Duplicate from katanaml-org/sparrow-ml

Co-authored-by: Andrej Baranovskij <[email protected]>

.gitattributes

@@ -0,0 +1,34 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,2 @@
+
+.DS_Store

Dockerfile

@@ -0,0 +1,20 @@
+FROM python:3.7-slim
+
+WORKDIR /code
+
+COPY requirements-fastapi.txt ./
+
+RUN pip install --no-cache-dir --upgrade -r /code/requirements-fastapi.txt
+
+RUN useradd -m -u 1000 user
+
+USER user
+
+ENV HOME=/home/user \
+	PATH=/home/user/.local/bin:$PATH
+
+WORKDIR $HOME/app
+
+COPY --chown=user . $HOME/app/
+
+CMD ["uvicorn", "endpoints:app", "--host", "0.0.0.0", "--port", "7860"]

README.md

@@ -0,0 +1,12 @@
+---
+title: Sparrow ML
+emoji: 🌍
+colorFrom: purple
+colorTo: indigo
+sdk: docker
+pinned: false
+license: mit
+duplicated_from: katanaml-org/sparrow-ml
+---
+
+Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

config.py

@@ -0,0 +1,19 @@
+from pydantic import BaseSettings
+import os
+
+
+class Settings(BaseSettings):
+    huggingface_key: str = os.environ.get("huggingface_key")
+    sparrow_key: str = os.environ.get("sparrow_key")
+    processor: str = "katanaml-org/invoices-donut-model-v1"
+    model: str = "katanaml-org/invoices-donut-model-v1"
+    dataset: str = "katanaml-org/invoices-donut-data-v1"
+    base_config: str = "naver-clova-ix/donut-base"
+    base_processor: str = "naver-clova-ix/donut-base"
+    base_model: str = "naver-clova-ix/donut-base"
+    inference_stats_file: str = "data/donut_inference_stats.json"
+    training_stats_file: str = "data/donut_training_stats.json"
+    evaluate_stats_file: str = "data/donut_evaluate_stats.json"
+
+
+settings = Settings()

data/donut_evaluate_stats.json

@@ -0,0 +1 @@
+[[498.8315510749817, {"accuracies": [0.9903846153846154, 0.6744186046511628, 0.9920948616600791, 0.981675392670157, 0.9728555917480999, 0.9886363636363636, 0.9846796657381616, 0.9858156028368794, 0.996031746031746, 0.9768115942028985, 0.9777227722772277, 0.9886363636363636, 0.9955555555555555, 0.9894847528916929, 0.9964028776978417, 0.9970238095238095, 0.975609756097561, 0.9941176470588236, 0.9921259842519685, 0.9533898305084746, 0.8410174880763116, 0.9781609195402299, 0.8535825545171339, 0.9862595419847329, 0.9906868451688009, 0.9929478138222849], "mean_accuracy": 0.9633126365834221}, 0.9633126365834221, "katanaml-org/invoices-donut-model-v1", "2023-05-15 16:16:59"], [480.58880615234375, {"accuracies": [0.9903846153846154, 0.6744186046511628, 0.9920948616600791, 0.981675392670157, 0.9728555917480999, 0.9886363636363636, 0.9846796657381616, 0.9858156028368794, 0.996031746031746, 0.9768115942028985, 0.9777227722772277, 0.9886363636363636, 0.9955555555555555, 0.9894847528916929, 0.9964028776978417, 0.9970238095238095, 0.975609756097561, 0.9941176470588236, 0.9921259842519685, 0.9533898305084746, 0.8410174880763116, 0.9781609195402299, 0.8535825545171339, 0.9862595419847329, 0.9906868451688009, 0.9929478138222849], "mean_accuracy": 0.9633126365834221}, 0.9633126365834221, "katanaml-org/invoices-donut-model-v1", "2023-05-15 16:29:24"], [496.27668499946594, {"accuracies": [0.9903846153846154, 0.6744186046511628, 0.9920948616600791, 0.981675392670157, 0.9728555917480999, 0.9886363636363636, 0.9846796657381616, 0.9858156028368794, 0.996031746031746, 0.9768115942028985, 0.9777227722772277, 0.9886363636363636, 0.9955555555555555, 0.9894847528916929, 0.9964028776978417, 0.9970238095238095, 0.975609756097561, 0.9941176470588236, 0.9921259842519685, 0.9533898305084746, 0.8410174880763116, 0.9781609195402299, 0.8535825545171339, 0.9862595419847329, 0.9906868451688009, 0.9929478138222849], "mean_accuracy": 0.9633126365834221}, 0.9633126365834221, "katanaml-org/invoices-donut-model-v1", "2023-05-17 11:26:54"], [496.5165719985962, {"accuracies": [0.9903846153846154, 0.6744186046511628, 0.9920948616600791, 0.981675392670157, 0.9728555917480999, 0.9886363636363636, 0.9846796657381616, 0.9858156028368794, 0.996031746031746, 0.9768115942028985, 0.9777227722772277, 0.9886363636363636, 0.9955555555555555, 0.9894847528916929, 0.9964028776978417, 0.9970238095238095, 0.975609756097561, 0.9941176470588236, 0.9921259842519685, 0.9533898305084746, 0.8410174880763116, 0.9781609195402299, 0.8535825545171339, 0.9862595419847329, 0.9906868451688009, 0.9929478138222849], "mean_accuracy": 0.9633126365834221}, 0.9633126365834221, "katanaml-org/invoices-donut-model-v1", "2023-05-17 11:40:15"], [528.6264460086823, {"accuracies": [0.9903846153846154, 0.6744186046511628, 0.9920948616600791, 0.981675392670157, 0.9728555917480999, 0.9886363636363636, 0.9846796657381616, 0.9858156028368794, 0.996031746031746, 0.9768115942028985, 0.9777227722772277, 0.9886363636363636, 0.9955555555555555, 0.9894847528916929, 0.9964028776978417, 0.9970238095238095, 0.975609756097561, 0.9941176470588236, 0.9921259842519685, 0.9533898305084746, 0.8410174880763116, 0.9781609195402299, 0.8535825545171339, 0.9862595419847329, 0.9906868451688009, 0.9929478138222849], "mean_accuracy": 0.9633126365834221}, 0.9633126365834221, "katanaml-org/invoices-donut-model-v1", "2023-05-17 22:22:59"]]

