Update app.py

app.py

@@ -10,6 +10,10 @@ from collections import Counter
 from itertools import chain
 from nltk.translate.bleu_score import sentence_bleu, SmoothingFunction
 import math
+import markdown
+from .text import doctree_from_url, get_selectors_for_class, split_by_heading, DocTree
+from .optimization import ngrams, count_ngrams, self_bleu, dist_n, perplexity, js_divergence
+
 
 model_name = 'philipp-zettl/t5-small-long-qa'
 qa_model = AutoModelForSeq2SeqLM.from_pretrained(model_name)
@@ -31,55 +35,6 @@ max_answers = 1
 max_elem_value = 100
 
 
-
-def ngrams(sequence, n):
-    return [tuple(sequence[i:i+n]) for i in range(len(sequence)-n+1)]
-
-def count_ngrams(sequence, max_n):
-    counts = Counter()
-    for n in range(1, max_n + 1):
-        counts.update(ngrams(sequence, n))
-    return counts
-
-def self_bleu(outputs):
-    smoothing_function = SmoothingFunction().method1
-    scores = []
-    for i in range(len(outputs)):
-        references = outputs[:i] + outputs[i+1:]
-        # Avoid calculating BLEU score for empty references
-        if references:
-            scores.append(sentence_bleu(references, outputs[i], smoothing_function=smoothing_function))
-    # If all references are empty, return a default value
-    if not scores:
-        return 0
-    return sum(scores) / len(scores)
-
-def dist_n(outputs, n):
-    all_ngrams = list(chain(*[ngrams(output, n) for output in outputs]))
-    unique_ngrams = set(all_ngrams)
-    return len(unique_ngrams) / len(all_ngrams) if all_ngrams else 0
-
-def perplexity(model, tokenizer, texts):
-    encodings = tokenizer(texts, return_tensors='pt', padding=True, truncation=True)
-    max_length = model.config.n_positions
-    stride = 512
-    lls = []
-    for i in range(0, encodings.input_ids.size(1), stride):
-        begin_loc = max(i + stride - max_length, 0)
-        end_loc = i + stride
-        trg_len = end_loc - i
-        input_ids = encodings.input_ids[:, begin_loc:end_loc].to(model.device)
-        target_ids = input_ids.clone()
-        target_ids[:, :-trg_len] = -100
-
-        with torch.no_grad():
-            outputs = model(input_ids, labels=target_ids)
-            log_likelihood = outputs.loss * trg_len
-        lls.append(log_likelihood)
-
-    ppl = torch.exp(torch.stack(lls).sum() / end_loc)
-    return ppl.item()
-
 def embedding_similarity(inputs, outputs):
     global embedding_model, embedding_tokenizer, device
     def embed(texts):
@@ -94,16 +49,6 @@ def embedding_similarity(inputs, outputs):
     similarities = pairwise_distances(input_embeddings, output_embeddings, metric='cosine')
     return sum(similarities) / len(similarities)
 
-def js_divergence(p, q):
-    def kl_divergence(p, q):
-        return sum(p[i] * math.log(p[i] / q[i]) for i in range(len(p)) if p[i] != 0 and q[i] != 0)
-    
-    p_norm = [float(i)/sum(p) for i in p]
-    q_norm = [float(i)/sum(q) for i in q]
-    
-    m = [(p_norm[i] + q_norm[i]) / 2 for i in range(len(p_norm))]
-    
-    return (kl_divergence(p_norm, m) + kl_divergence(q_norm, m)) / 2
 
 def evaluate_model(num_beams, num_beam_groups, model, tokenizer, eval_data, max_length=85):
     generated_outputs = []
@@ -150,6 +95,7 @@ def evaluate_model(num_beams, num_beam_groups, model, tokenizer, eval_data, max_
         "jsd_score": jsd_score
 }
 
+
 def find_best_parameters(eval_data, model, tokenizer, max_length=85):
 
     # Parameter ranges
@@ -184,8 +130,6 @@ def find_best_parameters(eval_data, model, tokenizer, max_length=85):
     return best_params
 
