I'm trying to apply transfer learning to MNIST using MobileNet weights in Keras. Keras documentation to use MobileNet https://keras.io/applications/#mobilenet
Mobilenet accepts 224x224x3 as input but MNIST is 28x28x1. I'm creating a Lambda layer which can convert 28x28x1 image into 224x224x3 and send it as input to MobileNet. The following code causes
TypeError: Input layers to a Model must be InputLayer objects. Received inputs: Tensor("lambda_2/ResizeNearestNeighbor:0", shape=(?, 224, 224, 3), dtype=float32). Input 0 (0-based) originates from layer type Lambda.
height = 28
width = 28
input_image = Input(shape=(height,width,1))
def resize_image_to_inception(x):
x = K.repeat_elements(x, 3, axis=3)
x = K.resize_images(x, 8, 8, data_format="channels_last")
return x
input_image_ = Lambda(resize_image_to_inception, output_shape=(224, 224, 3))(input_image)
print(type(input_image_))
base_model = MobileNet(input_tensor=input_image_, weights='imagenet', include_top=False)
Related
I'm trying to build a Sequential model with tensorflow.
import tensorflow as tf
import keras
from tensorflow.keras import layers
from keras import optimizers
import numpy as np
model = keras.Sequential (name="model")
model.add(keras.Input(shape=(786,)))
model.add(layers.Dense(2048, activation="relu", name="layer1"))
model.add(layers.Dense(786, activation="relu", name="layer2"))
model.add(layers.Dense(786, activation="relu", name="layer3"))
output = model.add(layers.Dense(786, activation="relu", name="output"))
model.summary()
model.compile(
optimizer=tf.optimizers.Adam(), # Optimizer
loss=keras.losses.CategoricalCrossentropy(),
metrics=[keras.metrics.SparseCategoricalAccuracy()],
)
history = model.fit(
x_train,
y_train,
batch_size=1,
epochs=5,
)
The input shape is a vector with length of 768 (so the input shape is (768,) right?), representing a chess board:
def get_dataset():
container = np.load('/content/drive/MyDrive/test_data_vector.npz')
b, v = container['arr_0'], container['arr_1']
v = np.asarray(v / abs(v).max() / 2 + 0.5, dtype=np.float32) # normalization (0 - 1)
return b, v
xtrain, ytrain = get_dataset()
print(xtrain.shape)
print(ytrain.shape)
>> (37, 786) #there are 37 samples
>> (37, 786)
But I always get the error:
ValueError: Input 0 of layer model is incompatible with the layer: expected axis -1 of input shape to have value 786 but received input with shape (1, 1, 768)
I tried with np.expand_dims(), which ended in the same Error.
The error is just a typo, as the user mentioned the issue is resolved by changing the output shape from 786 to 768 and the issue is resolved.
One suggestion based on the model structure.
The number of units are not related to your input shape, you don't have to match that number.
The number of units like 2048 and 786 in dense layer is too large and this may not help the model to learn better.
Try with smaller numbers like 32,64 etc, you can refer some of the examples in the tensorflow document.
when feeding a tf.data.Dataset to train EfficientnetB0 model I get the following error:
ValueError: in converted code:
C:\Users\fconrad\AppData\Local\Continuum\anaconda3\envs\venv_spielereien\lib\site-packages\tensorflow_core\python\keras\engine\training_v2.py:677 map_fn
batch_size=None)
C:\Users\fconrad\AppData\Local\Continuum\anaconda3\envs\venv_spielereien\lib\site-packages\tensorflow_core\python\keras\engine\training.py:2410 _standardize_tensors
exception_prefix='input')
C:\Users\fconrad\AppData\Local\Continuum\anaconda3\envs\venv_spielereien\lib\site-packages\tensorflow_core\python\keras\engine\training_utils.py:573 standardize_input_data
'with shape ' + str(data_shape))
ValueError: Error when checking input: expected efficientnet-b0_input to have 4 dimensions, but got array with shape (224, 224, 3)
I realy wonder why this happens, since when I create a batch from my Dataset:
train_generator = (tf.data.Dataset
.from_tensor_slices((train_imgs, train_labels))
.map(read_img)
.map(flip_img)
.map(brightness)
.map(blur)
.map(noise)
.map(rotate_90)
.repeat()
.shuffle(512)
.batch(BATCH_SIZE)
.prefetch(True))
validation_generator = (tf.data.Dataset
.from_tensor_slices((validation_imgs, validation_labels))
.map(read_img)
)
print(train_generator.__iter__().__next__()[0].shape)
I get the expected result (64, 224, 224, 3).
But after creating the model the error above raises when I start training:
effn = tfkeras.EfficientNetB0(include_top=False, input_shape=img_shape, classes=4)
effn_model = tf.keras.Sequential()
effn_model.add(effn)
effn_model.add(tf.keras.layers.GlobalAveragePooling2D())
effn_model.add(tf.keras.layers.Dense(4, 'softmax'))
effn_model.compile(optimizer= 'adam', loss='categorical_crossentropy', metrics= ['categorical_accuracy'])
effn_model.fit(train_generator,
epochs=20,
steps_per_epoch=train_imgs.shape[0] // BATCH_SIZE,
validation_data= validation_generator)
Does anyone know why the slices from dataset have shape (64,224,224,3) but the model doesnt recognize the batch dimension? when I try to train a keras.application model, everything works fine.
I use tensorflow 2.1 and the pip install of efficientnet. Thanks
as explained here keras.io/api/applications/efficientnet/
input_shape: Optional shape tuple, only to be specified if include_top is False. It should have exactly 3 inputs channels.
as so try this->
from tensorflow.keras.applications.efficientnet import EfficientNetB0, EfficientNetB5
mm = EfficientNetB0(include_top=True, weights=None, input_tensor=None, input_shape=(128, 128, 3), pooling=None, classes=2, classifier_activation="sigmoid")
mm.summary()
note the input_shape=(128, 128, 3) It has 3 channels.
