Duplicate from abxhr/design-project

Co-authored-by: Abshar Mohammed Aslam <abxhr@users.noreply.huggingface.co>

README.md

@@ -0,0 +1,10 @@
+---
+title: Arabic NLP Demo
+emoji: ⌨
+colorFrom: purple
+colorTo: green
+sdk: streamlit
+app_file: app.py
+pinned: true
+duplicated_from: abxhr/design-project
+---

app.py

@@ -0,0 +1,38 @@
+import awesome_streamlit as ast
+import streamlit as st
+
+from backend.utils import get_current_ram_usage, ga
+
+import backend.aragpt
+import backend.home
+import backend.processor
+import backend.sa
+import backend.qa
+
+st.set_page_config(
+    page_title="TEST", page_icon="📖", initial_sidebar_state="expanded", layout="wide"
+)
+
+ga(st.__file__)
+
+PAGES = {
+    "Home": backend.home,
+    "Arabic Sentiment Analysis": backend.sa,
+    "Arabic Question Answering": backend.qa,
+}
+
+
+st.sidebar.title("Navigation")
+selection = st.sidebar.radio("Pages", list(PAGES.keys()))
+
+page = PAGES[selection]
+# with st.spinner(f"Loading {selection} ..."):
+ast.shared.components.write_page(page)
+
+st.sidebar.header("Info")
+st.sidebar.write("Arabic NLP by [**Abshar Mohammed Aslam** (*2019A7PS0233U*)](https://github.com/abxhr)")
+
+st.sidebar.write("Submitted to *Dr. Sujala D. Shetty*")
+# if st.sidebar.checkbox("Show RAM usage"):
+#     ram = get_current_ram_usage()
+#     st.sidebar.write("Ram usage: {:.2f}/{:.2f} GB".format(ram[0], ram[1]))

backend/aragpt.py

@@ -0,0 +1,189 @@
+import streamlit as st
+from .services import TextGeneration
+from tokenizers import Tokenizer
+from functools import lru_cache
+
+# @st.cache(allow_output_mutation=False, hash_funcs={Tokenizer: str})
+@lru_cache(maxsize=1)
+def load_text_generator():
+    generator = TextGeneration()
+    generator.load()
+    return generator
+
+
+generator = load_text_generator()
+
+qa_prompt = """
+            أجب عن السؤال التالي:
+            """
+qa_prompt_post = """ الجواب هو  """
+qa_prompt_post_year = """ في سنة: """
+
+
+def write():
+    st.markdown(
+        """
+        <h1 style="text-align:left;">Arabic Language Generation</h1>
+        """,
+        unsafe_allow_html=True,
+    )
+
+    # Sidebar
+
+    # Taken from https://huggingface.co/spaces/flax-community/spanish-gpt2/blob/main/app.py
+    st.sidebar.subheader("Configurable parameters")
+
+    model_name = st.sidebar.selectbox(
+        "Model Selector",
+        options=[
+            "AraGPT2-Base",
+            # "AraGPT2-Medium",
+            # "Aragpt2-Large",
+            "AraGPT2-Mega",
+        ],
+        index=0,
+    )
+
+    max_new_tokens = st.sidebar.number_input(
+        "Maximum length",
+        min_value=0,
+        max_value=1024,
+        value=100,
+        help="The maximum length of the sequence to be generated.",
+    )
+    temp = st.sidebar.slider(
+        "Temperature",
+        value=1.0,
+        min_value=0.1,
+        max_value=100.0,
+        help="The value used to module the next token probabilities.",
+    )
+    top_k = st.sidebar.number_input(
+        "Top k",
+        value=10,
+        help="The number of highest probability vocabulary tokens to keep for top-k-filtering.",
+    )
+    top_p = st.sidebar.number_input(
+        "Top p",
+        value=0.95,
+        help=" If set to float < 1, only the most probable tokens with probabilities that add up to top_p or higher are kept for generation.",
+    )
+    do_sample = st.sidebar.selectbox(
+        "Sampling?",
+        (True, False),
+        help="Whether or not to use sampling; use greedy decoding otherwise.",
+    )
+    num_beams = st.sidebar.number_input(
+        "Number of beams",
+        min_value=1,
+        max_value=10,
+        value=3,
+        help="The number of beams to use for beam search.",
+    )
+    repetition_penalty = st.sidebar.number_input(
+        "Repetition Penalty",
+        min_value=0.0,
+        value=3.0,
+        step=0.1,
+        help="The parameter for repetition penalty. 1.0 means no penalty",
+    )
+    no_repeat_ngram_size = st.sidebar.number_input(
+        "No Repeat N-Gram Size",
+        min_value=0,
+        value=3,
+        help="If set to int > 0, all ngrams of that size can only occur once.",
+    )
+
+    st.write("#")
+
+    col = st.columns(2)
+
+    col[0].image("images/AraGPT2.png", width=200)
+
+    st.markdown(
+        """
+
+        <h3 style="text-align:left;">AraGPT2 is GPT2 model trained from scratch on 77GB of Arabic text.</h3>
+        <h4 style="text-align:left;"> More details in our <a href="https://github.com/aub-mind/arabert/tree/master/aragpt2">repo</a>.</h4>
+
+        <p style="text-align:left;"><p>
+        <p style="text-align:left;">Use the generation paramters on the sidebar to adjust generation quality.</p>
+        <p style="text-align:right;"><p>
+        """,
+        unsafe_allow_html=True,
+    )
+
+    # col[0].write(
+    #     "AraGPT2 is trained from screatch on 77GB of Arabic text. More details in our [repo](https://github.com/aub-mind/arabert/tree/master/aragpt2)."
+    # )
+    # st.write("## Generate Arabic Text")
+
+    st.markdown(
+        """
+        <style>
+        p, div, input, label, textarea{
+        text-align: right;
+        }
+        </style>
+        """,
+        unsafe_allow_html=True,
+    )
+
+    prompt = st.text_area(
+        "Prompt",
+        "يحكى أن مزارعا مخادعا قام ببيع بئر الماء الموجود في أرضه لجاره مقابل مبلغ كبير من المال",
+    )
+    if st.button("Generate"):
+        with st.spinner("Generating..."):
+            generated_text = generator.generate(
+                prompt=prompt,
+                model_name=model_name,
+                max_new_tokens=max_new_tokens,
+                temperature=temp,
+                top_k=top_k,
+                top_p=top_p,
+                repetition_penalty=repetition_penalty,
+                do_sample=do_sample,
+                num_beams=num_beams,
+                no_repeat_ngram_size=no_repeat_ngram_size,
+            )
+            st.write(generated_text)
+
+    st.markdown("---")
+    st.subheader("")
+    st.markdown(
+        """
+        <p style="text-align:left;"><p>
+        <h2 style="text-align:left;">Zero-Shot Question Answering</h2>
+
+        <p style="text-align:left;">Adjust the maximum length to closely match the expected output length. Setting the Sampling paramter to False is recommended</p>
+        <p style="text-align:left;"><p>
+        """,
+        unsafe_allow_html=True,
+    )
+
+    question = st.text_input(
+        "Question", "من كان رئيس ألمانيا النازية في الحرب العالمية الثانية ؟"
+    )
+    is_date = st.checkbox("Help the model: Is the answer a date?")
+    if st.button("Answer"):
+
+        prompt2 = qa_prompt + question + qa_prompt_post
+        if is_date:
+            prompt2 += qa_prompt_post_year
+        else:
+            prompt2 += " : "
+        with st.spinner("Thinking..."):
+            answer = generator.generate(
+                prompt=prompt2,
+                model_name=model_name,
+                max_new_tokens=max_new_tokens,
+                temperature=temp,
+                top_k=top_k,
+                top_p=top_p,
+                repetition_penalty=repetition_penalty,
+                do_sample=do_sample,
+                num_beams=num_beams,
+                no_repeat_ngram_size=no_repeat_ngram_size,
+            )
+            st.write(answer)

backend/home.py

@@ -0,0 +1,20 @@
+import streamlit as st
+import awesome_streamlit as ast
+
+
+def write():
+    st.markdown(
+        """
+    # Arabic Natural Language Processing
+
+    Design project for **Arabic Natural Language Processing**, by [**Abshar Mohammed Aslam**](https://github.com/abxhr).
+    """
+    )
+    st.markdown("#")
+
+
+    st.markdown(
+        """
+
+        """
+    )