 
-
-
 def run_model(inputs, tokenizer, model, num_beams=2, num_beam_groups=2, temperature=0.5, num_return_sequences=1, max_length=85, seed=42069):
     all_outputs = []
     torch.manual_seed(seed)
@@ -198,29 +142,13 @@ def run_model(inputs, tokenizer, model, num_beams=2, num_beam_groups=2, temperat
                 sample_output = model.generate(
                     input_ids[:1],
                     max_length=max_length,
-                    #temperature=temperature,
-                    #do_sample=True,
                     num_return_sequences=num_return_sequences,
                     low_memory=True,
-                    #top_p=temperature,
-                    #num_beams=max(2, num_return_sequences),
                     use_cache=True,
-                    # Contrastive search
-                    #penalty_alpha=0.6,
-                    #top_k=4,
-                    # Multi-nomial sampling
-                    #do_sample=True,
-                    #num_beams=1,
-                    # Beam search
-                    #num_beams=5,
-                    # Beam search multinomial sampling
-                    #num_beams=5,
-                    #do_sample=True,
                     # Diverse Beam search decoding
                     num_beams=max(2, num_return_sequences),
                     num_beam_groups=max(2, num_return_sequences),
                     diversity_penalty=temperature,
-                    #do_sample=True,
 
                 )
                 for i, sample_output in enumerate(sample_output):
@@ -311,38 +239,26 @@ def create_file_download(qnas):
     return 'qnas.tsv'
 
 
-with gr.Blocks() as demo:
-    with gr.Tab(label='Description'):
-        with gr.Row(equal_height=True):
-            with gr.Column():
-                gr.Markdown(
-                """
-                # QA-Generator
-                A combination of fine-tuned flan-T5(-small) models chained into sequence
-                to generate:
-                
-                a) a versatile set of questions  
-                b) an accurate set of matching answers
-                
-                according to a given piece of text content.""")
-            with gr.Column():
-                gr.Markdown(
-                """ 
-                The idea is simple:
-                
-                1. Add your content
-                2. Select the amount of questions you want to generate
-                3. (optional) Select the amount of answers you want to generate per goven question
-                4. Press generate
-                5. ???
-                6. Profit
-                """)
-        with gr.Row(equal_height=True):
-            gr.Markdown("""
-            If you're satisfied with the generated data set, you can export it as TSV
-            to edit or import it into your favourite tool.
-            """)
-        with gr.Row(equal_height=True):
+def main():
+    with gr.Tab(label='QA Generator'):
+        with gr.Tab(label='Explanation'):
+            gr.Markdown(
+                '''
+                # QA Generator
+                This tab allows you to generate questions and answers from a given piece of text content.
+
+                ## How to use
+                1. Enter the text content you want to generate questions and answers from.
+                2. Adjust the diversity penalty for question generation and answer generation.
+                3. Set the maximum length of the generated questions and answers.
+                4. Choose the number of questions and answers you want to generate.
+                5. Click on the "Generate" button.
+
+                The next section will give you insights into the generated questions and answers.
+
+                If you're satisfied with the generated questions and answers, you can download them as a TSV file.
+                '''
+            )
             with gr.Accordion(label='Optimization', open=False):
                 gr.Markdown("""
                 For optimization of the question generation we apply the following combined score:
@@ -359,48 +275,156 @@ with gr.Blocks() as demo:
 
