Add app

.gitignore

@@ -0,0 +1,4 @@
+venv
+.idea
+__pycache__
+*~

.streamlit/config.toml

@@ -0,0 +1,8 @@
+[server]
+headless = true
+
+[theme]
+base="dark"
+primaryColor="#139ace"
+secondaryBackgroundColor="#2b2b39"
+textColor="#cdd8d3"

README.md

@@ -1,11 +1,12 @@
 ---
-title: Netherator
-emoji: 💩
-colorFrom: yellow
-colorTo: yellow
+title: Netherator - teller of tales from the Netherlands
+emoji: 🧙
+colorFrom: gray
+colorTo: indigo
 sdk: streamlit
 app_file: app.py
-pinned: false
+pinned: true
+sdk_version: 1.0.0
 ---
 
 # Configuration

app.py

@@ -0,0 +1,245 @@
+import json
+import os
+import pprint
+import time
+from random import randint
+
+import psutil
+import streamlit as st
+import torch
+from transformers import (AutoModelForCausalLM, AutoTokenizer, pipeline,
+                          set_seed)
+
+device = torch.cuda.device_count() - 1
+
+
+@st.cache(suppress_st_warning=True, allow_output_mutation=True)
+def load_model(model_name):
+    os.environ["TOKENIZERS_PARALLELISM"] = "false"
+    try:
+        if not os.path.exists(".streamlit/secrets.toml"):
+            raise FileNotFoundError
+        access_token = st.secrets.get("netherator")
+    except FileNotFoundError:
+        access_token = os.environ.get("HF_ACCESS_TOKEN", None)
+    tokenizer = AutoTokenizer.from_pretrained(model_name, use_auth_token=access_token)
+    model = AutoModelForCausalLM.from_pretrained(
+        model_name, use_auth_token=access_token
+    )
+    if device != -1:
+        model.to(f"cuda:{device}")
+    return tokenizer, model
+
+
+class StoryGenerator:
+    def __init__(self, model_name):
+        self.model_name = model_name
+        self.tokenizer = None
+        self.model = None
+        self.generator = None
+        self.model_loaded = False
+
+    def load(self):
+        if not self.model_loaded:
+            self.tokenizer, self.model = load_model(self.model_name)
+            self.generator = pipeline(
+                "text-generation",
+                model=self.model,
+                tokenizer=self.tokenizer,
+                device=device,
+            )
+            self.model_loaded = True
+
+    def get_text(self, text: str, **generate_kwargs) -> str:
+        return self.generator(text, **generate_kwargs)
+
+
+STORY_GENERATORS = [
+    {
+        "model_name": "yhavinga/gpt-neo-125M-dutch-nedd",
+        "desc": "Dutch GPTNeo Small",
+        "story_generator": None,
+    },
+    {
+        "model_name": "yhavinga/gpt2-medium-dutch-nedd",
+        "desc": "Dutch GPT2 Medium",
+        "story_generator": None,
+    },
+    # {
+    #     "model_name": "yhavinga/gpt-neo-125M-dutch",
+    #     "desc": "Dutch GPTNeo Small",
+    #     "story_generator": None,
+    # },
+    # {
+    #     "model_name": "yhavinga/gpt2-medium-dutch",
+    #     "desc": "Dutch GPT2 Medium",
+    #     "story_generator": None,
+    # },
+]
+
+
+def instantiate_models():
+    for sg in STORY_GENERATORS:
+        sg["story_generator"] = StoryGenerator(sg["model_name"])
+        with st.spinner(text=f"Loading the model {sg['desc']} ..."):
+            sg["story_generator"].load()
+
+
+def set_new_seed():
+    seed = randint(0, 2 ** 32 - 1)
+    set_seed(seed)
+    return seed
+
+
+def main():
+    st.set_page_config(  # Alternate names: setup_page, page, layout
+        page_title="Netherator",  # String or None. Strings get appended with "• Streamlit".
+        layout="wide",  # Can be "centered" or "wide". In the future also "dashboard", etc.
+        initial_sidebar_state="expanded",  # Can be "auto", "expanded", "collapsed"
+        page_icon="📚",  # String, anything supported by st.image, or None.
+    )
+    instantiate_models()
+
+    with open("style.css") as f:
+        st.markdown(f"<style>{f.read()}</style>", unsafe_allow_html=True)
+
+    st.sidebar.image("demon-reading-Stewart-Orr.png", width=200)
+
+    st.sidebar.markdown(
+        """# Netherator
+Teller of tales from the Netherlands"""
+    )
+
+    model_desc = st.sidebar.selectbox(
+        "Model", [sg["desc"] for sg in STORY_GENERATORS], index=1
+    )
+
+    st.sidebar.title("Parameters:")
+
+    if "prompt_box" not in st.session_state:
+        st.session_state["prompt_box"] = "Het was een koude winterdag"
+
+    st.session_state["text"] = st.text_area("Enter text", st.session_state.prompt_box)
+
+    # min_length = st.sidebar.number_input(
+    #     "Min length", min_value=10, max_value=150, value=75
+    # )
+    max_length = st.sidebar.number_input(
+        "Lengte van de tekst",
+        value=300,
+        max_value=512,
+    )
+    no_repeat_ngram_size = st.sidebar.number_input(
+        "No-repeat NGram size", min_value=1, max_value=5, value=3
+    )
+    repetition_penalty = st.sidebar.number_input(
+        "Repetition penalty", min_value=0.0, max_value=5.0, value=1.2, step=0.1
+    )
+    num_return_sequences = st.sidebar.number_input(
+        "Num return sequences", min_value=1, max_value=5, value=1
+    )
+
+    if sampling_mode := st.sidebar.selectbox(
+        "select a Mode", index=0, options=["Top-k Sampling", "Beam Search"]
+    ):
+        if sampling_mode == "Beam Search":
+            num_beams = st.sidebar.number_input(
+                "Num beams", min_value=1, max_value=10, value=4
+            )
+            length_penalty = st.sidebar.number_input(
+                "Length penalty", min_value=0.0, max_value=5.0, value=1.5, step=0.1
+            )
+            params = {
+                "max_length": max_length,
+                "no_repeat_ngram_size": no_repeat_ngram_size,
+                "repetition_penalty": repetition_penalty,
+                "num_return_sequences": num_return_sequences,
+                "num_beams": num_beams,
+                "early_stopping": True,
+                "length_penalty": length_penalty,
+            }
+        else:
+            top_k = st.sidebar.number_input(
+                "Top K", min_value=0, max_value=100, value=50
+            )
+            top_p = st.sidebar.number_input(
+                "Top P", min_value=0.0, max_value=1.0, value=0.95, step=0.05
+            )
+            temperature = st.sidebar.number_input(
+                "Temperature", min_value=0.05, max_value=1.0, value=0.8, step=0.05
+            )
+            params = {
+                "max_length": max_length,
+                "no_repeat_ngram_size": no_repeat_ngram_size,
+                "repetition_penalty": repetition_penalty,
+                "num_return_sequences": num_return_sequences,
+                "do_sample": True,
+                "top_k": top_k,
+                "top_p": top_p,
+                "temperature": temperature,
+            }
+
+    st.sidebar.markdown(
+        """For an explanation of the parameters, head over to the [Huggingface blog post about text generation](https://huggingface.co/blog/how-to-generate)
+and the [Huggingface text generation interface doc](https://huggingface.co/transformers/main_classes/model.html?highlight=generate#transformers.generation_utils.GenerationMixin.generate).
+"""
+    )
+
+    if st.button("Run"):
+        estimate = max_length / 18
+        if device == -1:
+            ## cpu
+            estimate = estimate * (1 + 0.7 * (num_return_sequences - 1))
+            if sampling_mode == "Beam Search":
+                estimate = estimate * (1.1 + 0.3 * (num_beams - 1))
+        else:
+            ## gpu
+            estimate = estimate * (1 + 0.1 * (num_return_sequences - 1))
+            estimate = 0.5 + estimate / 5
+            if sampling_mode == "Beam Search":
+                estimate = estimate * (1.0 + 0.1 * (num_beams - 1))
+        estimate = int(estimate)
+
+        with st.spinner(
+            text=f"Please wait ~ {estimate} second{'s' if estimate != 1 else ''} while getting results ..."
+        ):
+            memory = psutil.virtual_memory()
+            story_generator = next(
+                (
+                    x["story_generator"]
+                    for x in STORY_GENERATORS
+                    if x["desc"] == model_desc
+                ),
+                None,
+            )
+            seed = set_new_seed()
+            time_start = time.time()
+            result = story_generator.get_text(text=st.session_state.text, **params)
+            time_end = time.time()
+            time_diff = time_end - time_start
+
+            st.subheader("Result")
+            for text in result:
+                st.write(text.get("generated_text").replace("\n", "  \n"))
+
+            # st.text("*Translation*")
+            # translation = translate(result, "en", "nl")
+            # st.write(translation.replace("\n", "  \n"))
+            #
+            info = f"""
+            ---
+            *Memory: {memory.total / (1024 * 1024 * 1024):.2f}GB, used: {memory.percent}%, available: {memory.available / (1024 * 1024 * 1024):.2f}GB*        
+            *Text generated using seed {seed} in {time_diff:.5} seconds*
+            """
+            st.write(info)
+
+            params["seed"] = seed
+            params["prompt"] = st.session_state.text
+            params["model"] = story_generator.model_name
+            params_text = json.dumps(params)
+            print(params_text)
+            st.json(params_text)
+
+
+if __name__ == "__main__":
+    main()

