def load_model(cls, path: str):
"""
Deserializes a VGSL model from a CoreML file.
Args:
path (str): CoreML file
Returns:
A TorchVGSLModel instance.
Raises:
KrakenInvalidModelException if the model data is invalid (not a
string, protobuf file, or without appropriate metadata).
FileNotFoundError if the path doesn't point to a file.
"""
try:
mlmodel = MLModel(path)
except TypeError as e:
raise KrakenInvalidModelException(str(e))
except DecodeError as e:
raise KrakenInvalidModelException('Failure parsing model protobuf: {}'.format(str(e)))
if 'vgsl' not in mlmodel.user_defined_metadata:
raise KrakenInvalidModelException('No VGSL spec in model metadata')
vgsl_spec = mlmodel.user_defined_metadata['vgsl']
nn = cls(vgsl_spec)
for name, layer in nn.nn.named_children():
layer.deserialize(name, mlmodel.get_spec())
if 'codec' in mlmodel.user_defined_metadata:
nn.add_codec(PytorchCodec(json.loads(mlmodel.user_defined_metadata['codec'])))
nn.user_metadata = {'accuracy': [], 'seg_type': 'bbox', 'one_channel_mode': '1', 'model_type': None, 'hyper_params': {}} # type: dict[str, str]
if 'kraken_meta' in mlmodel.user_defined_metadata:
nn.user_metadata.update(json.loads(mlmodel.user_defined_metadata['kraken_meta']))
return nn
def load_pronn_model(cls, path: str):
"""
Loads an pronn model to VGSL.
"""
with open(path, 'rb') as fp:
net = pyrnn_pb2.pyrnn()
try:
net.ParseFromString(fp.read())
except Exception:
raise KrakenInvalidModelException('File does not contain valid proto msg')
if not net.IsInitialized():
raise KrakenInvalidModelException('Model incomplete')
# extract codec
codec = PytorchCodec(net.codec)
input = net.ninput
hidden = net.fwdnet.wgi.dim[0]
# extract weights
weightnames = ('wgi', 'wgf', 'wci', 'wgo', 'wip', 'wfp', 'wop')
fwd_w = []
rev_w = []
for w in weightnames:
fwd_ar = getattr(net.fwdnet, w)
rev_ar = getattr(net.revnet, w)
fwd_w.append(torch.Tensor(fwd_ar.value).view(list(fwd_ar.dim)))
rev_w.append(torch.Tensor(rev_ar.value).view(list(rev_ar.dim)))
t = torch.cat(fwd_w[:4])
weight_ih_l0 = t[:, :input+1]
weight_hh_l0 = t[:, input+1:]
t = torch.cat(rev_w[:4])
weight_ih_l0_rev = t[:, :input+1]
weight_hh_l0_rev = t[:, input+1:]
weight_lin = torch.Tensor(net.softmax.w2.value).view(list(net.softmax.w2.dim))
# build vgsl spec and set weights
nn = cls('[1,1,0,{} Lbxo{} O1ca{}]'.format(input, hidden, len(net.codec)))
nn.nn.L_0.layer.weight_ih_l0 = torch.nn.Parameter(weight_ih_l0)
nn.nn.L_0.layer.weight_hh_l0 = torch.nn.Parameter(weight_hh_l0)
nn.nn.L_0.layer.weight_ih_l0_reverse = torch.nn.Parameter(weight_ih_l0_rev)
nn.nn.L_0.layer.weight_hh_l0_reverse = torch.nn.Parameter(weight_hh_l0_rev)
nn.nn.L_0.layer.weight_ip_l0 = torch.nn.Parameter(fwd_w[4])
nn.nn.L_0.layer.weight_fp_l0 = torch.nn.Parameter(fwd_w[5])
nn.nn.L_0.layer.weight_op_l0 = torch.nn.Parameter(fwd_w[6])
nn.nn.L_0.layer.weight_ip_l0_reverse = torch.nn.Parameter(rev_w[4])
nn.nn.L_0.layer.weight_fp_l0_reverse = torch.nn.Parameter(rev_w[5])
nn.nn.L_0.layer.weight_op_l0_reverse = torch.nn.Parameter(rev_w[6])
nn.nn.O_1.lin.weight = torch.nn.Parameter(weight_lin)
nn.add_codec(codec)
return nn
def load_model(cls, path: Union[str, pathlib.Path]):
"""
Deserializes a VGSL model from a CoreML file.
Args:
path: CoreML file
Returns:
A TorchVGSLModel instance.
Raises:
KrakenInvalidModelException if the model data is invalid (not a
string, protobuf file, or without appropriate metadata).
FileNotFoundError if the path doesn't point to a file.
"""
if isinstance(path, pathlib.Path):
path = path.as_posix()
try:
mlmodel = MLModel(path)
except TypeError as e:
raise KrakenInvalidModelException(str(e))
except DecodeError as e:
raise KrakenInvalidModelException(
'Failure parsing model protobuf: {}'.format(str(e)))
if 'vgsl' not in mlmodel.user_defined_metadata:
raise KrakenInvalidModelException('No VGSL spec in model metadata')
vgsl_spec = mlmodel.user_defined_metadata['vgsl']
nn = cls(vgsl_spec)
def _deserialize_layers(name, layer):
logger.debug(f'Deserializing layer {name} with type {type(layer)}')
if type(layer) in (layers.MultiParamParallel,
layers.MultiParamSequential):
for name, l in layer.named_children():
_deserialize_layers(name, l)
else:
layer.deserialize(name, mlmodel.get_spec())
_deserialize_layers('', nn.nn)
if 'codec' in mlmodel.user_defined_metadata:
nn.add_codec(
PytorchCodec(json.loads(
mlmodel.user_defined_metadata['codec'])))
nn.user_metadata = {
'accuracy': [],
'seg_type': 'bbox',
'one_channel_mode': '1',
'model_type': None,
'hyper_params': {}
} # type: dict[str, str]
if 'kraken_meta' in mlmodel.user_defined_metadata:
nn.user_metadata.update(
json.loads(mlmodel.user_defined_metadata['kraken_meta']))
return nn
def encode(self, codec: Optional[PytorchCodec] = None) -> None:
"""
Adds a codec to the dataset and encodes all text lines.
