keras to CoreML with custom loss function

[hollance]

keras to CoreML with custom loss function

[hollance]

The Keras model only needs to have that loss for training, not for making predictions.

[hollance]

You should convert base_model instead of model. That’s all.

[hollance]

Your Core ML model should not include the loss function as that is used only for training, which is not possible with Core ML.

[un-lock-me]

Thank you so much for replying back:)
you mean that we can not convert any model which has custom loss to core ML?
if so, do you have any idea how can I convert the ocr code to any other model understandable by ios?
Most OCR keras code I explored have the lambda layer for their loss.

Thanks again for taking the time

[un-lock-me]

this is my code:

import coremltools

def convert_lambda(layer):
    # Only convert this Lambda layer if it is for our swish function.
    if layer.function == ctc_lambda_func:
        params = NeuralNetwork_pb2.CustomLayerParams()

        # The name of the Swift or Obj-C class that implements this layer.
        params.className = "x"

        # The desciption is shown in Xcode's mlmodel viewer.
        params.description = "A fancy new loss"

        return params
    else:
        return None


print("\nConverting the model:")

# Convert the model to Core ML.
coreml_model = coremltools.converters.keras.convert(
    model,
    # 'weightswithoutstnlrchangedbackend.best.hdf5',
    input_names="image",
    image_input_names="image",
    output_names="output",
    add_custom_layers=True,
    custom_conversion_functions={"Lambda": convert_lambda},
    )

and this is the error:

Converting the model:
Traceback (most recent call last):
  File "/home/sgnbx/Downloads/projects/CRNN-with-STN-master/CRNN_with_STN.py", line 201, in <module>
    custom_conversion_functions={"Lambda": convert_lambda},
  File "/home/sgnbx/anaconda3/envs/tf_gpu/lib/python3.6/site-packages/coremltools/converters/keras/_keras_converter.py", line 760, in convert
    custom_conversion_functions=custom_conversion_functions)
  File "/home/sgnbx/anaconda3/envs/tf_gpu/lib/python3.6/site-packages/coremltools/converters/keras/_keras_converter.py", line 556, in convertToSpec
    custom_objects=custom_objects)
  File "/home/sgnbx/anaconda3/envs/tf_gpu/lib/python3.6/site-packages/coremltools/converters/keras/_keras2_converter.py", line 255, in _convert
    if input_names[idx] in input_name_shape_dict:
IndexError: list index out of range
Input name length mismatch

[hollance]

The easiest thing to do is to remove this lambda layer from the Keras model first.

[un-lock-me]

Isnt it obligatory to have lambda layer here as the model need to use a custom version of the loss rather than the already built in loss in the keras?
I have looked at a couple of the code with crrn architecture, all had the same custom loss.

[un-lock-me]

Can you please link me to a blog, source whatever?
I cant understand your saying honestly. as I just use the custom loss in the training. then saved the model, then use the saved model to convert to CoreMl.
Which part I am missing that I can not get you. sorry for the many questions and taking your time.

[hollance]

After training, remove the lambda layer from the Keras model and save this model. Then use coremltools to convert that Keras model (without the lambda layer) to Core ML.

[un-lock-me]

So I can not get your approach.
Do you mind have a quick look on the code and say that which part I have to change?(myconfusion is that, I can save the trained model, how can I save a model without training, what will be saved on the model, in terms of weight or other parameter?)
I am really sorry for taking your time and much appreciated.
(if you know a link regarding your approach I will be happy to explore it and do by myself but unfortunately I can not get your saying)

from coremltools.proto import NeuralNetwork_pb2
from keras import backend ,optimizers
from keras.callbacks import *
from keras.layers import *
from keras.models import *
from keras.optimizers import SGD
from keras.utils import *

from keras.callbacks import ModelCheckpoint
from keras.callbacks import TensorBoard

# from STN.spatial_transformer import SpatialTransformer

from Batch_Generator import img_gen, img_gen_val

from config import learning_rate, load_model_path, width, height, characters, label_len, label_classes, \
    cp_save_path, base_model_path, tb_log_dir