requirements.txt

@@ -0,0 +1,7 @@
+-f https://download.pytorch.org/whl/torch_stable.html
+streamlit==1.4.0
+torch==1.6.0+cpu
+torchvision==0.7.0+cpu
+transformers>=4.13.0
+mtranslate
+psutil

style.css

@@ -0,0 +1,42 @@
+body {
+    background-color: #eee;
+}
+/*.fullScreenFrame > div {*/
+/*    display: flex;*/
+/*    justify-content: center;*/
+/*}*/
+/*.stButton>button {*/
+/*    color: #4F8BF9;*/
+/*    border-radius: 50%;*/
+/*    height: 3em;*/
+/*    width: 3em;*/
+/*}*/
+
+.stTextInput>div>div>input {
+    color: #4F8BF9;
+}
+.stTextArea>div>div>input {
+    color: #4F8BF9;
+    min-height: 500px;
+}
+
+
+/*.st-cj {*/
+/*    min-height: 500px;*/
+/*    spellcheck="false";*/
+/*    color: #4F8BF9;*/
+/*}*/
+/*.st-ch {*/
+/*    min-height: 500px;*/
+/*    spellcheck="false";*/
+/*    color: #4F8BF9;*/
+/*}*/
+/*.st-bb {*/
+/*    min-height: 500px;*/
+/*    spellcheck="false";*/
+/*    color: #4F8BF9;*/
+/*}*/
+
+/*body {*/
+/*    background-color: #f1fbff*/
+/*}*/