Has to be run before sampling from the dataset.
"""
if codec:
self.codec = codec
else:
self.codec = PytorchCodec(''.join(self.alphabet.keys()))
self.training_set = [] # type: List[Tuple[Union[Image, torch.Tensor], torch.Tensor]]
for im, gt in zip(self._images, self._gt):
self.training_set.append((im, self.codec.encode(gt)))
def load_model(cls, path: str):
"""
Deserializes a VGSL model from a CoreML file.
Args:
path (str): CoreML file
"""
mlmodel = MLModel(path)
if 'vgsl' not in mlmodel.user_defined_metadata:
raise ValueError('No VGSL spec in model metadata')
vgsl_spec = mlmodel.user_defined_metadata['vgsl']
nn = cls(vgsl_spec)
for name, layer in nn.nn.named_children():
layer.deserialize(name, mlmodel.get_spec())
if 'codec' in mlmodel.user_defined_metadata:
nn.add_codec(
PytorchCodec(json.loads(
mlmodel.user_defined_metadata['codec'])))
return nn
def load_clstm_model(cls, path: str):
"""
Loads an CLSTM model to VGSL.
"""
net = clstm_pb2.NetworkProto()
with open(path, 'rb') as fp:
try:
net.ParseFromString(fp.read())
except Exception:
raise KrakenInvalidModelException('File does not contain valid proto msg')
if not net.IsInitialized():
raise KrakenInvalidModelException('Model incomplete')
input = net.ninput
attrib = {a.key: a.value for a in list(net.attribute)}
# mainline clstm model
if len(attrib) > 1:
mode = 'clstm'
else:
mode = 'clstm_compat'
# extract codec
codec = PytorchCodec([''] + [chr(x) for x in net.codec[1:]])
# separate layers
nets = {}
nets['softm'] = [n for n in list(net.sub) if n.kind == 'SoftmaxLayer'][0]
parallel = [n for n in list(net.sub) if n.kind == 'Parallel'][0]
nets['lstm1'] = [n for n in list(parallel.sub) if n.kind.startswith('NPLSTM')][0]
rev = [n for n in list(parallel.sub) if n.kind == 'Reversed'][0]
nets['lstm2'] = rev.sub[0]
hidden = int(nets['lstm1'].attribute[0].value)
weights = {} # type: Dict[str, torch.Tensor]
for n in nets:
weights[n] = {}
for w in list(nets[n].weights):
weights[n][w.name] = torch.Tensor(w.value).view(list(w.dim))
if mode == 'clstm_compat':
weightnames = ('.WGI', '.WGF', '.WCI', '.WGO')
weightname_softm = '.W'
else:
weightnames = ('WGI', 'WGF', 'WCI', 'WGO')
weightname_softm = 'W1'
# input hidden and hidden-hidden weights are in one matrix. also
# CLSTM/ocropy likes 1-augmenting every other tensor so the ih weights
# are input+1 in one dimension.
t = torch.cat(list(w for w in [weights['lstm1'][wn] for wn in weightnames]))
weight_ih_l0 = t[:, :input+1]
weight_hh_l0 = t[:, input+1:]
t = torch.cat(list(w for w in [weights['lstm2'][wn] for wn in weightnames]))
weight_ih_l0_rev = t[:, :input+1]
weight_hh_l0_rev = t[:, input+1:]
weight_lin = weights['softm'][weightname_softm]
if mode == 'clstm_compat':
weight_lin = torch.cat([torch.zeros(len(weight_lin), 1), weight_lin], 1)
# build vgsl spec and set weights
nn = cls('[1,1,0,{} Lbxc{} O1ca{}]'.format(input, hidden, len(net.codec)))
nn.nn.L_0.layer.weight_ih_l0 = torch.nn.Parameter(weight_ih_l0)
nn.nn.L_0.layer.weight_hh_l0 = torch.nn.Parameter(weight_hh_l0)
nn.nn.L_0.layer.weight_ih_l0_reverse = torch.nn.Parameter(weight_ih_l0_rev)
nn.nn.L_0.layer.weight_hh_l0_reverse = torch.nn.Parameter(weight_hh_l0_rev)
nn.nn.O_1.lin.weight = torch.nn.Parameter(weight_lin)
nn.add_codec(codec)
return nn
def load_pyrnn_model(cls, path: str):
"""
Loads an pyrnn model to VGSL.
"""
if not PY2:
raise KrakenInvalidModelException('Loading pickle models is not supported on python 3')
import cPickle
def find_global(mname, cname):
aliases = {
'lstm.lstm': kraken.lib.lstm,
'ocrolib.lstm': kraken.lib.lstm,
'ocrolib.lineest': kraken.lib.lineest,
}
if mname in aliases:
return getattr(aliases[mname], cname)
return getattr(sys.modules[mname], cname)
of = io.open
if path.endswith('.gz'):
of = gzip.open
with io.BufferedReader(of(path, 'rb')) as fp:
unpickler = cPickle.Unpickler(fp)
unpickler.find_global = find_global
try:
net = unpickler.load()
except Exception as e:
raise KrakenInvalidModelException(str(e))
if not isinstance(net, kraken.lib.lstm.SeqRecognizer):
raise KrakenInvalidModelException('Pickle is %s instead of '
'SeqRecognizer' %
type(net).__name__)
# extract codec
codec = PytorchCodec({k: [v] for k, v in net.codec.char2code.items()})
input = net.Ni
parallel, softmax = net.lstm.nets
fwdnet, revnet = parallel.nets
revnet = revnet.net
hidden = fwdnet.WGI.shape[0]
# extract weights
weightnames = ('WGI', 'WGF', 'WCI', 'WGO', 'WIP', 'WFP', 'WOP')
fwd_w = []
rev_w = []
for w in weightnames:
fwd_w.append(torch.Tensor(getattr(fwdnet, w)))
rev_w.append(torch.Tensor(getattr(revnet, w)))
t = torch.cat(fwd_w[:4])
weight_ih_l0 = t[:, :input+1]
weight_hh_l0 = t[:, input+1:]
t = torch.cat(rev_w[:4])
weight_ih_l0_rev = t[:, :input+1]
weight_hh_l0_rev = t[:, input+1:]
weight_lin = torch.Tensor(softmax.W2)
# build vgsl spec and set weights
nn = cls('[1,1,0,{} Lbxo{} O1ca{}]'.format(input, hidden, len(net.codec.code2char)))
nn.nn.L_0.layer.weight_ih_l0 = torch.nn.Parameter(weight_ih_l0)
nn.nn.L_0.layer.weight_hh_l0 = torch.nn.Parameter(weight_hh_l0)
nn.nn.L_0.layer.weight_ih_l0_reverse = torch.nn.Parameter(weight_ih_l0_rev)
nn.nn.L_0.layer.weight_hh_l0_reverse = torch.nn.Parameter(weight_hh_l0_rev)
nn.nn.L_0.layer.weight_ip_l0 = torch.nn.Parameter(fwd_w[4])
nn.nn.L_0.layer.weight_fp_l0 = torch.nn.Parameter(fwd_w[5])
nn.nn.L_0.layer.weight_op_l0 = torch.nn.Parameter(fwd_w[6])
nn.nn.L_0.layer.weight_ip_l0_reverse = torch.nn.Parameter(rev_w[4])
nn.nn.L_0.layer.weight_fp_l0_reverse = torch.nn.Parameter(rev_w[5])
nn.nn.L_0.layer.weight_op_l0_reverse = torch.nn.Parameter(rev_w[6])
nn.nn.O_1.lin.weight = torch.nn.Parameter(weight_lin)
nn.add_codec(codec)
return nn
def train(ctx, pad, output, spec, append, load, freq, quit, epochs, lag,
min_delta, device, optimizer, lrate, momentum, weight_decay,
schedule, partition, normalization, normalize_whitespace, codec,
resize, reorder, training_files, evaluation_files, preload, threads,
ground_truth):
"""
Trains a model from image-text pairs.