backend/modeling_gpt2.py

@@ -0,0 +1,1599 @@
+# coding=utf-8
+# Copyright 2018 The OpenAI Team Authors and HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+PyTorch OpenAI GPT-2 model.
+Adapted from https://github.com/huggingface/transformers/blob/v4.0.1/src/transformers/models/gpt2/modeling_gpt2.py
+and https://github.com/ghosthamlet/gpt2-ml-torch/blob/master/gpt2_ml_torch/modeling_gpt2.py
+"""
+
+
+import logging
+import os
+from dataclasses import dataclass
+from typing import List, Optional, Tuple
+
+import torch
+import torch.nn as nn
+from torch.nn import CrossEntropyLoss, MSELoss
+from transformers import CONFIG_NAME, WEIGHTS_NAME, GPT2Config, GPT2Model
+from transformers.activations import ACT2FN
+from transformers.file_utils import (
+    ModelOutput,
+    add_code_sample_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    replace_return_docstrings,
+)
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPastAndCrossAttentions,
+    CausalLMOutputWithCrossAttentions,
+    SequenceClassifierOutputWithPast,
+    TokenClassifierOutput,
+)
+from transformers.modeling_utils import (
+    Conv1D,
+    PreTrainedModel,
+    SequenceSummary,
+    find_pruneable_heads_and_indices,
+    prune_conv1d_layer,
+)
+from transformers.utils.model_parallel_utils import assert_device_map, get_device_map
+
+# THe Difference from Transformers is code under _USE_GROVER
+_USE_GROVER = True
+
+logger = logging.getLogger(__name__)
+
+_CONFIG_FOR_DOC = "GPT2Config"
+_TOKENIZER_FOR_DOC = "GPT2Tokenizer"
+
+GPT2_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "gpt2",
+    "gpt2-medium",
+    "gpt2-large",
+    "gpt2-xl",
+    "distilgpt2",
+    # See all GPT-2 models at https://huggingface.co/models?filter=gpt2
+]
+
+logger.setLevel(logging.INFO)
+console = logging.StreamHandler()
+console.setLevel(logging.INFO)
+logger.addHandler(console)
+
+_GPT2_ML_TF_TO_TORCH = {
+    "LayerNorm_embed_norm": "emb_norm",
+    "pos_embed": "wpe.weight",
+    "word_embed": "wte.weight",
+    "layer": "h",
+    # Most importently This two layer norm must be put on the same position as gpt2-ml
+    # or generated data is bad, just repeat the last token
+    "LayerNorm_mlp_ln0": "ln_1",
+    "LayerNorm_mlp_ln1": "ln_2",
+    "intermediate": "mlp.c_fc",
+    "output": "mlp.c_proj",
+    "query_layer": "attn.c_attn",
+    "key_layer": "attn.c_attn",
+    "value_layer": "attn.c_attn",
+    "context_projection_layer": "attn.c_proj",
+    "gamma": "weight",
+    "kernel": "weight",
+    "beta": "bias",
+    "bias": "bias",
+}
+
+
+def convert_gpt2_checkpoint_to_pytorch(
+    gpt2_checkpoint_path, gpt2_config_file, pytorch_dump_folder_path
+):
+    # Construct model
+    if gpt2_config_file == "":
+        config = GPT2Config()
+    else:
+        config = GPT2Config.from_json_file(gpt2_config_file)
+    model = GPT2Model(config)
+
+    # Load weights from numpy
+    load_tf_weights_in_gpt2(model, config, gpt2_checkpoint_path)
+
+    # Save pytorch-model
+    pytorch_weights_dump_path = pytorch_dump_folder_path + "/" + WEIGHTS_NAME
+    pytorch_config_dump_path = pytorch_dump_folder_path + "/" + CONFIG_NAME
+    print("Save PyTorch model to {}".format(pytorch_weights_dump_path))
+    torch.save(model.state_dict(), pytorch_weights_dump_path)
+    print("Save configuration file to {}".format(pytorch_config_dump_path))
+    with open(pytorch_config_dump_path, "w", encoding="utf-8") as f:
+        f.write(config.to_json_string())
+
+
+# XXX: MUST do like: convert_gpt2_checkpoint_to_pytorch('./model.ckpt-100000', './mega.json', './')
+#      https://github.com/tensorflow/models/issues/2675#issuecomment-516595597
+def load_tf_weights_in_gpt2(model, config, gpt2_checkpoint_path):
+    """Load tf checkpoints in a pytorch model"""
+    try:
+        import re
+
+        import tensorflow as tf
+    except ImportError:
+        logger.error(
+            "Loading a TensorFlow model in PyTorch, requires TensorFlow to be installed. Please see "
+            "https://www.tensorflow.org/install/ for installation instructions."
+        )
+        raise
+    tf_path = os.path.abspath(gpt2_checkpoint_path)
+    logger.info("Converting TensorFlow checkpoint from {}".format(tf_path))
+    # Load weights from TF model
+    init_vars = tf.train.list_variables(tf_path)
+    names = []
+    arrays = []
+    for name, shape in init_vars:
+        logger.info("Loading TF weight {} with shape {}".format(name, shape))
+        array = tf.train.load_variable(tf_path, name)
+        names.append(name)
+        arrays.append(array.squeeze())
+
+    import copy
+
+    orig_model = copy.deepcopy(model)
+
+    for name, array in zip(names, arrays):
+        name = name[6:]  # skip "model/"
+        name = name.split("/")
+        pointer = model
+
+        attn_layer = ""
+        for m_name in name:
+            if re.fullmatch(r"[A-Za-z]+\d+", m_name):
+                scope_names = re.split(r"(\d+)", m_name)
+            else:
+                scope_names = [m_name]
+            sname = scope_names[0]
+
+            if sname == "" or sname == "embeddings":
+                continue
+            elif sname not in _GPT2_ML_TF_TO_TORCH:
+                print("=========================================================")
+                logger.info("Skip var name {}".format(scope_names))
+                pointer = None
+                break
+            else:
+                tname = _GPT2_ML_TF_TO_TORCH[sname]
+                if "." in tname:
+                    parent, child = tname.split(".")
+                    pointer = getattr(pointer, parent)
+                    pointer = getattr(pointer, child)
+                else:
+                    pointer = getattr(pointer, tname)
+
+                if tname == "attn.c_attn":
+                    attn_layer = sname
+
+            if len(scope_names) >= 2:
+                num = int(scope_names[1])
+                pointer = pointer[num]
+
+        if pointer is None:
+            continue
+        if attn_layer == "":
+            try:
+                assert pointer.shape == array.shape
+            except AssertionError as e:
+                e.args += (pointer.shape, array.shape)
+                raise
+        logger.info(
+            "Initialize PyTorch weight {}, {}, {}".format(
+                name, array.mean(), pointer.mean()
+            )
+        )
+        if attn_layer == "":
+            pointer.data = torch.from_numpy(array)
+        else:
+            shape = pointer.shape
+            d = torch.from_numpy(array)
+            is_bias = len(shape) == 1
+            end = int(shape[0 if is_bias else 1] / 3)
+            m = dict(
+                query_layer=0,
+                key_layer=end,
+                value_layer=end * 2,
+            )
+            start = m[attn_layer]
+            end = start + end
+            if is_bias:
+                pointer.data[start:end] = d
+            else:
+                pointer.data[:, start:end] = d
+        logger.info(
+            "Initialize PyTorch weight {}, {}, {}".format(
+                name, array.mean(), pointer.mean()
+            )
+        )
+
+    for name, params in orig_model.named_parameters():
+        for n, p in model.named_parameters():
+            if name == n:
+                if params.equal(p):
+                    print("--------------------------")
+                    print(" %s not changed!" % n)
+    return model
+
+
+class Attention(nn.Module):
+    def __init__(self, nx, n_ctx, config, scale=False, is_cross_attention=False):
+        super().__init__()
+
+        n_state = nx  # in Attention: n_state=768 (nx=n_embd)
+        # [switch nx => n_state from Block to Attention to keep identical to TF implem]
+        assert n_state % config.n_head == 0
+        self.register_buffer(
+            "bias",
+            torch.tril(torch.ones((n_ctx, n_ctx), dtype=torch.uint8)).view(
+                1, 1, n_ctx, n_ctx
+            ),
+        )
+        self.register_buffer("masked_bias", torch.tensor(-1e4))
+        self.n_head = config.n_head
+        self.split_size = n_state
+        self.scale = scale
+        self.is_cross_attention = is_cross_attention
+        if self.is_cross_attention:
+            self.c_attn = Conv1D(2 * n_state, nx)
+            self.q_attn = Conv1D(n_state, nx)
+        else:
+            self.c_attn = Conv1D(3 * n_state, nx)
+        self.c_proj = Conv1D(n_state, nx)
+        self.attn_dropout = nn.Dropout(config.attn_pdrop)
+        self.resid_dropout = nn.Dropout(config.resid_pdrop)
+        self.pruned_heads = set()
+
+    def prune_heads(self, heads):
+        if len(heads) == 0:
+            return
+        heads, index = find_pruneable_heads_and_indices(
+            heads, self.n_head, self.split_size // self.n_head, self.pruned_heads
+        )
+        index_attn = torch.cat(
+            [index, index + self.split_size, index + (2 * self.split_size)]
+        )
+
+        # Prune conv1d layers
+        self.c_attn = prune_conv1d_layer(self.c_attn, index_attn, dim=1)
+        self.c_proj = prune_conv1d_layer(self.c_proj, index, dim=0)
+
+        # Update hyper params
+        self.split_size = (self.split_size // self.n_head) * (self.n_head - len(heads))
+        self.n_head = self.n_head - len(heads)
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    def _attn(
+        self, q, k, v, attention_mask=None, head_mask=None, output_attentions=False
+    ):
+        w = torch.matmul(q, k)
+        if self.scale:
+            w = w / (float(v.size(-1)) ** 0.5)
+        nd, ns = w.size(-2), w.size(-1)
+
+        if not self.is_cross_attention:
+            # if only "normal" attention layer implements causal mask
+            mask = self.bias[:, :, ns - nd : ns, :ns]
+            w = torch.where(mask.bool(), w, self.masked_bias.to(w.dtype))
+
+        if attention_mask is not None:
+            # Apply the attention mask
+            w = w + attention_mask
+
+        w = nn.Softmax(dim=-1)(w)
+        w = self.attn_dropout(w)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            w = w * head_mask
+
+        outputs = [torch.matmul(w, v)]
+        if output_attentions:
+            outputs.append(w)
+        return outputs
+
+    def merge_heads(self, x):
+        x = x.permute(0, 2, 1, 3).contiguous()
+        new_x_shape = x.size()[:-2] + (x.size(-2) * x.size(-1),)
+        return x.view(*new_x_shape)  # in Tensorflow implem: fct merge_states
+
+    def split_heads(self, x, k=False):
+        new_x_shape = x.size()[:-1] + (self.n_head, x.size(-1) // self.n_head)
+        x = x.view(*new_x_shape)  # in Tensorflow implem: fct split_states
+        if k:
+            return x.permute(0, 2, 3, 1)  # (batch, head, head_features, seq_length)
+        else:
+            return x.permute(0, 2, 1, 3)  # (batch, head, seq_length, head_features)
+
+    def forward(
+        self,
+        hidden_states,
+        layer_past=None,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        use_cache=False,
+        output_attentions=False,
+    ):
+        if encoder_hidden_states is not None:
+            assert hasattr(
+                self, "q_attn"
+            ), "If class is used as cross attention, the weights `q_attn` have to be defined. Please make sure to instantiate class with `Attention(..., is_cross_attention=True)`."
+            query = self.q_attn(hidden_states)
+            key, value = self.c_attn(encoder_hidden_states).split(
+                self.split_size, dim=2
+            )
+            attention_mask = encoder_attention_mask
+        else:
+            query, key, value = self.c_attn(hidden_states).split(self.split_size, dim=2)
+
+        query = self.split_heads(query)
+        key = self.split_heads(key, k=True)
+        value = self.split_heads(value)
+        if layer_past is not None:
+            past_key, past_value = (
+                layer_past[0].transpose(-2, -1),
+                layer_past[1],
+            )  # transpose back cf below
+            key = torch.cat((past_key, key), dim=-1)
+            value = torch.cat((past_value, value), dim=-2)
+
+        if use_cache is True:
+            present = torch.stack(
+                (key.transpose(-2, -1), value)
+            )  # transpose to have same shapes for stacking
+        else:
+            present = (None,)
+
+        attn_outputs = self._attn(
+            query, key, value, attention_mask, head_mask, output_attentions
+        )
+        a = attn_outputs[0]
+
+        a = self.merge_heads(a)
+        a = self.c_proj(a)
+        a = self.resid_dropout(a)
+
+        outputs = [a, present] + attn_outputs[1:]
+        return outputs  # a, present, (attentions)
+
+
+class MLP(nn.Module):
+    def __init__(self, n_state, config):  # in MLP: n_state=3072 (4 * n_embd)
+        super().__init__()
+        nx = config.n_embd
+        self.c_fc = Conv1D(n_state, nx)
+        self.c_proj = Conv1D(nx, n_state)
+        self.act = ACT2FN[config.activation_function]
+        self.dropout = nn.Dropout(config.resid_pdrop)
+
+    def forward(self, x):
+        h = self.act(self.c_fc(x))
+        h2 = self.c_proj(h)
+        return self.dropout(h2)
+
+
+class Block(nn.Module):
+    def __init__(self, n_ctx, config, scale=False):
+        super().__init__()
+        hidden_size = config.n_embd
+        inner_dim = config.n_inner if config.n_inner is not None else 4 * hidden_size
+        self.ln_1 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
+        self.attn = Attention(hidden_size, n_ctx, config, scale)
+        self.ln_2 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
+        if config.add_cross_attention:
+            self.crossattention = Attention(
+                hidden_size, n_ctx, config, scale, is_cross_attention=True
+            )
+            self.ln_cross_attn = nn.LayerNorm(
+                hidden_size, eps=config.layer_norm_epsilon
+            )
+        self.mlp = MLP(inner_dim, config)
+
+    def forward(
+        self,
+        hidden_states,
+        layer_past=None,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        use_cache=False,
+        output_attentions=False,
+    ):
+        attn_outputs = self.attn(
+            hidden_states,
+            layer_past=layer_past,
+            attention_mask=attention_mask,
+            head_mask=head_mask,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+        )
+        attn_output = attn_outputs[0]  # output_attn: a, present, (attentions)
+        outputs = attn_outputs[1:]
+        # residual connection
+        hidden_states = attn_output + hidden_states
+
+        if encoder_hidden_states is not None:
+            # add one self-attention block for cross-attention
+            assert hasattr(
+                self, "crossattention"
+            ), f"If `encoder_hidden_states` are passed, {self} has to be instantiated with cross-attention layers by setting `config.add_cross_attention=True`"
+            cross_attn_outputs = self.crossattention(
+                self.ln_cross_attn(hidden_states),
+                attention_mask=attention_mask,
+                head_mask=head_mask,
+                encoder_hidden_states=encoder_hidden_states,
+                encoder_attention_mask=encoder_attention_mask,
+                output_attentions=output_attentions,
+            )
+            attn_output = cross_attn_outputs[0]
+            # residual connection
+            hidden_states = hidden_states + attn_output
+            outputs = (
+                outputs + cross_attn_outputs[2:]
+            )  # add cross attentions if we output attention weights
+
+        feed_forward_hidden_states = self.mlp(self.ln_1(hidden_states))
+        # residual connection
+        hidden_states = hidden_states + feed_forward_hidden_states
+
+        hidden_states = self.ln_2(hidden_states)
+
+        outputs = [hidden_states] + outputs
+        return outputs  # hidden_states, present, (attentions, cross_attentions)
+
+
+class GPT2PreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = GPT2Config
+    load_tf_weights = load_tf_weights_in_gpt2
+    base_model_prefix = "transformer"
+    is_parallelizable = True
+
+    def __init__(self, *inputs, **kwargs):
+        super().__init__(*inputs, **kwargs)
+
+    def _init_weights(self, module):
+        """Initialize the weights."""
+        if isinstance(module, (nn.Linear, nn.Embedding, Conv1D)):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if isinstance(module, (nn.Linear, Conv1D)) and module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+
+
+@dataclass
+class GPT2DoubleHeadsModelOutput(ModelOutput):
+    """
+    Base class for outputs of models predicting if two sentences are consecutive or not.
+
+    Args:
+        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when ``labels`` is provided):
+            Language modeling loss.
+        mc_loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`mc_labels` is provided):
+            Multiple choice classification loss.
+        logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, num_choices, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        mc_logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, num_choices)`):
+            Prediction scores of the multiple choice classification head (scores for each choice before SoftMax).
+        past_key_values (:obj:`List[torch.FloatTensor]`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
+            List of :obj:`torch.FloatTensor` of length :obj:`config.n_layers`, with each tensor of shape :obj:`(2,
+            batch_size, num_heads, sequence_length, embed_size_per_head)`).
+
+            Contains pre-computed hidden-states (key and values in the attention blocks) that can be used (see
+            :obj:`past_key_values` input) to speed up sequential decoding.
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    mc_loss: Optional[torch.FloatTensor] = None
+    logits: torch.FloatTensor = None
+    mc_logits: torch.FloatTensor = None
+    past_key_values: Optional[List[torch.FloatTensor]] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+GPT2_START_DOCSTRING = r"""
+
+    This model inherits from :class:`~transformers.PreTrainedModel`. Check the superclass documentation for the generic
+    methods the library implements for all its model (such as downloading or saving, resizing the input embeddings,
+    pruning heads etc.)
+
+    This model is also a PyTorch `torch.nn.Module <https://pytorch.org/docs/stable/nn.html#torch.nn.Module>`__
