Machine Learning & NLP Bug Fix Expert
Budget: $25 – $50 USD
The project is already completed, but some steps are inaccurate, therefore the final output is inaccurate. Your task is to correct the mistakes and ensure the project fully aligns with the methods used in the GitHub repository while ensuring it’s adaptable to different datasets.
The graphs showing the predicted values vs the true values are inaccurate and misaligned. the predicted values in the tables are therefore inaccurate, so are the attention weight analysis results, and maybe even the MSE results for all the datasets. Double check everything, and that it each step fully aligns with the original code as much as possible, if you need to change something that is completely different from the original code in GitHub, let me know so we can discuss. I’m looking for quality output.
The direction of the project is correct, in the sense that I want to see the MSE, both types of attention weights with percentages, the graphs and the tables, but they must all be accurate. So the requirement is for you to review and correct the mistakes of the project for a quality output.
I’m attaching the notebooks, and I’m also including the GitHub link so you can refer to the original GitHub repository for you to see how the coders did their projects, and I’m including a screenshot of one of their outputs for you to get an idea of the kind of quality work that I want.
I’m also including the dataset to make things easier.
GitHub Links:
- Transformers Diviner Model: [Diviner-Nonstationary-time-series-forecasting](https://github.com/CapricornGuang/Diviner-Nonstationary-time-series-forecasting/blob/main/README.md)
The original project description was: I would like to know how the Transformers diviner model provided below from Github can be used to forecast various time series in the original scale, and make sense of them. The idea is to stick to the methods of the original coders from GitHub as much as possible while making sure the code is adaptable to different datasets. Disregard all the time series provided in the original coding like ETT, WTH, etc, and from that rather just focus on the Exchange/gold prices data 60 day gold prices, and the new data that I provide. Use good visuals for the graphs.
1. Forecasting Gold Prices:
- After obtaining the MSE/RMSE (already provided in the code), provide code to create a table that displays the data, actual value and predicted values over the next 60 days for each test sample, in the original scale.
- Create a table with columns for date, "true" prices, and "pred" prices for a 60-day period for the entire test sample. For the predictions, I want to see each prediction in the time step until the 60 day period.
Generate graph for each dataset in the original scale, that shows the timestep1 “pred” values vs the actual value” and another two graphs that shows two different test samples with each a 60 day pred values vs 60 days of actual prices, just like in the home page README section of the GitHub repository for WTH for example.
Create a table for each dataset: Date, True, Pred as column headers, and underneath them, have the exact date the test sample, have the original price of the specific date, and have the forecasted values for each timestep until the 60th day. (This should be reflected in the graphs).
So I would know, e.g Jul 10th 2021, the actual price was $1211.9, I want to know the the forecasted price is for each time step until we reach 60 days later, and how that compares with the actual prices, for each time step. I also prefer to see the date of each time step so I can track it easily. So I know that if 60 days from today for example in Sept 10, the model is telling me the price of gold will be $1510 (this is just a an example), I want to see how that model actually performed when we actually get to that specific date. Both in the tables, and graph, that is the main part of the project. The gold prices dataset is in the GitHub repository under data, under Exchange/gold prices data 60 day gold prices.
2. Data Collection and Forecasting for Multiple Assets:
- Show how you’re pulling data from Excel file (csv) that contains the date, TARGET FEATURE, and THE VARIABLES NEXT TO IT, and HOW YOU are UPLOADing THAT INTO THE MODEL AND THE MODEL FORECASTED BASED ON ALL THOSE INFORMATION.
- Daily Apple stock prices excel (alognside its P/E ratio and Earnings per Share for a 10 year period).
3. Attention Weights Analysis:
- Clearly code and Display attention weights for the features/variables used to rank by importance in percentage terms.
Use both the traditional attention weight for all layers, and separately the InverseDeepthDifferenceBlock (ddblock) so I can see how they are both ranking features. (Clearly show/make notes how you are interpreting that visual. They should have dark backgrounds.)
4. Repeat all these steps for EUR/USD spot price 10 years of data as target feature along with two variables, U.S. interest rates and U.S. trade balance (net trade deficit for a 10 year period).
5. Repeat steps 1 to 3 for 10 years of data for US Real GDP as target feature along with three variables, U.S. interest rates and U.S. inflation CPI, and China’s Real GDP
6. Use the model to forecast for 60 days each of these 4 target features simultaneously, just like in the original mode it simultaneously forecasts ETT, WTH, etc
7. Implementation Details:
- Ensure coding compatibility with Google Colab, utilizing GPU
8. Include clear notes in the code for ease of understanding and replication, because eventually I’ll be using different assets with different kinds of variables.