"""
if not load and append:
raise click.BadOptionUsage(
'append', 'append option requires loading an existing model')
if resize != 'fail' and not load:
raise click.BadOptionUsage(
'resize', 'resize option requires loading an existing model')
import re
import torch
import shutil
import numpy as np
from torch.utils.data import DataLoader
from kraken.lib import models, vgsl, train
from kraken.lib.util import make_printable
from kraken.lib.train import EarlyStopping, EpochStopping, TrainStopper, TrainScheduler, add_1cycle
from kraken.lib.codec import PytorchCodec
from kraken.lib.dataset import GroundTruthDataset, generate_input_transforms
logger.info('Building ground truth set from {} line images'.format(
len(ground_truth) + len(training_files)))
completed_epochs = 0
# load model if given. if a new model has to be created we need to do that
# after data set initialization, otherwise to output size is still unknown.
nn = None
#hyper_fields = ['freq', 'quit', 'epochs', 'lag', 'min_delta', 'optimizer', 'lrate', 'momentum', 'weight_decay', 'schedule', 'partition', 'normalization', 'normalize_whitespace', 'reorder', 'preload', 'completed_epochs', 'output']
if load:
logger.info('Loading existing model from {} '.format(load))
message('Loading existing model from {}'.format(load), nl=False)
nn = vgsl.TorchVGSLModel.load_model(load)
#if nn.user_metadata and load_hyper_parameters:
# for param in hyper_fields:
# if param in nn.user_metadata:
# logger.info('Setting \'{}\' to \'{}\''.format(param, nn.user_metadata[param]))
# message('Setting \'{}\' to \'{}\''.format(param, nn.user_metadata[param]))
# locals()[param] = nn.user_metadata[param]
message('\u2713', fg='green', nl=False)
# preparse input sizes from vgsl string to seed ground truth data set
# sizes and dimension ordering.
if not nn:
spec = spec.strip()
if spec[0] != '[' or spec[-1] != ']':
raise click.BadOptionUsage(
'spec', 'VGSL spec {} not bracketed'.format(spec))
blocks = spec[1:-1].split(' ')
m = re.match(r'(\d+),(\d+),(\d+),(\d+)', blocks[0])
if not m:
raise click.BadOptionUsage(
'spec', 'Invalid input spec {}'.format(blocks[0]))
batch, height, width, channels = [int(x) for x in m.groups()]
else:
batch, channels, height, width = nn.input
try:
transforms = generate_input_transforms(batch, height, width, channels,
pad)
except KrakenInputException as e:
raise click.BadOptionUsage('spec', str(e))
# disable automatic partition when given evaluation set explicitly
if evaluation_files:
partition = 1
ground_truth = list(ground_truth)
# merge training_files into ground_truth list
if training_files:
ground_truth.extend(training_files)
if len(ground_truth) == 0:
raise click.UsageError(
'No training data was provided to the train command. Use `-t` or the `ground_truth` argument.'
)
np.random.shuffle(ground_truth)
if len(ground_truth) > 2500 and not preload:
logger.info(
'Disabling preloading for large (>2500) training data set. Enable by setting --preload parameter'
)
preload = False
# implicit preloading enabled for small data sets
if preload is None:
preload = True
tr_im = ground_truth[:int(len(ground_truth) * partition)]
if evaluation_files:
logger.debug('Using {} lines from explicit eval set'.format(
len(evaluation_files)))
te_im = evaluation_files
else:
te_im = ground_truth[int(len(ground_truth) * partition):]
logger.debug('Taking {} lines from training for evaluation'.format(
len(te_im)))
# set multiprocessing tensor sharing strategy
if 'file_system' in torch.multiprocessing.get_all_sharing_strategies():
logger.debug(
'Setting multiprocessing tensor sharing strategy to file_system')
torch.multiprocessing.set_sharing_strategy('file_system')
gt_set = GroundTruthDataset(normalization=normalization,
whitespace_normalization=normalize_whitespace,
reorder=reorder,
im_transforms=transforms,
preload=preload)
with log.progressbar(tr_im, label='Building training set') as bar:
for im in bar:
logger.debug('Adding line {} to training set'.format(im))
try:
gt_set.add(im)
except FileNotFoundError as e:
logger.warning('{}: {}. Skipping.'.format(
e.strerror, e.filename))
except KrakenInputException as e:
logger.warning(str(e))
val_set = GroundTruthDataset(normalization=normalization,
whitespace_normalization=normalize_whitespace,
reorder=reorder,
im_transforms=transforms,
preload=preload)
with log.progressbar(te_im, label='Building validation set') as bar:
for im in bar:
logger.debug('Adding line {} to validation set'.format(im))
try:
val_set.add(im)
except FileNotFoundError as e:
logger.warning('{}: {}. Skipping.'.format(
e.strerror, e.filename))
except KrakenInputException as e:
logger.warning(str(e))
logger.info(
'Training set {} lines, validation set {} lines, alphabet {} symbols'.