data/donut_inference_stats.json

@@ -0,0 +1 @@
+[[14.571558952331543, 21, "invoice_10.jpg", "katanaml-org/invoices-donut-model-v1", "2023-04-13 21:45:30"], [14.510485887527466, 16, "docs/inference/invoice_0_16823599391530209.jpg", "katanaml-org/invoices-donut-model-v1", "2023-04-24 21:12:37"]]

data/donut_training_stats.json

@@ -0,0 +1 @@
+[[112.83321595191956, "katanaml-org/invoices-donut-model-v1", "2023-05-17 22:05:20"], [47.31714415550232, "katanaml-org/invoices-donut-model-v1", "2023-05-17 22:06:31"]]

endpoints.py

@@ -0,0 +1,25 @@
+from fastapi import FastAPI
+from fastapi.middleware.cors import CORSMiddleware
+from routers import inference, training
+from huggingface_hub import login
+from config import settings
+
+login(settings.huggingface_key)
+
+app = FastAPI(openapi_url="/api/v1/sparrow-ml/openapi.json", docs_url="/api/v1/sparrow-ml/docs")
+
+app.add_middleware(
+    CORSMiddleware,
+    allow_origins=["*"],
+    allow_methods=["*"],
+    allow_headers=["*"],
+    allow_credentials=True,
+)
+
+app.include_router(inference.router, prefix="/api-inference/v1/sparrow-ml", tags=["Inference"])
+app.include_router(training.router, prefix="/api-training/v1/sparrow-ml", tags=["Training"])
+
+
+@app.get("/")
+async def root():
+    return {"message": "Sparrow ML API"}

requirements-fastapi.txt

@@ -0,0 +1,10 @@
+transformers
+datasets
+sentencepiece
+tensorboard
+pytorch-lightning
+Pillow
+donut-python
+fastapi==0.95.2
+uvicorn[standard]
+python-multipart