+    subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to
+    general usage and behavior.
+
+    Parameters:
+        config (:class:`~transformers.GPT2Config`): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the
+            configuration. Check out the :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model
+            weights.
+"""
+
+GPT2_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, input_ids_length)`):
+            :obj:`input_ids_length` = ``sequence_length`` if :obj:`past_key_values` is ``None`` else
+            ``past_key_values[0].shape[-2]`` (``sequence_length`` of input past key value states). Indices of input
+            sequence tokens in the vocabulary.
+
+            If :obj:`past_key_values` is used, only ``input_ids`` that do not have their past calculated should be
+            passed as ``input_ids``.
+
+            Indices can be obtained using :class:`~transformers.GPT2Tokenizer`. See
+            :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__` for
+            details.
+
+            `What are input IDs? <../glossary.html#input-ids>`__
+        past_key_values (:obj:`List[torch.FloatTensor]` of length :obj:`config.n_layers`):
+            Contains precomputed hidden-states (key and values in the attention blocks) as computed by the model (see
+            :obj:`past_key_values` output below). Can be used to speed up sequential decoding. The ``input_ids`` which
+            have their past given to this model should not be passed as ``input_ids`` as they have already been
+            computed.
+        attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Mask to avoid performing attention on padding token indices. Mask values selected in ``[0, 1]``:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            `What are attention masks? <../glossary.html#attention-mask>`__
+        token_type_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, input_ids_length)`, `optional`):
+            Segment token indices to indicate first and second portions of the inputs. Indices are selected in ``[0,
+            1]``:
+
+            - 0 corresponds to a `sentence A` token,
+            - 1 corresponds to a `sentence B` token.
+
+            `What are token type IDs? <../glossary.html#token-type-ids>`_
+        position_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range ``[0,
+            config.max_position_embeddings - 1]``.
+
+            `What are position IDs? <../glossary.html#position-ids>`_
+        head_mask (:obj:`torch.FloatTensor` of shape :obj:`(num_heads,)` or :obj:`(num_layers, num_heads)`, `optional`):
+            Mask to nullify selected heads of the self-attention modules. Mask values selected in ``[0, 1]``:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
+            Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert :obj:`input_ids` indices into associated
+            vectors than the model's internal embedding lookup matrix.
+
+            If :obj:`past_key_values` is used, optionally only the last :obj:`inputs_embeds` have to be input (see
+            :obj:`past_key_values`).
+        use_cache (:obj:`bool`, `optional`):
+            If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
+            decoding (see :obj:`past_key_values`).
+        output_attentions (:obj:`bool`, `optional`):
+            Whether or not to return the attentions tensors of all attention layers. See ``attentions`` under returned
+            tensors for more detail.
+        output_hidden_states (:obj:`bool`, `optional`):
+            Whether or not to return the hidden states of all layers. See ``hidden_states`` under returned tensors for
+            more detail.
+        return_dict (:obj:`bool`, `optional`):
+            Whether or not to return a :class:`~transformers.file_utils.ModelOutput` instead of a plain tuple.
+"""
+
+PARALLELIZE_DOCSTRING = r"""
+    This is an experimental feature and is a subject to change at a moment's notice.
+
+    Uses a device map to distribute attention modules of the model across several devices. If no device map is given,
+    it will evenly distribute blocks across all devices.
+
+    Args:
+        device_map (:obj:`Dict[int, list]`, optional, defaults to None):
+            A dictionary that maps attention modules to devices. Note that the embedding module and LMHead are always
+            automatically mapped to the first device (for esoteric reasons). That means that the first device should
+            have fewer attention modules mapped to it than other devices. For reference, the gpt2 models have the
+            following number of attention modules:
+
+                - gpt2: 12
+                - gpt2-medium: 24
+                - gpt2-large: 36
+                - gpt2-xl: 48
+
+    Example::
+
+            # Here is an example of a device map on a machine with 4 GPUs using gpt2-xl, which has a total of 48 attention modules:
+            model = GPT2LMHeadModel.from_pretrained('gpt2-xl')
+            device_map = {0: [0, 1, 2, 3, 4, 5, 6, 7, 8],
+
+                          1: [9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21],
+                          2: [22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34],
+                          3: [35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47]}
+            model.parallelize(device_map)
+"""
+DEPARALLELIZE_DOCSTRING = r"""
+    Moves the model to cpu from a model parallel state.
+
+    Example::
+
+        # On a 4 GPU machine with gpt2-large:
+        model = GPT2LMHeadModel.from_pretrained('gpt2-large')
+        device_map = {0: [0, 1, 2, 3, 4, 5, 6, 7],
+
+                    1: [8, 9, 10, 11, 12, 13, 14, 15],
+                    2: [16, 17, 18, 19, 20, 21, 22, 23],
+                    3: [24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35]}
+        model.parallelize(device_map) # Splits the model across several devices
+        model.deparallelize() # Put the model back on cpu and cleans memory by calling torch.cuda.empty_cache()
+"""
+
+
+@add_start_docstrings(
+    "The bare GPT2 Model transformer outputting raw hidden-states without any specific head on top.",
+    GPT2_START_DOCSTRING,
+)
+class GPT2Model(GPT2PreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.wte = nn.Embedding(config.vocab_size, config.n_embd)
+        self.wpe = nn.Embedding(config.n_positions, config.n_embd)
+        if _USE_GROVER:
+            self.emb_norm = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
+
+        self.drop = nn.Dropout(config.embd_pdrop)
+        self.h = nn.ModuleList(
+            [Block(config.n_ctx, config, scale=True) for _ in range(config.n_layer)]
+        )
+        if not _USE_GROVER:
+            self.ln_f = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
+
+        self.init_weights()
+
+        # Model parallel
+        self.model_parallel = False
+        self.device_map = None
+
+    @add_start_docstrings(PARALLELIZE_DOCSTRING)
+    def parallelize(self, device_map=None):
+        # Check validity of device_map
+        self.device_map = (
+            get_device_map(len(self.h), range(torch.cuda.device_count()))
+            if device_map is None
+            else device_map
+        )
+        assert_device_map(self.device_map, len(self.h))
+        self.model_parallel = True
+        self.first_device = (
+            "cpu"
+            if "cpu" in self.device_map.keys()
+            else "cuda:" + str(min(self.device_map.keys()))
+        )
+        self.last_device = "cuda:" + str(max(self.device_map.keys()))
+        self.wte = self.wte.to(self.first_device)
+        self.wpe = self.wpe.to(self.first_device)
+        # Load onto devices
+        for k, v in self.device_map.items():
+            for block in v:
+                cuda_device = "cuda:" + str(k)
+                self.h[block] = self.h[block].to(cuda_device)
+        # ln_f to last
+        self.ln_f = self.ln_f.to(self.last_device)
+
+    @add_start_docstrings(DEPARALLELIZE_DOCSTRING)
+    def deparallelize(self):
+        self.model_parallel = False
+        self.device_map = None
+        self.first_device = "cpu"
+        self.last_device = "cpu"
+        self.wte = self.wte.to("cpu")
+        self.wpe = self.wpe.to("cpu")
+        for index in range(len(self.h)):
+            self.h[index] = self.h[index].to("cpu")
+        self.ln_f = self.ln_f.to("cpu")
+        torch.cuda.empty_cache()
+
+    def get_input_embeddings(self):
+        return self.wte
+
+    def set_input_embeddings(self, new_embeddings):
+        self.wte = new_embeddings
+
+    def _prune_heads(self, heads_to_prune):
+        """
+        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
+        """
+        for layer, heads in heads_to_prune.items():
+            self.h[layer].attn.prune_heads(heads)
+
+    @add_start_docstrings_to_model_forward(GPT2_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="gpt2",
+        output_type=BaseModelOutputWithPastAndCrossAttentions,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids=None,
+        past_key_values=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        output_attentions = (
+            output_attentions
+            if output_attentions is not None
+            else self.config.output_attentions
+        )
+        output_hidden_states = (
+            output_hidden_states
+            if output_hidden_states is not None
+            else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError(
+                "You cannot specify both input_ids and inputs_embeds at the same time"
+            )
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+            input_ids = input_ids.view(-1, input_shape[-1])
+            batch_size = input_ids.shape[0]
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+            batch_size = inputs_embeds.shape[0]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        if token_type_ids is not None:
+            token_type_ids = token_type_ids.view(-1, input_shape[-1])
+        if position_ids is not None:
+            position_ids = position_ids.view(-1, input_shape[-1])
+
+        if past_key_values is None:
+            past_length = 0
+            past_key_values = [None] * len(self.h)
+        else:
+            past_length = past_key_values[0][0].size(-2)
+        if position_ids is None:
+            device = input_ids.device if input_ids is not None else inputs_embeds.device
+            position_ids = torch.arange(
+                past_length,
+                input_shape[-1] + past_length,
+                dtype=torch.long,
+                device=device,
+            )
+            position_ids = position_ids.unsqueeze(0).view(-1, input_shape[-1])
+
+        # Attention mask.
+        if attention_mask is not None:
+            if batch_size <= 0:
+                raise ValueError("batch_size has to be defined and > 0")
+            attention_mask = attention_mask.view(batch_size, -1)
+            # We create a 3D attention mask from a 2D tensor mask.
+            # Sizes are [batch_size, 1, 1, to_seq_length]
+            # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
+            # this attention mask is more simple than the triangular masking of causal attention
+            # used in OpenAI GPT, we just need to prepare the broadcast dimension here.
+            attention_mask = attention_mask[:, None, None, :]
+
+            # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
+            # masked positions, this operation will create a tensor which is 0.0 for
+            # positions we want to attend and -10000.0 for masked positions.
+            # Since we are adding it to the raw scores before the softmax, this is
+            # effectively the same as removing these entirely.
+            attention_mask = attention_mask.to(dtype=self.dtype)  # fp16 compatibility
+            attention_mask = (1.0 - attention_mask) * -10000.0
+
+        # If a 2D ou 3D attention mask is provided for the cross-attention
+        # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
+        if self.config.add_cross_attention and encoder_hidden_states is not None:
+            (
+                encoder_batch_size,
+                encoder_sequence_length,
+                _,
+            ) = encoder_hidden_states.size()
+            encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
+            if encoder_attention_mask is None:
+                encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
+            encoder_attention_mask = self.invert_attention_mask(encoder_attention_mask)
+        else:
+            encoder_attention_mask = None
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # head_mask has shape n_layer x batch x n_heads x N x N
+        head_mask = self.get_head_mask(head_mask, self.config.n_layer)
+
+        if inputs_embeds is None:
+            inputs_embeds = self.wte(input_ids)
+        position_embeds = self.wpe(position_ids)
+        hidden_states = inputs_embeds + position_embeds
+
+        if token_type_ids is not None:
+            token_type_embeds = self.wte(token_type_ids)
+            hidden_states = hidden_states + token_type_embeds
+
+        hidden_states = self.drop(hidden_states)
+        if _USE_GROVER:
+            hidden_states = self.emb_norm(hidden_states)
+        output_shape = input_shape + (hidden_states.size(-1),)
+
+        presents = () if use_cache else None
+        all_self_attentions = () if output_attentions else None
+        all_cross_attentions = (
+            () if output_attentions and self.config.add_cross_attention else None
+        )
+        all_hidden_states = () if output_hidden_states else None
+        for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)):
+
+            # Model parallel
+            if self.model_parallel:
+                torch.cuda.set_device(hidden_states.device)
+                # Ensure layer_past is on same device as hidden_states (might not be correct)
+                if layer_past is not None:
+                    layer_past = tuple(
+                        past_state.to(hidden_states.device) for past_state in layer_past
+                    )
+                # Ensure that attention_mask is always on the same device as hidden_states
+                if attention_mask is not None:
+                    attention_mask = attention_mask.to(hidden_states.device)
+                if isinstance(head_mask, torch.Tensor):
+                    head_mask = head_mask.to(hidden_states.device)
+
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (
+                    hidden_states.view(*output_shape),
+                )
+
+            if getattr(self.config, "gradient_checkpointing", False):
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        # checkpointing only works with tuple returns, not with lists
+                        return tuple(
+                            output
+                            for output in module(*inputs, use_cache, output_attentions)
+                        )
+
+                    return custom_forward
+
+                outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states,
+                    layer_past,
+                    attention_mask,
+                    head_mask[i],
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                )
+            else:
+                outputs = block(
+                    hidden_states,
+                    layer_past=layer_past,
+                    attention_mask=attention_mask,
+                    head_mask=head_mask[i],
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_attention_mask=encoder_attention_mask,
+                    use_cache=use_cache,
+                    output_attentions=output_attentions,
+                )
+
+            hidden_states, present = outputs[:2]
+            if use_cache is True:
+                presents = presents + (present,)
+
+            if output_attentions:
+                all_self_attentions = all_self_attentions + (
+                    outputs[2 if use_cache else 1],
+                )
+                if self.config.add_cross_attention:
+                    all_cross_attentions = all_cross_attentions + (
+                        outputs[3 if use_cache else 2],
+                    )
+
+            # Model Parallel: If it's the last layer for that device, put things on the next device
+            if self.model_parallel:
+                for k, v in self.device_map.items():
+                    if i == v[-1] and "cuda:" + str(k) != self.last_device:
+                        hidden_states = hidden_states.to("cuda:" + str(k + 1))
+
+        if not _USE_GROVER:
+            hidden_states = self.ln_f(hidden_states)
+
+        hidden_states = hidden_states.view(*output_shape)
+        # Add last hidden state
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(
+                v
+                for v in [
+                    hidden_states,
+                    presents,
+                    all_hidden_states,
+                    all_self_attentions,
+                    all_cross_attentions,
+                ]
+                if v is not None
+            )
+
+        return BaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            past_key_values=presents,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+            cross_attentions=all_cross_attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    The GPT2 Model transformer with a language modeling head on top (linear layer with weights tied to the input
+    embeddings).
+    """,
+    GPT2_START_DOCSTRING,
+)
+class GPT2LMHeadModel(GPT2PreTrainedModel):
+    _keys_to_ignore_on_load_missing = [r"h\.\d+\.attn\.masked_bias", r"lm_head\.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.transformer = GPT2Model(config)
+        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
+
+        self.init_weights()
+
+        # Model parallel
+        self.model_parallel = False
+        self.device_map = None
+
+    @add_start_docstrings(PARALLELIZE_DOCSTRING)
+    def parallelize(self, device_map=None):
+        self.device_map = (
+            get_device_map(len(self.transformer.h), range(torch.cuda.device_count()))
+            if device_map is None
+            else device_map
+        )
+        assert_device_map(self.device_map, len(self.transformer.h))
+        self.transformer.parallelize(self.device_map)
+        self.lm_head = self.lm_head.to(self.transformer.first_device)
+        self.model_parallel = True