format(len(gt_set._images), len(val_set._images),
len(gt_set.alphabet)))
alpha_diff = set(gt_set.alphabet).symmetric_difference(
set(val_set.alphabet))
if alpha_diff:
logger.warn('alphabet mismatch {}'.format(alpha_diff))
logger.info('grapheme\tcount')
for k, v in sorted(gt_set.alphabet.items(),
key=lambda x: x[1],
reverse=True):
char = make_printable(k)
if char == k:
char = '\t' + char
logger.info(u'{}\t{}'.format(char, v))
logger.debug('Encoding training set')
# use model codec when given
if append:
# is already loaded
nn = cast(vgsl.TorchVGSLModel, nn)
gt_set.encode(codec)
message('Slicing and dicing model ', nl=False)
# now we can create a new model
spec = '[{} O1c{}]'.format(spec[1:-1], gt_set.codec.max_label() + 1)
logger.info('Appending {} to existing model {} after {}'.format(
spec, nn.spec, append))
nn.append(append, spec)
nn.add_codec(gt_set.codec)
message('\u2713', fg='green')
logger.info('Assembled model spec: {}'.format(nn.spec))
elif load:
# is already loaded
nn = cast(vgsl.TorchVGSLModel, nn)
# prefer explicitly given codec over network codec if mode is 'both'
codec = codec if (codec and resize == 'both') else nn.codec
try:
gt_set.encode(codec)
except KrakenEncodeException as e:
message('Network codec not compatible with training set')
alpha_diff = set(gt_set.alphabet).difference(set(codec.c2l.keys()))
if resize == 'fail':
logger.error(
'Training data and model codec alphabets mismatch: {}'.
format(alpha_diff))
ctx.exit(code=1)
elif resize == 'add':
message('Adding missing labels to network ', nl=False)
logger.info(
'Resizing codec to include {} new code points'.format(
len(alpha_diff)))
codec.c2l.update({
k: [v]
for v, k in enumerate(alpha_diff,
start=codec.max_label() + 1)
})
nn.add_codec(PytorchCodec(codec.c2l))
logger.info(
'Resizing last layer in network to {} outputs'.format(
codec.max_label() + 1))
nn.resize_output(codec.max_label() + 1)
gt_set.encode(nn.codec)
message('\u2713', fg='green')
elif resize == 'both':
message('Fitting network exactly to training set ', nl=False)
logger.info(
'Resizing network or given codec to {} code sequences'.
format(len(gt_set.alphabet)))
gt_set.encode(None)
ncodec, del_labels = codec.merge(gt_set.codec)
logger.info(
'Deleting {} output classes from network ({} retained)'.
format(len(del_labels),
len(codec) - len(del_labels)))
gt_set.encode(ncodec)
nn.resize_output(ncodec.max_label() + 1, del_labels)
message('\u2713', fg='green')
else:
raise click.BadOptionUsage(
'resize', 'Invalid resize value {}'.format(resize))
else:
gt_set.encode(codec)
logger.info('Creating new model {} with {} outputs'.format(
spec,
gt_set.codec.max_label() + 1))
spec = '[{} O1c{}]'.format(spec[1:-1], gt_set.codec.max_label() + 1)
nn = vgsl.TorchVGSLModel(spec)
# initialize weights
message('Initializing model ', nl=False)
nn.init_weights()
nn.add_codec(gt_set.codec)
# initialize codec
message('\u2713', fg='green')
# half the number of data loading processes if device isn't cuda and we haven't enabled preloading
if device == 'cpu' and not preload:
loader_threads = threads // 2
else:
loader_threads = threads
train_loader = DataLoader(gt_set,
batch_size=1,
shuffle=True,
num_workers=loader_threads,
pin_memory=True)
threads -= loader_threads
# don't encode validation set as the alphabets may not match causing encoding failures
val_set.training_set = list(zip(val_set._images, val_set._gt))
logger.debug('Constructing {} optimizer (lr: {}, momentum: {})'.format(
optimizer, lrate, momentum))
# set mode to trainindg
nn.train()
# set number of OpenMP threads
logger.debug('Set OpenMP threads to {}'.format(threads))
nn.set_num_threads(threads)
logger.debug('Moving model to device {}'.format(device))
optim = getattr(torch.optim, optimizer)(nn.nn.parameters(), lr=0)
if 'accuracy' not in nn.user_metadata:
nn.user_metadata['accuracy'] = []
tr_it = TrainScheduler(optim)
if schedule == '1cycle':
add_1cycle(tr_it, int(len(gt_set) * epochs), lrate, momentum,
momentum - 0.10, weight_decay)
else:
# constant learning rate scheduler
tr_it.add_phase(1, (lrate, lrate), (momentum, momentum), weight_decay,
train.annealing_const)
if quit == 'early':
st_it = EarlyStopping(min_delta, lag)
elif quit == 'dumb':
st_it = EpochStopping(epochs - completed_epochs)
else:
raise click.BadOptionUsage(
'quit', 'Invalid training interruption scheme {}'.format(quit))
#for param in hyper_fields:
# logger.debug('Setting \'{}\' to \'{}\' in model metadata'.format(param, locals()[param]))
# nn.user_metadata[param] = locals()[param]
trainer = train.KrakenTrainer(model=nn,
optimizer=optim,
device=device,
filename_prefix=output,
event_frequency=freq,
train_set=train_loader,
val_set=val_set,
stopper=st_it)
trainer.add_lr_scheduler(tr_it)
with log.progressbar(label='stage {}/{}'.format(
1, trainer.stopper.epochs if trainer.stopper.epochs > 0 else '∞'),
length=trainer.event_it,
show_pos=True) as bar:
def _draw_progressbar():
bar.update(1)
def _print_eval(epoch, accuracy, chars, error):
message('Accuracy report ({}) {:0.4f} {} {}'.format(
epoch, accuracy, chars, error))
# reset progress bar
bar.label = 'stage {}/{}'.format(
epoch + 1,
trainer.stopper.epochs if trainer.stopper.epochs > 0 else '∞')
bar.pos = 0
bar.finished = False
trainer.run(_print_eval, _draw_progressbar)
if quit == 'early':
message('Moving best model {0}_{1}.mlmodel ({2}) to {0}_best.mlmodel'.
