jiang-du / blazepose-tensorflow Goto Github PK

预训练模型没有

Hello, I was looking through the implementation, and I was curious about how you made choices regarding what to use for the architecture. Looking at the original paper here, I saw conv filter sizes and channel numbers listed, but I don't see any listings of the number of blaze blocks to use per layer and such parameters listed.

I used this paper as a reference to understand the structure of BlazeBlocks, but is that what you based your implementation on? Or did you get your architecture parameters from a different paper, or did you just experiment with it yourself?

Hello,i train by the train.py but not converge,the results is wrong.My environment is ubuntu16.04 and python3.6.2,tensorflow2.4,could you give me some advice?

Epoch 000: Train Loss: 1.061, Accuracy: 30.17884%
2020-11-19 17:06:08
Epoch 001: Train Loss: 0.715, Accuracy: 23.60993%
2020-11-19 17:07:15
Epoch 002: Train Loss: 0.621, Accuracy: 20.01843%
2020-11-19 17:08:42
Epoch 003: Train Loss: 0.481, Accuracy: 13.48590%
2020-11-19 17:09:47
Epoch 004: Train Loss: 0.266, Accuracy: 5.80015%
Epoch 004, Validation accuracy: 2.96213%
2020-11-19 17:11:15
Epoch 005: Train Loss: 0.164, Accuracy: 1.45889%
2020-11-19 17:12:21
Epoch 006: Train Loss: 0.123, Accuracy: 1.06765%
2020-11-19 17:13:27
Epoch 007: Train Loss: 0.091, Accuracy: 1.22532%
2020-11-19 17:14:33
Epoch 008: Train Loss: 0.089, Accuracy: 1.19700%
2020-11-19 17:15:39
Epoch 009: Train Loss: 0.078, Accuracy: 0.89985%
Epoch 009, Validation accuracy: 0.81636%
2020-11-19 17:17:06
Epoch 010: Train Loss: 0.073, Accuracy: 0.73148%
2020-11-19 17:18:12
Epoch 011: Train Loss: 0.071, Accuracy: 0.69256%
2020-11-19 17:19:18
Epoch 012: Train Loss: 0.068, Accuracy: 0.71409%
2020-11-19 17:20:24
Epoch 013: Train Loss: 0.065, Accuracy: 0.70186%
2020-11-19 17:21:30
Epoch 014: Train Loss: 0.064, Accuracy: 0.67790%
Epoch 014, Validation accuracy: 0.67563%

你好！可否分享一个已经train好的模型？

It seems the dataset at https://sam.johnson.io/research/lsp_dataset.zip is no longer available. Do you have a copy of it that can be shared?

Hello! I saw your great TF realization. I would be thank from your help. Now, I'm realizing the similar task for BlazePose model in TF, I want to add-learn this model additionally based on my new dataset in the specific area (make a fine-tuning), but I cannot understand to which layer I need to freeze the model. You have a freezing by "for layer in model.layers[0:16]" - so, it's a feature extractor (FE) in this range of layers, and in your case fine-tuning is started after FE by indicating a number of block (16). Thank you for your work. At the same time, I have find other implementation of this model with more number of blocks, and it confused me.
What's differences between both models, why they both have different number of blocks ?

I'm using this model with pre-trained weights by this graph.
Thus, I need a number of layer, from which I need to freeze a model. Could you help with these issues. Thank you in advance.

Thanks for your work!

How about real time？Have you tested how much FPS can reach?And what about the accuracy on online camera?

Do you have a plan to achieve the part of pose tracking?

Thank you for open this project.

1. In paper ** The gradients from the re- gression encoder are not propagated back to the heatmap- trained features (note the gradient-stopping connections in Figure 4).** However, not see this operate in your train step.
2. Offset also mention in paper.
   Look forward your reply.

The website of dataset is not valid

did you evaluate your network on the lsp dataset? if yes can you kindly share how many epochs for pre-training and training or other parameters?

Hello！Sorry to interrupt you. I have also paid attention to this article recently. Thank you for your work. However, some parts of the article confuse me. I would like to consult that what exactly is pose track and what is the difference between it and general pose estimation（such as openpose）? , Thank you very much for your answers

Thanks you for providing this excellent implemention for us, but I encounter some error when I train the model.
When I train the model, it has the following error "tensorflow.python.framework.errors_impl.ResourceExhaustedError: Exception encountered when calling layer "depthwise_conv2d_43" " f"(type DepthwiseConv2D). "
I don't know what goint wrong.

It looks like the LSP dataset is not available anymore, do you have an alternative?

Thank you!

hello ～I have a question
Data picture, do I need to keep the human body facing up? For example, head down, legs up
thanks

1. train_mode=0
ValueError: Dimensions must be equal, but are 14 and 128 for '{{node mean_squared_error/SquaredDifference}} = SquaredDifference[T=DT_FLOAT](blaze_pose/sequential_76/reshape/Reshape, IteratorGetNext:1)' with input shapes: [?,14,3], [?,128,128,14].
2. train_mode=1,output is [ 8.4 16.7 29.8 27.8 16.1 9.3 6.3 9.9 25.5 21.4 14.2 7. 34.7 23.7] Average = 17.914286%

Hi there,

Am trying to run the demp.py file after training, I downloaded the Yolov4 weights and the yolo "coco.names" files, kept in the folder yolov4, but as I execute the file demo.py, I get an error in ---> set_memory_growth raise RuntimeError(.....
Could you please explain what do you mean when you say train your dataset in the 2nd point of the demo.py instructions?

Hi author,

Looks like there's no data augmentation all all, any reasons for that?

Training the joints model is overfitting to training data. Training is converging, but validation accuracy hovers around 10000% according to mse training metric. Is this a known issue? If so, any guidance on how to address it?

如题

Hello！Sorry to interrupt you. I am comfused about the tracker part on blazepose. I just understand the whole process is that detect the first frame,then tracker the next frame according to the pre-frame's skeleton position.But I don't understand the tracker part on blazepose.Can you explain the tracker part on blazepose? Thank you very much!

Dear friend!

Thank you for your work! It's really impressive!
I have several questions concerning your model. I would be very grateful for your answers!

1. I tried to run the model on Colab Pro with 25GB RAM. When I run the training process on 200 photos, I get a message about memory overflow. Is this behavior normal? What could be the reason of such memory consumption?
2. Could you please specify how much time one epoch takes for you?
3. Why did you divide the LSP dataset into two equal parts for training and testing? Usually the data is divided in a ratio of about 80% to 20%.
4. Why don't you use data normalization before training?

how can we train our model using custom dataset with your implementation and it can be happen than is it efficent or not