routers/donut_evaluate.py

@@ -0,0 +1,90 @@
+from transformers import DonutProcessor, VisionEncoderDecoderModel
+import locale
+
+import re
+import json
+import torch
+from tqdm.auto import tqdm
+import numpy as np
+from donut import JSONParseEvaluator
+from datasets import load_dataset
+from functools import lru_cache
+import os
+import time
+from config import settings
+
+locale.getpreferredencoding = lambda: "UTF-8"
+
+
+@lru_cache(maxsize=1)
+def prepare_model():
+    processor = DonutProcessor.from_pretrained(settings.processor)
+    model = VisionEncoderDecoderModel.from_pretrained(settings.model)
+
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+
+    model.eval()
+    model.to(device)
+
+    dataset = load_dataset(settings.dataset, split="test")
+
+    return processor, model, device, dataset
+
+
+def run_evaluate_donut():
+    worker_pid = os.getpid()
+    print(f"Handling evaluation request with worker PID: {worker_pid}")
+
+    start_time = time.time()
+
+    output_list = []
+    accs = []
+
+    processor, model, device, dataset = prepare_model()
+
+    for idx, sample in tqdm(enumerate(dataset), total=len(dataset)):
+        # prepare encoder inputs
+        pixel_values = processor(sample["image"].convert("RGB"), return_tensors="pt").pixel_values
+        pixel_values = pixel_values.to(device)
+        # prepare decoder inputs
+        task_prompt = "<s_cord-v2>"
+        decoder_input_ids = processor.tokenizer(task_prompt, add_special_tokens=False, return_tensors="pt").input_ids
+        decoder_input_ids = decoder_input_ids.to(device)
+
+        # autoregressively generate sequence
+        outputs = model.generate(
+            pixel_values,
+            decoder_input_ids=decoder_input_ids,
+            max_length=model.decoder.config.max_position_embeddings,
+            early_stopping=True,
+            pad_token_id=processor.tokenizer.pad_token_id,
+            eos_token_id=processor.tokenizer.eos_token_id,
+            use_cache=True,
+            num_beams=1,
+            bad_words_ids=[[processor.tokenizer.unk_token_id]],
+            return_dict_in_generate=True,
+        )
+
+        # turn into JSON
+        seq = processor.batch_decode(outputs.sequences)[0]
+        seq = seq.replace(processor.tokenizer.eos_token, "").replace(processor.tokenizer.pad_token, "")
+        seq = re.sub(r"<.*?>", "", seq, count=1).strip()  # remove first task start token
+        seq = processor.token2json(seq)
+
+        ground_truth = json.loads(sample["ground_truth"])
+        ground_truth = ground_truth["gt_parse"]
+        evaluator = JSONParseEvaluator()
+        score = evaluator.cal_acc(seq, ground_truth)
+
+        accs.append(score)
+        output_list.append(seq)
+
+    end_time = time.time()
+    processing_time = end_time - start_time
+
+    scores = {"accuracies": accs, "mean_accuracy": np.mean(accs)}
+    print(scores, f"length : {len(accs)}")
+    print("Mean accuracy:", np.mean(accs))
+    print(f"Evaluation done, worker PID: {worker_pid}")
+
+    return scores, np.mean(accs), processing_time

routers/donut_inference.py

@@ -0,0 +1,60 @@
+import re
+import time
+import torch
+from transformers import DonutProcessor, VisionEncoderDecoderModel
+from config import settings
+from functools import lru_cache
+import os
+
+
+@lru_cache(maxsize=1)
+def load_model():
+    processor = DonutProcessor.from_pretrained(settings.processor)
+    model = VisionEncoderDecoderModel.from_pretrained(settings.model)
+
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    model.to(device)
+
+    return processor, model, device
+
+
+def process_document_donut(image):
+    worker_pid = os.getpid()
+    print(f"Handling inference request with worker PID: {worker_pid}")
+
+    start_time = time.time()
+
+    processor, model, device = load_model()
+
+    # prepare encoder inputs
+    pixel_values = processor(image, return_tensors="pt").pixel_values
+
+    # prepare decoder inputs
+    task_prompt = "<s_cord-v2>"
+    decoder_input_ids = processor.tokenizer(task_prompt, add_special_tokens=False, return_tensors="pt").input_ids
+
+    # generate answer
+    outputs = model.generate(
+        pixel_values.to(device),
+        decoder_input_ids=decoder_input_ids.to(device),
+        max_length=model.decoder.config.max_position_embeddings,
+        early_stopping=True,
+        pad_token_id=processor.tokenizer.pad_token_id,
+        eos_token_id=processor.tokenizer.eos_token_id,
+        use_cache=True,
+        num_beams=1,
+        bad_words_ids=[[processor.tokenizer.unk_token_id]],
+        return_dict_in_generate=True,
+    )
+
+    # postprocess
+    sequence = processor.batch_decode(outputs.sequences)[0]
+    sequence = sequence.replace(processor.tokenizer.eos_token, "").replace(processor.tokenizer.pad_token, "")
+    sequence = re.sub(r"<.*?>", "", sequence, count=1).strip()  # remove first task start token
+
+    end_time = time.time()
+    processing_time = end_time - start_time
+
+    print(f"Inference done, worker PID: {worker_pid}")
+
+    return processor.token2json(sequence), processing_time