format(output, trainer.stopper.best_epoch,
trainer.stopper.best_loss))
logger.info(
'Moving best model {0}_{1}.mlmodel ({2}) to {0}_best.mlmodel'.
format(output, trainer.stopper.best_epoch,
trainer.stopper.best_loss))
shutil.copy('{}_{}.mlmodel'.format(output, trainer.stopper.best_epoch),
'{}_best.mlmodel'.format(output))
def recognition_train_gen(
cls,
hyper_params: Dict = default_specs.RECOGNITION_HYPER_PARAMS,
progress_callback: Callable[[str, int], Callable[
[None], None]] = lambda string, length: lambda: None,
message: Callable[[str], None] = lambda *args, **kwargs: None,
output: str = 'model',
spec: str = default_specs.RECOGNITION_SPEC,
append: Optional[int] = None,
load: Optional[str] = None,
device: str = 'cpu',
reorder: bool = True,
training_data: Sequence[Dict] = None,
evaluation_data: Sequence[Dict] = None,
preload: Optional[bool] = None,
threads: int = 1,
load_hyper_parameters: bool = False,
repolygonize: bool = False,
force_binarization: bool = False,
format_type: str = 'path',
codec: Optional[Dict] = None,
resize: str = 'fail',
augment: bool = False):
"""
This is an ugly constructor that takes all the arguments from the command
line driver, finagles the datasets, models, and hyperparameters correctly
and returns a KrakenTrainer object.
Setup parameters (load, training_data, evaluation_data, ....) are named,
model hyperparameters (everything in
kraken.lib.default_specs.RECOGNITION_HYPER_PARAMS) are in in the
`hyper_params` argument.
Args:
hyper_params (dict): Hyperparameter dictionary containing all fields
from
kraken.lib.default_specs.RECOGNITION_HYPER_PARAMS
progress_callback (Callable): Callback for progress reports on various
computationally expensive processes. A
human readable string and the process
length is supplied. The callback has to
return another function which will be
executed after each step.
message (Callable): Messaging printing method for above log but below
warning level output, i.e. infos that should
generally be shown to users.
**kwargs: Setup parameters, i.e. CLI parameters of the train() command.
Returns:
A KrakenTrainer object.
"""
# load model if given. if a new model has to be created we need to do that
# after data set initialization, otherwise to output size is still unknown.
nn = None
if load:
logger.info(f'Loading existing model from {load} ')
message(f'Loading existing model from {load} ', nl=False)
nn = vgsl.TorchVGSLModel.load_model(load)
if load_hyper_parameters:
hyper_params.update(nn.hyper_params)
nn.hyper_params = hyper_params
message('\u2713', fg='green', nl=False)
DatasetClass = GroundTruthDataset
valid_norm = True
if format_type and format_type != 'path':
logger.info(
f'Parsing {len(training_data)} XML files for training data')
if repolygonize:
message('Repolygonizing data')
training_data = preparse_xml_data(training_data, format_type,
repolygonize)
evaluation_data = preparse_xml_data(evaluation_data, format_type,
repolygonize)
DatasetClass = PolygonGTDataset
valid_norm = False
elif format_type == 'path':
if force_binarization:
logger.warning(
'Forced binarization enabled in `path` mode. Will be ignored.'
)
force_binarization = False
if repolygonize:
logger.warning(
'Repolygonization enabled in `path` mode. Will be ignored.'
)
training_data = [{'image': im} for im in training_data]
if evaluation_data:
evaluation_data = [{'image': im} for im in evaluation_data]
valid_norm = True
# format_type is None. Determine training type from length of training data entry
else:
if len(training_data[0]) >= 4:
DatasetClass = PolygonGTDataset
valid_norm = False
else:
if force_binarization:
logger.warning(
'Forced binarization enabled with box lines. Will be ignored.'
)
force_binarization = False
if repolygonize:
logger.warning(
'Repolygonization enabled with box lines. Will be ignored.'
)
# preparse input sizes from vgsl string to seed ground truth data set
# sizes and dimension ordering.
if not nn:
spec = spec.strip()
if spec[0] != '[' or spec[-1] != ']':
raise click.BadOptionUsage(
'spec', 'VGSL spec {} not bracketed'.format(spec))
blocks = spec[1:-1].split(' ')
m = re.match(r'(\d+),(\d+),(\d+),(\d+)', blocks[0])
if not m:
raise click.BadOptionUsage('spec',
f'Invalid input spec {blocks[0]}')
batch, height, width, channels = [int(x) for x in m.groups()]
else:
batch, channels, height, width = nn.input
try:
transforms = generate_input_transforms(batch, height, width,
channels,
hyper_params['pad'],
valid_norm,
force_binarization)
except KrakenInputException as e:
raise click.BadOptionUsage('spec', str(e))
if len(training_data) > 2500 and not preload:
logger.info(
'Disabling preloading for large (>2500) training data set. Enable by setting --preload parameter'
)
preload = False
# implicit preloading enabled for small data sets
if preload is None:
preload = True
# set multiprocessing tensor sharing strategy
if 'file_system' in torch.multiprocessing.get_all_sharing_strategies():
logger.debug(
'Setting multiprocessing tensor sharing strategy to file_system'
)
torch.multiprocessing.set_sharing_strategy('file_system')
gt_set = DatasetClass(
normalization=hyper_params['normalization'],
whitespace_normalization=hyper_params['normalize_whitespace'],
reorder=reorder,
im_transforms=transforms,
preload=preload,
augmentation=hyper_params['augment'])
bar = progress_callback('Building training set', len(training_data))
for im in training_data:
logger.debug(f'Adding line {im} to training set')
try:
gt_set.add(**im)
bar()
except FileNotFoundError as e:
logger.warning(f'{e.strerror}: {e.filename}. Skipping.')
except KrakenInputException as e:
logger.warning(str(e))
val_set = DatasetClass(
normalization=hyper_params['normalization'],
whitespace_normalization=hyper_params['normalize_whitespace'],
reorder=reorder,
im_transforms=transforms,
preload=preload)
bar = progress_callback('Building validation set',
len(evaluation_data))
for im in evaluation_data:
logger.debug(f'Adding line {im} to validation set')
try:
val_set.add(**im)
bar()
except FileNotFoundError as e:
logger.warning(f'{e.strerror}: {e.filename}. Skipping.')
except KrakenInputException as e:
logger.warning(str(e))
if len(gt_set._images) == 0:
logger.error(
'No valid training data was provided to the train command. Please add valid XML or line data.'
