update

README.md

@@ -1,5 +1,5 @@
 ---
-title: GT4SD - HuggingFace transformers
+title: GT4SD - Patent Generative Transformers
 emoji: 💡
 colorFrom: green
 colorTo: blue

app.py

@@ -2,14 +2,11 @@ import logging
 import pathlib
 import gradio as gr
 import pandas as pd
-from gt4sd.algorithms.generation.hugging_face import (
-    HuggingFaceCTRLGenerator,
-    HuggingFaceGenerationAlgorithm,
-    HuggingFaceGPT2Generator,
-    HuggingFaceTransfoXLGenerator,
-    HuggingFaceOpenAIGPTGenerator,
-    HuggingFaceXLMGenerator,
-    HuggingFaceXLNetGenerator,
+from gt4sd.algorithms.generation.pgt import (
+    PGT,
+    PGTCoherenceChecker,
+    PGTEditor,
+    PGTGenerator,
 )
 from gt4sd.algorithms.registry import ApplicationsRegistry
 
@@ -18,42 +15,36 @@ logger = logging.getLogger(__name__)
 logger.addHandler(logging.NullHandler())
 
 MODEL_FN = {
-    "HuggingFaceCTRLGenerator": HuggingFaceCTRLGenerator,
-    "HuggingFaceGPT2Generator": HuggingFaceGPT2Generator,
-    "HuggingFaceTransfoXLGenerator": HuggingFaceTransfoXLGenerator,
-    "HuggingFaceOpenAIGPTGenerator": HuggingFaceOpenAIGPTGenerator,
-    "HuggingFaceXLMGenerator": HuggingFaceXLMGenerator,
-    "HuggingFaceXLNetGenerator": HuggingFaceXLNetGenerator,
+    "PGTGenerator": PGTGenerator,
+    "PGTEditor": PGTEditor,
+    "PGTCoherenceChecker": PGTCoherenceChecker,
 }
 
 
 def run_inference(
     model_type: str,
+    generator_task: str,
+    editor_task: str,
+    checker_task: str,
     prompt: str,
-    length: float,
-    temperature: float,
-    prefix: str,
-    k: float,
+    second_prompt: str,
+    length: int,
+    k: int,
     p: float,
-    repetition_penalty: float,
 ):
-    model = model_type.split("_")[0]
-    version = model_type.split("_")[1]
-
-    if model not in MODEL_FN.keys():
-        raise ValueError(f"Model type {model} not supported")
-    config = MODEL_FN[model](
-        algorithm_version=version,
-        prompt=prompt,
-        length=length,
-        temperature=temperature,
-        repetition_penalty=repetition_penalty,
-        k=k,
-        p=p,
-        prefix=prefix,
-    )
 
-    model = HuggingFaceGenerationAlgorithm(config)
+    kwargs = {"max_length": length, "top_k": k, "top_p": p}
+
+    if model_type == "PGTGenerator":
+        config = PGTGenerator(task=generator_task, input_text=prompt, **kwargs)
+    elif model_type == "PGTEditor":
+        config = PGTEditor(input_type=editor_task, input_text=prompt, **kwargs)
+    elif model_type == "PGTCoherenceChecker":
+        config = PGTCoherenceChecker(
+            coherence_type=checker_task, input_a=prompt, input_b=second_prompt, **kwargs
+        )
+
+    model = PGT(config)
     text = list(model.sample(1))[0]
 
     return text
@@ -64,8 +55,8 @@ if __name__ == "__main__":
     # Preparation (retrieve all available algorithms)
     all_algos = ApplicationsRegistry.list_available()
     algos = [
-        x["algorithm_application"] + "_" + x["algorithm_version"]
-        for x in list(filter(lambda x: "HuggingFace" in x["algorithm_name"], all_algos))
+        x["algorithm_application"]
+        for x in list(filter(lambda x: "PGT" in x["algorithm_name"], all_algos))
     ]
 
     # Load metadata
@@ -81,30 +72,40 @@ if __name__ == "__main__":
     with open(metadata_root.joinpath("description.md"), "r") as f:
         description = f.read()
 