    
# functions
class Evaluate(Callback):

    def on_epoch_end(self, epoch, logs=None):
        acc = evaluate(base_model)
        print('')
        print('acc:'+str(acc)+"%")

evaluator = Evaluate()


def evaluate(input_model):
    correct_prediction = 0
    generator = img_gen_val()

    x_test, y_test = next(generator)
    # print(" ")
    y_pred = input_model.predict(x_test) 
    shape = y_pred[:, 2:, :].shape 
    ctc_decode = backend.ctc_decode(y_pred[:, 2:, :], input_length=np.ones(shape[0])*shape[1])[0][0]
    out = backend.get_value(ctc_decode)[:, :label_len]

    for m in range(1000):
        result_str = ''.join([characters[k] for k in out[m]])
        result_str = result_str.replace('-', '')
        if result_str == y_test[m]:
            correct_prediction += 1
            # print(m)
        else:
            print(result_str, y_test[m])

    return correct_prediction*1.0/10

def ctc_lambda_func(args):
    iy_pred, ilabels, iinput_length, ilabel_length = args
    # the 2 is critical here since the first couple outputs of the RNN
    # tend to be garbage:
    iy_pred = iy_pred[:, 2:, :]  # no such influence
    return backend.ctc_batch_cost(ilabels, iy_pred, iinput_length, ilabel_length)


# initial bias_initializer
def loc_net(input_shape):
    b = np.zeros((2, 3), dtype='float32')
    b[0, 0] = 1
    b[1, 1] = 1
    w = np.zeros((64, 6), dtype='float32')
    weights = [w, b.flatten()]

    loc_input = Input(input_shape)

    loc_conv_1 = Conv2D(16, (5, 5), padding='same', activation='relu')(loc_input)
    loc_conv_2 = Conv2D(32, (5, 5), padding='same', activation='relu')(loc_conv_1)
    loc_fla = Flatten()(loc_conv_2)
    loc_fc_1 = Dense(64, activation='relu')(loc_fla)
    loc_fc_2 = Dense(6, weights=weights)(loc_fc_1)

    output = Model(inputs=loc_input, outputs=loc_fc_2)

    return output


# build model
inputShape = Input((width, height, 3))  # base on Tensorflow backend
conv_1 = Conv2D(64, (3, 3), activation='relu', padding='same')(inputShape)
batchnorm_1 = BatchNormalization()(conv_1)

conv_2 = Conv2D(128, (3, 3), activation='relu', padding='same')(conv_1)
conv_3 = Conv2D(256, (3, 3), activation='relu', padding='same')(conv_2)
batchnorm_3 = BatchNormalization()(conv_3)
pool_3 = MaxPooling2D(pool_size=(2, 2))(batchnorm_3)

conv_4 = Conv2D(256, (3, 3), activation='relu', padding='same')(pool_3)
conv_5 = Conv2D(512, (3, 3), activation='relu', padding='same')(conv_4)
batchnorm_5 = BatchNormalization()(conv_5)
pool_5 = MaxPooling2D(pool_size=(2, 2))(batchnorm_5)

conv_6 = Conv2D(512, (3, 3), activation='relu', padding='same')(pool_5)
conv_7 = Conv2D(512, (3, 3), activation='relu', padding='same')(conv_6)
batchnorm_7 = BatchNormalization()(conv_7)

bn_shape = batchnorm_7.get_shape()  # (?, {dimension}50, {dimension}12, {dimension}256)

'''----------------------STN-------------------------'''
# you can run the model without this STN part by commenting out the STN lines then connecting batchnorm_7 to x_reshape,
# which may bring you a higher accuracy
# stn_input_shape = batchnorm_7.get_shape()
# loc_input_shape = (stn_input_shape[1].value, stn_input_shape[2].value, stn_input_shape[3].value)
# stn = SpatialTransformer(localization_net=loc_net(loc_input_shape),
#                          output_size=(loc_input_shape[0], loc_input_shape[1]))(batchnorm_7)
'''----------------------STN-------------------------'''

print(bn_shape)  # (?, 50, 7, 512)