                 4. **jsd**: This is the Jensen-Shannon Divergence between the n-gram distributions of the generated outputs and the reference data. <u>**Lower values indicate greater similarity between distributions.**</u>
                 """, latex_delimiters=[{'display': False, 'left': '$$', 'right': '$$'}])
-    with gr.Tab(label='QA Generator'):
-        with gr.Row(equal_height=True):
-            with gr.Group("Content"):
-                content = gr.Textbox(label='Content', lines=15, placeholder='Enter text here', max_lines=10_000)
-            with gr.Group("Settings"):
-                temperature_qg = gr.Slider(label='Diversity Penalty QG', value=0.2, minimum=0, maximum=1, step=0.01)
-                temperature_qa = gr.Slider(label='Diversity Penalty QA', value=0.5, minimum=0, maximum=1, step=0.01)
-                max_length = gr.Number(label='Max Length', value=85, minimum=1, step=1, maximum=512)
-                num_return_sequences_qg = gr.Number(label='Number Questions', value=max_questions, minimum=1, step=1, maximum=max(max_questions, max_elem_value))
-                num_return_sequences_qa = gr.Number(label="Number Answers", value=max_answers, minimum=1, step=1, maximum=max(max_questions, max_elem_value))
-                seed = gr.Number(label="seed", value=42069)
-                optimize_questions = gr.Checkbox(label="Optimize questions?", value=False)
-
-        with gr.Row():
-            gen_btn = gr.Button("Generate")
-
-        @gr.render(
-            inputs=[
-                content, temperature_qg, temperature_qa, num_return_sequences_qg, num_return_sequences_qa,
-                max_length, seed, optimize_questions
-            ],
-            triggers=[gen_btn.click]
-        )
-        def render_results(content, temperature_qg, temperature_qa, num_return_sequences_qg, num_return_sequences_qa, max_length, seed, optimize_questions):
-            if not content.strip():
-                raise gr.Error('Please enter some content to generate questions and answers.')
-            qnas = gen(
-                content, temperature_qg, temperature_qa, num_return_sequences_qg, num_return_sequences_qa,
-                max_length, seed, optimize_questions
+        with gr.Tab(label='Generate QA'):
+            with gr.Row(equal_height=True):
+                with gr.Group("Content"):
+                    content = gr.Textbox(label='Content', lines=15, placeholder='Enter text here', max_lines=10_000)
+                with gr.Group("Settings"):
+                    temperature_qg = gr.Slider(label='Diversity Penalty QG', value=0.2, minimum=0, maximum=1, step=0.01)
+                    temperature_qa = gr.Slider(label='Diversity Penalty QA', value=0.5, minimum=0, maximum=1, step=0.01)
+                    max_length = gr.Number(label='Max Length', value=85, minimum=1, step=1, maximum=512)
+                    num_return_sequences_qg = gr.Number(label='Number Questions', value=max_questions, minimum=1, step=1, maximum=max(max_questions, max_elem_value))
+                    num_return_sequences_qa = gr.Number(label="Number Answers", value=max_answers, minimum=1, step=1, maximum=max(max_questions, max_elem_value))
+                    seed = gr.Number(label="seed", value=42069)
+                    optimize_questions = gr.Checkbox(label="Optimize questions?", value=False)
+
+            with gr.Row():
+                gen_btn = gr.Button("Generate")
+
+            @gr.render(
+                inputs=[
+                    content, temperature_qg, temperature_qa, num_return_sequences_qg, num_return_sequences_qa,
+                    max_length, seed, optimize_questions
+                ],
+                triggers=[gen_btn.click]
             )
-            df = gr.Dataframe(
-                value=[u.values() for u in qnas],
-                headers=['Question', 'Answer'],
-                col_count=2,
-                wrap=True
+            def render_results(content, temperature_qg, temperature_qa, num_return_sequences_qg, num_return_sequences_qa, max_length, seed, optimize_questions):
+                if not content.strip():
+                    raise gr.Error('Please enter some content to generate questions and answers.')
+                qnas = gen(
+                    content, temperature_qg, temperature_qa, num_return_sequences_qg, num_return_sequences_qa,
+                    max_length, seed, optimize_questions
+                )
+                df = gr.Dataframe(
+                    value=[u.values() for u in qnas],
+                    headers=['Question', 'Answer'],
+                    col_count=2,
+                    wrap=True
+                )
+                pd_df = pd.DataFrame([u.values() for u in qnas], columns=['Question', 'Answer'])
+
+                download = gr.DownloadButton(label='Download (without headers)', value=create_file_download(pd_df))
+
+            content.change(lambda x: x.strip(), content)
+
+
+def new_main():
+    with gr.Tab('Content extraction from URL'):
+        with gr.Tab(label='Explanation'):
+            gr.Markdown(
+                '''