routers/donut_training.py

@@ -0,0 +1,393 @@
+# !pip install -q git+https://github.com/huggingface/transformers.git datasets sentencepiece
+# !pip install -q pytorch-lightning==1.9.5 wandb
+
+from config import settings
+from datasets import load_dataset
+from transformers import VisionEncoderDecoderConfig
+from transformers import DonutProcessor, VisionEncoderDecoderModel
+
+import json
+import random
+from typing import Any, List, Tuple
+
+import torch
+from torch.utils.data import Dataset
+
+from torch.utils.data import DataLoader
+
+import re
+from nltk import edit_distance
+import numpy as np
+import os
+import time
+
+import pytorch_lightning as pl
+from functools import lru_cache
+
+from pytorch_lightning.loggers import WandbLogger
+from pytorch_lightning.callbacks import Callback
+from config import settings
+
+added_tokens = []
+
+dataset_name = settings.dataset
+base_config_name = settings.base_config
+base_processor_name = settings.base_processor
+base_model_name = settings.base_model
+model_name = settings.model
+
+@lru_cache(maxsize=1)
+def prepare_job():
+    print("Preparing job...")
+
+    dataset = load_dataset(dataset_name)
+
+    max_length = 768
+    image_size = [1280, 960]
+
+    # update image_size of the encoder
+    # during pre-training, a larger image size was used
+    config = VisionEncoderDecoderConfig.from_pretrained(base_config_name)
+    config.encoder.image_size = image_size  # (height, width)
+    # update max_length of the decoder (for generation)
+    config.decoder.max_length = max_length
+    # TODO we should actually update max_position_embeddings and interpolate the pre-trained ones:
+    # https://github.com/clovaai/donut/blob/0acc65a85d140852b8d9928565f0f6b2d98dc088/donut/model.py#L602
+
+    processor = DonutProcessor.from_pretrained(base_processor_name)
+    model = VisionEncoderDecoderModel.from_pretrained(base_model_name, config=config)
+
+    return model, processor, dataset, config, image_size, max_length
+
+
+class DonutDataset(Dataset):
+    """
+    DonutDataset which is saved in huggingface datasets format. (see details in https://huggingface.co/docs/datasets)
+    Each row, consists of image path(png/jpg/jpeg) and gt data (json/jsonl/txt),
+    and it will be converted into input_tensor(vectorized image) and input_ids(tokenized string).
+    Args:
+        dataset_name_or_path: name of dataset (available at huggingface.co/datasets) or the path containing image files and metadata.jsonl
+        max_length: the max number of tokens for the target sequences
+        split: whether to load "train", "validation" or "test" split
+        ignore_id: ignore_index for torch.nn.CrossEntropyLoss
+        task_start_token: the special token to be fed to the decoder to conduct the target task
+        prompt_end_token: the special token at the end of the sequences
+        sort_json_key: whether or not to sort the JSON keys
+    """
+
+    def __init__(
+            self,
+            dataset_name_or_path: str,
+            max_length: int,
+            split: str = "train",
+            ignore_id: int = -100,
+            task_start_token: str = "<s>",
+            prompt_end_token: str = None,
+            sort_json_key: bool = True,
+    ):
+        super().__init__()
+
+        model, processor, dataset, config, image_size, p1 = prepare_job()
+
+        self.max_length = max_length
+        self.split = split
+        self.ignore_id = ignore_id
+        self.task_start_token = task_start_token
+        self.prompt_end_token = prompt_end_token if prompt_end_token else task_start_token
+        self.sort_json_key = sort_json_key
+
+        self.dataset = load_dataset(dataset_name_or_path, split=self.split)
+        self.dataset_length = len(self.dataset)
+
+        self.gt_token_sequences = []
+        for sample in self.dataset:
+            ground_truth = json.loads(sample["ground_truth"])
+            if "gt_parses" in ground_truth:  # when multiple ground truths are available, e.g., docvqa
+                assert isinstance(ground_truth["gt_parses"], list)
+                gt_jsons = ground_truth["gt_parses"]
+            else:
+                assert "gt_parse" in ground_truth and isinstance(ground_truth["gt_parse"], dict)
+                gt_jsons = [ground_truth["gt_parse"]]
+
+            self.gt_token_sequences.append(
+                [
+                    self.json2token(
+                        gt_json,
+                        update_special_tokens_for_json_key=self.split == "train",
+                        sort_json_key=self.sort_json_key,
+                    )
+                    + processor.tokenizer.eos_token
+                    for gt_json in gt_jsons  # load json from list of json
+                ]
+            )
+
+        self.add_tokens([self.task_start_token, self.prompt_end_token])
+        self.prompt_end_token_id = processor.tokenizer.convert_tokens_to_ids(self.prompt_end_token)
+
+    def json2token(self, obj: Any, update_special_tokens_for_json_key: bool = True, sort_json_key: bool = True):
+        """
+        Convert an ordered JSON object into a token sequence
+        """
+        if type(obj) == dict:
+            if len(obj) == 1 and "text_sequence" in obj:
+                return obj["text_sequence"]
+            else:
+                output = ""
+                if sort_json_key:
+                    keys = sorted(obj.keys(), reverse=True)
+                else:
+                    keys = obj.keys()
+                for k in keys:
+                    if update_special_tokens_for_json_key:
+                        self.add_tokens([fr"<s_{k}>", fr"</s_{k}>"])