# reshape to (batch_size, width, height*dim)
# x_reshape = Reshape(target_shape=(int(bn_shape[1]), int(bn_shape[2] * bn_shape[3])))(stn_7)
x_reshape = Reshape(target_shape=(int(bn_shape[1]), int(bn_shape[2] * bn_shape[3])))(batchnorm_7)

fc_1 = Dense(128, activation='relu')(x_reshape)  # (?, 50, 128)

print(x_reshape.get_shape())  # (?, 50, 3584)
print(fc_1.get_shape())  # (?, 50, 128)

rnn_1 = LSTM(128, kernel_initializer="he_normal", return_sequences=True)(fc_1)
rnn_1b = LSTM(128, kernel_initializer="he_normal", go_backwards=True, return_sequences=True)(fc_1)
rnn1_merged = add([rnn_1, rnn_1b])

rnn_2 = LSTM(128, kernel_initializer="he_normal", return_sequences=True)(rnn1_merged)
rnn_2b = LSTM(128, kernel_initializer="he_normal", go_backwards=True, return_sequences=True)(rnn1_merged)
rnn2_merged = concatenate([rnn_2, rnn_2b])

drop_1 = Dropout(0.25)(rnn2_merged)

fc_2 = Dense(label_classes, kernel_initializer='he_normal', activation='softmax')(drop_1)

# model setting
base_model = Model(inputs=inputShape, outputs=fc_2)  # the model for prediecting

labels = Input(name='the_labels', shape=[label_len], dtype='float32')
input_length = Input(name='input_length', shape=[1], dtype='int64')
label_length = Input(name='label_length', shape=[1], dtype='int64')

loss_out = Lambda(ctc_lambda_func, output_shape=(1,), name='ctc')([fc_2, labels, input_length, label_length])

model = Model(inputs=[inputShape, labels, input_length, label_length], outputs=[loss_out])  # the model for trainning

# clipnorm seems to speeds up convergence
sgd = SGD(lr=learning_rate, decay=1e-6, momentum=0.9, nesterov=True, clipnorm=5)
adam = optimizers.Adam()

model.compile(loss={'ctc': lambda y_true, y_pred: y_pred}, optimizer=sgd)

model.summary()  # print a summary representation of your model.
# plot_model(model, to_file='CRNN_with_STN.png', show_shapes=True)  # save a image which is the architecture of the model
cp_save_path = "weightswithoutstnlrchanged.best.hdf5"
checkpoint = ModelCheckpoint(cp_save_path, monitor='loss', verbose=1, save_best_only=True, mode='min')

# if you want to load a trained model weights, just fill the load_model_path in config.py, it will automaticlly add into the
# new trainning. if you want to train a new model, just set load_model_path = ''.


# if len(load_model_path) > 5:
#     savedmodel = load_model(load_model_path, custom_objects={"bknd": backend,'loss':loss_out})


# try your own fit_generator() settings, you may get a better result
model.fit_generator(img_gen(input_shape=bn_shape), steps_per_epoch=100, epochs=1, verbose=1,
                    callbacks=[evaluator,
                               checkpoint,
                               TensorBoard(log_dir=tb_log_dir)])

base_model.save(base_model_path)

import coremltools

def convert_lambda(layer):
    # Only convert this Lambda layer if it is for our swish function.
    if layer.function == ctc_lambda_func:
        params = NeuralNetwork_pb2.CustomLayerParams()

        # The name of the Swift or Obj-C class that implements this layer.
        params.className = "x"

        # The desciption is shown in Xcode's mlmodel viewer.
        params.description = "A fancy new loss"

        return params
    else:
        return None


print("\nConverting the model:")

# Convert the model to Core ML.
coreml_model = coremltools.converters.keras.convert(
    model,
    # 'weightswithoutstnlrchangedbackend.best.hdf5',
    input_names="image",
    image_input_names="image",
    output_names="output",
    add_custom_layers=True,
    custom_conversion_functions={"Lambda": convert_lambda},