Models: InvalidArgumentError: logits and labels must be broadcastable: logits_size=[64,48] labels_size=[32,48] [[node softmax_cross_entropy_loss/xentropy (defined at <ipython-input-51-b7b564720801>:112) = SoftmaxCrossEntropyWithLogits[T=DT_FLOAT, _device="/job:localhost/replica:0/task:0/device:GPU:0"](softmax_cross_entropy_loss/xentropy/Reshape, softmax_cross_entropy_loss/xentropy/Reshape_1)]]

[ramadevsaiteja]

batch_size = 32
total_batch = len(train_batches)
num_epochs = 50

tf.reset_default_graph()
x = tf.placeholder(tf.float32, [None, 784])
y_ = tf.placeholder(tf.float32, [None, 48], name="truth")

Set the weights for the network

xavier = tf.contrib.layers.xavier_initializer_conv2d()
conv1_weights = tf.get_variable(name="c1", initializer=xavier, shape=[3, 3, 1, 32])
conv1_biases = tf.Variable(tf.zeros([32]))
conv2_weights = tf.get_variable(name="c2", initializer=xavier, shape=[3, 3, 32, 32])
conv2_biases = tf.Variable(tf.zeros([32]))
conv3_weights = tf.get_variable(name="c3", initializer=xavier, shape=[3, 3, 32, 32])
conv3_biases = tf.Variable(tf.zeros([32]))
conv4_weights = tf.get_variable(name="c4", initializer=xavier, shape=[3, 3, 32, 32])
conv4_biases = tf.Variable(tf.zeros([32]))
conv5_weights = tf.get_variable(name="c5", initializer=xavier, shape=[3, 3, 32, 32])
conv5_biases = tf.Variable(tf.zeros([32]))

conv6_weights = tf.get_variable(name="c6", initializer=xavier, shape=[3, 3, 32, 64])
conv6_biases = tf.Variable(tf.zeros([64]))
conv7_weights = tf.get_variable(name="c7", initializer=xavier, shape=[3, 3, 64, 64])
conv7_biases = tf.Variable(tf.zeros([64]))
conv8_weights = tf.get_variable(name="c8", initializer=xavier, shape=[3, 3, 64, 64])
conv8_biases = tf.Variable(tf.zeros([64]))
conv9_weights = tf.get_variable(name="c9", initializer=xavier, shape=[3, 3, 64, 64])
conv9_biases = tf.Variable(tf.zeros([64]))

conv10_weights = tf.get_variable(name="c10", initializer=xavier, shape=[3, 3, 64, 128])
conv10_biases = tf.Variable(tf.zeros([128]))
conv11_weights = tf.get_variable(name="c11", initializer=xavier, shape=[3, 3, 128, 128])
conv11_biases = tf.Variable(tf.zeros([128]))
conv12_weights = tf.get_variable(name="c12", initializer=xavier, shape=[3, 3, 128, 128])
conv12_biases = tf.Variable(tf.zeros([128]))
conv13_weights = tf.get_variable(name="c13", initializer=xavier, shape=[3, 3, 128, 128])
conv13_biases = tf.Variable(tf.zeros([128]))

conv14_weights = tf.get_variable(name="c14", initializer=xavier, shape=[3, 3, 128, 256])
conv14_biases = tf.Variable(tf.zeros([256]))
conv15_weights = tf.get_variable(name="c15", initializer=xavier, shape=[3, 3, 256, 256])
conv15_biases = tf.Variable(tf.zeros([256]))
conv16_weights = tf.get_variable(name="c16", initializer=xavier, shape=[3, 3, 256, 256])
conv16_biases = tf.Variable(tf.zeros([256]))

fc1_weights = tf.Variable(tf.truncated_normal([1024, 200], stddev=0.1))
fc1_biases = tf.Variable(tf.zeros([200]))
fc2_weights = tf.Variable(tf.truncated_normal([200, 100], stddev=0.1))
fc2_biases = tf.Variable(tf.zeros([100]))

out_weights = tf.Variable(tf.truncated_normal([100, 48], stddev=0.1))
out_biases = tf.Variable(tf.zeros([48]))

Stack the Layers

reshaped_input = tf.reshape(x, [-1, 28, 28, 1], name="absolute_input")

Stack 1

conv1 = tf.nn.conv2d(reshaped_input, conv1_weights, strides=[1, 1, 1, 1], padding='SAME')
relu1 = tf.nn.relu(tf.nn.bias_add(conv1, conv1_biases))
conv2 = tf.nn.conv2d(relu1, conv2_weights, strides=[1, 1, 1, 1], padding='SAME')
relu2 = tf.nn.relu(tf.nn.bias_add(conv2, conv2_biases))
conv3 = tf.nn.conv2d(relu2, conv3_weights, strides=[1, 1, 1, 1], padding='SAME')
relu3 = tf.nn.relu(tf.nn.bias_add(conv3, conv3_biases))
conv4 = tf.nn.conv2d(relu3, conv4_weights, strides=[1, 1, 1, 1], padding='SAME')
relu4 = tf.nn.relu(tf.nn.bias_add(conv4, conv4_biases))
conv5 = tf.nn.conv2d(relu4, conv5_weights, strides=[1, 1, 1, 1], padding='SAME')
relu5 = tf.nn.relu(tf.nn.bias_add(conv5, conv5_biases))
pool1 = tf.nn.max_pool(relu5, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
print(pool1.shape)