Skill : Machine Learning, Python, PyTorch, Computer Vision, Django
The graphs showing the predicted values vs the true values are inaccurate and misaligned. the predicted values in the tables are therefore inaccurate, so are the attention weight analysis results, and maybe even the MSE results for all the datasets. Double check everything, and that it each step fully aligns with the original code as much as possible, if you need to change something that is completely different from the original code in GitHub, let me know so we can discuss. I’m looking for quality output.
The direction of the project is correct, in the sense that I want to see the MSE, both types of attention weights with percentages, the graphs and the tables, but they must all be accurate. So the requirement is for you to review and correct the mistakes of the project for a quality output.
I’m attaching the notebooks, and I’m also including the GitHub link so you can refer to the original GitHub repository for you to see how the coders did their projects, and I’m including a screenshot of one of their outputs for you to get an idea of the kind of quality work that I want.
I’m also including the dataset to make things easier.
GitHub Links:
- Transformers Diviner Model: [Diviner-Nonstationary-time-series-forecasting](https://github.com/CapricornGuang/Diviner-Nonstationary-time-series-forecasting/blob/main/README.md)
The original project description was: I would like to know how the Transformers diviner model provided below from Github can be used to forecast various time series in the original scale, and make sense of them. The idea is to stick to the methods of the original coders from GitHub as much as possible while making sure the code is adaptable to different datasets. Disregard all the time series provided in the original coding like ETT, WTH, etc, and from that rather just focus on the Exchange/gold prices data 60 day gold prices, and the new data that I provide. Use good visuals for the graphs.
1. Forecasting Gold Prices:
- After obtaining the MSE/RMSE (already provided in the code), provide code to create a table that displays the data, actual value and predicted values over the next 60 days for each test sample, in the original scale.
- Create a table with columns for date, "true" prices, and "pred" prices for a 60-day period for the entire test sample. For the predictions, I want to see each prediction in the time step until the 60 day period.
Generate graph for each dataset in the original scale, that shows the timestep1 “pred” values vs the actual value” and another two graphs that shows two different test samples with each a 60 day pred values vs 60 days of actual prices, just like in the home page README section of the GitHub repository for WTH for example.
Create a table for each dataset: Date, True, Pred as column headers, and underneath them, have the exact date the test sample, have the original price of the specific date, and have the forecasted values for each timestep until the 60th day. (This should be reflected in the graphs).
So I would know, e.g Jul 10th 2021, the actual price was $1211.9, I want to know the the forecasted price is for each time step until we reach 60 days later, and how that compares with the actual prices, for each time step. I also prefer to see the date of each time step so I can track it easily. So I know that if 60 days from today for example in Sept 10, the model is telling me the price of gold will be $1510 (this is just a an example), I want to see how that model actually performed when we actually get to that specific date. Both in the tables, and graph, that is the main part of the project. The gold prices dataset is in the GitHub repository under data, under Exchange/gold prices data 60 day gold prices.
2. Data Collection and Forecasting for Multiple Assets:
- Show how you’re pulling data from Excel file (csv) that contains the date, TARGET FEATURE, and THE VARIABLES NEXT TO IT, and HOW YOU are UPLOADing THAT INTO THE MODEL AND THE MODEL FORECASTED BASED ON ALL THOSE INFORMATION.
- Daily Apple stock prices excel (alognside its P/E ratio and Earnings per Share for a 10 year period).
3. Attention Weights Analysis:
- Clearly code and Display attention weights for the features/variables used to rank by importance in percentage terms.
Use both the traditional attention weight for all layers, and separately the InverseDeepthDifferenceBlock (ddblock) so I can see how they are both ranking features. (Clearly show/make notes how you are interpreting that visual. They should have dark backgrounds.)
4. Repeat all these steps for EUR/USD spot price 10 years of data as target feature along with two variables, U.S. interest rates and U.S. trade balance (net trade deficit for a 10 year period).
5. Repeat steps 1 to 3 for 10 years of data for US Real GDP as target feature along with three variables, U.S. interest rates and U.S. inflation CPI, and China’s Real GDP
6. Use the model to forecast for 60 days each of these 4 target features simultaneously, just like in the original mode it simultaneously forecasts ETT, WTH, etc
7. Implementation Details:
- Ensure coding compatibility with Google Colab, utilizing GPU
8. Include clear notes in the code for ease of understanding and replication, because eventually I’ll be using different assets with different kinds of variables.
Skill : Machine Learning, Python, PyTorch, Computer Vision, Django