using keras-vis with tf.keras

jashshah commented 5 years ago

[x] Check that you are up-to-date with the master branch of keras-vis. You can update with: pip install git+git://github.com/raghakot/keras-vis.git --upgrade --no-deps
[x] If running on TensorFlow, check that you are up-to-date with the latest version. The installation instructions can be found here.
[x] If running on Theano, check that you are up-to-date with the master branch of Theano. You can update with: pip install git+git://github.com/Theano/Theano.git --upgrade --no-deps
[x] Provide a link to a GitHub Gist of a Python script that can reproduce your issue (or just copy the script here if it is short).

Hi,

Kudos to the work done with the library so far. However when using it with tf.keras I am facing the following error:

ValueError: Unknown initializer: GlorotUniform

Here is the code I used to train the model:


from tensorflow.keras.layers import Conv3D, MaxPool3D, Flatten, Dense
from tensorflow.keras.layers import Dropout, Input, BatchNormalization
from tensorflow.keras.layers import AvgPool3D
from tensorflow.keras import Model
from tensorflow.keras.optimizers import Adam
from tensorflow.keras.layers import ReLU
from tensorflow.keras.layers import GlobalAveragePooling3D
from tensorflow.keras.layers import Add

def block_a(inputs, block_name, filters):
    with tf.name_scope(block_name):
        bn = BatchNormalization()(inputs)
        relu = ReLU()(bn)
        conv1 = Conv3D(filters=filters, strides=(1, 1, 1), kernel_size=(3, 3, 3), padding='same')(relu)
        bn1 = BatchNormalization()(conv1)
        relu1 = ReLU()(bn1)
        conv2 = Conv3D(filters=filters, strides=(1, 1, 1), kernel_size=(3, 3, 3), padding='same')(relu1)
    return conv2

def block_a(inputs, block_name, filters):
    with tf.name_scope(block_name):
        bn = BatchNormalization()(inputs)
        relu = ReLU()(bn)
        conv1 = Conv3D(filters=filters, strides=(1, 1, 1), kernel_size=(3, 3, 3), padding='same')(relu)
        bn1 = BatchNormalization()(conv1)
        relu1 = ReLU()(bn1)
        conv2 = Conv3D(filters=filters, strides=(1, 1, 1), kernel_size=(3, 3, 3), padding='same')(relu1)
    return conv2

def block_c(inputs, block_name, filters):

    with tf.name_scope(block_name):
        conv1a = Conv3D(filters=filters, strides=(1, 1, 1), kernel_size=(3, 3, 3), padding='same')(inputs)
        bn1a = BatchNormalization()(conv1a)
        relu1a = ReLU()(bn1a)
        conv1b = Conv3D(filters=filters, strides=(1, 1, 1), kernel_size=(3, 3, 3), padding='same')(relu1a)
        bn1b = BatchNormalization()(conv1b)
        relu1b = ReLU()(bn1b)

    return relu1b

inputs = Input(shape=(110, 110, 110, 1), name='input')
block1 = block_c(inputs, 'block_c', 32)
stepdown = Conv3D(filters=64, strides=(2, 2, 2), kernel_size=(3, 3, 3), padding='same')(block1)
block2 = block_a(stepdown, 'block_a1', 64)
addblock2 = Add()([block2, stepdown])
block3 = block_a(addblock2, 'block_a2', 64)
addblock3 = Add()([addblock2, block3])
block4 = block_b(addblock3, 'block_b1', 64)
block5 = block_a(block4, 'block_a3', 64)
addblock5 = Add()([block5, block4])
block6 = block_a(addblock5, 'block_a4', 64)
addblock6 = Add()([block6, addblock5])
block7 = block_b(addblock6, 'block_b2', 128)
block8 = block_a(block7, 'block_a5', 128)
addblock8 = Add()([block8, block7])
block9 = block_a(addblock8, 'block_a6', 128)
addblock9 = Add()([block9, addblock8])
pool = GlobalAveragePooling3D()(addblock9)
dense = Dense(128, activation='relu')(pool)
output = Dense(1, activation='softmax')(dense)

model = Model(inputs=inputs, outputs=output)
model.compile(optimizer=Adam(lr=1e-3), loss='binary_crossentropy', metrics=['accuracy'])

Getting the cam visualizations:

from vis.visualization import visualize_cam
from vis.utils import utils
from tensorflow.keras import activations

layer_idx = utils.find_layer_idx(model, 'dense_1')

model.layers[layer_idx].activation = activations.linear

model = utils.apply_modifications(model)

Here is the error that I get:

ValueError: Unknown initializer: GlorotUniform

And when I change the utils.apply_modifications(model) to utils.apply_modifications(model, custom_objects={"GlorotUniform": tf.keras.initializers.glorot_uniform}) I get

TypeError: tuple indices must be integers or slices, not list

Using tensorflow 1.12.0

Any help will be appreciated.

keisen commented 5 years ago

Hi, @jashshah . The problem seems to be caused by tf.keras or something else. At least, I believe keras-vis don't have the cause of the problem, because utils.apply_modifications(model) just call keras.models.Model.save() and keras.models.load_model() internally.

https://github.com/raghakot/keras-vis/blob/668b0e11dab93f3487f23c17e07f40554a8939e9/vis/utils/utils.py#L112-L113

I suggest you ask StackOverFlow or Tensorflow's community this problem.

B-Yassine commented 5 years ago

try this line of code it worked for me: with CustomObjectScope({'GlorotUniform': glorot_uniform()}): .............................................