I trying to build a deep learning model with VGG16 on top. I have implemented it in Keras using following code:
image_input = Input(shape=(224, 224, 3))
model = VGG16(input_tensor=image_input, include_top=True,weights='imagenet')
model.summary()
fc7 = model.get_layer('fc2').output
conv1d = Conv1D(1,5,activation='relu', name="conv1d",input_shape=(1,4096)) (fc7) #error appears here
# flat = Flatten()(conv1d)
fc8 = Dense(512, activation='relu', name="fc8")(conv1d)
#x= Flatten(name='flatten')(last_layer)
out = Dense(num_classes, activation='softmax', name='output')(fc8)
custom_vgg_model = Model(image_input, out)
custom_vgg_model.summary()
I am getting the following error:
ValueError: Input 0 is incompatible with layer conv1d: expected ndim=3, found ndim=2
Why can't we do the consecutive feature vectors 1d convolution like in the image below?
enter link description here
A fully connected layer in a VGG is 2D, and a 1D convolutional layer expects 3D data.
At the point where VGG adds a Dense layer, it destroys the image format (4D) with a flatten or a global pooling, transforming it into plain data (2D). You no longer have dimensions to use convolutions.
If you try to explain why you want a Conv1D, what do you expect from it, then we could think of an alternative.
Example model:
movie_data = any_data_with_shape((number_of_videos, frames, 224, 224, 3))
movie_input = Input((None,224,224,3)) #None means any number of frames
vgg = VGG16(include_top=True,weights='imagenet')
This part is only necessary if you're getting intermediary outputs from vgg:
vgg_in = vgg.input
vgg_out = vgg.get_layer('fc2').output #make sure this layer exists
vgg = Model(vgg_in, vgg_out)
Continue:
vgg_outs = TimeDistributed(vgg)(movie_input) #out shape (None, frames, fc2_units)
outs = Conv1D(.....)(vgg_outs)
outs = GlobalAveragePooling1D()(outs)
outs = Dense(....)(outs)
.....
your_model = model(move_input, outs)
I am using VGG16 for transfer learning. My images are grayscale. So, I need to change the input channel shape of Vgg16 from (224, 224, 3) to (224, 224, 1). I tried the following code and got error:
TypeError: build() takes from 1 to 2 positional arguments but 4 were given
Can anyone help me where Am I doing it wrong?
vgg16_model= load_model('Fetched_VGG.h5')
vgg16_model.summary()
# transform the model to Sequential
model= Sequential()
for layer in vgg16_model.layers[1:-1]:
model.add(layer)
# Freezing the layers (Oppose weights to be updated)
for layer in model.layers:
layer.trainable = False
model.build(224,224,1)
model.add(Dense(2, activation='softmax', name='predictions'))
you can't, even if you get rid of the input layer, this model has a graph that has already been compiled and your first conv layer expects an input with 3 channels. I don't think there is really an easy work around to make it accept 1 channel if there is any at all.
you need to repeat your data in third dimension and have the same grayscale image in all 3 bands instead of RGB, that works just fine.
if your image has the shape of : (224,224,1):
import numpy as np
gray_image_3band = np.repeat(gray_img, repeats = 3, axis = -1)
if your image has the shape of : (224,224)
gray_image_3band = np.repeat(gray_img[..., np.newaxis], repeats = 3, axis = -1)
you don't need to call the model.build() anymore this way, keep the input layer. but if you ever wanted to call it you need to pass the shape as a tuple like this:
model.build( (224, 224, 1) ) # this is correct, notice the parentheses
I have Keras' image_dim_ordering property set to 'tf', so I define my models as this:
model = Sequential()
model.add(ZeroPadding2D((1, 1), input_shape=(224, 224, 3)))
model.add(Convolution2D(64, 3, 3, activation='relu'))
But when I call load_weights method, it crashes because my model was saved using "th" format:
Exception: Layer weight shape (3, 3, 3, 64) not compatible with provided weight shape (64, 3, 3, 3)
How can I load these weights and automatically transpose them to fix Tensorflow's format?
I asked Francois Chollet about this (he doesn't have an SO account) and he kindly passed along this reply:
"th" format means that the convolutional kernels will have the shape (depth, input_depth, rows, cols)
"tf" format means that the convolutional kernels will have the shape (rows, cols, input_depth, depth)
Therefore you can convert from the former to the later via np.transpose(x, (2, 3, 1, 0)) where x is the value of the convolution kernel.
Here's some code to do the conversion:
from keras import backend as K
K.set_image_dim_ordering('th')
# build model in TH mode, as th_model
th_model = ...
# load weights that were saved in TH mode into th_model
th_model.load_weights(...)
K.set_image_dim_ordering('tf')
# build model in TF mode, as tf_model
tf_model = ...
# transfer weights from th_model to tf_model
for th_layer, tf_layer in zip(th_model.layers, tf_model.layers):
if th_layer.__class__.__name__ == 'Convolution2D':
kernel, bias = layer.get_weights()
kernel = np.transpose(kernel, (2, 3, 1, 0))
tf_layer.set_weights([kernel, bias])
else:
tf_layer.set_weights(tf_layer.get_weights())
In case the model contains Dense layers downstream of the Convolution2D layers, then the weight matrix of the first Dense layer would need to be shuffled as well.
You can Use This Script which auto translates theano/tensorflow backend trained model weights directly into the other 3 possible combinations of backend / dim ordering.