)
return None
logger.info(
f'Training set {len(gt_set._images)} lines, validation set {len(val_set._images)} lines, alphabet {len(gt_set.alphabet)} symbols'
)
alpha_diff_only_train = set(gt_set.alphabet).difference(
set(val_set.alphabet))
alpha_diff_only_val = set(val_set.alphabet).difference(
set(gt_set.alphabet))
if alpha_diff_only_train:
logger.warning(
f'alphabet mismatch: chars in training set only: {alpha_diff_only_train} (not included in accuracy test during training)'
)
if alpha_diff_only_val:
logger.warning(
f'alphabet mismatch: chars in validation set only: {alpha_diff_only_val} (not trained)'
)
logger.info('grapheme\tcount')
for k, v in sorted(gt_set.alphabet.items(),
key=lambda x: x[1],
reverse=True):
char = make_printable(k)
if char == k:
char = '\t' + char
logger.info(f'{char}\t{v}')
logger.debug('Encoding training set')
# use model codec when given
if append:
# is already loaded
nn = cast(vgsl.TorchVGSLModel, nn)
gt_set.encode(codec)
message('Slicing and dicing model ', nl=False)
# now we can create a new model
spec = '[{} O1c{}]'.format(spec[1:-1],
gt_set.codec.max_label() + 1)
logger.info(
f'Appending {spec} to existing model {nn.spec} after {append}')
nn.append(append, spec)
nn.add_codec(gt_set.codec)
message('\u2713', fg='green')
logger.info(f'Assembled model spec: {nn.spec}')
elif load:
# is already loaded
nn = cast(vgsl.TorchVGSLModel, nn)
# prefer explicitly given codec over network codec if mode is 'both'
codec = codec if (codec and resize == 'both') else nn.codec
try:
gt_set.encode(codec)
except KrakenEncodeException:
message('Network codec not compatible with training set')
alpha_diff = set(gt_set.alphabet).difference(
set(codec.c2l.keys()))
if resize == 'fail':
logger.error(
f'Training data and model codec alphabets mismatch: {alpha_diff}'
)
return None
elif resize == 'add':
message('Adding missing labels to network ', nl=False)
logger.info(
f'Resizing codec to include {len(alpha_diff)} new code points'
)
codec.c2l.update({
k: [v]
for v, k in enumerate(alpha_diff,
start=codec.max_label() + 1)
})
nn.add_codec(PytorchCodec(codec.c2l))
logger.info(
f'Resizing last layer in network to {codec.max_label()+1} outputs'
)
nn.resize_output(codec.max_label() + 1)
gt_set.encode(nn.codec)
message('\u2713', fg='green')
elif resize == 'both':
message('Fitting network exactly to training set ',
nl=False)
logger.info(
f'Resizing network or given codec to {gt_set.alphabet} code sequences'
)
gt_set.encode(None)
ncodec, del_labels = codec.merge(gt_set.codec)
logger.info(
f'Deleting {len(del_labels)} output classes from network ({len(codec)-len(del_labels)} retained)'
)
gt_set.encode(ncodec)
nn.resize_output(ncodec.max_label() + 1, del_labels)
message('\u2713', fg='green')
else:
logger.error(f'invalid resize parameter value {resize}')
return None
else:
gt_set.encode(codec)
logger.info(
f'Creating new model {spec} with {gt_set.codec.max_label()+1} outputs'
)
spec = '[{} O1c{}]'.format(spec[1:-1],
gt_set.codec.max_label() + 1)
nn = vgsl.TorchVGSLModel(spec)
# initialize weights
message('Initializing model ', nl=False)
nn.init_weights()
nn.add_codec(gt_set.codec)
# initialize codec
message('\u2713', fg='green')
if nn.one_channel_mode and gt_set.im_mode != nn.one_channel_mode:
logger.warning(
f'Neural network has been trained on mode {nn.one_channel_mode} images, training set contains mode {gt_set.im_mode} data. Consider setting `force_binarization`'
)
if format_type != 'path' and nn.seg_type == 'bbox':
logger.warning(
'Neural network has been trained on bounding box image information but training set is polygonal.'