+
+    @add_start_docstrings(DEPARALLELIZE_DOCSTRING)
+    def deparallelize(self):
+        self.transformer.deparallelize()
+        self.transformer = self.transformer.to("cpu")
+        self.lm_head = self.lm_head.to("cpu")
+        self.model_parallel = False
+        torch.cuda.empty_cache()
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def prepare_inputs_for_generation(self, input_ids, past=None, **kwargs):
+        token_type_ids = kwargs.get("token_type_ids", None)
+        # only last token for inputs_ids if past is defined in kwargs
+        if past:
+            input_ids = input_ids[:, -1].unsqueeze(-1)
+            if token_type_ids is not None:
+                token_type_ids = token_type_ids[:, -1].unsqueeze(-1)
+
+        attention_mask = kwargs.get("attention_mask", None)
+        position_ids = kwargs.get("position_ids", None)
+
+        if attention_mask is not None and position_ids is None:
+            # create position_ids on the fly for batch generation
+            position_ids = attention_mask.long().cumsum(-1) - 1
+            position_ids.masked_fill_(attention_mask == 0, 1)
+            if past:
+                position_ids = position_ids[:, -1].unsqueeze(-1)
+        else:
+            position_ids = None
+        return {
+            "input_ids": input_ids,
+            "past_key_values": past,
+            "use_cache": kwargs.get("use_cache"),
+            "position_ids": position_ids,
+            "attention_mask": attention_mask,
+            "token_type_ids": token_type_ids,
+        }
+
+    @add_start_docstrings_to_model_forward(GPT2_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="gpt2",
+        output_type=CausalLMOutputWithCrossAttentions,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids=None,
+        past_key_values=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        labels=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
+            ``labels = input_ids`` Indices are selected in ``[-100, 0, ..., config.vocab_size]`` All labels set to
+            ``-100`` are ignored (masked), the loss is only computed for labels in ``[0, ..., config.vocab_size]``
+        """
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        transformer_outputs = self.transformer(
+            input_ids,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        hidden_states = transformer_outputs[0]
+
+        # Set device for model parallelism
+        if self.model_parallel:
+            torch.cuda.set_device(self.transformer.first_device)
+            hidden_states = hidden_states.to(self.lm_head.weight.device)
+
+        lm_logits = self.lm_head(hidden_states)
+
+        loss = None
+        if labels is not None:
+            # Shift so that tokens < n predict n
+            shift_logits = lm_logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(
+                shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1)
+            )
+
+        if not return_dict:
+            output = (lm_logits,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return CausalLMOutputWithCrossAttentions(
+            loss=loss,
+            logits=lm_logits,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+            cross_attentions=transformer_outputs.cross_attentions,
+        )
+
+    @staticmethod
+    def _reorder_cache(
+        past: Tuple[Tuple[torch.Tensor]], beam_idx: torch.Tensor
+    ) -> Tuple[Tuple[torch.Tensor]]:
+        """
+        This function is used to re-order the :obj:`past_key_values` cache if
+        :meth:`~transformers.PreTrainedModel.beam_search` or :meth:`~transformers.PreTrainedModel.beam_sample` is
+        called. This is required to match :obj:`past_key_values` with the correct beam_idx at every generation step.
+        """
+        return tuple(
+            tuple(
+                past_state.index_select(0, beam_idx.to(past_state.device))
+                for past_state in layer_past
+            )
+            for layer_past in past
+        )
+
+
+@add_start_docstrings(
+    """
+The GPT2 Model transformer with a language modeling and a multiple-choice classification head on top e.g. for
+RocStories/SWAG tasks. The two heads are two linear layers. The language modeling head has its weights tied to the
+input embeddings, the classification head takes as input the input of a specified classification token index in the
+input sequence).
+""",
+    GPT2_START_DOCSTRING,
+)
+class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        config.num_labels = 1
+        self.transformer = GPT2Model(config)
+        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
+        self.multiple_choice_head = SequenceSummary(config)
+
+        self.init_weights()
+
+        # Model parallel
+        self.model_parallel = False
+        self.device_map = None
+
+    @add_start_docstrings(PARALLELIZE_DOCSTRING)
+    def parallelize(self, device_map=None):
+        self.device_map = (
+            get_device_map(len(self.transformer.h), range(torch.cuda.device_count()))
+            if device_map is None
+            else device_map
+        )
+        assert_device_map(self.device_map, len(self.transformer.h))
+        self.transformer.parallelize(self.device_map)
+        self.lm_head = self.lm_head.to(self.transformer.first_device)
+        self.multiple_choice_head = self.multiple_choice_head.to(
+            self.transformer.first_device
+        )
+        self.model_parallel = True
+
+    @add_start_docstrings(DEPARALLELIZE_DOCSTRING)
+    def deparallelize(self):
+        self.transformer.deparallelize()
+        self.transformer = self.transformer.to("cpu")
+        self.lm_head = self.lm_head.to("cpu")
+        self.multiple_choice_head = self.multiple_choice_head.to("cpu")
+        self.model_parallel = False
+        torch.cuda.empty_cache()
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def prepare_inputs_for_generation(self, input_ids, past=None, **kwargs):
+        token_type_ids = kwargs.get("token_type_ids", None)
+        # only last token for inputs_ids if past is defined in kwargs
+        if past:
+            input_ids = input_ids[:, -1].unsqueeze(-1)
+            if token_type_ids is not None:
+                token_type_ids = token_type_ids[:, -1].unsqueeze(-1)
+
+        attention_mask = kwargs.get("attention_mask", None)
+        position_ids = kwargs.get("position_ids", None)
+
+        if attention_mask is not None and position_ids is None:
+            # create position_ids on the fly for batch generation
+            position_ids = attention_mask.long().cumsum(-1) - 1
+            position_ids.masked_fill_(attention_mask == 0, 1)
+            if past:
+                position_ids = position_ids[:, -1].unsqueeze(-1)
+        else:
+            position_ids = None
+
+        return {
+            "input_ids": input_ids,
+            "past_key_values": past,
+            "use_cache": kwargs.get("use_cache"),
+            "position_ids": position_ids,
+            "attention_mask": attention_mask,
+            "token_type_ids": token_type_ids,
+        }
+
+    @add_start_docstrings_to_model_forward(GPT2_INPUTS_DOCSTRING)
+    @replace_return_docstrings(
+        output_type=GPT2DoubleHeadsModelOutput, config_class=_CONFIG_FOR_DOC
+    )
+    def forward(
+        self,
+        input_ids=None,
+        past_key_values=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        mc_token_ids=None,
+        labels=None,
+        mc_labels=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+        **kwargs,
+    ):
+        r"""
+        mc_token_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, num_choices)`, `optional`, default to index of the last token of the input):
+            Index of the classification token in each input sequence. Selected in the range ``[0, input_ids.size(-1) -
+            1[``.
+        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
+            ``labels = input_ids`` Indices are selected in ``[-1, 0, ..., config.vocab_size]`` All labels set to
+            ``-100`` are ignored (masked), the loss is only computed for labels in ``[0, ..., config.vocab_size]``
+        mc_labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size)`, `optional`):
+            Labels for computing the multiple choice classification loss. Indices should be in ``[0, ...,
+            num_choices]`` where `num_choices` is the size of the second dimension of the input tensors. (see
+            `input_ids` above)
+
+        Return:
+
+        Example::
+
+            >>> import torch
+            >>> from transformers import GPT2Tokenizer, GPT2DoubleHeadsModel
+
+            >>> tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
+            >>> model = GPT2DoubleHeadsModel.from_pretrained('gpt2')
+
+            >>> # Add a [CLS] to the vocabulary (we should train it also!)
+            >>> num_added_tokens = tokenizer.add_special_tokens({'cls_token': '[CLS]'})
+
+            >>> embedding_layer = model.resize_token_embeddings(len(tokenizer))  # Update the model embeddings with the new vocabulary size
+
+            >>> choices = ["Hello, my dog is cute [CLS]", "Hello, my cat is cute [CLS]"]
+            >>> encoded_choices = [tokenizer.encode(s) for s in choices]
+            >>> cls_token_location = [tokens.index(tokenizer.cls_token_id) for tokens in encoded_choices]
+
+            >>> input_ids = torch.tensor(encoded_choices).unsqueeze(0)  # Batch size: 1, number of choices: 2
+            >>> mc_token_ids = torch.tensor([cls_token_location])  # Batch size: 1
+
+            >>> outputs = model(input_ids, mc_token_ids=mc_token_ids)
+            >>> lm_logits = outputs.lm_logits
+            >>> mc_logits = outputs.mc_logits
+
+        """
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        transformer_outputs = self.transformer(
+            input_ids,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = transformer_outputs[0]
+
+        # Set device for model parallelism
+        if self.model_parallel:
+            torch.cuda.set_device(self.transformer.first_device)
+            hidden_states = hidden_states.to(self.lm_head.weight.device)
+
+        lm_logits = self.lm_head(hidden_states)
+        mc_logits = self.multiple_choice_head(hidden_states, mc_token_ids).squeeze(-1)
+
+        mc_loss = None
+        if mc_labels is not None:
+            loss_fct = CrossEntropyLoss()
+            mc_loss = loss_fct(
+                mc_logits.view(-1, mc_logits.size(-1)), mc_labels.view(-1)
+            )
+        lm_loss = None
+        if labels is not None:
+            shift_logits = lm_logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            loss_fct = CrossEntropyLoss()
+            lm_loss = loss_fct(
+                shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1)
+            )
+
+        if not return_dict:
+            output = (lm_logits, mc_logits) + transformer_outputs[1:]
+            if mc_loss is not None:
+                output = (mc_loss,) + output
+            return ((lm_loss,) + output) if lm_loss is not None else output
+
+        return GPT2DoubleHeadsModelOutput(
+            loss=lm_loss,
+            mc_loss=mc_loss,
+            logits=lm_logits,
+            mc_logits=mc_logits,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
+
+    @staticmethod
+    def _reorder_cache(
+        past: Tuple[Tuple[torch.Tensor]], beam_idx: torch.Tensor
+    ) -> Tuple[Tuple[torch.Tensor]]:
+        """
+        This function is used to re-order the :obj:`past_key_values` cache if
+        :meth:`~transformers.PreTrainedModel.beam_search` or :meth:`~transformers.PreTrainedModel.beam_sample` is
+        called. This is required to match :obj:`past_key_values` with the correct beam_idx at every generation step.
+        """
+        return tuple(
+            tuple(
+                past_state.index_select(0, beam_idx.to(past_state.device))
+                for past_state in layer_past
+            )
+            for layer_past in past
+        )
+
+
+@add_start_docstrings(
+    """
+    The GPT2 Model transformer with a sequence classification head on top (linear layer).
+
+    :class:`~transformers.GPT2ForSequenceClassification` uses the last token in order to do the classification, as
+    other causal models (e.g. GPT-1) do.
+
+    Since it does classification on the last token, it requires to know the position of the last token. If a
+    :obj:`pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each
+    row. If no :obj:`pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot
+    guess the padding tokens when :obj:`inputs_embeds` are passed instead of :obj:`input_ids`, it does the same (take
+    the last value in each row of the batch).
+    """,
+    GPT2_START_DOCSTRING,
+)
+class GPT2ForSequenceClassification(GPT2PreTrainedModel):
+    _keys_to_ignore_on_load_missing = [r"h\.\d+\.attn\.masked_bias", r"lm_head\.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.transformer = GPT2Model(config)
+        self.score = nn.Linear(config.n_embd, self.num_labels, bias=False)
+
+        self.init_weights()
+
+        # Model parallel
+        self.model_parallel = False
+        self.device_map = None
+
+    @add_start_docstrings_to_model_forward(GPT2_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="microsoft/dialogrpt",
+        output_type=SequenceClassifierOutputWithPast,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids=None,
+        past_key_values=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        labels=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
+            Labels for computing the sequence classification/regression loss. Indices should be in :obj:`[0, ...,
+            config.num_labels - 1]`. If :obj:`config.num_labels == 1` a regression loss is computed (Mean-Square loss),
+            If :obj:`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        transformer_outputs = self.transformer(
+            input_ids,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        hidden_states = transformer_outputs[0]
+        logits = self.score(hidden_states)
+
+        if input_ids is not None:
+            batch_size, sequence_length = input_ids.shape[:2]
+        else:
+            batch_size, sequence_length = inputs_embeds.shape[:2]
+
+        assert (
+            self.config.pad_token_id is not None or batch_size == 1
+        ), "Cannot handle batch sizes > 1 if no padding token is defined."
+        if self.config.pad_token_id is None:
+            sequence_lengths = -1
+        else:
+            if input_ids is not None:
+                sequence_lengths = (
+                    torch.ne(input_ids, self.config.pad_token_id).sum(-1) - 1
+                )
+            else:
+                sequence_lengths = -1
+                logger.warning(
+                    f"{self.__class__.__name__} will not detect padding tokens in `inputs_embeds`. Results may be "
+                    f"unexpected if using padding tokens in conjunction with `inputs_embeds.`"
+                )
+
+        pooled_logits = logits[range(batch_size), sequence_lengths]
+
+        loss = None
+        if labels is not None:
+            if self.num_labels == 1:
+                #  We are doing regression
+                loss_fct = MSELoss()
+                loss = loss_fct(pooled_logits.view(-1), labels.to(self.dtype).view(-1))
+            else:
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(
+                    pooled_logits.view(-1, self.num_labels), labels.view(-1)
+                )
+
+        if not return_dict:
+            output = (pooled_logits,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return SequenceClassifierOutputWithPast(
+            loss=loss,
+            logits=pooled_logits,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    GPT2 Model with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for
+    Named-Entity-Recognition (NER) tasks.
+    """,
+    GPT2_START_DOCSTRING,
+)
+class GPT2ForTokenClassification(GPT2PreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.transformer = GPT2Model(config)
+        if (
+            hasattr(config, "classifier_dropout")
+            and config.classifier_dropout is not None
+        ):
+            classifier_dropout = config.classifier_dropout
+        elif hasattr(config, "hidden_dropout") and config.hidden_dropout is not None:
+            classifier_dropout = config.hidden_dropout
+        else:
+            classifier_dropout = 0.1
+        self.dropout = nn.Dropout(classifier_dropout)
+        self.classifier = nn.Linear(config.hidden_size, config.num_labels)
+
+        self.init_weights()
+
+        # Model parallel
+        self.model_parallel = False
+        self.device_map = None
+
+    @add_start_docstrings_to_model_forward(GPT2_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="microsoft/DialogRPT-updown",
+        output_type=TokenClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids=None,
+        past_key_values=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        labels=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
+            Labels for computing the sequence classification/regression loss. Indices should be in :obj:`[0, ...,
+            config.num_labels - 1]`. If :obj:`config.num_labels == 1` a regression loss is computed (Mean-Square loss),
+            If :obj:`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        transformer_outputs = self.transformer(
+            input_ids,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = transformer_outputs[0]
+        hidden_states = self.dropout(hidden_states)
+        logits = self.classifier(hidden_states)
+
+        loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            # Only keep active parts of the loss
+            if attention_mask is not None:
+                active_loss = attention_mask.view(-1) == 1
+                active_logits = logits.view(-1, self.num_labels)
+                active_labels = torch.where(
+                    active_loss,
+                    labels.view(-1),
+                    torch.tensor(loss_fct.ignore_index).type_as(labels),
+                )
+                loss = loss_fct(active_logits, active_labels)
+            else:
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+
+        if not return_dict:
+            output = (logits,) + transformer_outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TokenClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )

