CODEWITHSUNDEEP
pythontensorflowtensorflow2.0
Roman
Roman

Reputation: 3241

How to solve, tensorflow.python.framework.errors_impl.InvalidArgumentError?

import tensorflow as tf
import numpy as np
from sklearn.model_selection import train_test_split

np.random.seed(4213)

data = np.random.randint(low=1,high=29, size=(500, 160, 160, 10)) 
labels = np.random.randint(low=0,high=5, size=(500, 160, 160)) 
nclass = len(np.unique(labels))
print (nclass)

samples, width, height, nbands = data.shape


X_train, X_test, y_train, y_test = train_test_split(data, labels, test_size=0.25, random_state=421)

print (X_train.shape)
print (y_train.shape)

arch = tf.keras.applications.VGG16(input_shape=[width, height, nbands],
                      include_top=False,
                      weights=None)

model = tf.keras.Sequential()
model.add(arch)
model.add(tf.keras.layers.Flatten())
model.add(tf.keras.layers.Dense(nclass))

model.compile(optimizer = tf.keras.optimizers.Adam(0.0001),
              loss=tf.keras.losses.SparseCategoricalCrossentropy(),
              metrics=[tf.keras.metrics.SparseCategoricalAccuracy()])
    

model.fit(X_train,
          y_train,                                 
          epochs=3,
          batch_size=32,
          verbose=2)


res = model.predict(X_test)
print(res.shape)

When running the above code for semantic segmentation I get Exception has occurred:

InvalidArgumentError
 Incompatible shapes: [32,160,160] vs. [32]
     [[node Equal (defined at c...:38) ]] [Op:__inference_train_function_1815]


tensorflow.python.framework.errors_impl.InvalidArgumentError

Upvotes: 2

Views: 2271

Answers (1)

ibarrond
ibarrond

Reputation: 7641

Your issue comes from the size of the last layer (to avoid these mistakes it is always desirable to use python constants for N_IMAGES, WIDTH, HEIGHT, N_CHANNELS and N_CLASSES):

For image classification

You should assign one single label to each image. Try switching labels:

import tensorflow as tf
import numpy as np
from sklearn.model_selection import train_test_split

np.random.seed(4213)

N_IMAGES, WIDTH, HEIGHT, N_CHANNELS = (500, 160, 160, 10)
N_CLASSES  = 5

data = np.random.randint(low=1,high=29, size=(N_IMAGES, WIDTH, HEIGHT, N_CHANNELS)) 
labels = np.random.randint(low=0,high=N_CLASSES, size=(N_IMAGES)) 
#...

For semantic segmentation

Make sure your classifier (last layers of the network) is sized accordingly. In this case you need 1 class per pixel:

#...
model = tf.keras.Sequential()
model.add(arch)
model.add(tf.keras.layers.Flatten())
model.add(tf.keras.layers.Dense(width * height))
model.add(tf.keras.layers.Reshape([width , height]))
#...

This is the simplest you can get. Instead, you can set up multiple deconvolution layers to act as classifier, or you can even flip the arch architecture over and use it to generate the classification results. Orthogohally, you can perform one_hot encoding on the labels and thus expand them by a factor of N_CLASSES, effectively multiplying the number of neurons in the last layer.

Upvotes: 1

Related Questions