+    gen_tasks = [
+        "title-to-abstract",
+        "abstract-to-title",
+        "abstract-to-claim",
+        "claim-to-abstract",
+    ]
+
     demo = gr.Interface(
         fn=run_inference,
-        title="HuggingFace language models",
+        title="Patent Generative Transformer",
         inputs=[
+            gr.Dropdown(algos, label="Model type", value="PGTGenerator"),
+            gr.Dropdown(gen_tasks, label="Generator task", value="title-to-abstract"),
+            gr.Dropdown(["abstract", "claim"], label="Editor task", value="abstract"),
             gr.Dropdown(
-                algos,
-                label="Language model",
-                value="HuggingFaceGPT2Generator_gpt2",
+                ["title-abstract", "title-claim", "abstract-claim"],
+                label="Checker task",
+                value="title-abstract",
             ),
             gr.Textbox(
                 label="Text prompt",
-                placeholder="I'm a stochastic parrot.",
-                lines=1,
-            ),
-            gr.Slider(minimum=5, maximum=100, value=20, label="Maximal length", step=1),
-            gr.Slider(
-                minimum=0.6, maximum=1.5, value=1.1, label="Decoding temperature"
+                placeholder="Artificial intelligence and machine learning infrastructure",
+                lines=5,
             ),
             gr.Textbox(
-                label="Prefix", placeholder="Some prefix (before the prompt)", lines=1
+                label="Secondary text prompt (only for coherence checker)",
+                placeholder="",
+                lines=1
+            ),
+            gr.Slider(
+                minimum=5, maximum=1024, value=512, label="Maximal length", step=1
             ),
             gr.Slider(minimum=2, maximum=500, value=50, label="Top-k", step=1),
-            gr.Slider(minimum=0.5, maximum=1, value=1.0, label="Decoding-p", step=1),
-            gr.Slider(minimum=0.5, maximum=5, value=1.0, label="Repetition penalty"),
+            gr.Slider(minimum=0.5, maximum=1, value=1.0, label="Top-p", step=1),
         ],
         outputs=gr.Textbox(label="Output"),
         article=article,

model_cards/examples.csv

@@ -1,2 +1,5 @@
-HuggingFaceGPT2Generator_gpt2, The role of generative models is,20,1.1,,50,1,1
-HuggingFaceOpenAIGPTGenerator_openai-gpt, The best country in the world is,10,0.9,,50,1,1
+PGTGenerator|title-to-abstract|||Artificial intelligence and machine learning infrastructure||512|50|1.0
+PGTGenerator|title-to-abstract|||Artificial intelligence and machine learning infrastructure||756|20|0.95
+PGTEditor||abstract||In one step of a method for infusing an [MASK], the infusion fluid is pumped through a fluid delivery line of an infusion system. In another step, measurements are taken with at least one sensor connected to the infusion system. In an additional step, an air determination is determined with at least one processor. The air determination is related to air in the fluid delivery line. The air determination is based on the measurements taken by the at least one sensor. The air determination is further based on: (1) [MASK] information regarding the infusion of the infusion fluid; or (2) multi-channel filtering of the measurements from the at least one sensor or non-linear mapping of the measurements from the at least one sensor; and statistical process control charts applied to the multi-channel filtered measurements or applied to the non-linear mapped measurements.|512|50|1
+PGTCoherenceChecker|||title-abstract|Artificial intelligence and machine learning infrastructure|An artificial intelligence and machine learning infrastructure system, including: one or more storage systems comprising, respectively, one or more storage devices; and one or more graphical processing units, wherein the graphical processing units are configured to communicate with the one or more storage systems over a communication fabric; where the one or more storage systems, the one or more graphical processing units, and the communication fabric are implemented within a single chassis.|512|50|1
+PGTCoherenceChecker|||title-abstract|Analog image processing|An artificial intelligence and machine learning infrastructure system for image classification, including: one or more storage systems comprising, respectively, one or more storage devices; and one or more graphical processing units, wherein the graphical processing units are configured to communicate with the one or more storage systems over a communication fabric; where the one or more storage systems, the one or more graphical processing units, and the communication fabric are implemented within a single chassis.|512|50|1