backend/preprocess.py

@@ -0,0 +1,736 @@
+import html
+import logging
+import re
+from typing import List
+from farasa.segmenter import FarasaSegmenter
+import emoji
+
+import pyarabic.araby as araby
+
+ACCEPTED_MODELS = [
+    "bert-base-arabertv01",
+    "bert-base-arabert",
+    "bert-base-arabertv02",
+    "bert-base-arabertv2",
+    "bert-large-arabertv02",
+    "bert-large-arabertv2",
+    "araelectra-base",
+    "araelectra-base-discriminator",
+    "araelectra-base-generator",
+    "araelectra-base-artydiqa",
+    "aragpt2-base",
+    "aragpt2-medium",
+    "aragpt2-large",
+    "aragpt2-mega",
+]
+
+SEGMENTED_MODELS = [
+    "bert-base-arabert",
+    "bert-base-arabertv2",
+    "bert-large-arabertv2",
+]
+
+SECOND_GEN_MODELS = [
+    "bert-base-arabertv02",
+    "bert-base-arabertv2",
+    "bert-large-arabertv02",
+    "bert-large-arabertv2",
+    "araelectra-base",
+    "araelectra-base-discriminator",
+    "araelectra-base-generator",
+    "araelectra-base-artydiqa",
+    "aragpt2-base",
+    "aragpt2-medium",
+    "aragpt2-large",
+    "aragpt2-mega",
+]
+
+farasa_segmenter = FarasaSegmenter(interactive=True)
+
+
+class ArabertPreprocessor:
+    """
+    A Preprocessor class that cleans and preprocesses text for all models in the AraBERT repo.
+    It also can unprocess the text ouput of the generated text
+
+    Args:
+
+        model_name (:obj:`str`): model name from the HuggingFace Models page without
+        the aubmindlab tag. Will default to a base Arabic preprocessor if model name was not found.
+        Current accepted models are:
+
+            - "bert-base-arabertv01": No farasa segmentation.
+            - "bert-base-arabert": with farasa segmentation.
+            - "bert-base-arabertv02": No farasas egmentation.
+            - "bert-base-arabertv2": with farasa segmentation.
+            - "bert-large-arabertv02": No farasas egmentation.
+            - "bert-large-arabertv2": with farasa segmentation.
+            - "araelectra-base": No farasa segmentation.
+            - "araelectra-base-discriminator": No farasa segmentation.
+            - "araelectra-base-generator": No farasa segmentation.
+            - "aragpt2-base": No farasa segmentation.
+            - "aragpt2-medium": No farasa segmentation.
+            - "aragpt2-large": No farasa segmentation.
+            - "aragpt2-mega": No farasa segmentation.
+
+
+        keep_emojis(:obj:`bool`, `optional`, defaults to :obj:`False`): don't remove emojis while preprocessing.
+
+        remove_html_markup(:obj: `bool`, `optional`, defaults to :obj:`True`): Whether to remove html artfacts,
+        should be set to False when preprocessing TyDi QA.
+
+        replace_urls_emails_mentions(:obj:`bool`, `optional`, defaults to :obj:`True`): Whether to replace email urls
+        and mentions by special tokens.
+
+        strip_tashkeel(:obj:`bool`, `optional`, defaults to :obj:`True`): remove diacritics (FATHATAN, DAMMATAN, KASRATAN, FATHA, DAMMA,
+        KASRA, SUKUN, SHADDA).
+
+        strip_tatweel(:obj:`bool`, `optional`, defaults to :obj:`True`): remove tatweel '\\u0640'.
+
+        insert_white_spaces(:obj:`bool`, `optional`, defaults to :obj:`True`): insert whitespace before and after all non Arabic digits
+        or English digits or Arabic and English Alphabet or the 2 brackets, then inserts whitespace
+        between words and numbers or numbers and words.
+
+        remove_non_digit_repetition(:obj:`bool`, `optional`, defaults to :obj:`True`): replace repetition of more than 2 non-digit character with
+        2 of this character.
+
+        replace_slash_with_dash(:obj:`bool`, `optional`, defaults to :obj:`None`): Will be automatically set to True in AraBERTv02,
+        AraELECTRA and AraGPT2.
+        Set to False to force disable, and True to force enable. Replaces the "/"  with "-",
+        since "/" is missing from AraBERTv2, AraELECTRA and ARAGPT2 vocabulary.
+
+        map_hindi_numbers_to_arabic(:obj:`bool`, `optional`, defaults to :obj:`None`): Will be automatically set to True in
+        AraBERTv02, AraELECTRA and AraGPT2.Set to False to force disable, and True to force enable.
+        Replaces hindi numbers with the corresponding Arabic one. ex: "١٩٩٥" --> "1995".
+        This is behavior is present by default in AraBERTv1 and v2 (with pre-segmentation),
+        and fixes the issue of caused by a bug when inserting white spaces.
+
+        apply_farasa_segmentation(:obj:`bool`, `optional`, defaults to :obj:`None`): Will be automatically set to True in
+        AraBERTv2, and AraBERTv1. Set to False to force disable, and True to force enable.
+
+
+
+    Returns:
+
+        ArabertPreprocessor: A preprocessor instance
+
+    Example:
+
+        from preprocess import ArabertPreprocessor
+
+        arabert_prep = ArabertPreprocessor("aubmindlab/bert-base-arabertv2")
+
+        arabert_prep.preprocess("SOME ARABIC TEXT")
+    """
+
+    def __init__(
+        self,
+        model_name: str,
+        keep_emojis: bool = False,
+        remove_html_markup: bool = True,
+        replace_urls_emails_mentions: bool = True,
+        strip_tashkeel: bool = True,
+        strip_tatweel: bool = True,
+        insert_white_spaces: bool = True,
+        remove_non_digit_repetition: bool = True,
+        replace_slash_with_dash: bool = None,
+        map_hindi_numbers_to_arabic: bool = None,
+        apply_farasa_segmentation: bool = None,
+    ):
+
+        model_name = model_name.replace("aubmindlab/", "").replace("wissamantoun/", "")
+
+        if model_name not in ACCEPTED_MODELS:
+            logging.warning(
+                """Model provided is not in the accepted model list. Preprocessor will default to a base Arabic preprocessor"""
+            )
+            self.model_name = "bert-base-arabertv02"
+        else:
+            self.model_name = model_name
+
+        if apply_farasa_segmentation is None:
+            if self.model_name in SEGMENTED_MODELS:
+                self.apply_farasa_segmentation = True
+            else:
+                self.apply_farasa_segmentation = False
+        else:
+            if (
+                apply_farasa_segmentation == False
+                and self.model_name in SEGMENTED_MODELS
+            ):
+                logging.warning(
+                    "The selected model_name requires Farasa pre-segmentation, but apply_farasa_segmentation was set to False!"
+                )
+
+            self.apply_farasa_segmentation = apply_farasa_segmentation
+
+        self.keep_emojis = keep_emojis
+        self.remove_html_markup = remove_html_markup
+        self.replace_urls_emails_mentions = replace_urls_emails_mentions
+        self.strip_tashkeel = strip_tashkeel
+        self.strip_tatweel = strip_tatweel
+        self.insert_white_spaces = insert_white_spaces
+        self.remove_non_digit_repetition = remove_non_digit_repetition
+
+        if replace_slash_with_dash is None:
+            if self.model_name in SECOND_GEN_MODELS:
+                self.replace_slash_with_dash = True
+            else:
+                self.replace_slash_with_dash = False
+        else:
+            self.replace_slash_with_dash = replace_slash_with_dash
+
+        if map_hindi_numbers_to_arabic is None:
+            if self.model_name in SECOND_GEN_MODELS:
+                self.map_hindi_numbers_to_arabic = True
+            else:
+                self.map_hindi_numbers_to_arabic = False
+        else:
+            self.map_hindi_numbers_to_arabic = map_hindi_numbers_to_arabic
+
+    def preprocess(self, text: str) -> str:
+        """
+        Preprocess takes an input text line an applies the same preprocessing used in AraBERT
+                            pretraining, or according to settings
+
+        Args:
+
+            text (:obj:`str`): inout text string
+
+        Returns:
+
+            string: A preprocessed string depending on which model was selected
+        """
+        if (
+            self.model_name == "bert-base-arabert"
+            or self.model_name == "bert-base-arabertv01"
+        ):
+            return self._preprocess_v1(
+                text,
+                do_farasa_tokenization=self.apply_farasa_segmentation,
+            )
+
+        if self.model_name in SECOND_GEN_MODELS:
+            return self._preprocess_v2(text)
+
+        return self._preprocess_v3(text)
+
+    def unpreprocess(self, text: str, desegment: bool = True) -> str:
+        """Re-formats the text to a classic format where punctuations, brackets, parenthesis are not seperated by whitespaces.
+        The objective is to make the generated text of any model appear natural and not preprocessed.
+
+        Args:
+            text (:obj:`str`): input text to be un-preprocessed
+            desegment (:obj:`bool`, optional): [whether or not to remove farasa pre-segmentation before]..
+
+        Returns:
+            str: The unpreprocessed (and possibly Farasa-desegmented) text.
+        """
+
+        if self.apply_farasa_segmentation and desegment:
+            text = self.desegment(text)
+
+        # removes the spaces around quotation marks ex: i " ate " an apple --> i "ate" an apple
+        # https://stackoverflow.com/a/53436792/5381220
+        text = re.sub(white_spaced_double_quotation_regex, '"' + r"\1" + '"', text)
+        text = re.sub(white_spaced_single_quotation_regex, "'" + r"\1" + "'", text)
+        text = re.sub(white_spaced_back_quotation_regex, "\`" + r"\1" + "\`", text)
+        text = re.sub(white_spaced_back_quotation_regex, "\—" + r"\1" + "\—", text)
+
+        # during generation, sometimes the models don't put a space after the dot, this handles it
+        text = text.replace(".", " . ")
+        text = " ".join(text.split())
+
+        # handle decimals
+        text = re.sub(r"(\d+) \. (\d+)", r"\1.\2", text)
+        text = re.sub(r"(\d+) \, (\d+)", r"\1,\2", text)
+
+        text = re.sub(left_and_right_spaced_chars, r"\1", text)
+        text = re.sub(left_spaced_chars, r"\1", text)
+        text = re.sub(right_spaced_chars, r"\1", text)
+
+        return text
+
+    def desegment(self, text: str) -> str:
+        """
+        Use this function if sentence tokenization was done using
+        `from arabert.preprocess_arabert import preprocess` with Farasa enabled
+        AraBERT segmentation using Farasa adds a space after the '+' for prefixes,
+        and after before the '+' for suffixes
+
+        Example:
+        >>> desegment('ال+ دراس +ات')
+        الدراسات
+        """
+        text = text.replace("+ ", "+")
+        text = text.replace(" +", "+")
+        text = " ".join([self._desegmentword(word) for word in text.split(" ")])
+        return text
+
+    def _desegmentword(self, orig_word: str) -> str:
+        """
+        Word segmentor that takes a Farasa Segmented Word and removes the '+' signs
+
+        Example:
+        >>> _desegmentword("ال+يومي+ة")
+        اليومية
+        """
+        word = orig_word.replace("ل+ال+", "لل")
+        if "ال+ال" not in orig_word:
+            word = word.replace("ل+ال", "لل")
+        word = word.replace("+", "")
+        word = word.replace("للل", "لل")
+        return word
+
+    def _preprocess_v3(self, text: str) -> str:
+        text = str(text)
+        text = html.unescape(text)
+        if self.strip_tashkeel:
+            text = araby.strip_tashkeel(text)
+        if self.strip_tatweel:
+            text = araby.strip_tatweel(text)
+
+        if self.replace_urls_emails_mentions:
+            # replace all possible URLs
+            for reg in url_regexes:
+                text = re.sub(reg, " [رابط] ", text)
+            # REplace Emails with [بريد]
+            for reg in email_regexes:
+                text = re.sub(reg, " [بريد] ", text)
+            # replace mentions with [مستخدم]
+            text = re.sub(user_mention_regex, " [مستخدم] ", text)
+
+        if self.remove_html_markup:
+            # remove html line breaks
+            text = re.sub("<br />", " ", text)
+            # remove html markup
+            text = re.sub("</?[^>]+>", " ", text)
+
+        if self.map_hindi_numbers_to_arabic:
+            text = text.translate(hindi_to_arabic_map)
+
+        # remove repeated characters >2
+        if self.remove_non_digit_repetition:
+            text = self._remove_non_digit_repetition(text)
+
+        # insert whitespace before and after all non Arabic digits or English Digits and Alphabet and the 2 brackets
+        if self.insert_white_spaces:
+            text = re.sub(
+                "([^0-9\u0621-\u063A\u0641-\u064A\u0660-\u0669a-zA-Z ])",
+                r" \1 ",
+                text,
+            )
+
+            # re-fix brackets
+            text = text.replace("[ رابط ]", "[رابط]")
+            text = text.replace("[ بريد ]", "[بريد]")
+            text = text.replace("[ مستخدم ]", "[مستخدم]")
+
+            # insert whitespace between words and numbers or numbers and words
+            text = re.sub(
+                "(\d+)([\u0621-\u063A\u0641-\u064A\u066A-\u066C\u0654-\u0655]+)",
+                r" \1 \2 ",
+                text,
+            )
+            text = re.sub(
+                "([\u0621-\u063A\u0641-\u064A\u066A-\u066C\u0654-\u0655]+)(\d+)",
+                r" \1 \2 ",
+                text,
+            )
+
+        # remove unwanted characters
+        if self.keep_emojis:
+            emoji_regex = "".join(list(emoji.UNICODE_EMOJI["en"].keys()))
+            rejected_chars_regex2 = "[^%s%s]" % (chars_regexv2, emoji_regex)
+            text = re.sub(rejected_chars_regex2, " ", text)
+        else:
+            text = re.sub(rejected_chars_regexv2, " ", text)
+
+        # remove extra spaces
+        text = " ".join(text.replace("\uFE0F", "").split())
+
+        if self.apply_farasa_segmentation:
+            if self.keep_emojis:
+                new_text = []
+                for word in text.split():
+                    if word in list(emoji.UNICODE_EMOJI["en"].keys()):
+                        new_text.append(word)
+                    else:
+                        new_text.append(farasa_segmenter.segment(word))
+                text = " ".join(new_text)
+            else:
+                text = farasa_segmenter.segment(text)
+            return self._farasa_segment(text)
+
+        # ALl the other models dont require Farasa Segmentation
+        return text
+
+    def _preprocess_v2(self, text: str) -> str:
+        text = str(text)
+        text = html.unescape(text)
+        if self.strip_tashkeel:
+            text = araby.strip_tashkeel(text)
+        if self.strip_tatweel:
+            text = araby.strip_tatweel(text)
+
+        if self.replace_urls_emails_mentions:
+            # replace all possible URLs
+            for reg in url_regexes:
+                text = re.sub(reg, " [رابط] ", text)
+            # REplace Emails with [بريد]
+            for reg in email_regexes:
+                text = re.sub(reg, " [بريد] ", text)
+            # replace mentions with [مستخدم]
+            text = re.sub(user_mention_regex, " [مستخدم] ", text)
+
+        if self.remove_html_markup:
+            # remove html line breaks
+            text = re.sub("<br />", " ", text)
+            # remove html markup
+            text = re.sub("</?[^>]+>", " ", text)
+
+        if self.map_hindi_numbers_to_arabic:
+            text = text.translate(hindi_to_arabic_map)
+
+        # remove repeated characters >2
+        if self.remove_non_digit_repetition:
+            text = self._remove_non_digit_repetition(text)
+
+        # insert whitespace before and after all non Arabic digits or English Digits and Alphabet and the 2 brackets
+        if self.insert_white_spaces:
+            text = re.sub(
+                "([^0-9\u0621-\u063A\u0641-\u064A\u0660-\u0669a-zA-Z\[\]])",
+                r" \1 ",
+                text,
+            )
+
+            # insert whitespace between words and numbers or numbers and words
+            text = re.sub(
+                "(\d+)([\u0621-\u063A\u0641-\u064A\u0660-\u066C]+)", r" \1 \2 ", text
+            )
+            text = re.sub(
+                "([\u0621-\u063A\u0641-\u064A\u0660-\u066C]+)(\d+)", r" \1 \2 ", text
+            )
+
+        if self.replace_slash_with_dash:
+            text = text.replace("/", "-")
+
+        # remove unwanted characters
+        if self.keep_emojis:
+            emoji_regex = "".join(list(emoji.UNICODE_EMOJI["en"].keys()))
+            rejected_chars_regex2 = "[^%s%s]" % (chars_regex, emoji_regex)
+            text = re.sub(rejected_chars_regex2, " ", text)
+        else:
+            text = re.sub(rejected_chars_regex, " ", text)
+
+        # remove extra spaces
+        text = " ".join(text.replace("\uFE0F", "").split())
+
+        if (
+            self.model_name == "bert-base-arabertv2"
+            or self.model_name == "bert-large-arabertv2"
+        ):
+            if self.keep_emojis:
+                new_text = []
+                for word in text.split():
+                    if word in list(emoji.UNICODE_EMOJI["en"].keys()):
+                        new_text.append(word)
+                    else:
+                        new_text.append(farasa_segmenter.segment(word))
+                text = " ".join(new_text)
+            else:
+                text = farasa_segmenter.segment(text)
+            return self._farasa_segment(text)
+
+        # ALl the other models dont require Farasa Segmentation
+        return text
+
+    def _preprocess_v1(self, text: str, do_farasa_tokenization: bool) -> str:
+        """
+        AraBERTv1 preprocessing Function
+        """
+        text = str(text)
+        if self.strip_tashkeel:
+            text = araby.strip_tashkeel(text)
+
+        text = re.sub(r"\d+\/[ء-ي]+\/\d+\]", "", text)
+        text = re.sub("ـ", "", text)
+        text = re.sub("[«»]", ' " ', text)
+
+        if self.replace_urls_emails_mentions:
+            # replace the [رابط] token with space if you want to clean links
+            text = re.sub(regex_url_step1, "[رابط]", text)
+            text = re.sub(regex_url_step2, "[رابط]", text)
+            text = re.sub(regex_url, "[رابط]", text)
+            text = re.sub(regex_email, "[بريد]", text)
+            text = re.sub(regex_mention, "[مستخدم]", text)
+        text = re.sub("…", r"\.", text).strip()
+        text = self._remove_redundant_punct(text)
+
+        if self.replace_urls_emails_mentions:
+            text = re.sub(r"\[ رابط \]|\[ رابط\]|\[رابط \]", " [رابط] ", text)
+            text = re.sub(r"\[ بريد \]|\[ بريد\]|\[بريد \]", " [بريد] ", text)
+            text = re.sub(r"\[ مستخدم \]|\[ مستخدم\]|\[مستخدم \]", " [مستخدم] ", text)
+
+        if self.remove_non_digit_repetition:
+            text = self._remove_non_digit_repetition(text)
+
+        if self.insert_white_spaces:
+            text = re.sub(
+                "([^0-9\u0621-\u063A\u0641-\u0669\u0671-\u0673a-zA-Z\[\]])",
+                r" \1 ",
+                text,
+            )
+        if do_farasa_tokenization:
+            text = self._tokenize_arabic_words_farasa(text)
+
+        text = " ".join(text.split())
+
+        return text
+
+    def _farasa_segment(self, text: str) -> str:
+        line_farasa = text.split()
+        segmented_line = []
+        for index, word in enumerate(line_farasa):
+            if word in ["[", "]"]:
+                continue
+            if word in ["رابط", "بريد", "مستخدم"] and line_farasa[index - 1] in [
+                "[",
+                "]",
+            ]:
+                segmented_line.append("[" + word + "]")
+                continue
+            if "+" not in word:
+                segmented_line.append(word)
+                continue
+            segmented_word = self._split_farasa_output(word)
+            segmented_line.extend(segmented_word)
+
+        return " ".join(segmented_line)
+
+    def _split_farasa_output(self, word: str) -> str:
+        segmented_word = []
+        temp_token = ""
+        for i, c in enumerate(word):
+            if c == "+":
+                # if the token is KAF, it could be a suffix or prefix
+                if temp_token == "ك":
+                    # if we are at the second token, then KAF is surely a prefix
+                    if i == 1:
+                        segmented_word.append(temp_token + "+")
+                        temp_token = ""
+                    # If the KAF token is between 2 tokens
+                    elif word[i - 2] == "+":
+                        # if the previous token is prefix, then this KAF must be a prefix
+                        if segmented_word[-1][-1] == "+":
+                            segmented_word.append(temp_token + "+")
+                            temp_token = ""
+                        # else it is a suffix, this KAF could not be a second suffix
+                        else:
+                            segmented_word.append("+" + temp_token)
+                            temp_token = ""
+                    # if Kaf is at the end, this is handled with the statement after the loop
+                elif temp_token in prefix_list:
+                    segmented_word.append(temp_token + "+")
+                    temp_token = ""
+                elif temp_token in suffix_list:
+                    segmented_word.append("+" + temp_token)
+                    temp_token = ""
+                else:
+                    segmented_word.append(temp_token)
+                    temp_token = ""
+                continue
+            temp_token += c
+        if temp_token != "":
+            if temp_token in suffix_list:
+                segmented_word.append("+" + temp_token)
+            else:
+                segmented_word.append(temp_token)
+        return segmented_word
+
+    def _tokenize_arabic_words_farasa(self, line_input: str) -> str:
+
+        if self.keep_emojis:
+            # insert whitespace before and after all non Arabic digits or English Digits and Alphabet and the 2 brackets
+            line_farasa = []
+            for word in line_input.split():
+                if word in list(emoji.UNICODE_EMOJI["en"].keys()):
+                    line_farasa.append(word)
+                else:
+                    line_farasa.append(farasa_segmenter.segment(word))
+        else:
+            line_farasa = farasa_segmenter.segment(line_input).split()
+
+        segmented_line = []
+        for index, word in enumerate(line_farasa):
+            if word in ["[", "]"]:
+                continue
+            if word in ["رابط", "بريد", "مستخدم"] and line_farasa[index - 1] in [
+                "[",
+                "]",
+            ]:
+                segmented_line.append("[" + word + "]")
+                continue
+            segmented_word = []
+            for token in word.split("+"):
+                if token in prefix_list:
+                    segmented_word.append(token + "+")
+                elif token in suffix_list:
+                    segmented_word.append("+" + token)
+                else:
+                    segmented_word.append(token)
+            segmented_line.extend(segmented_word)
+        return " ".join(segmented_line)
+
+    def _remove_non_digit_repetition(self, text: str) -> str:
+        """
+        :param text:  the input text to remove elongation
+        :return: delongated text
+        """
+        # loop over the number of times the regex matched the text
+        # OLD
+        # for index_ in range(len(re.findall(regex_tatweel, text))):
+        #     elongation = re.search(regex_tatweel, text)
+        #     if elongation:
+        #         elongation_pattern = elongation.group()
+        #         elongation_replacement = elongation_pattern[0]
+        #         elongation_pattern = re.escape(elongation_pattern)
+        #         text = re.sub(
+        #             elongation_pattern, elongation_replacement, text, flags=re.MULTILINE
+        #         )
+        #     else:
+        #         break
+
+        # New
+        text = multiple_char_pattern.sub(r"\1\1", text)
+        return text
+
+    def _remove_redundant_punct(self, text: str) -> str:
+        text_ = text
+        result = re.search(redundant_punct_pattern, text)
+        dif = 0
+        while result:
+            sub = result.group()
+            sub = sorted(set(sub), key=sub.index)
+            sub = " " + "".join(list(sub)) + " "
+            text = "".join(
+                (text[: result.span()[0] + dif], sub, text[result.span()[1] + dif :])
+            )
+            text_ = "".join(
+                (text_[: result.span()[0]], text_[result.span()[1] :])
+            ).strip()
+            dif = abs(len(text) - len(text_))
+            result = re.search(redundant_punct_pattern, text_)
+        text = re.sub(r"\s+", " ", text)
+        return text.strip()
+
+
+prefix_list = [
+    "ال",
+    "و",
+    "ف",
+    "ب",
+    "ك",
+    "ل",
+    "لل",
+    "\u0627\u0644",
+    "\u0648",
+    "\u0641",
+    "\u0628",
+    "\u0643",
+    "\u0644",
+    "\u0644\u0644",
+    "س",
+]
+suffix_list = [
+    "ه",
+    "ها",
+    "ك",
+    "ي",
+    "هما",
+    "كما",
+    "نا",
+    "كم",
+    "هم",
+    "هن",
+    "كن",
+    "ا",
+    "ان",
+    "ين",
+    "ون",
+    "وا",
+    "ات",
+    "ت",
+    "ن",
+    "ة",
+    "\u0647",
+    "\u0647\u0627",
+    "\u0643",
+    "\u064a",
+    "\u0647\u0645\u0627",
+    "\u0643\u0645\u0627",
+    "\u0646\u0627",
+    "\u0643\u0645",
+    "\u0647\u0645",
+    "\u0647\u0646",
+    "\u0643\u0646",
+    "\u0627",
+    "\u0627\u0646",
+    "\u064a\u0646",
+    "\u0648\u0646",
+    "\u0648\u0627",
+    "\u0627\u062a",
+    "\u062a",
+    "\u0646",
+    "\u0629",
+]
+other_tokens = ["[رابط]", "[مستخدم]", "[بريد]"]
+
+# the never_split list is ussed with the transformers library
+prefix_symbols = [x + "+" for x in prefix_list]
+suffix_symblos = ["+" + x for x in suffix_list]
+never_split_tokens = list(set(prefix_symbols + suffix_symblos + other_tokens))
+
+url_regexes = [
+    r"(http(s)?:\/\/.)?(www\.)?[-a-zA-Z0-9@:%._\+~#=]{2,256}\.[a-z]{2,6}\b([-a-zA-Z0-9@:%_\+.~#?&//=]*)",
+    r"@(https?|ftp)://(-\.)?([^\s/?\.#-]+\.?)+(/[^\s]*)?$@iS",
+    r"http[s]?://[a-zA-Z0-9_\-./~\?=%&]+",
+    r"www[a-zA-Z0-9_\-?=%&/.~]+",
+    r"[a-zA-Z]+\.com",
+    r"(?=http)[^\s]+",
+    r"(?=www)[^\s]+",
+    r"://",
+]
+user_mention_regex = r"@[\w\d]+"
+email_regexes = [r"[\w-]+@([\w-]+\.)+[\w-]+", r"\S+@\S+"]
+redundant_punct_pattern = (
+    r"([!\"#\$%\'\(\)\*\+,\.:;\-<=·>?@\[\\\]\^_ـ`{\|}~—٪’،؟`୍“؛”ۚ【»؛\s+«–…‘]{2,})"
+)
+
+regex_tatweel = r"(\D)\1{2,}"
+multiple_char_pattern = re.compile(r"(\D)\1{2,}", re.DOTALL)
+
+rejected_chars_regex = r"[^0-9\u0621-\u063A\u0640-\u066C\u0671-\u0674a-zA-Z\[\]!\"#\$%\'\(\)\*\+,\.:;\-<=·>?@\[\\\]\^_ـ`{\|}~—٪’،؟`୍“؛”ۚ»؛\s+«–…‘]"
+rejected_chars_regexv2 = r"[^0-9\u0621-\u063A\u0641-\u066C\u0671-\u0674a-zA-Z\[\]!\"#\$%\'\(\)\*\+,\.:;\-<=·>?@\[\\\]\^_ـ`{\|}~—٪’،؟`୍“؛”ۚ»؛\s+«–…‘/]"
+
+regex_url_step1 = r"(?=http)[^\s]+"
+regex_url_step2 = r"(?=www)[^\s]+"
+regex_url = r"(http(s)?:\/\/.)?(www\.)?[-a-zA-Z0-9@:%._\+~#=]{2,256}\.[a-z]{2,6}\b([-a-zA-Z0-9@:%_\+.~#?&//=]*)"
+regex_mention = r"@[\w\d]+"
+regex_email = r"\S+@\S+"
+
+chars_regex = r"0-9\u0621-\u063A\u0640-\u066C\u0671-\u0674a-zA-Z\[\]!\"#\$%\'\(\)\*\+,\.:;\-<=·>?@\[\\\]\^_ـ`{\|}~—٪’،؟`୍“؛”ۚ»؛\s+«–…‘"
+chars_regexv2 = r"0-9\u0621-\u063A\u0640-\u066C\u0671-\u0674a-zA-Z\[\]!\"#\$%\'\(\)\*\+,\.:;\-<=·>?@\[\\\]\^_ـ`{\|}~—٪’،؟`୍“؛”ۚ»؛\s+«–…‘/"
+
+white_spaced_double_quotation_regex = r'\"\s+([^"]+)\s+\"'
+white_spaced_single_quotation_regex = r"\'\s+([^']+)\s+\'"
+white_spaced_back_quotation_regex = r"\`\s+([^`]+)\s+\`"
+white_spaced_em_dash = r"\—\s+([^—]+)\s+\—"
+
+left_spaced_chars = r" ([\]!#\$%\),\.:;\?}٪’،؟”؛…»·])"
+right_spaced_chars = r"([\[\(\{“«‘*\~]) "
+left_and_right_spaced_chars = r" ([\+\-\<\=\>\@\\\^\_\|\–]) "
+
+hindi_nums = "٠١٢٣٤٥٦٧٨٩"
+arabic_nums = "0123456789"
+hindi_to_arabic_map = str.maketrans(hindi_nums, arabic_nums)

