3D Convolutional Neural Network input shape

Question

I'm having a problem feeding a 3D CNN using Keras and Python to classify 3D shapes. I have a folder with some models in JSON format. I read those models into a Numpy Array. The models are 25*25*25 and represent the occupancy grid of the voxelized model (each position represents if the voxel in position (i,j,k) has points in it or no), so I only have 1 channel of input, like grayscale images in 2D images. The code that I have is the following:

import numpy as np
from keras.models import Sequential
from keras.layers import Dense, Dropout, Activation, Flatten
from keras.layers import Convolution3D, MaxPooling3D
from keras.optimizers import SGD
from keras.utils import np_utils
from keras import backend as K

# Number of Classes and Epochs of Training
nb_classes = 3 # cube, cone or sphere
nb_epoch = 100
batch_size = 2

# Input Image Dimensions
img_rows, img_cols, img_depth = 25, 25, 25

# Number of Convolutional Filters to use
nb_filters = 32

# Convolution Kernel Size
kernel_size = [5,5,5]

X_train, Y_train = [], []

# Read from File
import os
import json

i=0
for filename in os.listdir(os.path.join(os.getcwd(), 'models')):
    with open(os.path.join(os.getcwd(), 'models', filename)) as f:
        file = f.readlines()
        json_file = '\n'.join(file)
        content = json.loads(json_file)
        occupancy = content['model']['occupancy']
        form = []
        for value in occupancy:
            form.append(int(value))
        final_model = [ [ [ 0 for i in range(img_rows) ]
                              for j in range(img_cols) ]
                              for k in range(img_depth) ]
        a = 0
        for i in range(img_rows):
            for j in range(img_cols):
                for k in range(img_depth):
                    final_model[i][j][k] = form[a]
                    a = a + 1
        X_train.append(final_model)
        Y_train.append(content['model']['label'])

X_train = np.array(X_train)
Y_train = np.array(Y_train)

# (1 channel, 25 rows, 25 cols, 25 of depth)
input_shape = (1, img_rows, img_cols, img_depth)

# Init
model = Sequential()

# 3D Convolution layer
model.add(Convolution3D(nb_filters, kernel_size[0], kernel_size[1], kernel_size[2],
                        input_shape=input_shape,
                        activation='relu'))

# Fully Connected layer
model.add(Flatten())
model.add(Dense(128,
          init='normal',
          activation='relu'))
model.add(Dropout(0.5))

# Softmax Layer
model.add(Dense(nb_classes,
                init='normal'))
model.add(Activation('softmax'))

# Compile
model.compile(loss='categorical_crossentropy',
              optimizer=SGD())

# Fit network
model.fit(X_train, Y_train, nb_epoch=nb_epoch,
         verbose=1)

After this, I get the following error

Using TensorFlow backend. Traceback (most recent call last): File "/usr/local/lib/python3.6/site-packages/tensorflow/python/framework/common_shapes.py", line 670, in _call_cpp_shape_fn_impl status) File "/usr/local/Cellar/python3/3.6.0/Frameworks/Python.framework/Versions/3.6/lib/python3.6/contextlib.py", line 89, in exit next(self.gen) File "/usr/local/lib/python3.6/site-packages/tensorflow/python/framework/errors_impl.py", line 469, in raise_exception_on_not_ok_status pywrap_tensorflow.TF_GetCode(status)) tensorflow.python.framework.errors_impl.InvalidArgumentError: Negative dimension size caused by subtracting 5 from 1 for 'Conv3D' (op: 'Conv3D') with input shapes: [?,1,25,25,25], [5,5,5,25,32].

During handling of the above exception, another exception occurred:

