Keras: ValueError: negative dimensions are not allowed

[nishantprateek]

I am trying to implement a VGG-16 convnet for classification on the Clothing Attribute dataset. I am getting the following error:

Traceback (most recent call last):
File "vgg_16.py", line 84, in
model.add(Dense())#fc6
File "/usr/local/lib/python2.7/dist-packages/keras/layers/containers.py", line 32, in add
self.layers[-1].set_previous(self.layers[-2])
File "/usr/local/lib/python2.7/dist-packages/keras/layers/core.py", line 39, in set_previous
self.build()
File "/usr/local/lib/python2.7/dist-packages/keras/layers/core.py", line 654, in build
self.W = self.init((input_dim, self.output_dim))
File "/usr/local/lib/python2.7/dist-packages/keras/initializations.py", line 40, in glorot_uniform
return uniform(shape, s)
File "/usr/local/lib/python2.7/dist-packages/keras/initializations.py", line 13, in uniform
return K.variable(np.random.uniform(low=-scale, high=scale, size=shape))
File "mtrand.pyx", line 1177, in mtrand.RandomState.uniform (numpy/random/mtrand/mtrand.c:9148)
File "mtrand.pyx", line 203, in mtrand.cont2_array_sc (numpy/random/mtrand/mtrand.c:2371)
ValueError: negative dimensions are not allowed

My code:

import os
import numpy as np
import scipy.io

from skimage import io, transform
import random
from keras.models import Sequential
from keras.layers.core import Dense, Dropout, Activation, Flatten
from keras.layers.convolutional import Convolution2D, MaxPooling2D
from keras.optimizers import SGD
from keras.utils import np_utils, generic_utils
from keras.callbacks import ModelCheckpoint

images = os.listdir('ClothingAttributeDataset/images')
images.sort()

print images

mat = scipy.io.loadmat('./ClothingAttributeDataset/labels/category_GT.mat')
mat = mat['GT']
print mat.shape
os.chdir('ClothingAttributeDataset/images')

im = np.zeros((1856, 266, 400, 3), dtype='float32')
old_labels = np.zeros((1856, 1), dtype='int')
j = 0
trash = 0
for i in images:
inp = io.imread(i)
inp = transform.resize(inp, (266, 400))
im[j] = inp
try:
old_labels[j] = int(mat[j][0]-1)
except:
old_labels[j] = random.randint(1,7)-1
trash+=1
print old_labels[j]
j+=1

print trash
n_classes = 8
batch_size = 2
nb_epoch = 300

labels = np_utils.to_categorical(old_labels, n_classes)

print labels.shape
model = Sequential()

model.add(Convolution2D(32, 3, 3, border_mode='valid', input_shape=(266, 400, 3), dim_ordering='tf')) #conv1_!
model.add(Activation('relu')) #relu1_!
model.add(Convolution2D(64, 3, 3))#conv1_2
model.add(Activation('relu'))#relu1_2
model.add(MaxPooling2D(pool_size=(2,2), strides=(2, 2)))#pool1

model.add(Convolution2D(64, 3, 3))#conv2_!
model.add(Activation('relu'))#relu2_!
model.add(Convolution2D(128, 3, 3))#conv2_2
model.add(Activation('relu'))#relu2_2
model.add(MaxPooling2D(pool_size=(2,2), strides = (2,2)))#pool2

model.add(Convolution2D(128, 3, 3)) #conv3_1
model.add(Activation('relu'))#relu3_1
model.add(Convolution2D(256, 3, 3)) #conv3_2
model.add(Activation('relu'))#relu3_2
model.add(Convolution2D(256, 3, 3))#conv3_3
model.add(Activation('relu'))#relu3_3
model.add(MaxPooling2D(pool_size=(2,2), strides=(2,2)))#pool3

model.add(Convolution2D(256, 3, 3))#conv4_1
model.add(Activation('relu'))#relu4_1
model.add(Convolution2D(512, 3, 3))#conv4_2
model.add(Activation('relu'))#relu4_2
model.add(Convolution2D(512, 3, 3))#conv4_3
model.add(Activation('relu'))#relu4_3
model.add(MaxPooling2D(pool_size=(2, 2), strides=(2,2)))#pool4

model.add(Convolution2D(256, 3, 3)) #conv5_1
model.add(Activation('relu'))#relu5_1
model.add(Convolution2D(512, 3, 3)) #conv5_2
model.add(Activation('relu'))#relu5_2
model.add(Convolution2D(512, 3, 3)) #conv5_3
model.add(Activation('relu'))#relu5_3
model.add(MaxPooling2D(pool_size=(2, 2), strides=(2,2)))#pool5

model.add(Flatten())
model.add(Dense(4096))#fc6
model.add(Activation('relu'))#relu6
model.add(Dropout(0.5))#drop6
model.add(Dense(4096))#fc7
model.add(Activation('relu'))#relu7
model.add(Dropout(0.5))#drop7
model.add(Dense(1000))#fc8
model.add(Activation('softmax'))#prob

sgd = SGD(lr=0.1, decay=1e-6, momentum=0.9, nesterov=True)
model.compile(loss='categorical_crossentropy', optimizer=sgd)

checkpointer = ModelCheckpoint(filepath="weights/weights.{epoch:02d}.hdf5", verbose=1, save_best_only=False)
print("training starts...")
model.fit(im, labels, batch_size=batch_size, nb_epoch=nb_epoch, callbacks=[checkpointer])

[Sebubu]

reshape your input like https://github.com/fchollet/keras/blob/master/examples/mnist_cnn.py

[jbr00000]

I have the same problem with you, so do you solve this problem?

[joelthchao]

@nishantprateek You miss argument dim_ordering='tf' in later layer, e.g. Convolution2D and MaxPooling2D. Therefore in some layer filter size larger than input size.
@jbr00000 Please make sure your input shape is valid for next layer. You can use model.summary() to list the output shape of each layer.

[huitian-jiao]

In my case, changing "model.add(Convolution2D(32, 3, 3, border_mode='valid', input_shape=(266, 400, 3), dim_ordering='tf')) " to "model.add(Convolution2D(32, 3, 3, border_mode='valid', input_shape=(266, 400, 3), dim_ordering='th')) " fixes my problem.

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs, but feel free to re-open it if needed.

[pengcao]

ValueError: Dimension 2 in both shapes must be equal, but are 300 and 3. Shapes are [3,3,300,64] and [3,3,3,64]. for 'Assign' (op: 'Assign') with input shapes: [3,3,300,64], [3,3,3,64].

[pengcao]

ValueError: Dimension 2 in both shapes must be equal, but are 300 and 3. Shapes are [3,3,300,64] and [3,3,3,64]. for 'Assign' (op: 'Assign') with input shapes: [3,3,300,64], [3,3,3,64].这个issue有解决吗？