backend/processor.py

@@ -0,0 +1,183 @@
+import streamlit as st
+import awesome_streamlit as ast
+from .preprocess import (
+    ArabertPreprocessor,
+    white_spaced_back_quotation_regex,
+    white_spaced_double_quotation_regex,
+    white_spaced_em_dash,
+    white_spaced_single_quotation_regex,
+    left_and_right_spaced_chars,
+    left_spaced_chars,
+    right_spaced_chars,
+)
+import re
+
+MODELS_to_SELECT = [
+    "None",
+    "bert-base-arabertv01",
+    "bert-base-arabert",
+    "bert-base-arabertv02",
+    "bert-base-arabertv2",
+    "bert-large-arabertv02",
+    "bert-large-arabertv2",
+    "araelectra-base",
+    "araelectra-base-discriminator",
+    "araelectra-base-generator",
+    "araelectra-base-artydiqa",
+    "aragpt2-base",
+    "aragpt2-medium",
+    "aragpt2-large",
+    "aragpt2-mega",
+]
+
+
+def unpreprocess(text: str) -> str:
+    """Re-formats the text to a classic format where punctuations, brackets, parenthesis are not seperated by whitespaces.
+    The objective is to make the generated text of any model appear natural and not preprocessed.
+
+    Args:
+        text (:obj:`str`): input text to be un-preprocessed
+        desegment (:obj:`bool`, optional): [whether or not to remove farasa pre-segmentation before]..
+
+    Returns:
+        str: The unpreprocessed (and possibly Farasa-desegmented) text.
+    """
+
+    text = desegment(text)
+
+    # removes the spaces around quotation marks ex: i " ate " an apple --> i "ate" an apple
+    # https://stackoverflow.com/a/53436792/5381220
+    text = re.sub(white_spaced_double_quotation_regex, '"' + r"\1" + '"', text)
+    text = re.sub(white_spaced_single_quotation_regex, "'" + r"\1" + "'", text)
+    text = re.sub(white_spaced_back_quotation_regex, "\`" + r"\1" + "\`", text)
+    text = re.sub(white_spaced_back_quotation_regex, "\—" + r"\1" + "\—", text)
+
+    # during generation, sometimes the models don't put a space after the dot, this handles it
+    text = text.replace(".", " . ")
+    text = " ".join(text.split())
+
+    # handle decimals
+    text = re.sub(r"(\d+) \. (\d+)", r"\1.\2", text)
+    text = re.sub(r"(\d+) \, (\d+)", r"\1,\2", text)
+
+    text = re.sub(left_and_right_spaced_chars, r"\1", text)
+    text = re.sub(left_spaced_chars, r"\1", text)
+    text = re.sub(right_spaced_chars, r"\1", text)
+
+    return text
+
+
+def desegment(text: str) -> str:
+    """
+    Use this function if sentence tokenization was done using
+    `from arabert.preprocess_arabert import preprocess` with Farasa enabled
+    AraBERT segmentation using Farasa adds a space after the '+' for prefixes,
+    and after before the '+' for suffixes
+
+    Example:
+    >>> desegment('ال+ دراس +ات')
+    الدراسات
+    """
+    text = text.replace("+ ", "+")
+    text = text.replace(" +", "+")
+    text = " ".join([_desegmentword(word) for word in text.split(" ")])
+    return text
+
+
+def _desegmentword(orig_word: str) -> str:
+    """
+    Word segmentor that takes a Farasa Segmented Word and removes the '+' signs
+
+    Example:
+    >>> _desegmentword("ال+يومي+ة")
+    اليومية
+    """
+    word = orig_word.replace("ل+ال+", "لل")
+    if "ال+ال" not in orig_word:
+        word = word.replace("ل+ال", "لل")
+    word = word.replace("+", "")
+    word = word.replace("للل", "لل")
+    return word
+
+
+def write():
+
+    st.markdown(
+        """
+        <h1 style="text-align:left;">Arabic Text Pre-Processor</h1>
+        """,
+        unsafe_allow_html=True,
+    )
+    st.markdown(
+        """
+        <style>
+        p, div, input, label {
+        text-align: right;
+        }
+        </style>
+        """,
+        unsafe_allow_html=True,
+    )
+    input_text = st.text_input(
+        "Text to Pre-Process",
+        value="ولن نبالغ إذا قلنا: إن 'هاتف' أو 'كمبيوتر المكتب' في زمننا هذا ضروري",
+    )
+
+    st.sidebar.title("Model Selector")
+    model_selector = st.sidebar.selectbox(
+        """Select None to enable further filters""", options=MODELS_to_SELECT, index=3
+    )
+    if model_selector == "None":
+        keep_emojis = st.sidebar.checkbox("Keep emojis", False)
+        remove_html_markup = st.sidebar.checkbox("Remove html markup", True)
+        strip_tashkeel = st.sidebar.checkbox("Strip tashkeel", True)
+        replace_urls_emails_mentions = st.sidebar.checkbox(
+            "Replace urls and emails", True
+        )
+        strip_tatweel = st.sidebar.checkbox("Strip tatweel", True)
+        insert_white_spaces = st.sidebar.checkbox("Insert white spaces", True)
+        remove_non_digit_repetition = st.sidebar.checkbox(
+            "Remove non-digit repetition", True
+        )
+        replace_slash_with_dash = st.sidebar.checkbox("Replace slash with dash", None)
+        map_hindi_numbers_to_arabic = st.sidebar.checkbox(
+            "Map hindi numbers to arabic", None
+        )
+        apply_farasa_segmentation = st.sidebar.checkbox(
+            "Apply farasa segmentation", None
+        )
+
+    run_preprocessor = st.button("Run Pre-Processor")
+
+    prep_text = None
+    if run_preprocessor:
+        if model_selector == "None":
+            arabert_preprocessor = ArabertPreprocessor(
+                model_selector,
+                keep_emojis,
+                remove_html_markup,
+                replace_urls_emails_mentions,
+                strip_tashkeel,
+                strip_tatweel,
+                insert_white_spaces,
+                remove_non_digit_repetition,
+                replace_slash_with_dash,
+                map_hindi_numbers_to_arabic,
+                apply_farasa_segmentation,
+            )
+        else:
+            arabert_preprocessor = ArabertPreprocessor(model_name=model_selector)
+        prep_text = arabert_preprocessor._preprocess_v3(input_text)
+        st.write(prep_text)
+
+    st.write("-----")
+    input_text_unprep = st.text_input(
+        "Text to Undo the Pre-Processing",
+        value=prep_text
+        if prep_text
+        else "و+ لن نبالغ إذا قل +نا : إن ' هاتف ' أو ' كمبيوتر ال+ مكتب ' في زمن +نا هذا ضروري",
+    )
+    run_unpreprocessor = st.button("Run Un-Pre-Processor")
+
+    if run_unpreprocessor:
+        st.write(unpreprocess(input_text_unprep))

backend/qa.py

@@ -0,0 +1,45 @@
+import streamlit as st
+
+from .qa_utils import annotate_answer
+from .services import get_qa_answers
+
+
+def write():
+    _, col1, _ = st.columns(3)
+
+    with col1:
+        st.title("Ask any question!")
+
+    st.markdown(
+        """
+    <style>
+    p, div, input, label {
+    text-align: right;
+    }
+    </style>
+        """,
+        unsafe_allow_html=True,
+    )
+
+    st.sidebar.write("\n")
+    n_answers = st.sidebar.slider(
+        "Max. number of answers", min_value=1, max_value=10, value=2, step=1
+    )
+
+    question = st.text_input("", value="من هو جو بايدن؟")
+    if "؟" not in question:
+        question += "؟"
+
+    run_query = st.button("Find answers")
+    if run_query:
+        # https://discuss.streamlit.io/t/showing-a-gif-while-st-spinner-runs/5084
+        with st.spinner("Searching..."):
+            results_dict = get_qa_answers(question)
+
+        if len(results_dict) > 0:
+            st.write("## Answers:")
+            for result in results_dict["results"][:n_answers]:
+                annotate_answer(result)
+                f"[**Source**](<{result['link']}>)"
+        else:
+            st.write("## 😞 No results found.")

backend/qa_utils.py

@@ -0,0 +1,163 @@
+import streamlit.components.v1
+
+from htbuilder import HtmlElement, div, span, styles
+from htbuilder.units import px, rem, em
+
+
+def annotation(body, label="", background="#ddd", color="#333", **style):
+    """Build an HtmlElement span object with the given body and annotation label.
+
+    The end result will look something like this:
+
+        [body | label]
+
+    Parameters
+    ----------
+    body : string
+        The string to put in the "body" part of the annotation.
+    label : string
+        The string to put in the "label" part of the annotation.
+    background : string
+        The color to use for the background "chip" containing this annotation.
+    color : string
+        The color to use for the body and label text.
+    **style : dict
+        Any CSS you want to use to customize the containing "chip".
+
+    Examples
+    --------
+
+    Produce a simple annotation with default colors:
+
+    >>> annotation("apple", "fruit")
+
+    Produce an annotation with custom colors:
+
+    >>> annotation("apple", "fruit", background="#FF0", color="black")
+
+    Produce an annotation with crazy CSS:
+
+    >>> annotation("apple", "fruit", background="#FF0", border="1px dashed red")
+
+    """
+
+    if "font_family" not in style:
+        style["font_family"] = "sans-serif"
+
+    return span(
+        style=styles(
+            background=background,
+            border_radius=rem(0.33),
+            color=color,
+            padding=(rem(0.17), rem(0.67)),
+            display="inline-flex",
+            justify_content="center",
+            align_items="center",
+            **style,
+        )
+    )(
+        body,
+        span(
+            style=styles(
+                color=color,
+                font_size=em(0.67),
+                opacity=0.5,
+                padding_left=rem(0.5),
+                text_transform="uppercase",
+                margin_bottom=px(-2),
+            )
+        )(label),
+    )
+
+
+def annotated_text(*args, **kwargs):
+    """Writes test with annotations into your Streamlit app.
+
+    Parameters
+    ----------
+    *args : str, tuple or htbuilder.HtmlElement
+        Arguments can be:
+        - strings, to draw the string as-is on the screen.
+        - tuples of the form (main_text, annotation_text, background, color) where
+          background and foreground colors are optional and should be an CSS-valid string such as
+          "#aabbcc" or "rgb(10, 20, 30)"
+        - HtmlElement objects in case you want to customize the annotations further. In particular,
+          you can import the `annotation()` function from this module to easily produce annotations
+          whose CSS you can customize via keyword arguments.
+
+    Examples
+    --------
+
+    >>> annotated_text(
+    ...     "This ",
+    ...     ("is", "verb", "#8ef"),
+    ...     " some ",
+    ...     ("annotated", "adj", "#faa"),
+    ...     ("text", "noun", "#afa"),
+    ...     " for those of ",
+    ...     ("you", "pronoun", "#fea"),
+    ...     " who ",
+    ...     ("like", "verb", "#8ef"),
+    ...     " this sort of ",
+    ...     ("thing", "noun", "#afa"),
+    ... )
+
+    >>> annotated_text(
+    ...     "Hello ",
+    ...     annotation("world!", "noun", color="#8ef", border="1px dashed red"),
+    ... )
+
+    """
+    out = div(
+        style=styles(
+            font_family="sans-serif",
+            line_height="1.45",
+            font_size=px(16),
+            text_align="right",
+        )
+    )
+
+    for arg in args:
+        if isinstance(arg, str):
+            out(arg)
+
+        elif isinstance(arg, HtmlElement):
+            out(arg)
+
+        elif isinstance(arg, tuple):
+            out(annotation(*arg))
+
+        else:
+            raise Exception("Oh noes!")
+
+    streamlit.components.v1.html(str(out), **kwargs)
+
+
+def shorten_text(text, n, reverse=False):
+    if text.isspace() or text == "":
+        return text
+    if reverse:
+        text = text[::-1]
+    words = iter(text.split())
+    lines, current = [], next(words)
+    for word in words:
+        if len(current) + 1 + len(word) > n:
+            break
+        else:
+            current += " " + word
+    lines.append(current)
+    if reverse:
+        return lines[0][::-1]
+    return lines[0]
+
+
+def annotate_answer(result):
+    annotated_text(
+        shorten_text(
+            result["original"][: result["new_start"]],
+            500,
+            reverse=True,
+        ),
+        (result["new_answer"], "جواب", "#8ef"),
+        shorten_text(result["original"][result["new_end"] :], 500) + " ...... إلخ",
+    )

backend/sa.py

@@ -0,0 +1,36 @@
+import streamlit as st
+from .services import SentimentAnalyzer
+from functools import lru_cache
+
+# @st.cache(allow_output_mutation=False, hash_funcs={Tokenizer: str})
+@lru_cache(maxsize=1)
+def load_text_generator():
+    predictor = SentimentAnalyzer()
+    return predictor
+
+
+predictor = load_text_generator()
+
+
+def write():
+    st.markdown(
+        """
+        # Arabic Sentiment Analysis
+
+        """
+    )
+    
+    input_text = st.text_input(
+        "Enter your text here:",
+    )
+    if st.button("Predict"):
+        with st.spinner("Predicting..."):
+            prediction, score, all_score = predictor.predict([input_text])
+            st.write(f"Result: {prediction[0]}")
+            detailed_score = {
+                "Positive": all_score[0][0],
+                "Neutral": all_score[0][1],
+                "Negative": all_score[0][2],
+            }
+            st.write("All scores:")
+            st.write(detailed_score)