+                # Content extraction from URL
+                This tab allows you to extract content from a URL and chunk it into sections.
+
+                ## How to use
+                1. Enter the URL of the webpage you want to extract content from.
+                2. Select the element class and class name of the HTML element you want to extract content from.
+                3. Click on the "Extract content" button.
+
+                The next section will give you insights into the extracted content.
+
+                This was done to give you the possibility to look at the extracted content, as well as manipulate it further.
+
+                Once you extract the content, you can choose the depth level to chunk the content into sections.
+                1. Enter the depth level you want to chunk the content into. **Note: <u>This is based on the HTML structure of the webpage, we're utilizing heading tags for this purpose</u>**
+                2. Click on the "Chunk content" button.
+                '''
+            )
+        with gr.Tab(label='Extract content'):
+            url = gr.Textbox(label='URL', placeholder='Enter URL here', lines=1, max_lines=1)
+            elem_class = gr.Dropdown(label='CSS element class', choices=['div', 'p', 'span'], value='div')
+            class_name = gr.Dropdown(label='CSS class name', choices=[], allow_custom_value=True)
+
+            extract_btn = gr.Button('Extract content')
+
+            with gr.Group():
+                content_state = gr.State(None)
+                final_content = gr.Textbox(value='', show_copy_button=True, label='Final content', interactive=True)
+                with gr.Accordion('Reveal original input', open=False):
+                    og_content = gr.Textbox(value='', label='OG HTML content')
+
+            with gr.Group(visible=False) as step_2_group:
+                depth_level = gr.Number(label='Depth level', value=1, minimum=0, step=1, maximum=6)
+                continue_btn = gr.Button('Chunk content')
+
+            def render_results(url, elem_class_, class_name_):
+                if not url.strip():
+                    raise gr.Error('Please enter a URL to extract content.')
+                content = doctree_from_url(url, elem_class_, class_name_)
+                return [
+                    content,
+                    content.content,
+                    content.as_markdown(content.merge_sections(content.get_sections(0))),
+                    gr.Group(visible=True)
+                ]
+
+            def get_class_options(url, elem_class):
+                if not url.strip():
+                    raise gr.Error('Please enter a URL to extract content.')
+
+                return gr.Dropdown(label='CSS class name', choices=list(set(get_selectors_for_class(url, elem_class))))
+
+            def update_content_state_on_final_change(final_content):
+                html_content = markdown.markdown(final_content)
+                return DocTree(split_by_heading(html_content, 1))
+
+            @gr.render(inputs=[content_state, depth_level], triggers=[continue_btn.click])
+            def select_content(content, depth_level):
+                if not content:
+                    raise gr.Error('Please extract content first.')
+
+                sections = content.get_sections_by_depth(depth_level)
+                print(f'Found {len(sections)} sections')
+                ds = []
+                for idx, section in enumerate(sections):
+                    ds.append([idx, content.as_markdown(content.merge_sections(section))])
+                gr.Dataframe(value=ds, headers=['Section #', 'Content'], interactive=True, wrap=True)
+
+            url.change(
+                get_class_options,
+                inputs=[url, elem_class],
+                outputs=[class_name]
             )
-            pd_df = pd.DataFrame([u.values() for u in qnas], columns=['Question', 'Answer'])
 
-            download = gr.DownloadButton(label='Download (without headers)', value=create_file_download(pd_df))
+            extract_btn.click(
+                render_results,
+                inputs=[
+                    url, elem_class, class_name,
+                ],
+                outputs=[
+                    content_state, og_content, final_content, step_2_group
+                ]
+            )
+            final_content.change(update_content_state_on_final_change, inputs=[final_content], outputs=[content_state])
 
-        content.change(lambda x: x.strip(), content)
+
+with gr.Blocks() as demo:
+    gr.Markdown(
+        '''
+        # QA-Generator
+        A tool to build FAQs or QnAs from a given piece of text content.
+
+        ## How to use
+        We provide you two major functionalities:
+        1. **Content extraction from URL**: Extract content from a URL and chunk it into sections.
+        2. **QA Generator**: Generate questions and answers from a given text content.
+
+        Select the tab you want to use and follow the instructions.
+        '''
+    )
+    new_main()
+    main()
 
 
 demo.queue()
-demo.launch()
+demo.launch()