to converge using the params in FPN paper. such as 800 pixels of the shorter side, including anchor boxes that are outside the image.