Hi yoshall,

Thanks for you contribution. I have some question about the code and work. I hope you could give me a hand.

Part 1: According to the sample_data, I think there are 35 nodes, each node generate 19 time series. However, I am a little confusing about the meaning of all input files:

1. I notice that the code reads 7 files and process them to generate inputs for training. While the mearning for local_inputs, external_inputs and decoder_gts are obvious, I don't know what is global_atten_state and local_atten_state, why not generate them from raw data (for example, local inputs and global inputs)?
2. What is global_inputs.npy? Why it's shape is 500*35?
3. Is global_attn_state_indics the same as global_inputs_indics? I noticed that in get_batch_feed_dict() function, train_global_inp = training_data[1] which in fact is global_attn_index.

Part 2: Besides I think the model generate prediction for each node separately. Do you train the model for each node separately or train a unified model?

Part 3: a suggestion: I hope you could add some explanation to the input files as they are different to raw inputs and maybe also publish the code process your raw data (http://urban-computing.com/data/Data-1.zip).

Looking forward to your answer and thanks for you patient.

我是用的是JD lab的Data_1的空气质量数据，里面没有POI的数据，而且我主要是想知道，POI是只在external module中使用的吗？是对全北京的POI数据的分类统计数据吗？那这样的话不是只是固定的值吗？跟PM2.5的预测会有关系吗？
总是是想知道具体关于POI是怎么处理和使用的。

原始数据集http://urban-computing.com/data/Data-1.zip
看代码您有5个npy文件，想问下您方便提供一下从上面的原始数据集转变为npy文件的代码么？谢谢

Hi, yoshall:

  I've installed tensorflow-gpu 1.2.1 and cuda8.0. An error occurred in line 30 while **train_model.py** was running. Line 30 is  **hps.override_from_dict(hps_dict)** and the error is **'HParams' object has no attribute 'override_from_dict'**. Thus, I‘ve tried some other versions of tensorflow, but the error still exists. Can you give me advice on how to solve this problem ?

Jack

Hi yoshall,

I'd like to implement you code using sample_data. However, I got some troubles. You can see this information:
total parameters: 554054 ----------epoch 0----------- ----------epoch 1----------- ----------epoch 2----------- ----------epoch 3----------- ----------epoch 4----------- ----------epoch 5----------- ----------epoch 6----------- ----------epoch 7----------- ----------epoch 8----------- ----------epoch 9----------- ----------epoch 10----------- ----------epoch 11----------- ----------epoch 12----------- ----------epoch 13----------- ----------epoch 14----------- stop training !!!
May be not run this code:
if valid_loss < min(valid_losses[:-1]): print('iter {}\tvalid_loss = {:.6f}\tmodel saved!!'.format( iter, valid_loss)) saver.save(sess, model_dir + 'model_{}.ckpt'.format(iter)) saver.save(sess, model_dir + 'final_model.ckpt') else: print('iter {}\tvalid_loss = {:.6f}\t'.format( iter, valid_loss))

So,the previously saved model cannot be loaded when running the test_model.py file.
NotFoundError (see above for traceback): Unsuccessful TensorSliceReader constructor: Failed to find any matching files for ./logs/GeoMAN-12-6-2-128-0.30-0.001/saved_models/final_model.ckpt [[Node: save/RestoreV2 = RestoreV2[dtypes=[DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, ..., DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_INT32, DT_FLOAT], _device="/job:localhost/replica:0/task:0/device:CPU:0"](_arg_save/Const_0_0, save/RestoreV2/tensor_names, save/RestoreV2/shape_and_slices)]]
Look forward to receiving your kind reply soon. Thank you very much.
best wishes,
wendong

In the GeoMAN.py, Line 262 to Line 273:

multiply attention weights with the original input

local_x = local_attn * local_inp
global_x = global_attn * global_inp

Run the BasicLSTM with the newly input

cell_output, state = cell(tf.concat([local_x, global_x], axis=1), state)

Run the attention mechanism.

with tf.variable_scope('local_spatial_attn'):
local_attn = local_attention(state)
with tf.variable_scope('global_spatial_attn'):
global_attn = global_attention(state)
attn_weights.append((local_attn, global_attn))

Does this only consider cell state when calculating attention? In the paper Equation (1), cell state and hidden state are concatenated?

Hi Yoshall,

I'd like to implement you project using my own data. However, I got some troubles about the meta data of input files. Following your uploaded codes, I can only see the input file names in GeoMAN/utils.py, could you please upload a small sample of your input files for reference.

Look forward to receiving your kind reply soon. Thank you very much.

Gabby

Hello, I was curios when do you plan to release the datasets, I'm interested in using them as well.

Thank you in advance.

Hi yoshall，

We implemented your project with our own dataset and had a problem with the test errors. We set
the n_encoder_steps and the n_decoder_steps T=τ=12 during the training phase to make predictions. Generally，when testing the model, there‘ll be an increase of test errors over the 12 test timesteps. However，the errors decreased and the 12th timestep got the least test error.

Looking forward to your reply and thank you very much.