)
# half the number of data loading processes if device isn't cuda and we haven't enabled preloading
if device == 'cpu' and not preload:
loader_threads = threads // 2
else:
loader_threads = threads
train_loader = InfiniteDataLoader(
gt_set,
batch_size=hyper_params['batch_size'],
shuffle=True,
num_workers=loader_threads,
pin_memory=True,
collate_fn=collate_sequences)
threads = max(threads - loader_threads, 1)
# don't encode validation set as the alphabets may not match causing encoding failures
val_set.no_encode()
val_loader = DataLoader(val_set,
batch_size=hyper_params['batch_size'],
num_workers=loader_threads,
pin_memory=True,
collate_fn=collate_sequences)
logger.debug('Constructing {} optimizer (lr: {}, momentum: {})'.format(
hyper_params['optimizer'], hyper_params['lrate'],
hyper_params['momentum']))
# set model type metadata field
nn.model_type = 'recognition'
# set mode to trainindg
nn.train()
# set number of OpenMP threads
logger.debug(f'Set OpenMP threads to {threads}')
nn.set_num_threads(threads)
optim = getattr(torch.optim,
hyper_params['optimizer'])(nn.nn.parameters(), lr=0)
if 'seg_type' not in nn.user_metadata:
nn.user_metadata[
'seg_type'] = 'baselines' if format_type != 'path' else 'bbox'
tr_it = TrainScheduler(optim)
if hyper_params['schedule'] == '1cycle':
add_1cycle(tr_it, int(len(gt_set) * hyper_params['epochs']),
hyper_params['lrate'], hyper_params['momentum'],
hyper_params['momentum'] - 0.10,
hyper_params['weight_decay'])
elif hyper_params['schedule'] == 'exponential':
add_exponential_decay(tr_it,
int(len(gt_set) * hyper_params['epochs']),
len(gt_set), hyper_params['lrate'], 0.95,
hyper_params['momentum'],
hyper_params['weight_decay'])
else:
# constant learning rate scheduler
tr_it.add_phase(1, 2 * (hyper_params['lrate'], ),
2 * (hyper_params['momentum'], ),
hyper_params['weight_decay'], annealing_const)
if hyper_params['quit'] == 'early':
st_it = EarlyStopping(hyper_params['min_delta'],
hyper_params['lag'])
elif hyper_params['quit'] == 'dumb':
st_it = EpochStopping(hyper_params['epochs'] -
hyper_params['completed_epochs'])
else:
logger.error(f'Invalid training interruption scheme {quit}')
return None
trainer = cls(model=nn,
optimizer=optim,
device=device,
filename_prefix=output,
event_frequency=hyper_params['freq'],
train_set=train_loader,
val_set=val_loader,
stopper=st_it)
trainer.add_lr_scheduler(tr_it)
return trainer
def train(ctx, pad, output, spec, append, load, savefreq, report, quit, epochs,
lag, min_delta, device, optimizer, lrate, momentum, weight_decay,
schedule, partition, normalization, codec, resize, reorder,
training_files, evaluation_files, preload, threads, ground_truth):
"""
Trains a model from image-text pairs.
"""
if not load and append:
raise click.BadOptionUsage(
'append', 'append option requires loading an existing model')
if resize != 'fail' and not load:
raise click.BadOptionUsage(
'resize', 'resize option requires loading an existing model')
import re
import torch
import shutil
import numpy as np
from torch.utils.data import DataLoader
from kraken.lib import models, vgsl, train
from kraken.lib.util import make_printable
from kraken.lib.train import EarlyStopping, EpochStopping, TrainStopper, TrainScheduler, add_1cycle
from kraken.lib.codec import PytorchCodec
from kraken.lib.dataset import GroundTruthDataset, compute_error, generate_input_transforms
logger.info('Building ground truth set from {} line images'.format(
len(ground_truth) + len(training_files)))
# load model if given. if a new model has to be created we need to do that
# after data set initialization, otherwise to output size is still unknown.
nn = None
if load:
logger.info('Loading existing model from {} '.format(load))
message('Loading model {}'.format(load), nl=False)
nn = vgsl.TorchVGSLModel.load_model(load)
message('\u2713', fg='green', nl=False)
# preparse input sizes from vgsl string to seed ground truth data set
# sizes and dimension ordering.
if not nn:
spec = spec.strip()
if spec[0] != '[' or spec[-1] != ']':
raise click.BadOptionUsage(
'spec', 'VGSL spec {} not bracketed'.format(spec))
blocks = spec[1:-1].split(' ')
m = re.match(r'(\d+),(\d+),(\d+),(\d+)', blocks[0])
if not m:
raise click.BadOptionUsage(
'spec', 'Invalid input spec {}'.format(blocks[0]))
batch, height, width, channels = [int(x) for x in m.groups()]
else:
batch, channels, height, width = nn.input
try:
transforms = generate_input_transforms(batch, height, width, channels,
pad)
except KrakenInputException as e:
raise click.BadOptionUsage('spec', str(e))
# disable automatic partition when given evaluation set explicitly
if evaluation_files:
partition = 1
ground_truth = list(ground_truth)
# merge training_files into ground_truth list
if training_files:
ground_truth.extend(training_files)
if len(ground_truth) == 0:
raise click.UsageError(
'No training data was provided to the train command. Use `-t` or the `ground_truth` argument.'
)
np.random.shuffle(ground_truth)
if len(ground_truth) > 2500 and not preload:
logger.info(
'Disabling preloading for large (>2500) training data set. Enable by setting --preload parameter'
)
preload = False
# implicit preloading enabled for small data sets
if preload is None:
preload = True
tr_im = ground_truth[:int(len(ground_truth) * partition)]
if evaluation_files:
logger.debug('Using {} lines from explicit eval set'.format(
len(evaluation_files)))
te_im = evaluation_files
else:
te_im = ground_truth[int(len(ground_truth) * partition):]
logger.debug('Taking {} lines from training for evaluation'.format(
len(te_im)))
gt_set = GroundTruthDataset(normalization=normalization,
reorder=reorder,
im_transforms=transforms,
preload=preload)
with log.progressbar(tr_im, label='Building training set') as bar:
for im in bar:
logger.debug('Adding line {} to training set'.format(im))
try:
gt_set.add(im)
except FileNotFoundError as e:
logger.warning('{}: {}. Skipping.'.format(
e.strerror, e.filename))
except KrakenInputException as e:
logger.warning(str(e))
val_set = GroundTruthDataset(normalization=normalization,
reorder=reorder,
im_transforms=transforms,
preload=preload)
with log.progressbar(te_im, label='Building validation set') as bar:
for im in bar:
logger.debug('Adding line {} to validation set'.format(im))
try:
val_set.add(im)
except FileNotFoundError as e:
logger.warning('{}: {}. Skipping.'.format(
e.strerror, e.filename))
except KrakenInputException as e:
logger.warning(str(e))
logger.info(
'Training set {} lines, validation set {} lines, alphabet {} symbols'.
