xuannianz / keras-gaussianyolov3 Goto Github PK

您好，我在下载权重的过程中发现，百度网盘的链接需要提取码，请问可以提供一下么？谢谢您

Hi, I want to know where does the introduction of gaussian parameters for uncertainty calculation?
Thanks.

When I implement with pytorch, the validation stage costs much more time than origin YOLOv3...it seems like NMS occupies a lot of time. I wonder if your validation stage is also long?

Regress the mean and variance of X, y, W, h, and finally how to regress the value of X, y, W, H???

Hi, my project needs a fast detector but I'm not good at C, so I want to use repo in python. Is this repo as fast as claimed on the paper?
Using pytorch_GaussianYOLOv3, I save the resized image which is 416x416 and test on them, getting 21 fps, which is a little bit low.
Many thanks!

@xuannianz hi, thank you for this good project!
I have noticed that in your implementation, the gaussian loss is
pi = tf.constant(np.pi)
Z = (2 * pi * (sigma + sigma_const) ** 2) ** 0.5
probability_density = tf.exp(-0.5 * (x - mu) ** 2 / ((sigma + sigma_const) ** 2)) / Z
nll = -tf.log(probability_density + 1e-7)
and want to know how the mu and sigma will change during the training.
by my inference， maybe the mu should approach to the true value of x,y,w,h, and the sigmas should be reduced to 0. Is this right？

hi, if i want to train my own data, how to do. I don;'t know how to make may data, and the entry of data to train. Please help me.
Thanks.

# in loss.py
   y_true_delta_xy = raw_y_true[..., :2] * grid_shapes[l][::-1] - grid
# ----------------------------------------------------------------------
# in util_graph.py : def y_pred_graph:
    ...
    y_pred_delta_xy = K.sigmoid(raw_y_pred[..., :2])
    y_pred_log_wh = raw_y_pred[..., 2:4]
    y_pred_sigma = K.sigmoid(raw_y_pred[..., 4:8])

raw_y_true is in (0, 1) and grid is in (0, raw_y_pred_shape_wh), so y_true_delta_xy is not in (0, 1).
but y_pred_delta_xy was sigmoided, it must be in (0, 1).
that confuse me very much...
help, pls

Can you explain to me how to implement loss function Gaussian yolov3 in your code? I want to reimplement it on pytorch

hi, @xuannianz I try to train the VOC data, and get this error
File "/home/bigdata/chenfangxiong/keras-GaussianYOLOv3/train.py", line 341, in
main()
File "/home/bigdata/chenfangxiong/keras-GaussianYOLOv3/train.py", line 295, in main
model, prediction_model = yolo_body(anchors, num_classes=num_classes)
File "/home/bigdata/chenfangxiong/keras-GaussianYOLOv3/model.py", line 130, in yolo_body
[y1, y2, y3, fm_13_input, fm_26_input, fm_52_input])
File "/home/bigdata/.conda/envs/fcon/lib/python3.6/site-packages/keras/engine/base_layer.py", line 451, in call
output = self.call(inputs, **kwargs)
File "/home/bigdata/.conda/envs/fcon/lib/python3.6/site-packages/keras/layers/core.py", line 716, in call
return self.function(inputs, **arguments)
File "/home/bigdata/chenfangxiong/keras-GaussianYOLOv3/loss.py", line 46, in yolo_loss
raw_y_pred = K.reshape(yolo_output, [-1, grid_shape[0], grid_shape[1], num_anchors_per_layer, num_classes + 9])
File "/home/bigdata/.conda/envs/fcon/lib/python3.6/site-packages/keras/backend/tensorflow_backend.py", line 2189, in reshape
return tf.reshape(x, shape)
File "/home/bigdata/.conda/envs/fcon/lib/python3.6/site-packages/tensorflow/python/ops/gen_array_ops.py", line 6197, in reshape
"Reshape", tensor=tensor, shape=shape, name=name)
File "/home/bigdata/.conda/envs/fcon/lib/python3.6/site-packages/tensorflow/python/framework/op_def_library.py", line 513, in _apply_op_helper
raise err
File "/home/bigdata/.conda/envs/fcon/lib/python3.6/site-packages/tensorflow/python/framework/op_def_library.py", line 510, in _apply_op_helper
preferred_dtype=default_dtype)
File "/home/bigdata/.conda/envs/fcon/lib/python3.6/site-packages/tensorflow/python/framework/ops.py", line 1107, in internal_convert_to_tensor
ret = conversion_func(value, dtype=dtype, name=name, as_ref=as_ref)
File "/home/bigdata/.conda/envs/fcon/lib/python3.6/site-packages/tensorflow/python/framework/constant_op.py", line 217, in _constant_tensor_conversion_function
return constant(v, dtype=dtype, name=name)
File "/home/bigdata/.conda/envs/fcon/lib/python3.6/site-packages/tensorflow/python/framework/constant_op.py", line 196, in constant
value, dtype=dtype, shape=shape, verify_shape=verify_shape))
File "/home/bigdata/.conda/envs/fcon/lib/python3.6/site-packages/tensorflow/python/framework/tensor_util.py", line 436, in make_tensor_proto
_AssertCompatible(values, dtype)
File "/home/bigdata/.conda/envs/fcon/lib/python3.6/site-packages/tensorflow/python/framework/tensor_util.py", line 344, in _AssertCompatible
raise TypeError("List of Tensors when single Tensor expected")
TypeError: List of Tensors when single Tensor expected

How many interations do we need to get mAP50 0.8535 accuracy in VOC ?

HI, I download your code and just run it. But in several hundreds steps, the train loss is nan. I never any setting, what's the problems? I use the VOC dataset. Please help me!
Thanks!

你好，我现在想将普通YOLOv3转换为GaussianYOLOv3，现在已经参照xuannianz的keras-GaussianYOLOv3修改了模型的yolo_body、yolo_loss函数，增加了null_loss函数，以及修改convert文件中函数。但是执行convert将Gaussian_yolov3_BDD.weights转换为h5时会出现如下报错：
buffer=weights_file.read(weights_size * 4))
TypeError: buffer is too small for requested array
这是怎么导致的，还需要更改什么函数或者配置文件吗，请帮我解答，谢谢

HI, could you please complete the KITTI evaluation code, or add explanation of how to conduct model on KITTI test dataset, just like the original paper?

how to run on CPU？
What configuration should I change？
Thanks！

Parsing section convolutional_57
conv2d bn leaky (1024, 512, 3, 3)
Parsing section convolutional_58
conv2d linear (54, 1024, 1, 1)
Parsing section Gaussian_yolo_0
Traceback (most recent call last):
File "convert.py", line 250, in
main(parser.parse_args())
File "convert.py", line 208, in main
raise ValueError('Unsupported section header type: {}'.format(section))
ValueError: Unsupported section header type: Gaussian_yolo_0

i want to see the mAP curves on training process. thx

I trained with my own dataset and made the above mistakes

i want to know if why the yolov3.cfg not change into gaussian_yolo.cfg? if the network dont need change?
thank you

@xuannianz Hello, i am trying to implement guassian yolov3 in pytorch,
and i am using your guassian loss function.
During training, the guassian loss of x,y,w,h can reduce as normal, but in the test, all the test mAPs are 0. So i checked the training process, and have found that the mean of x,y,w,h didn't approach to the true value of GT while the sigmas can be rightly reduced to zero.
Can you give some advice? thanks a lot :）