backend/sa_utils.py

@@ -0,0 +1,510 @@
+import re
+from contextlib import contextmanager
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from fuzzysearch import find_near_matches
+from pyarabic import araby
+from torch import nn
+from transformers import AutoTokenizer, BertModel, BertPreTrainedModel, pipeline
+from transformers.modeling_outputs import SequenceClassifierOutput
+
+from .preprocess import ArabertPreprocessor, url_regexes, user_mention_regex
+
+multiple_char_pattern = re.compile(r"(.)\1{2,}", re.DOTALL)
+
+# ASAD-NEW_AraBERT_PREP-Balanced
+class NewArabicPreprocessorBalanced(ArabertPreprocessor):
+    def __init__(
+        self,
+        model_name: str,
+        keep_emojis: bool = False,
+        remove_html_markup: bool = True,
+        replace_urls_emails_mentions: bool = True,
+        strip_tashkeel: bool = True,
+        strip_tatweel: bool = True,
+        insert_white_spaces: bool = True,
+        remove_non_digit_repetition: bool = True,
+        replace_slash_with_dash: bool = None,
+        map_hindi_numbers_to_arabic: bool = None,
+        apply_farasa_segmentation: bool = None,
+    ):
+        if "UBC-NLP" in model_name or "CAMeL-Lab" in model_name:
+            keep_emojis = True
+            remove_non_digit_repetition = True
+        super().__init__(
+            model_name=model_name,
+            keep_emojis=keep_emojis,
+            remove_html_markup=remove_html_markup,
+            replace_urls_emails_mentions=replace_urls_emails_mentions,
+            strip_tashkeel=strip_tashkeel,
+            strip_tatweel=strip_tatweel,
+            insert_white_spaces=insert_white_spaces,
+            remove_non_digit_repetition=remove_non_digit_repetition,
+            replace_slash_with_dash=replace_slash_with_dash,
+            map_hindi_numbers_to_arabic=map_hindi_numbers_to_arabic,
+            apply_farasa_segmentation=apply_farasa_segmentation,
+        )
+        self.true_model_name = model_name
+
+    def preprocess(self, text):
+        if "UBC-NLP" in self.true_model_name:
+            return self.ubc_prep(text)
+
+    def ubc_prep(self, text):
+        text = re.sub("\s", " ", text)
+        text = text.replace("\\n", " ")
+        text = text.replace("\\r", " ")
+        text = araby.strip_tashkeel(text)
+        text = araby.strip_tatweel(text)
+        # replace all possible URLs
+        for reg in url_regexes:
+            text = re.sub(reg, " URL ", text)
+        text = re.sub("(URL\s*)+", " URL ", text)
+        # replace mentions with USER
+        text = re.sub(user_mention_regex, " USER ", text)
+        text = re.sub("(USER\s*)+", " USER ", text)
+        # replace hashtags with HASHTAG
+        # text = re.sub(r"#[\w\d]+", " HASH TAG ", text)
+        text = text.replace("#", " HASH ")
+        text = text.replace("_", " ")
+        text = " ".join(text.split())
+        # text = re.sub("\B\\[Uu]\w+", "", text)
+        text = text.replace("\\U0001f97a", "🥺")
+        text = text.replace("\\U0001f928", "🤨")
+        text = text.replace("\\U0001f9d8", "😀")
+        text = text.replace("\\U0001f975", "😥")
+        text = text.replace("\\U0001f92f", "😲")
+        text = text.replace("\\U0001f92d", "🤭")
+        text = text.replace("\\U0001f9d1", "😐")
+        text = text.replace("\\U000e0067", "")
+        text = text.replace("\\U000e006e", "")
+        text = text.replace("\\U0001f90d", "♥")
+        text = text.replace("\\U0001f973", "🎉")
+        text = text.replace("\\U0001fa79", "")
+        text = text.replace("\\U0001f92b", "🤐")
+        text = text.replace("\\U0001f9da", "🦋")
+        text = text.replace("\\U0001f90e", "♥")
+        text = text.replace("\\U0001f9d0", "🧐")
+        text = text.replace("\\U0001f9cf", "")
+        text = text.replace("\\U0001f92c", "😠")
+        text = text.replace("\\U0001f9f8", "😸")
+        text = text.replace("\\U0001f9b6", "💩")
+        text = text.replace("\\U0001f932", "🤲")
+        text = text.replace("\\U0001f9e1", "🧡")
+        text = text.replace("\\U0001f974", "☹")
+        text = text.replace("\\U0001f91f", "")
+        text = text.replace("\\U0001f9fb", "💩")
+        text = text.replace("\\U0001f92a", "🤪")
+        text = text.replace("\\U0001f9fc", "")
+        text = text.replace("\\U000e0065", "")
+        text = text.replace("\\U0001f92e", "💩")
+        text = text.replace("\\U000e007f", "")
+        text = text.replace("\\U0001f970", "🥰")
+        text = text.replace("\\U0001f929", "🤩")
+        text = text.replace("\\U0001f6f9", "")
+        text = text.replace("🤍", "♥")
+        text = text.replace("🦠", "😷")
+        text = text.replace("🤢", "مقرف")
+        text = text.replace("🤮", "مقرف")
+        text = text.replace("🕠", "⌚")
+        text = text.replace("🤬", "😠")
+        text = text.replace("🤧", "😷")
+        text = text.replace("🥳", "🎉")
+        text = text.replace("🥵", "🔥")
+        text = text.replace("🥴", "☹")
+        text = text.replace("🤫", "🤐")
+        text = text.replace("🤥", "كذاب")
+        text = text.replace("\\u200d", " ")
+        text = text.replace("u200d", " ")
+        text = text.replace("\\u200c", " ")
+        text = text.replace("u200c", " ")
+        text = text.replace('"', "'")
+        text = text.replace("\\xa0", "")
+        text = text.replace("\\u2066", " ")
+        text = re.sub("\B\\\[Uu]\w+", "", text)
+        text = super(NewArabicPreprocessorBalanced, self).preprocess(text)
+
+        text = " ".join(text.split())
+        return text
+
+
+"""CNNMarbertArabicPreprocessor"""
+# ASAD-CNN_MARBERT
+class CNNMarbertArabicPreprocessor(ArabertPreprocessor):
+    def __init__(
+        self,
+        model_name,
+        keep_emojis=False,
+        remove_html_markup=True,
+        replace_urls_emails_mentions=True,
+        remove_elongations=True,
+    ):
+        if "UBC-NLP" in model_name or "CAMeL-Lab" in model_name:
+            keep_emojis = True
+            remove_elongations = False
+        super().__init__(
+            model_name,
+            keep_emojis,
+            remove_html_markup,
+            replace_urls_emails_mentions,
+            remove_elongations,
+        )
+        self.true_model_name = model_name
+
+    def preprocess(self, text):
+        if "UBC-NLP" in self.true_model_name:
+            return self.ubc_prep(text)
+
+    def ubc_prep(self, text):
+        text = re.sub("\s", " ", text)
+        text = text.replace("\\n", " ")
+        text = araby.strip_tashkeel(text)
+        text = araby.strip_tatweel(text)
+        # replace all possible URLs
+        for reg in url_regexes:
+            text = re.sub(reg, " URL ", text)
+        text = re.sub("(URL\s*)+", " URL ", text)
+        # replace mentions with USER
+        text = re.sub(user_mention_regex, " USER ", text)
+        text = re.sub("(USER\s*)+", " USER ", text)
+        # replace hashtags with HASHTAG
+        # text = re.sub(r"#[\w\d]+", " HASH TAG ", text)
+        text = text.replace("#", " HASH ")
+        text = text.replace("_", " ")
+        text = " ".join(text.split())
+        text = super(CNNMarbertArabicPreprocessor, self).preprocess(text)
+        text = text.replace("\u200d", " ")
+        text = text.replace("u200d", " ")
+        text = text.replace("\u200c", " ")
+        text = text.replace("u200c", " ")
+        text = text.replace('"', "'")
+        # text = re.sub('[\d\.]+', ' NUM ', text)
+        # text = re.sub('(NUM\s*)+', ' NUM ', text)
+        text = multiple_char_pattern.sub(r"\1\1", text)
+        text = " ".join(text.split())
+        return text
+
+
+"""Trial5ArabicPreprocessor"""
+
+
+class Trial5ArabicPreprocessor(ArabertPreprocessor):
+    def __init__(
+        self,
+        model_name,
+        keep_emojis=False,
+        remove_html_markup=True,
+        replace_urls_emails_mentions=True,
+    ):
+        if "UBC-NLP" in model_name:
+            keep_emojis = True
+        super().__init__(
+            model_name, keep_emojis, remove_html_markup, replace_urls_emails_mentions
+        )
+        self.true_model_name = model_name
+
+    def preprocess(self, text):
+        if "UBC-NLP" in self.true_model_name:
+            return self.ubc_prep(text)
+
+    def ubc_prep(self, text):
+        text = re.sub("\s", " ", text)
+        text = text.replace("\\n", " ")
+        text = araby.strip_tashkeel(text)
+        text = araby.strip_tatweel(text)
+        # replace all possible URLs
+        for reg in url_regexes:
+            text = re.sub(reg, " URL ", text)
+        # replace mentions with USER
+        text = re.sub(user_mention_regex, " USER ", text)
+        # replace hashtags with HASHTAG
+        # text = re.sub(r"#[\w\d]+", " HASH TAG ", text)
+        text = text.replace("#", " HASH TAG ")
+        text = text.replace("_", " ")
+        text = " ".join(text.split())
+        text = super(Trial5ArabicPreprocessor, self).preprocess(text)
+        # text = text.replace("السلام عليكم"," ")
+        # text = text.replace(find_near_matches("السلام عليكم",text,max_deletions=3,max_l_dist=3)[0].matched," ")
+        return text
+
+
+"""SarcasmArabicPreprocessor"""
+
+
+class SarcasmArabicPreprocessor(ArabertPreprocessor):
+    def __init__(
+        self,
+        model_name,
+        keep_emojis=False,
+        remove_html_markup=True,
+        replace_urls_emails_mentions=True,
+    ):
+        if "UBC-NLP" in model_name:
+            keep_emojis = True
+        super().__init__(
+            model_name, keep_emojis, remove_html_markup, replace_urls_emails_mentions
+        )
+        self.true_model_name = model_name
+
+    def preprocess(self, text):
+        if "UBC-NLP" in self.true_model_name:
+            return self.ubc_prep(text)
+        else:
+            return super(SarcasmArabicPreprocessor, self).preprocess(text)
+
+    def ubc_prep(self, text):
+        text = re.sub("\s", " ", text)
+        text = text.replace("\\n", " ")
+        text = araby.strip_tashkeel(text)
+        text = araby.strip_tatweel(text)
+        # replace all possible URLs
+        for reg in url_regexes:
+            text = re.sub(reg, " URL ", text)
+        # replace mentions with USER
+        text = re.sub(user_mention_regex, " USER ", text)
+        # replace hashtags with HASHTAG
+        # text = re.sub(r"#[\w\d]+", " HASH TAG ", text)
+        text = text.replace("#", " HASH TAG ")
+        text = text.replace("_", " ")
+        text = text.replace('"', " ")
+        text = " ".join(text.split())
+        text = super(SarcasmArabicPreprocessor, self).preprocess(text)
+        return text
+
+
+"""NoAOAArabicPreprocessor"""
+
+
+class NoAOAArabicPreprocessor(ArabertPreprocessor):
+    def __init__(
+        self,
+        model_name,
+        keep_emojis=False,
+        remove_html_markup=True,
+        replace_urls_emails_mentions=True,
+    ):
+        if "UBC-NLP" in model_name:
+            keep_emojis = True
+        super().__init__(
+            model_name, keep_emojis, remove_html_markup, replace_urls_emails_mentions
+        )
+        self.true_model_name = model_name
+
+    def preprocess(self, text):
+        if "UBC-NLP" in self.true_model_name:
+            return self.ubc_prep(text)
+        else:
+            return super(NoAOAArabicPreprocessor, self).preprocess(text)
+
+    def ubc_prep(self, text):
+        text = re.sub("\s", " ", text)
+        text = text.replace("\\n", " ")
+        text = araby.strip_tashkeel(text)
+        text = araby.strip_tatweel(text)
+        # replace all possible URLs
+        for reg in url_regexes:
+            text = re.sub(reg, " URL ", text)
+        # replace mentions with USER
+        text = re.sub(user_mention_regex, " USER ", text)
+        # replace hashtags with HASHTAG
+        # text = re.sub(r"#[\w\d]+", " HASH TAG ", text)
+        text = text.replace("#", " HASH TAG ")
+        text = text.replace("_", " ")
+        text = " ".join(text.split())
+        text = super(NoAOAArabicPreprocessor, self).preprocess(text)
+        text = text.replace("السلام عليكم", " ")
+        text = text.replace("ورحمة الله وبركاته", " ")
+        matched = find_near_matches("السلام عليكم", text, max_deletions=3, max_l_dist=3)
+        if len(matched) > 0:
+            text = text.replace(matched[0].matched, " ")
+        matched = find_near_matches(
+            "ورحمة الله وبركاته", text, max_deletions=3, max_l_dist=3
+        )
+        if len(matched) > 0:
+            text = text.replace(matched[0].matched, " ")
+        return text
+
+
+class CnnBertForSequenceClassification(BertPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.config = config
+
+        self.bert = BertModel(config)
+
+        filter_sizes = [1, 2, 3, 4, 5]
+        num_filters = 32
+        self.convs1 = nn.ModuleList(
+            [nn.Conv2d(4, num_filters, (K, config.hidden_size)) for K in filter_sizes]
+        )
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.classifier = nn.Linear(len(filter_sizes) * num_filters, config.num_labels)
+
+        self.init_weights()
+
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        labels=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        outputs = self.bert(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        x = outputs[2][-4:]
+
+        x = torch.stack(x, dim=1)
+        x = [F.relu(conv(x)).squeeze(3) for conv in self.convs1]
+        x = [F.max_pool1d(i, i.size(2)).squeeze(2) for i in x]
+        x = torch.cat(x, 1)
+        x = self.dropout(x)
+        logits = self.classifier(x)
+
+        loss = None
+        if labels is not None:
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = "regression"
+                elif self.num_labels > 1 and (
+                    labels.dtype == torch.long or labels.dtype == torch.int
+                ):
+                    self.config.problem_type = "single_label_classification"
+                else:
+                    self.config.problem_type = "multi_label_classification"
+
+            if self.config.problem_type == "regression":
+                loss_fct = nn.MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(logits, labels)
+            elif self.config.problem_type == "single_label_classification":
+                loss_fct = nn.CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = nn.BCEWithLogitsLoss()
+                loss = loss_fct(logits, labels)
+        if not return_dict:
+            output = (logits,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return SequenceClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=None,
+            attentions=outputs.attentions,
+        )
+
+
+class CNNTextClassificationPipeline:
+    def __init__(self, model_path, device, return_all_scores=False):
+        self.model_path = model_path
+        self.model = CnnBertForSequenceClassification.from_pretrained(self.model_path)
+        # Special handling
+        self.device = torch.device("cpu" if device < 0 else f"cuda:{device}")
+        if self.device.type == "cuda":
+            self.model = self.model.to(self.device)
+        self.tokenizer = AutoTokenizer.from_pretrained(model_path)
+        self.return_all_scores = return_all_scores
+
+    @contextmanager
+    def device_placement(self):
+        """
+        Context Manager allowing tensor allocation on the user-specified device in framework agnostic way.
+        Returns:
+            Context manager
+        Examples::
+            # Explicitly ask for tensor allocation on CUDA device :0
+            pipe = pipeline(..., device=0)
+            with pipe.device_placement():
+                # Every framework specific tensor allocation will be done on the request device
+                output = pipe(...)
+        """
+
+        if self.device.type == "cuda":
+            torch.cuda.set_device(self.device)
+
+        yield
+
+    def ensure_tensor_on_device(self, **inputs):
+        """
+        Ensure PyTorch tensors are on the specified device.
+        Args:
+            inputs (keyword arguments that should be :obj:`torch.Tensor`): The tensors to place on :obj:`self.device`.
+        Return:
+            :obj:`Dict[str, torch.Tensor]`: The same as :obj:`inputs` but on the proper device.
+        """
+        return {
+            name: tensor.to(self.device) if isinstance(tensor, torch.Tensor) else tensor
+            for name, tensor in inputs.items()
+        }
+
+    def __call__(self, text):
+        """
+        Classify the text(s) given as inputs.
+        Args:
+            args (:obj:`str` or :obj:`List[str]`):
+                One or several texts (or one list of prompts) to classify.
+        Return:
+            A list or a list of list of :obj:`dict`: Each result comes as list of dictionaries with the following keys:
+            - **label** (:obj:`str`) -- The label predicted.
+            - **score** (:obj:`float`) -- The corresponding probability.
+            If ``self.return_all_scores=True``, one such dictionary is returned per label.
+        """
+        # outputs = super().__call__(*args, **kwargs)
+        inputs = self.tokenizer.batch_encode_plus(
+            text,
+            add_special_tokens=True,
+            max_length=64,
+            padding=True,
+            truncation="longest_first",
+            return_tensors="pt",
+        )
+
+        with torch.no_grad():
+            inputs = self.ensure_tensor_on_device(**inputs)
+            predictions = self.model(**inputs)[0].cpu()
+
+        predictions = predictions.numpy()
+
+        if self.model.config.num_labels == 1:
+            scores = 1.0 / (1.0 + np.exp(-predictions))
+        else:
+            scores = np.exp(predictions) / np.exp(predictions).sum(-1, keepdims=True)
+        if self.return_all_scores:
+            return [
+                [
+                    {"label": self.model.config.id2label[i], "score": score.item()}
+                    for i, score in enumerate(item)
+                ]
+                for item in scores
+            ]
+        else:
+            return [
+                {"label": self.inv_label_map[item.argmax()], "score": item.max().item()}
+                for item in scores
+            ]

backend/sarcasm.py

@@ -0,0 +1,21 @@
+import streamlit as st
+from .sa import predictor
+
+
+def write():
+    st.markdown(
+        """
+        # Arabic Sarcasm Detection
+
+        This is a simple sarcasm detection app that uses the [MARBERT](https://huggingface.co/UBC-NLP/MARBERT) model trained on [ArSarcasm](https://github.com/iabufarha/ArSarcasm)
+        """
+    )
+
+    input_text = st.text_input(
+        "Enter your text here:",
+    )
+    if st.button("Predict"):
+        with st.spinner("Predicting..."):
+            prediction, scores = predictor.get_preds_from_sarcasm([input_text])
+            st.write(f"Result: {prediction[0]}")
+            st.write(f"Score: {scores[0]}")