+                    output += (
+                            fr"<s_{k}>"
+                            + self.json2token(obj[k], update_special_tokens_for_json_key, sort_json_key)
+                            + fr"</s_{k}>"
+                    )
+                return output
+        elif type(obj) == list:
+            return r"<sep/>".join(
+                [self.json2token(item, update_special_tokens_for_json_key, sort_json_key) for item in obj]
+            )
+        else:
+            obj = str(obj)
+            if f"<{obj}/>" in added_tokens:
+                obj = f"<{obj}/>"  # for categorical special tokens
+            return obj
+
+    def add_tokens(self, list_of_tokens: List[str]):
+        """
+        Add special tokens to tokenizer and resize the token embeddings of the decoder
+        """
+        model, processor, dataset, config, image_size, p1 = prepare_job()
+
+        newly_added_num = processor.tokenizer.add_tokens(list_of_tokens)
+        if newly_added_num > 0:
+            model.decoder.resize_token_embeddings(len(processor.tokenizer))
+            added_tokens.extend(list_of_tokens)
+
+    def __len__(self) -> int:
+        return self.dataset_length
+
+    def __getitem__(self, idx: int) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """
+        Load image from image_path of given dataset_path and convert into input_tensor and labels
+        Convert gt data into input_ids (tokenized string)
+        Returns:
+            input_tensor : preprocessed image
+            input_ids : tokenized gt_data
+            labels : masked labels (model doesn't need to predict prompt and pad token)
+        """
+
+        model, processor, dataset, config, image_size, p1 = prepare_job()
+
+        sample = self.dataset[idx]
+
+        # inputs
+        pixel_values = processor(sample["image"], random_padding=self.split == "train",
+                                 return_tensors="pt").pixel_values
+        pixel_values = pixel_values.squeeze()
+
+        # targets
+        target_sequence = random.choice(self.gt_token_sequences[idx])  # can be more than one, e.g., DocVQA Task 1
+        input_ids = processor.tokenizer(
+            target_sequence,
+            add_special_tokens=False,
+            max_length=self.max_length,
+            padding="max_length",
+            truncation=True,
+            return_tensors="pt",
+        )["input_ids"].squeeze(0)
+
+        labels = input_ids.clone()
+        labels[labels == processor.tokenizer.pad_token_id] = self.ignore_id  # model doesn't need to predict pad token
+        # labels[: torch.nonzero(labels == self.prompt_end_token_id).sum() + 1] = self.ignore_id  # model doesn't need to predict prompt (for VQA)
+        return pixel_values, labels, target_sequence
+
+
+def build_data_loaders():
+    print("Building data loaders...")
+
+    model, processor, dataset, config, image_size, max_length = prepare_job()
+
+    # we update some settings which differ from pretraining; namely the size of the images + no rotation required
+    # source: https://github.com/clovaai/donut/blob/master/config/train_cord.yaml
+    processor.feature_extractor.size = image_size[::-1]  # should be (width, height)
+    processor.feature_extractor.do_align_long_axis = False
+
+    train_dataset = DonutDataset(dataset_name, max_length=max_length,
+                                 split="train", task_start_token="<s_cord-v2>", prompt_end_token="<s_cord-v2>",
+                                 sort_json_key=False,  # cord dataset is preprocessed, so no need for this
+                                 )
+
+    val_dataset = DonutDataset(dataset_name, max_length=max_length,
+                               split="validation", task_start_token="<s_cord-v2>", prompt_end_token="<s_cord-v2>",
+                               sort_json_key=False,  # cord dataset is preprocessed, so no need for this
+                               )
+
+    model.config.pad_token_id = processor.tokenizer.pad_token_id
+    model.config.decoder_start_token_id = processor.tokenizer.convert_tokens_to_ids(['<s_cord-v2>'])[0]
+
+    # feel free to increase the batch size if you have a lot of memory
+    # I'm fine-tuning on Colab and given the large image size, batch size > 1 is not feasible
+    # Set num_workers=4
+    train_dataloader = DataLoader(train_dataset, batch_size=1, shuffle=True, num_workers=4)
+    val_dataloader = DataLoader(val_dataset, batch_size=1, shuffle=False, num_workers=4)
+
+    return train_dataloader, val_dataloader, max_length
+
+
+class DonutModelPLModule(pl.LightningModule):
+    def __init__(self, config, processor, model):
+        super().__init__()
+        self.config = config
+        self.processor = processor
+        self.model = model
+
+        self.train_dataloader, self.val_dataloader, self.max_length = build_data_loaders()
+
+    def training_step(self, batch, batch_idx):
+        pixel_values, labels, _ = batch
+
+        outputs = self.model(pixel_values, labels=labels)
+        loss = outputs.loss
+        self.log_dict({"train_loss": loss}, sync_dist=True)
+        return loss
+
+    def validation_step(self, batch, batch_idx, dataset_idx=0):
+        pixel_values, labels, answers = batch
+        batch_size = pixel_values.shape[0]
+        # we feed the prompt to the model
+        decoder_input_ids = torch.full((batch_size, 1), self.model.config.decoder_start_token_id, device=self.device)
+
+        outputs = self.model.generate(pixel_values,
+                                      decoder_input_ids=decoder_input_ids,
+                                      max_length=self.max_length,