format(len(gt_set._images), len(val_set._images),
len(gt_set.alphabet)))
alpha_diff = set(gt_set.alphabet).symmetric_difference(
set(val_set.alphabet))
if alpha_diff:
logger.warn('alphabet mismatch {}'.format(alpha_diff))
logger.info('grapheme\tcount')
for k, v in sorted(gt_set.alphabet.items(),
key=lambda x: x[1],
reverse=True):
char = make_printable(k)
if char == k:
char = '\t' + char
logger.info(u'{}\t{}'.format(char, v))
logger.debug('Encoding training set')
# use model codec when given
if append:
# is already loaded
nn = cast(vgsl.TorchVGSLModel, nn)
gt_set.encode(codec)
message('Slicing and dicing model ', nl=False)
# now we can create a new model
spec = '[{} O1c{}]'.format(spec[1:-1], gt_set.codec.max_label() + 1)
logger.info('Appending {} to existing model {} after {}'.format(
spec, nn.spec, append))
nn.append(append, spec)
nn.add_codec(gt_set.codec)
message('\u2713', fg='green')
logger.info('Assembled model spec: {}'.format(nn.spec))
elif load:
# is already loaded
nn = cast(vgsl.TorchVGSLModel, nn)
# prefer explicitly given codec over network codec if mode is 'both'
codec = codec if (codec and resize == 'both') else nn.codec
try:
gt_set.encode(codec)
except KrakenEncodeException as e:
message('Network codec not compatible with training set')
alpha_diff = set(gt_set.alphabet).difference(set(codec.c2l.keys()))
if resize == 'fail':
logger.error(
'Training data and model codec alphabets mismatch: {}'.
format(alpha_diff))
ctx.exit(code=1)
elif resize == 'add':
message('Adding missing labels to network ', nl=False)
logger.info(
'Resizing codec to include {} new code points'.format(
len(alpha_diff)))
codec.c2l.update({
k: [v]
for v, k in enumerate(alpha_diff,
start=codec.max_label() + 1)
})
nn.add_codec(PytorchCodec(codec.c2l))
logger.info(
'Resizing last layer in network to {} outputs'.format(
codec.max_label() + 1))
nn.resize_output(codec.max_label() + 1)
message('\u2713', fg='green')
elif resize == 'both':
message('Fitting network exactly to training set ', nl=False)
logger.info(
'Resizing network or given codec to {} code sequences'.
format(len(gt_set.alphabet)))
gt_set.encode(None)
ncodec, del_labels = codec.merge(gt_set.codec)
logger.info(
'Deleting {} output classes from network ({} retained)'.
format(len(del_labels),
len(codec) - len(del_labels)))
gt_set.encode(ncodec)
nn.resize_output(ncodec.max_label() + 1, del_labels)
message('\u2713', fg='green')
else:
raise click.BadOptionUsage(
'resize', 'Invalid resize value {}'.format(resize))
else:
gt_set.encode(codec)
logger.info('Creating new model {} with {} outputs'.format(
spec,
gt_set.codec.max_label() + 1))
spec = '[{} O1c{}]'.format(spec[1:-1], gt_set.codec.max_label() + 1)
nn = vgsl.TorchVGSLModel(spec)
# initialize weights
message('Initializing model ', nl=False)
nn.init_weights()
nn.add_codec(gt_set.codec)
# initialize codec
message('\u2713', fg='green')
train_loader = DataLoader(gt_set,
batch_size=1,
shuffle=True,
pin_memory=True)
# don't encode validation set as the alphabets may not match causing encoding failures
val_set.training_set = list(zip(val_set._images, val_set._gt))
logger.debug('Constructing {} optimizer (lr: {}, momentum: {})'.format(
optimizer, lrate, momentum))
# set mode to trainindg
nn.train()
# set number of OpenMP threads
logger.debug('Set OpenMP threads to {}'.format(threads))
nn.set_num_threads(threads)
logger.debug('Moving model to device {}'.format(device))
rec = models.TorchSeqRecognizer(nn, train=True, device=device)
optim = getattr(torch.optim, optimizer)(nn.nn.parameters(), lr=0)
tr_it = TrainScheduler(optim)
if schedule == '1cycle':
add_1cycle(tr_it, epochs * len(gt_set), lrate, momentum,
momentum - 0.10, weight_decay)
else:
# constant learning rate scheduler
tr_it.add_phase(1, (lrate, lrate), (momentum, momentum), weight_decay,
train.annealing_const)
st_it = cast(TrainStopper, None) # type: TrainStopper
if quit == 'early':
st_it = EarlyStopping(train_loader, min_delta, lag)
elif quit == 'dumb':
st_it = EpochStopping(train_loader, epochs)
else:
raise click.BadOptionUsage(
'quit', 'Invalid training interruption scheme {}'.format(quit))
for epoch, loader in enumerate(st_it):
with log.progressbar(label='epoch {}/{}'.format(
epoch, epochs - 1 if epochs > 0 else '∞'),
length=len(loader),
show_pos=True) as bar:
acc_loss = torch.tensor(0.0).to(device, non_blocking=True)
for trial, (input, target) in enumerate(loader):
tr_it.step()
input = input.to(device, non_blocking=True)
target = target.to(device, non_blocking=True)
input = input.requires_grad_()
o = nn.nn(input)
# height should be 1 by now
if o.size(2) != 1:
raise KrakenInputException(
'Expected dimension 3 to be 1, actual {}'.format(
o.size(2)))
o = o.squeeze(2)
optim.zero_grad()
# NCW -> WNC
loss = nn.criterion(
o.permute(2, 0, 1), # type: ignore
target,
(o.size(2), ),
(target.size(1), ))
logger.info('trial {}'.format(trial))
if not torch.isinf(loss):
loss.backward()
optim.step()
else:
logger.debug('infinite loss in trial {}'.format(trial))
bar.update(1)
if not epoch % savefreq:
logger.info('Saving to {}_{}'.format(output, epoch))
try:
nn.save_model('{}_{}.mlmodel'.format(output, epoch))
except Exception as e:
logger.error('Saving model failed: {}'.format(str(e)))
if not epoch % report:
logger.debug('Starting evaluation run')
nn.eval()
chars, error = compute_error(rec, list(val_set))
nn.train()
accuracy = (chars - error) / chars
logger.info('Accuracy report ({}) {:0.4f} {} {}'.format(
epoch, accuracy, chars, error))
message('Accuracy report ({}) {:0.4f} {} {}'.format(
epoch, accuracy, chars, error))
st_it.update(accuracy)
if quit == 'early':
message('Moving best model {0}_{1}.mlmodel ({2}) to {0}_best.mlmodel'.
format(output, st_it.best_epoch, st_it.best_loss))
logger.info(
'Moving best model {0}_{1}.mlmodel ({2}) to {0}_best.mlmodel'.
format(output, st_it.best_epoch, st_it.best_loss))
shutil.copy('{}_{}.mlmodel'.format(output, st_it.best_epoch),
'{}_best.mlmodel'.format(output))