tanchongmin / tensorflow-implementations

TensorFlow-Implementations

This repository contains the projects I have done with TensorFlow, and includes a few tutorials.

Check out my Youtube channel which uses some of this material for lessons.

Other social media sites:

Youtube: https://www.youtube.com/channel/UCgRNCT8mrzKYebyG3Ao9DJA/videos

Twitch: https://www.twitch.tv/johncm99

AI Blog: https://delvingintotech.wordpress.com/

LinkedIn: https://www.linkedin.com/in/chong-min-tan-94652288/

Twitter: https://twitter.com/johntanchongmin

TensorFlow / Deep Learning Tutorial: These are the files I have created to teach others Deep Learning:

• MLP.pdf / MLP Part 2.pdf / CNN.pdf / RNN.pdf / Transformer.pdf / GNN.pdf
• AlphaGo:Zero.pdf
• Geometric Deep Learning.pdf (Summary and personal insights of Geometry Deep Learning proto-book https://geometricdeeplearning.com/ by Michael M. Bronstein, Joan Bruna, Taco Cohen, Petar Veličković)
• Linear Regression in TensorFlow.ipynb
• Fashion MNIST in TensorFlow.ipynb (Adapted from https://www.coursera.org/learn/introduction-tensorflow Week 1 Notebook)
• Fashion MNIST (CNN) in TensorFlow.ipynb
• GAT_GCN_node_classification.ipynb: Graph Neural Network implementation using TensorFlow on Cora Dataset
• Tutorial_RNN_Text_Generation.ipynb
• Custom Training Loop: Converts a simple Keras model.fit into an expanded customizable training loop using GradientTape. Includes graph plotting and model visualization utilities.
• GPT_Demo.ipynb: Uses huggingface GPT-2 model for text generation, as well as Bertviz for attention visualization
• ChatGPT - From Transformers to Reinforcement Learning from Human Feedback (RLHF).pdf: Introduction to how ChatGPT works

Paper_Reviews Folder:

• Overview: Documents Interesting Deep Learning / Machine Learning Papers
• Decision Transformers.pdf
• GATO.pdf
• Policy Graident Presentation.pdf
• AlphaTensor Presentation.pdf

RL Folder:

• Overview: Documents Reinforcement Learning Experiments and Tutorials
• RL Part 1.pdf, DQN.pdf
• MCTS.ipynb: Using Monte Carlo / Monte Carlo Tree Search / Value Estimates to solve the game of Nim
• Cart Pole.ipynb: Using rule-based and DQN to solve Cart Pole

Below are the projects I have done using TensorFlow (in Implementations folder):

• Text Generation: Compares RNN vs a Markov Chain method to generate text. Surprisingly, both are around the same, just by using like n=5 previous characters to predict the next one for the Markov Chain method.
• Text Generation for Code: Compares RNN vs a Markov Chain method to generate text for coding (and also news). Markov Chain methods generate better text, and that could be due to a very small dataset used. Datasets: news.txt, coding.txt
• If-Else (Part 1): Investigates if a neural network can model if-else statements
• Wordle: Solves the Wordle game using 5 different agents: Random (baseline), Minimize Worst Outcome, Maximize Entropy (Information Theory), Minimize Variance, Minimize Variance and Worst Outcome. Currently there are only algorithmic agents. In the future, there will also be Deep Q-Networks and Actor-Critic models.

tensorflow-implementations's Issues

error ValueError: too many values to unpack (expected 3) and wrong output for Step 2: Parse Knowledge Graph