+                                      early_stopping=True,
+                                      pad_token_id=self.processor.tokenizer.pad_token_id,
+                                      eos_token_id=self.processor.tokenizer.eos_token_id,
+                                      use_cache=True,
+                                      num_beams=1,
+                                      bad_words_ids=[[self.processor.tokenizer.unk_token_id]],
+                                      return_dict_in_generate=True, )
+
+        predictions = []
+        for seq in self.processor.tokenizer.batch_decode(outputs.sequences):
+            seq = seq.replace(self.processor.tokenizer.eos_token, "").replace(self.processor.tokenizer.pad_token, "")
+            seq = re.sub(r"<.*?>", "", seq, count=1).strip()  # remove first task start token
+            predictions.append(seq)
+
+        scores = list()
+        for pred, answer in zip(predictions, answers):
+            pred = re.sub(r"(?:(?<=>) | (?=</s_))", "", pred)
+            # NOT NEEDED ANYMORE
+            # answer = re.sub(r"<.*?>", "", answer, count=1)
+            answer = answer.replace(self.processor.tokenizer.eos_token, "")
+            scores.append(edit_distance(pred, answer) / max(len(pred), len(answer)))
+
+            if self.config.get("verbose", False) and len(scores) == 1:
+                print(f"Prediction: {pred}")
+                print(f"    Answer: {answer}")
+                print(f" Normed ED: {scores[0]}")
+
+        return scores
+
+    def validation_epoch_end(self, validation_step_outputs):
+        # I set this to 1 manually
+        # (previously set to len(self.config.dataset_name_or_paths))
+        num_of_loaders = 1
+        if num_of_loaders == 1:
+            validation_step_outputs = [validation_step_outputs]
+        assert len(validation_step_outputs) == num_of_loaders
+        cnt = [0] * num_of_loaders
+        total_metric = [0] * num_of_loaders
+        val_metric = [0] * num_of_loaders
+        for i, results in enumerate(validation_step_outputs):
+            for scores in results:
+                cnt[i] += len(scores)
+                total_metric[i] += np.sum(scores)
+            val_metric[i] = total_metric[i] / cnt[i]
+            val_metric_name = f"val_metric_{i}th_dataset"
+            self.log_dict({val_metric_name: val_metric[i]}, sync_dist=True)
+        self.log_dict({"val_metric": np.sum(total_metric) / np.sum(cnt)}, sync_dist=True)
+
+    def configure_optimizers(self):
+        # TODO add scheduler
+        optimizer = torch.optim.Adam(self.parameters(), lr=self.config.get("lr"))
+
+        return optimizer
+
+    def train_dataloader(self):
+        return self.train_dataloader
+
+    def val_dataloader(self):
+        return self.val_dataloader
+
+
+class PushToHubCallback(Callback):
+    def on_train_epoch_end(self, trainer, pl_module):
+        print(f"Pushing model to the hub, epoch {trainer.current_epoch}")
+        pl_module.model.push_to_hub(model_name,
+                                    commit_message=f"Training in progress, epoch {trainer.current_epoch}")
+
+    def on_train_end(self, trainer, pl_module):
+        print(f"Pushing model to the hub after training")
+        pl_module.processor.push_to_hub(model_name,
+                                        commit_message=f"Training done")
+        pl_module.model.push_to_hub(model_name,
+                                    commit_message=f"Training done")
+
+
+def run_training_donut(max_epochs_param, val_check_interval_param, warmup_steps_param):
+    worker_pid = os.getpid()
+    print(f"Handling training request with worker PID: {worker_pid}")
+
+    start_time = time.time()
+
+    # Set epochs = 30
+    # Set num_training_samples_per_epoch = training set size
+    # Set val_check_interval = 0.4
+    # Set warmup_steps: 425 / 8 = 54, 54 * 10 = 540, 540 * 0.15 = 81
+    config_params = {"max_epochs": max_epochs_param,
+                     "val_check_interval": val_check_interval_param,  # how many times we want to validate during an epoch
+                     "check_val_every_n_epoch": 1,
+                     "gradient_clip_val": 1.0,
+                     "num_training_samples_per_epoch": 425,
+                     "lr": 3e-5,
+                     "train_batch_sizes": [8],
+                     "val_batch_sizes": [1],
+                     # "seed":2022,
+                     "num_nodes": 1,
+                     "warmup_steps": warmup_steps_param,  # 425 / 8 = 54, 54 * 10 = 540, 540 * 0.15 = 81
+                     "result_path": "./result",
+                     "verbose": False,
+                     }
+
+    model, processor, dataset, config, image_size, p1 = prepare_job()
+
+    model_module = DonutModelPLModule(config, processor, model)
+
+    # wandb_logger = WandbLogger(project="sparrow", name="invoices-donut-v5")
+
+    # trainer = pl.Trainer(
+    #     accelerator="gpu",
+    #     devices=1,
+    #     max_epochs=config_params.get("max_epochs"),
+    #     val_check_interval=config_params.get("val_check_interval"),
+    #     check_val_every_n_epoch=config_params.get("check_val_every_n_epoch"),
+    #     gradient_clip_val=config_params.get("gradient_clip_val"),
+    #     precision=16,  # we'll use mixed precision
+    #     num_sanity_val_steps=0,
+    #     # logger=wandb_logger,
+    #     callbacks=[PushToHubCallback()],
+    # )
+
+    # trainer.fit(model_module)
+
+    end_time = time.time()
+    processing_time = end_time - start_time
+
+    print(f"Training done, worker PID: {worker_pid}")
+
+    return processing_time