backend/services.py

@@ -0,0 +1,519 @@
+import json
+import logging
+import os
+from functools import lru_cache
+from typing import List
+from urllib.parse import unquote
+
+import more_itertools
+import pandas as pd
+import requests
+import streamlit as st
+import wikipedia
+from codetiming import Timer
+from fuzzysearch import find_near_matches
+from googleapi import google
+from tqdm.auto import tqdm
+from transformers import (
+    AutoTokenizer,
+    GPT2LMHeadModel,
+    GPT2Tokenizer,
+    pipeline,
+    set_seed,
+)
+
+from .modeling_gpt2 import GPT2LMHeadModel as GROVERLMHeadModel
+from .preprocess import ArabertPreprocessor
+from .sa_utils import *
+from .utils import download_models, softmax
+
+logger = logging.getLogger(__name__)
+# Taken and Modified from https://huggingface.co/spaces/flax-community/chef-transformer/blob/main/app.py
+class TextGeneration:
+    def __init__(self):
+        self.debug = False
+        self.generation_pipline = {}
+        self.preprocessor = ArabertPreprocessor(model_name="aragpt2-mega")
+        self.tokenizer = GPT2Tokenizer.from_pretrained(
+            "aubmindlab/aragpt2-mega", use_fast=False
+        )
+        self.tokenizer.pad_token = self.tokenizer.eos_token
+        self.API_KEY = os.getenv("API_KEY")
+        self.headers = {"Authorization": f"Bearer {self.API_KEY}"}
+        # self.model_names_or_paths = {
+        #     "aragpt2-medium": "D:/ML/Models/aragpt2-medium",
+        #     "aragpt2-base": "D:/ML/Models/aragpt2-base",
+        # }
+        self.model_names_or_paths = {
+            # "aragpt2-medium": "aubmindlab/aragpt2-medium",
+            "aragpt2-base": "aubmindlab/aragpt2-base",
+            # "aragpt2-large": "aubmindlab/aragpt2-large",
+            "aragpt2-mega": "aubmindlab/aragpt2-mega",
+        }
+        set_seed(42)
+
+    def load_pipeline(self):
+        for model_name, model_path in self.model_names_or_paths.items():
+            if "base" in model_name or "medium" in model_name:
+                self.generation_pipline[model_name] = pipeline(
+                    "text-generation",
+                    model=GPT2LMHeadModel.from_pretrained(model_path),
+                    tokenizer=self.tokenizer,
+                    device=-1,
+                )
+            else:
+                self.generation_pipline[model_name] = pipeline(
+                    "text-generation",
+                    model=GROVERLMHeadModel.from_pretrained(model_path),
+                    tokenizer=self.tokenizer,
+                    device=-1,
+                )
+
+    def load(self):
+        if not self.debug:
+            self.load_pipeline()
+
+    def generate(
+        self,
+        model_name,
+        prompt,
+        max_new_tokens: int,
+        temperature: float,
+        top_k: int,
+        top_p: float,
+        repetition_penalty: float,
+        no_repeat_ngram_size: int,
+        do_sample: bool,
+        num_beams: int,
+    ):
+        logger.info(f"Generating with {model_name}")
+        prompt = self.preprocessor.preprocess(prompt)
+        return_full_text = False
+        return_text = True
+        num_return_sequences = 1
+        pad_token_id = 0
+        eos_token_id = 0
+        input_tok = self.tokenizer.tokenize(prompt)
+        max_length = len(input_tok) + max_new_tokens
+        if max_length > 1024:
+            max_length = 1024
+        if not self.debug:
+            generated_text = self.generation_pipline[model_name.lower()](
+                prompt,
+                max_length=max_length,
+                temperature=temperature,
+                top_k=top_k,
+                top_p=top_p,
+                repetition_penalty=repetition_penalty,
+                no_repeat_ngram_size=no_repeat_ngram_size,
+                pad_token_id=pad_token_id,
+                eos_token_id=eos_token_id,
+                return_full_text=return_full_text,
+                return_text=return_text,
+                do_sample=do_sample,
+                num_beams=num_beams,
+                num_return_sequences=num_return_sequences,
+            )[0]["generated_text"]
+        else:
+            generated_text = self.generate_by_query(
+                prompt,
+                model_name,
+                max_length=max_length,
+                temperature=temperature,
+                top_k=top_k,
+                top_p=top_p,
+                repetition_penalty=repetition_penalty,
+                no_repeat_ngram_size=no_repeat_ngram_size,
+                pad_token_id=pad_token_id,
+                eos_token_id=eos_token_id,
+                return_full_text=return_full_text,
+                return_text=return_text,
+                do_sample=do_sample,
+                num_beams=num_beams,
+                num_return_sequences=num_return_sequences,
+            )
+            # print(generated_text)
+            if isinstance(generated_text, dict):
+                if "error" in generated_text:
+                    if "is currently loading" in generated_text["error"]:
+                        return f"Model is currently loading, estimated time is {generated_text['estimated_time']}"
+                    return generated_text["error"]
+                else:
+                    return "Something happened 🤷‍♂️!!"
+            else:
+                generated_text = generated_text[0]["generated_text"]
+
+        logger.info(f"Prompt: {prompt}")
+        logger.info(f"Generated text: {generated_text}")
+        return self.preprocessor.unpreprocess(generated_text)
+
+    def query(self, payload, model_name):
+        data = json.dumps(payload)
+        url = (
+            "https://api-inference.huggingface.co/models/aubmindlab/"
+            + model_name.lower()
+        )
+        response = requests.request("POST", url, headers=self.headers, data=data)
+        return json.loads(response.content.decode("utf-8"))
+
+    def generate_by_query(
+        self,
+        prompt: str,
+        model_name: str,
+        max_length: int,
+        temperature: float,
+        top_k: int,
+        top_p: float,
+        repetition_penalty: float,
+        no_repeat_ngram_size: int,
+        pad_token_id: int,
+        eos_token_id: int,
+        return_full_text: int,
+        return_text: int,
+        do_sample: bool,
+        num_beams: int,
+        num_return_sequences: int,
+    ):
+        payload = {
+            "inputs": prompt,
+            "parameters": {
+                "max_length ": max_length,
+                "top_k": top_k,
+                "top_p": top_p,
+                "temperature": temperature,
+                "repetition_penalty": repetition_penalty,
+                "no_repeat_ngram_size": no_repeat_ngram_size,
+                "pad_token_id": pad_token_id,
+                "eos_token_id": eos_token_id,
+                "return_full_text": return_full_text,
+                "return_text": return_text,
+                "pad_token_id": pad_token_id,
+                "do_sample": do_sample,
+                "num_beams": num_beams,
+                "num_return_sequences": num_return_sequences,
+            },
+            "options": {
+                "use_cache": True,
+            },
+        }
+        return self.query(payload, model_name)
+
+
+class SentimentAnalyzer:
+    def __init__(self):
+        self.sa_models = [
+            "sa_trial5_1",
+            # "sa_no_aoa_in_neutral",
+            # "sa_cnnbert",
+            # "sa_sarcasm",
+            # "sar_trial10",
+            # "sa_no_AOA",
+        ]
+        download_models(self.sa_models)
+        # fmt: off
+        self.processors = {
+            "sa_trial5_1": Trial5ArabicPreprocessor(model_name='UBC-NLP/MARBERT'),
+            # "sa_no_aoa_in_neutral": NewArabicPreprocessorBalanced(model_name='UBC-NLP/MARBERT'),
+            # "sa_cnnbert": CNNMarbertArabicPreprocessor(model_name='UBC-NLP/MARBERT'),
+            # "sa_sarcasm": SarcasmArabicPreprocessor(model_name='UBC-NLP/MARBERT'),
+            # "sar_trial10": SarcasmArabicPreprocessor(model_name='UBC-NLP/MARBERT'),
+            # "sa_no_AOA": NewArabicPreprocessorBalanced(model_name='UBC-NLP/MARBERT'),
+        }
+
+        self.pipelines = {
+            "sa_trial5_1": [pipeline("sentiment-analysis", model="{}/train_{}/best_model".format("sa_trial5_1",i), device=-1,return_all_scores =True) for i in tqdm(range(0,5), desc=f"Loading pipeline for model: sa_trial5_1")],
+            # "sa_no_aoa_in_neutral": [pipeline("sentiment-analysis", model="{}/train_{}/best_model".format("sa_no_aoa_in_neutral",i), device=-1,return_all_scores =True) for i in tqdm(range(0,5), desc=f"Loading pipeline for model: sa_no_aoa_in_neutral")],
+            # "sa_cnnbert": [CNNTextClassificationPipeline("{}/train_{}/best_model".format("sa_cnnbert",i), device=-1, return_all_scores =True) for i in tqdm(range(0,5), desc=f"Loading pipeline for model: sa_cnnbert")],
+            # "sa_sarcasm": [pipeline("sentiment-analysis", model="{}/train_{}/best_model".format("sa_sarcasm",i), device=-1,return_all_scores =True) for i in tqdm(range(0,5), desc=f"Loading pipeline for model: sa_sarcasm")],
+            # "sar_trial10": [pipeline("sentiment-analysis", model="{}/train_{}/best_model".format("sar_trial10",i), device=-1,return_all_scores =True) for i in tqdm(range(0,5), desc=f"Loading pipeline for model: sar_trial10")],
+            # "sa_no_AOA": [pipeline("sentiment-analysis", model="{}/train_{}/best_model".format("sa_no_AOA",i), device=-1,return_all_scores =True) for i in tqdm(range(0,5), desc=f"Loading pipeline for model: sa_no_AOA")],
+        }
+        # fmt: on
+
+    def get_preds_from_sarcasm(self, texts):
+        prep = self.processors["sar_trial10"]
+        prep_texts = [prep.preprocess(x) for x in texts]
+
+        preds_df = pd.DataFrame([])
+        for i in range(0, 5):
+            preds = []
+            for s in more_itertools.chunked(list(prep_texts), 128):
+                preds.extend(self.pipelines["sar_trial10"][i](s))
+            preds_df[f"model_{i}"] = preds
+
+        final_labels = []
+        final_scores = []
+        for id, row in preds_df.iterrows():
+            pos_total = 0
+            neu_total = 0
+            for pred in row[:]:
+                pos_total += pred[0]["score"]
+                neu_total += pred[1]["score"]
+
+            pos_avg = pos_total / len(row[:])
+            neu_avg = neu_total / len(row[:])
+
+            final_labels.append(
+                self.pipelines["sar_trial10"][0].model.config.id2label[
+                    np.argmax([pos_avg, neu_avg])
+                ]
+            )
+            final_scores.append(np.max([pos_avg, neu_avg]))
+
+        return final_labels, final_scores
+
+    def get_preds_from_a_model(self, texts: List[str], model_name):
+        try:
+            prep = self.processors[model_name]
+
+            prep_texts = [prep.preprocess(x) for x in texts]
+            if model_name == "sa_sarcasm":
+                sarcasm_label, _ = self.get_preds_from_sarcasm(texts)
+                sarcastic_map = {"Not_Sarcastic": "غير ساخر", "Sarcastic": "ساخر"}
+                labeled_prep_texts = []
+                for t, l in zip(prep_texts, sarcasm_label):
+                    labeled_prep_texts.append(sarcastic_map[l] + " [SEP] " + t)
+
+            preds_df = pd.DataFrame([])
+            for i in range(0, 5):
+                preds = []
+                for s in more_itertools.chunked(list(prep_texts), 128):
+                    preds.extend(self.pipelines[model_name][i](s))
+                preds_df[f"model_{i}"] = preds
+
+            final_labels = []
+            final_scores = []
+            final_scores_list = []
+            for id, row in preds_df.iterrows():
+                pos_total = 0
+                neg_total = 0
+                neu_total = 0
+                for pred in row[2:]:
+                    pos_total += pred[0]["score"]
+                    neu_total += pred[1]["score"]
+                    neg_total += pred[2]["score"]
+
+                pos_avg = pos_total / 5
+                neu_avg = neu_total / 5
+                neg_avg = neg_total / 5
+
+                if model_name == "sa_no_aoa_in_neutral":
+                    final_labels.append(
+                        self.pipelines[model_name][0].model.config.id2label[
+                            np.argmax([neu_avg, neg_avg, pos_avg])
+                        ]
+                    )
+                else:
+                    final_labels.append(
+                        self.pipelines[model_name][0].model.config.id2label[
+                            np.argmax([pos_avg, neu_avg, neg_avg])
+                        ]
+                    )
+                final_scores.append(np.max([pos_avg, neu_avg, neg_avg]))
+                final_scores_list.append((pos_avg, neu_avg, neg_avg))
+        except RuntimeError as e:
+            if model_name == "sa_cnnbert":
+                return (
+                    ["Neutral"] * len(texts),
+                    [0.0] * len(texts),
+                    [(0.0, 0.0, 0.0)] * len(texts),
+                )
+            else:
+                raise RuntimeError(e)
+        return final_labels, final_scores, final_scores_list
+
+    def predict(self, texts: List[str]):
+        logger.info(f"Predicting for: {texts}")
+        # (
+        #     new_balanced_label,
+        #     new_balanced_score,
+        #     new_balanced_score_list,
+        # ) = self.get_preds_from_a_model(texts, "sa_no_aoa_in_neutral")
+        # (
+        #     cnn_marbert_label,
+        #     cnn_marbert_score,
+        #     cnn_marbert_score_list,
+        # ) = self.get_preds_from_a_model(texts, "sa_cnnbert")
+        trial5_label, trial5_score, trial5_score_list = self.get_preds_from_a_model(
+            texts, "sa_trial5_1"
+        )
+        # no_aoa_label, no_aoa_score, no_aoa_score_list = self.get_preds_from_a_model(
+        #     texts, "sa_no_AOA"
+        # )
+        # sarcasm_label, sarcasm_score, sarcasm_score_list = self.get_preds_from_a_model(
+        #     texts, "sa_sarcasm"
+        # )
+
+        id_label_map = {0: "Positive", 1: "Neutral", 2: "Negative"}
+
+        final_ensemble_prediction = []
+        final_ensemble_score = []
+        final_ensemble_all_score = []
+        for entry in zip(
+            # new_balanced_score_list,
+            # cnn_marbert_score_list,
+            trial5_score_list,
+            # no_aoa_score_list,
+            # sarcasm_score_list,
+        ):
+            pos_score = 0
+            neu_score = 0
+            neg_score = 0
+            for s in entry:
+                pos_score += s[0] * 1.57
+                neu_score += s[1] * 0.98
+                neg_score += s[2] * 0.93
+
+                # weighted 2
+                # pos_score += s[0]*1.67
+                # neu_score += s[1]
+                # neg_score += s[2]*0.95
+
+            final_ensemble_prediction.append(
+                id_label_map[np.argmax([pos_score, neu_score, neg_score])]
+            )
+            final_ensemble_score.append(np.max([pos_score, neu_score, neg_score]))
+            final_ensemble_all_score.append(
+                softmax(np.array([pos_score, neu_score, neg_score])).tolist()
+            )
+
+        logger.info(f"Result: {final_ensemble_prediction}")
+        logger.info(f"Score: {final_ensemble_score}")
+        logger.info(f"All Scores: {final_ensemble_all_score}")
+        return final_ensemble_prediction, final_ensemble_score, final_ensemble_all_score
+
+
+wikipedia.set_lang("ar")
+
+os.environ["TOKENIZERS_PARALLELISM"] = "false"
+
+preprocessor = ArabertPreprocessor("wissamantoun/araelectra-base-artydiqa")
+logger.info("Loading QA Pipeline...")
+tokenizer = AutoTokenizer.from_pretrained("wissamantoun/araelectra-base-artydiqa")
+qa_pipe = pipeline("question-answering", model="wissamantoun/araelectra-base-artydiqa")
+logger.info("Finished loading QA Pipeline...")
+
+
+@lru_cache(maxsize=100)
+def get_qa_answers(question):
+    logger.info("\n=================================================================")
+    logger.info(f"Question: {question}")
+
+    if "وسام أنطون" in question or "wissam antoun" in question.lower():
+        return {
+            "title": "Creator",
+            "results": [
+                {
+                    "score": 1.0,
+                    "new_start": 0,
+                    "new_end": 12,
+                    "new_answer": "My Creator 😜",
+                    "original": "My Creator 😜",
+                    "link": "https://github.com/WissamAntoun/",
+                }
+            ],
+        }
+    search_timer = Timer(
+        "search and wiki", text="Search and Wikipedia Time: {:.2f}", logger=logging.info
+    )
+    try:
+        search_timer.start()
+        search_results = google.search(
+            question + " site:ar.wikipedia.org", lang="ar", area="ar"
+        )
+        if len(search_results) == 0:
+            return {}
+
+        page_name = search_results[0].link.split("wiki/")[-1]
+        wiki_page = wikipedia.page(unquote(page_name))
+        wiki_page_content = wiki_page.content
+        search_timer.stop()
+    except:
+        return {}
+
+    sections = []
+    for section in re.split("== .+ ==[^=]", wiki_page_content):
+        if not section.isspace():
+            prep_section = tokenizer.tokenize(preprocessor.preprocess(section))
+            if len(prep_section) > 500:
+                subsections = []
+                for subsection in re.split("=== .+ ===", section):
+                    if subsection.isspace():
+                        continue
+                    prep_subsection = tokenizer.tokenize(
+                        preprocessor.preprocess(subsection)
+                    )
+                    subsections.append(subsection)
+                    # logger.info(f"Subsection found with length: {len(prep_subsection)}")
+                sections.extend(subsections)
+            else:
+                # logger.info(f"Regular Section with length: {len(prep_section)}")
+                sections.append(section)
+
+    full_len_sections = []
+    temp_section = ""
+    for section in sections:
+        if (
+            len(tokenizer.tokenize(preprocessor.preprocess(temp_section)))
+            + len(tokenizer.tokenize(preprocessor.preprocess(section)))
+            > 384
+        ):
+            if temp_section == "":
+                temp_section = section
+                continue
+            full_len_sections.append(temp_section)
+            # logger.info(
+            #     f"full section length: {len(tokenizer.tokenize(preprocessor.preprocess(temp_section)))}"
+            # )
+            temp_section = ""
+        else:
+            temp_section += " " + section + " "
+    if temp_section != "":
+        full_len_sections.append(temp_section)
+
+    reader_time = Timer("electra", text="Reader Time: {:.2f}", logger=logging.info)
+    reader_time.start()
+    results = qa_pipe(
+        question=[preprocessor.preprocess(question)] * len(full_len_sections),
+        context=[preprocessor.preprocess(x) for x in full_len_sections],
+    )
+
+    if not isinstance(results, list):
+        results = [results]
+
+    logger.info(f"Wiki Title: {unquote(page_name)}")
+    logger.info(f"Total Sections: {len(sections)}")
+    logger.info(f"Total Full Sections: {len(full_len_sections)}")
+
+    for result, section in zip(results, full_len_sections):
+        result["original"] = section
+        answer_match = find_near_matches(
+            " " + preprocessor.unpreprocess(result["answer"]) + " ",
+            result["original"],
+            max_l_dist=min(5, len(preprocessor.unpreprocess(result["answer"])) // 2),
+            max_deletions=0,
+        )
+        try:
+            result["new_start"] = answer_match[0].start
+            result["new_end"] = answer_match[0].end
+            result["new_answer"] = answer_match[0].matched
+            result["link"] = (
+                search_results[0].link + "#:~:text=" + result["new_answer"].strip()
+            )
+        except:
+            result["new_start"] = result["start"]
+            result["new_end"] = result["end"]
+            result["new_answer"] = result["answer"]
+            result["original"] = preprocessor.preprocess(result["original"])
+            result["link"] = search_results[0].link
+        logger.info(f"Answers: {preprocessor.preprocess(result['new_answer'])}")
+
+    sorted_results = sorted(results, reverse=True, key=lambda x: x["score"])
+
+    return_dict = {}
+    return_dict["title"] = unquote(page_name)
+    return_dict["results"] = sorted_results
+
+    reader_time.stop()
+    logger.info(f"Total time spent: {reader_time.last + search_timer.last}")
+    return return_dict

backend/utils.py

@@ -0,0 +1,64 @@
+import re
+import numpy as np
+import psutil
+import os
+from tqdm.auto import tqdm
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+def get_current_ram_usage():
+    ram = psutil.virtual_memory()
+    return ram.available / 1024 / 1024 / 1024, ram.total / 1024 / 1024 / 1024
+
+
+def download_models(models):
+    for model in tqdm(models, desc="Downloading models"):
+        logger.info(f"Downloading {model}")
+        for i in range(0, 5):
+            curr_dir = f"{model}/train_{i}/best_model/"
+            os.makedirs(curr_dir, exist_ok=True)
+            os.system(
+                f"wget -q https://huggingface.co/researchaccount/{model}/resolve/main/train_{i}/best_model/config.json -P {curr_dir}"
+            )
+            os.system(
+                f"wget -q https://huggingface.co/researchaccount/{model}/resolve/main/train_{i}/best_model/pytorch_model.bin -P {curr_dir}"
+            )
+            os.system(
+                f"wget -q https://huggingface.co/researchaccount/{model}/resolve/main/train_{i}/best_model/special_tokens_map.json -P {curr_dir}"
+            )
+            os.system(
+                f"wget -q https://huggingface.co/researchaccount/{model}/resolve/main/train_{i}/best_model/tokenizer_config.json -P {curr_dir}"
+            )
+            os.system(
+                f"wget -q https://huggingface.co/researchaccount/{model}/resolve/main/train_{i}/best_model/training_args.bin -P {curr_dir}"
+            )
+            os.system(
+                f"wget -q https://huggingface.co/researchaccount/{model}/resolve/main/train_{i}/best_model/vocab.txt -P {curr_dir}"
+            )
+
+
+def softmax(x):
+    return np.exp(x) / sum(np.exp(x))
+
+
+def ga(file):
+    code = """
+    <!-- Global site tag (gtag.js) - Google Analytics -->
+        <script async src="https://www.googletagmanager.com/gtag/js?id=G-NH9HWCW08F"></script>
+        <script>
+        window.dataLayer = window.dataLayer || [];
+        function gtag(){dataLayer.push(arguments);}
+        gtag('js', new Date());
+        gtag('config', 'G-NH9HWCW08F');
+    </script>
+    """
+
+    a = os.path.dirname(file) + "/static/index.html"
+    with open(a, "r") as f:
+        data = f.read()
+        if len(re.findall("G-", data)) == 0:
+            with open(a, "w") as ff:
+                newdata = re.sub("<head>", "<head>" + code, data)
+                ff.write(newdata)

packages.txt

@@ -0,0 +1,2 @@
+openjdk-11-jre
+curl

requirements.txt

@@ -0,0 +1,17 @@
+streamlit==0.84.2
+arabic-reshaper==2.1.3
+python-bidi==0.4.2
+PyArabic
+farasapy==0.0.14
+emoji==1.4.2
+awesome_streamlit
+torch==1.9.0
+transformers==4.10.0
+psutil==5.8.0
+fuzzysearch==0.7.3
+more-itertools==8.9.0
+cookiecutter
+git+https://github.com/dantru7/Google-Search-API
+codetiming==1.3.0
+htbuilder
+wikipedia==1.4.0

test.py

@@ -0,0 +1,10 @@
+#%%
+from transformers import GPT2Tokenizer
+
+# %%
+tok = GPT2Tokenizer.from_pretrained("D:/ML/Models/aragpt2-medium", use_fast=False)
+# %%
+tok.pad_token = tok.eos_token
+#%%
+tok.pad_token_id = [tok.eos_token_id]
+# %%