LLM with Knowledge Graphs.ipynb

I get error ValueError: too many values to unpack (expected 3) and wrong output for Step 2: Parse Knowledge Graph
https://github.com/tanchongmin/TensorFlow-Implementations/blob/main/Tutorial/LLM%20with%20Knowledge%20Graphs.ipynb

and

and

some updates in your code done to meet changes in updated packages
def strict_output(system_prompt, user_prompt, output_format, default_category = "", output_value_only = False,
model = 'gpt-3.5-turbo', temperature = 0, num_tries = 2, verbose = False):
''' Ensures that OpenAI will always adhere to the desired output json format.
Uses rule-based iterative feedback to ask GPT to self-correct.
Keeps trying up to num_tries it it does not. Returns empty json if unable to after num_tries iterations.
If output field is a list, will treat as a classification problem and output best classification category.
Text enclosed within < > will generated by GPT accordingly'''

# if the user input is in a list, we also process the output as a list of json
list_input = isinstance(user_prompt, list)
# if the output format contains dynamic elements of < or >, then add to the prompt to handle dynamic elements
dynamic_elements = '<' in str(output_format)
# if the output format contains list elements of [ or ], then we add to the prompt to handle lists
list_output = '[' in str(output_format)

# start off with no error message
error_msg = ''

for i in range(num_tries):
    
    output_format_prompt = f'''\nYou are to output the following in json format: {output_format}. 

Do not put quotation marks or escape character \ in the output fields.'''

    if list_output:
        output_format_prompt += f'''\nIf output field is a list, classify output into the best element of the list.'''
    
    # if output_format contains dynamic elements, process it accordingly
    if dynamic_elements: 
        output_format_prompt += f'''

Any text enclosed by < and > indicates you must generate content to replace it. Example input: Go to , Example output: Go to the garden
Any output key containing < and > indicates you must generate the key name to replace it. Example input: {{'': 'description of location'}}, Example output: {{'school': 'a place for education'}}'''

    # if input is in a list format, ask it to generate json in a list
    if list_input:
        output_format_prompt += '''\nGenerate a list of json, one json for each input element.'''
        
    # Use OpenAI to get a response
    response = openai.chat.completions.create(
      temperature = temperature,
      model=model,
      messages=[
        {"role": "system", "content": system_prompt + output_format_prompt + error_msg},
        {"role": "user", "content": str(user_prompt)}
      ]
    )

    #res = response['choices'][0]['message']['content'].replace('\'', '"')
    res = response.choices[0].message.content.replace('\'', '"')
    # ensure that we don't replace away aprostophes in text 
    res = re.sub(r"(\w)\"(\w)", r"\1'\2", res)

    if verbose:
        print('System prompt:', system_prompt + output_format_prompt + error_msg)
        print('\nUser prompt:', str(user_prompt))
        print('\nGPT response:', res)
    
    # try-catch block to ensure output format is adhered to
    try:
        output = json.loads(res)
        if isinstance(user_prompt, list):
            if not isinstance(output, list): raise Exception("Output format not in a list of json")
        else:
            output = [output]
            
        # check for each element in the output_list, the format is correctly adhered to
        for index in range(len(output)):
            for key in output_format.keys():
                # unable to ensure accuracy of dynamic output header, so skip it
                if '<' in key or '>' in key: continue
                # if output field missing, raise an error
                if key not in output[index]: raise Exception(f"{key} not in json output")
                # check that one of the choices given for the list of words is an unknown
                if isinstance(output_format[key], list):
                    choices = output_format[key]
                    # ensure output is not a list
                    if isinstance(output[index][key], list):
                        output[index][key] = output[index][key][0]
                    # output the default category (if any) if GPT is unable to identify the category
                    if output[index][key] not in choices and default_category:
                        output[index][key] = default_category
                    # if the output is a description format, get only the label
                    if ':' in output[index][key]:
                        output[index][key] = output[index][key].split(':')[0]
                        
            # if we just want the values for the outputs
            if output_value_only:
                output[index] = [value for value in output[index].values()]
                # just output without the list if there is only one element
                if len(output[index]) == 1:
                    output[index] = output[index][0]
                
        return output if list_input else output[0]

    except Exception as e:
        error_msg = f"\n\nResult: {res}\n\nError message: {str(e)}"
        print("An exception occurred:", str(e))
        print("Current invalid json format:", res)
     
return {}