Traceback (most recent call last): File "CNN_3D.py", line 76, in activation='relu')) File "/usr/local/lib/python3.6/site-packages/keras/models.py", line 299, in add layer.create_input_layer(batch_input_shape, input_dtype) File "/usr/local/lib/python3.6/site-packages/keras/engine/topology.py", line 401, in create_input_layer self(x) File "/usr/local/lib/python3.6/site-packages/keras/engine/topology.py", line 572, in call self.add_inbound_node(inbound_layers, node_indices, tensor_indices) File "/usr/local/lib/python3.6/site-packages/keras/engine/topology.py", line 635, in add_inbound_node Node.create_node(self, inbound_layers, node_indices, tensor_indices) File "/usr/local/lib/python3.6/site-packages/keras/engine/topology.py", line 166, in create_node output_tensors = to_list(outbound_layer.call(input_tensors[0], mask=input_masks[0])) File "/usr/local/lib/python3.6/site-packages/keras/layers/convolutional.py", line 1234, in call filter_shape=self.W_shape) File "/usr/local/lib/python3.6/site-packages/keras/backend/tensorflow_backend.py", line 2831, in conv3d x = tf.nn.conv3d(x, kernel, strides, padding) File "/usr/local/lib/python3.6/site-packages/tensorflow/python/ops/gen_nn_ops.py", line 522, in conv3d strides=strides, padding=padding, name=name) File "/usr/local/lib/python3.6/site-packages/tensorflow/python/framework/op_def_library.py", line 763, in apply_op op_def=op_def) File "/usr/local/lib/python3.6/site-packages/tensorflow/python/framework/ops.py", line 2397, in create_op set_shapes_for_outputs(ret) File "/usr/local/lib/python3.6/site-packages/tensorflow/python/framework/ops.py", line 1757, in set_shapes_for_outputs shapes = shape_func(op) File "/usr/local/lib/python3.6/site-packages/tensorflow/python/framework/ops.py", line 1707, in call_with_requiring return call_cpp_shape_fn(op, require_shape_fn=True) File "/usr/local/lib/python3.6/site-packages/tensorflow/python/framework/common_shapes.py", line 610, in call_cpp_shape_fn debug_python_shape_fn, require_shape_fn) File "/usr/local/lib/python3.6/site-packages/tensorflow/python/framework/common_shapes.py", line 675, in _call_cpp_shape_fn_impl raise ValueError(err.message) ValueError: Negative dimension size caused by subtracting 5 from 1 for 'Conv3D' (op: 'Conv3D') with input shapes: [?,1,25,25,25], [5,5,5,25,32].

What am I doing wrong to get this error?

Solution

I think that the problem is that you are setting the input shape in Theano ordering but you are using Keras with Tensorflow backend and Tensorflow img ordering. In addition the y_train array has to be converted to categorical labels.

Updated code:

from keras.utils import np_utils
from keras import backend as K

if K.image_dim_ordering() == 'th':
    X_train = X_train.reshape(X_train.shape[0], 1, img_rows, img_cols, img_depth)
    input_shape = (1, img_rows, img_cols, img_depth)
else:
    X_train = X_train.reshape(X_train.shape[0], img_rows, img_cols, img_depth, 1)
    input_shape = (img_rows, img_cols, img_depth, 1)

Y_train = np_utils.to_categorical(Y_train, nb_classes)

Adding this lines should fix it.

Comments

[Matt]

I think the shape of your training data is wrong. Tensorflow expects the data to be in the form of (sample, dim1, dim2, ..., channel). Given a list of regular 2D images, you can reshape like this: X_train.reshape((-1, WIDTH, HEIGHT, 1)). Adapting this to your case you could try X_train = X_train.reshape((-1, img_rows, img_cols, img_depth, 1)). And the input_shape should be (img_rows, img_cols, img_depth, 1).

[João Pedro Fontes]

I still get the same error. I can pass the creation of the layers when I add more channels (input_shape = (5, img_rows, img_cols, img_depth)) that overcome or equal the size of the convolutional filters. But I only have a channel of input. I think that the problem is in the definition of the Conv3D layer

[Braelyn B]

I am wanting to do exactly what you are doing, but I'm having trouble figuring out how to turn a 3D model file into an occupancy grid. How are you doing that? Thank you!

[João Pedro Fontes]

Using both codes almost had it. But now I get a new error: Using TensorFlow backend. Traceback (most recent call last): File "CNN_3D_2.py", line 86, in <module> verbose=1) ... ValueError: Error when checking model input: expected convolution3d_input_1 to have shape (None, 25, 25, 25, 1) but got array with shape (1, 25, 25, 25, 2)

[David de la Iglesia]

can you tell me your original (before reshaping) X_train.shape ??

[João Pedro Fontes]

(2, 25, 25, 25)

[David de la Iglesia]

I updated whe code with the reshape @JoãoPedroFontes. Try now

[João Pedro Fontes]

David, it is giving the error ValueError: Input 0 is incompatible with layer convolution3d_1: expected ndim=5, found ndim=4. The example that you gave me is not for 2D? In the input shape you only give rows and columns. I tried adding the img_depths and it gave ValueError: Error when checking model target: expected activation_1 to have shape (None, 3) but got array with shape (2, 1)

[David de la Iglesia]

@JoãoPedroFontes yes. It was a typo. Now is updated. The new error you are getting I assume that is now related to the y_train array. Could you tell me your y_train.shape?

[João Pedro Fontes]

I think it may have to be with the previous shape of the X_train array. It has (2, 25, 25, 25). I don't understand why it has that 2, because I only have 1 channel of input and 25*25*25 data

[João Pedro Fontes]

The Y_train array has (2,) shape

[David de la Iglesia]

maybe you could share on the question your models file so we can reproduce the code

[David de la Iglesia]

Okay @JoãoPedroFontes, You have to convert your y_train array.

[João Pedro Fontes]

here is an example of one of my files link