Prediction API Error

[Mageswaran1989]

I used cli to train on SROIE2019 dataset (original images are preprocessed into line images) with :

calamari-train \
--device.gpus 0 \
--trainer.gen SplitTrain \
--trainer.gen.validation_split_ratio=0.2  \
--trainer.output_dir /data/model_output \
--trainer.epochs 25 \
--early_stopping.frequency=1 \
--early_stopping.n_to_go=3 \
--train.images /data/*.jpg

Training went smooth and the logs are
train.log

After the training process, I am trying to load the model as mentioned here, however I get following error:

>>> predictor = Predictor.from_checkpoint(params=PredictorParams(), checkpoint='/data/model_output/best.ckpt')
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/usr/local/lib/python3.8/dist-packages/calamari_ocr/ocr/predict/predictor.py", line 31, in from_checkpoint
    keras.models.load_model(
  File "/usr/local/lib/python3.8/dist-packages/tensorflow/python/keras/saving/save.py", line 206, in load_model
    return hdf5_format.load_model_from_hdf5(filepath, custom_objects,
  File "/usr/local/lib/python3.8/dist-packages/tensorflow/python/keras/saving/hdf5_format.py", line 182, in load_model_from_hdf5
    model_config = json_utils.decode(model_config.decode('utf-8'))
AttributeError: 'str' object has no attribute 'decode'

I tried loading pretrainined model from antiqua_historical, and again I got the same error:

>>> predictor = Predictor.from_checkpoint(params=PredictorParams(), checkpoint='/data/model_output/antiqua_historical/0.ckpt')
/usr/local/lib/python3.8/dist-packages/paiargparse/dataclass_json_overrides.py:78: RuntimeWarning: `NoneType` object value of non-optional type tfaip_commit_hash detected when decoding CalamariScenarioParams.
  warnings.warn(f"`NoneType` object {warning}.", RuntimeWarning)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/usr/local/lib/python3.8/dist-packages/calamari_ocr/ocr/predict/predictor.py", line 26, in from_checkpoint
    ckpt = SavedCalamariModel(checkpoint, auto_update=auto_update_checkpoints)
  File "/usr/local/lib/python3.8/dist-packages/calamari_ocr/ocr/savedmodel/saved_model.py", line 31, in __init__
    self.update_checkpoint()
  File "/usr/local/lib/python3.8/dist-packages/calamari_ocr/ocr/savedmodel/saved_model.py", line 56, in update_checkpoint
    self._single_upgrade()
  File "/usr/local/lib/python3.8/dist-packages/calamari_ocr/ocr/savedmodel/saved_model.py", line 88, in _single_upgrade
    update_model(self.dict, self.ckpt_path)
  File "/usr/local/lib/python3.8/dist-packages/calamari_ocr/ocr/savedmodel/migrations/version3_4to5.py", line 22, in update_model
    pred_model.load_weights(path + ".h5")
  File "/usr/local/lib/python3.8/dist-packages/tensorflow/python/keras/engine/training.py", line 2234, in load_weights
    hdf5_format.load_weights_from_hdf5_group(f, self.layers)
  File "/usr/local/lib/python3.8/dist-packages/tensorflow/python/keras/saving/hdf5_format.py", line 662, in load_weights_from_hdf5_group
    original_keras_version = f.attrs['keras_version'].decode('utf8')
AttributeError: 'str' object has no attribute 'decode'

[ChWick]

This might be related to Tensorflow. Can you post your Tensorflow and h5py version? (tensorflow should be >= 2.3.0) I guess there is a mismatch.

[Mageswaran1989]

tf.version
'2.4.2'

I am using docker to build the model with only calamari installed.

best.ckpt.zip

[ChWick]

I just setup a fresh venv (current calamari master) with python 3.8:

virtualenv -p python3.8 venv
git clone https://github.com/Calamari-OCR/calamari.git
source venv/bin/activate
pip install -U pip
pip install -e calamari

I had not problems loading your provided model using the predict script (calamari-predict).
Using tensorflow==2.4.2 and h5py==2.10.0. Maybe your h5py version is already 3.x?

[Mageswaran1989]

Yup.... but I am installing only calamari using pip with Python version - 3.8.5 (default installation version.)

h5py 3.3.0

[Mageswaran1989]

Dockerfile

FROM nvidia/cuda:11.1-cudnn8-runtime-ubuntu20.04 as runtime-image

ARG DEBIAN_FRONTEND=noninteractive
RUN ln -snf /usr/share/zoneinfo/$CONTAINER_TIMEZONE /etc/localtime && echo $CONTAINER_TIMEZONE > /etc/timezone
RUN mkdir -p /usr/share/man/man1/

RUN --mount=type=cache,target=/var/cache/apt \
    apt-get update && \
    apt-get install --no-install-recommends --no-install-suggests -y \
    build-essential \
    curl \
    ca-certificates p11-kit \
    python3-dev \
    python3-distutils \
    python3-venv \
    openjdk-11-jre-headless \
    tesseract-ocr \
    libtesseract-dev \
    libpq-dev \
    python3-pip \
    libgl1-mesa-glx &&\
    apt clean && rm -rf /var/lib/apt/lists/*

RUN pip3 install calamari-ocr
COPY ops/docker/ocr/requirements.txt /
RUN --mount=type=cache,target=/root/.cache/pip3 pip3 install --no-cache-dir -r /requirements.txt
RUN ln -s /usr/local/cuda-11.1/targets/x86_64-linux/lib/libcusolver.so.11 /usr/local/cuda-11.1/targets/x86_64-linux/lib/libcusolver.so.10
ENV export LD_LIBRARY_PATH=/usr/local/cuda-11.1/lib64/:$LD_LIBRARY_PATH

[ChWick]

I guess the pip version is outdated (<21.x). Older pip version do not resolve version conflicts automatically, which is why a newer h5py version gets installed. When I install pip install calamari-ocr in a venv I receive h5py=2.10.0.

Please try to insert a

pip3 install -U pip setuptools

in your Dockerfile before you install calamari-ocr.

[Mageswaran1989]

Thanks a lot @ChWick

Its working with h5py=2.10.0 and the predictions are also quite good :)

However, keeping predictor alive is running into error:

>>> raw_predictor = predictor.raw().__enter()__  # you can also wrap the following lines in a `with`-block
  File "<stdin>", line 1
    raw_predictor = predictor.raw().__enter()__  # you can also wrap the following lines in a `with`-block
                                             ^
SyntaxError: invalid syntax

If you are Ok for an example, I can put out a small example to train on SROIE2019 and do predictions on whole image using craft text detection and calamari for OCR.

[ChWick]

oups, this is is a typo in the docs. Should be:

raw_predictor = predictor.raw().__enter__()

[Mageswaran1989]

>>> raw_predictor = predictor.raw().__enter__()
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
AttributeError: 'Predictor' object has no attribute 'raw'

[ChWick]

Ah, this particular interface is rather new and not yet included in the current pip-release, its only in the master and will be included in the next release.

[ChWick]

I can draft a quick minor release if you rely on this feature, though! Just let me know

[Mageswaran1989]

I hope there is not much difference between two methods, for now I am good with the working version.

It would be good to have in pip version in general :) if it is not much of a work. Thanks again.