FabianGroeger96 commented 5 years ago

I have the same problem

B-Yassine commented 5 years ago

Try this code if you used a custom loss for training your model, it also worked for me:

from keras.utils import CustomObjectScope
from keras.initializers import glorot_uniform
from tensorflow.keras.models import load_model

def myLoss(y_true, y_pred):
    ...
    return  Something

model = load_model('yourModel.hdf5', custom_objects={'myLoss': myLoss, "GlorotUniform": tf.keras.initializers.glorot_uniform})

GangaNagarajan commented 5 years ago

Query : can't use Keras-viz : from vis.visualization import visualize_cam There is a compatibility issue with scipy version. scipy.misc import imresize not available in colab

Solution : Check that you are up-to-date with the master branch of keras-vis. You can update with: pip install git+git://github.com/raghakot/keras-vis.git --upgrade --no-deps

Above solution resolves my issue. Thanks.

andrewt3000 commented 4 years ago

I have the same problem on TensorFlow 1.14.0.

When I call

    from tensorflow import keras
    model = utils.apply_modifications(model)

ValueError: Unknown initializer: GlorotUniform

I get a different second error... when I call

model = utils.apply_modifications(model, custom_objects={"GlorotUniform": keras.initializers.glorot_uniform})

I get error

TypeError: Unexpected keyword argument passed to optimizer: name

I also get a different error when from import keras

TypeError: glorot_uniform() got an unexpected keyword argument 'dtype'

andrewt3000 commented 4 years ago

I suspect the problem is... I am using tensorflow's version of keras

from tensorflow import keras

Whereas... this utils.py in keras-viz is using the keras library directly

from keras.models import load_model

Perhaps there is a mismatch between the two versions.

https://stackoverflow.com/questions/53183865/unknown-initializer-glorotuniform-when-loading-keras-model

khordoo commented 4 years ago

Installing TensorFlow version 2.0 and passing the custom objects solved the issue for me.

from tensorflow.python.keras.models import load_model
 model = load_model('model.h5',custom_objects={"adam": tf.keras.optimizers.Adam, 
 "mae":tf.keras.losses.mean_absolute_error})

andrewt3000 commented 4 years ago

Installing TensorFlow version 2.0 and passing the custom objects solved the issue for me.

Does this package support TF 2.0? When I run with TF 2.0, I get:

module 'tensorflow' has no attribute 'placeholder'

alessandro-montanari commented 4 years ago

I'm trying to use the library with Tensorflow 1.14 using tf.keras, however if I install it with pip install git+git://github.com/raghakot/keras-vis.git --upgrade --no-deps I get an error saying that the keras module has not been found. That's expected.

Instead, if I install it with pip install git+git://github.com/raghakot/keras-vis.git --upgrade I get TypeError: tuple indices must be integers or slices, not list when I run model = utils.apply_modifications(model)

[Edit] I have ported the library to tf.keras here https://github.com/alessandro-montanari/keras-vis/tree/tf-keras-support It seems to be working but has not been tested deeply yet.

keisen commented 4 years ago

Sorry for inconvenience. We've NOT been working for support Tensorflow 2.0 yet. For now, you guys can use alternative library that I developed for my own experiments. Would you please try it if it's okay.

https://github.com/keisen/tf-keras-vis

njerschow commented 4 years ago

Hey,

I'm getting one of two errors: Error 1:

(gcloud.beta.ai-platform.versions.create) Create Version failed. Bad model detected with error:  "Failed to load model: Unexpected error when loading the model: Unexpected keyword argument passed to optimizer: learning_rate (Error code: 0)"

Error 2:

(gcloud.beta.ai-platform.versions.create) Create Version failed. Bad model detected with error:  "Failed to load model: Unexpected error when loading the model: Unexpected keyword argument passed to optimizer: name (Error code: 0)"

(the only diff in the above to errors is that one complains about learning rate, and the other complains about name)

both errors pop up even though I have made no code changes. I am trying to publish a model onto the ai-platform following this guide exactly: https://cloud.google.com/ml-engine/docs/tensorflow/custom-prediction-routine-keras

I've tried the custom_objects every way I think possible. Does anyone know a solution?

neso613 commented 4 years ago

I'm trying to use the library with Tensorflow 1.14 using tf.keras, however if I install it with pip install git+git://github.com/raghakot/keras-vis.git --upgrade --no-deps I get an error saying that the keras module has not been found. That's expected.

Instead, if I install it with pip install git+git://github.com/raghakot/keras-vis.git --upgrade I get TypeError: tuple indices must be integers or slices, not list when I run model = utils.apply_modifications(model)

[Edit] I have ported the library to tf.keras here https://github.com/alessandro-montanari/keras-vis/tree/tf-keras-support It seems to be working but has not been tested deeply yet.

I am getting same errror

alexandraiov commented 4 years ago

try this line of code it worked for me: with CustomObjectScope({'GlorotUniform': glorot_uniform()}): .............................................

When I do that I get: TypeError: glorot_uniform() got an unexpected keyword argument 'dtype'

sauravmishra1710 commented 4 years ago

I am facing the same issue with Keras Vis.

ValueError: Unknown initializer: GlorotUniform

_model.layers[layer_index].activation = activations.linear model = utils.applymodifications(model)

I have tried all the possible approaches mentioned above. Also tried re-installing via pip install git+https://github.com/raghakot/keras-vis.git -U

Let me know if a solution is available for this.

andrewt3000 commented 4 years ago

I switched to tf explain.

https://github.com/sicara/tf-explain

raghakot / keras-vis

using keras-vis with tf.keras #160