routers/inference.py

@@ -0,0 +1,81 @@
+from fastapi import APIRouter, File, UploadFile, Form
+from typing import Optional
+from PIL import Image
+import urllib.request
+from io import BytesIO
+from config import settings
+import utils
+import os
+import json
+from routers.donut_inference import process_document_donut
+
+
+router = APIRouter()
+
+def count_values(obj):
+    if isinstance(obj, dict):
+        count = 0
+        for value in obj.values():
+            count += count_values(value)
+        return count
+    elif isinstance(obj, list):
+        count = 0
+        for item in obj:
+            count += count_values(item)
+        return count
+    else:
+        return 1
+
+
+@router.post("/inference")
+async def run_inference(file: Optional[UploadFile] = File(None), image_url: Optional[str] = Form(None),
+                        model_in_use: str = Form('donut'), sparrow_key: str = Form(None)):
+
+    if sparrow_key != settings.sparrow_key:
+        return {"error": "Invalid Sparrow key."}
+
+    result = []
+    if file:
+        # Ensure the uploaded file is a JPG image
+        if file.content_type not in ["image/jpeg", "image/jpg"]:
+            return {"error": "Invalid file type. Only JPG images are allowed."}
+
+        image = Image.open(BytesIO(await file.read()))
+        processing_time = 0
+        if model_in_use == 'donut':
+            result, processing_time = process_document_donut(image)
+        utils.log_stats(settings.inference_stats_file, [processing_time, count_values(result), file.filename, settings.model])
+        print(f"Processing time: {processing_time:.2f} seconds")
+    elif image_url:
+        # test image url: https://raw.githubusercontent.com/katanaml/sparrow/main/sparrow-data/docs/input/invoices/processed/images/invoice_10.jpg
+        with urllib.request.urlopen(image_url) as url:
+            image = Image.open(BytesIO(url.read()))
+
+        processing_time = 0
+        if model_in_use == 'donut':
+            result, processing_time = process_document_donut(image)
+        # parse file name from url
+        file_name = image_url.split("/")[-1]
+        utils.log_stats(settings.inference_stats_file, [processing_time, count_values(result), file_name, settings.model])
+        print(f"Processing time inference: {processing_time:.2f} seconds")
+    else:
+        result = {"info": "No input provided"}
+
+    return result
+
+
+@router.get("/statistics")
+async def get_statistics():
+    file_path = settings.inference_stats_file
+
+    # Check if the file exists, and read its content
+    if os.path.exists(file_path):
+        with open(file_path, 'r') as file:
+            try:
+                content = json.load(file)
+            except json.JSONDecodeError:
+                content = []
+    else:
+        content = []
+
+    return content

routers/training.py

@@ -0,0 +1,87 @@
+from fastapi import APIRouter, Form, BackgroundTasks
+from config import settings
+import os
+import json
+from routers.donut_evaluate import run_evaluate_donut
+from routers.donut_training import run_training_donut
+import utils
+
+
+router = APIRouter()
+
+
+def invoke_training(max_epochs, val_check_interval, warmup_steps, model_in_use, sparrow_key):
+    if sparrow_key != settings.sparrow_key:
+        return {"error": "Invalid Sparrow key."}
+
+    if model_in_use == 'donut':
+        processing_time = run_training_donut(max_epochs, val_check_interval, warmup_steps)
+        utils.log_stats(settings.training_stats_file, [processing_time, settings.model])
+        print(f"Processing time training: {processing_time:.2f} seconds")
+
+
+@router.post("/training")
+async def run_training(background_tasks: BackgroundTasks,
+                       max_epochs: int = Form(30),
+                       val_check_interval: float = Form(0.4),
+                       warmup_steps: int = Form(81),
+                       model_in_use: str = Form('donut'),
+                       sparrow_key: str = Form(None)):
+
+    background_tasks.add_task(invoke_training, max_epochs, val_check_interval, warmup_steps, model_in_use, sparrow_key)
+
+    return {"message": "Sparrow ML training started in the background"}
+
+
+def invoke_evaluate(model_in_use, sparrow_key):
+    if sparrow_key != settings.sparrow_key:
+        return {"error": "Invalid Sparrow key."}
+
+    if model_in_use == 'donut':
+        scores, accuracy, processing_time = run_evaluate_donut()
+        utils.log_stats(settings.evaluate_stats_file, [processing_time, scores, accuracy, settings.model])
+        print(f"Processing time evaluate: {processing_time:.2f} seconds")
+
+
+@router.post("/evaluate")
+async def run_evaluate(background_tasks: BackgroundTasks,
+                       model_in_use: str = Form('donut'),
+                       sparrow_key: str = Form(None)):
+
+    background_tasks.add_task(invoke_evaluate, model_in_use, sparrow_key)
+
+    return {"message": "Sparrow ML model evaluation started in the background"}
+
+
+@router.get("/statistics/training")
+async def get_statistics_training():
+    file_path = settings.training_stats_file
+
+    # Check if the file exists, and read its content
+    if os.path.exists(file_path):
+        with open(file_path, 'r') as file:
+            try:
+                content = json.load(file)
+            except json.JSONDecodeError:
+                content = []
+    else:
+        content = []
+
+    return content
+
+
+@router.get("/statistics/evaluate")
+async def get_statistics_evaluate():
+    file_path = settings.evaluate_stats_file
+
+    # Check if the file exists, and read its content
+    if os.path.exists(file_path):
+        with open(file_path, 'r') as file:
+            try:
+                content = json.load(file)
+            except json.JSONDecodeError:
+                content = []
+    else:
+        content = []
+
+    return content

utils.py

@@ -0,0 +1,29 @@
+import json
+import os
+from datetime import datetime
+
+
+def log_stats(file_path, new_data):
+    # Check if the file exists, and read its content
+    if os.path.exists(file_path):
+        with open(file_path, 'r') as file:
+            try:
+                content = json.load(file)
+            except json.JSONDecodeError:
+                content = []
+    else:
+        content = []
+
+    # Get the current date and time
+    now = datetime.now()
+    # Format the date and time as a string
+    date_time_string = now.strftime("%Y-%m-%d %H:%M:%S")
+    new_data.append(date_time_string)
+
+    # Append the new data to the content
+    content.append(new_data)
+
+    # Write the updated content back to the file
+    with open(file_path, 'w') as file:
+        json.dump(content, file)
+