Stack 2

conv6 = tf.nn.conv2d(pool1, conv6_weights, strides=[1, 1, 1, 1], padding='SAME')
relu6 = tf.nn.relu(tf.nn.bias_add(conv6, conv6_biases))
conv7 = tf.nn.conv2d(relu6, conv7_weights, strides=[1, 1, 1, 1], padding='SAME')
relu7 = tf.nn.relu(tf.nn.bias_add(conv7, conv7_biases))
conv8 = tf.nn.conv2d(relu7, conv8_weights, strides=[1, 1, 1, 1], padding='SAME')
relu8 = tf.nn.relu(tf.nn.bias_add(conv8, conv8_biases))
conv9 = tf.nn.conv2d(relu8, conv9_weights, strides=[1, 1, 1, 1], padding='SAME')
relu9 = tf.nn.relu(tf.nn.bias_add(conv9, conv9_biases))
pool2 = tf.nn.max_pool(relu9, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
print(pool2.shape)

Stack 3

conv10 = tf.nn.conv2d(pool2, conv10_weights, strides=[1, 1, 1, 1], padding='SAME')
relu10 = tf.nn.relu(tf.nn.bias_add(conv10, conv10_biases))
conv11 = tf.nn.conv2d(relu10, conv11_weights, strides=[1, 1, 1, 1], padding='SAME')
relu11 = tf.nn.relu(tf.nn.bias_add(conv11, conv11_biases))
conv12 = tf.nn.conv2d(relu11, conv12_weights, strides=[1, 1, 1, 1], padding='SAME')
relu12 = tf.nn.relu(tf.nn.bias_add(conv12, conv12_biases))
conv13 = tf.nn.conv2d(relu12, conv13_weights, strides=[1, 1, 1, 1], padding='SAME')
relu13 = tf.nn.relu(tf.nn.bias_add(conv13, conv13_biases))
pool3 = tf.nn.max_pool(relu13, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
print(pool3.shape)

Stack 4

conv14 = tf.nn.conv2d(pool3, conv14_weights, strides=[1, 1, 1, 1], padding='SAME')
relu14 = tf.nn.relu(tf.nn.bias_add(conv14, conv14_biases))
conv15 = tf.nn.conv2d(relu14, conv15_weights, strides=[1, 1, 1, 1], padding='SAME')
relu15 = tf.nn.relu(tf.nn.bias_add(conv15, conv15_biases))
conv16 = tf.nn.conv2d(relu15, conv16_weights, strides=[1, 1, 1, 1], padding='SAME')
relu16 = tf.nn.relu(tf.nn.bias_add(conv16, conv16_biases))
pool4 = tf.nn.max_pool(relu16, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
print(pool4.shape)

layer 4

pool_shape = pool4.get_shape().as_list()
reshaped = tf.reshape(pool3, [-1, pool_shape[1] * pool_shape[2] * pool_shape[3]])
fc1 = tf.nn.relu(tf.add(tf.matmul(reshaped, fc1_weights), fc1_biases))
fc2 = tf.nn.relu(tf.add(tf.matmul(fc1, fc2_weights), fc2_biases))
print(fc2.shape)
y = tf.add(tf.matmul(fc2, out_weights), out_biases, name="absolute_output")
print(y.shape,y_.shape)

Define loss and optimizer

cross_entropy = tf.reduce_mean(tf.losses.softmax_cross_entropy(y_,y))

train_step = tf.train.AdamOptimizer(learning_rate=0.001).minimize(cross_entropy)

Train the model

sess = tf.InteractiveSession()
tf.initialize_all_variables().run()
for i in tqdm(range(num_epochs)):
for i in range(total_batch):
batch_x, batch_y =train_batches[i]
print("batch",batch_x.shape,batch_y.shape)
sess.run(train_step, feed_dict={x: batch_x, y_: batch_y})

Test trained model

correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(y_, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
test_acc = []
train_acc = []
for i in tqdm(range(total_batch)):
batch_x, batch_y = train_batches[i]
test_acc.append(sess.run(accuracy, feed_dict={x: batch_x, y_: batch_y}))
for i in tqdm(range(len(test_batches))):
batch_x, batch_y = test_batches[i]
train_acc.append(sess.run(accuracy, feed_dict={x: batch_x, y_: batch_y}))
print(np.mean(train_acc), np.mean(test_acc))

I searched on Google but I didn't find any solution. It's saying logits have size [64,48] and labels have size [32,48]

My batch size is 32 and number of classes is 48

[ramadevsaiteja]

batch_size = 32
total_batch = len(train_batches)
num_epochs = 50

tf.reset_default_graph()
x = tf.placeholder(tf.float32, [None, 784])
y_ = tf.placeholder(tf.float32, [None, 48], name="truth")

Set the weights for the network

xavier = tf.contrib.layers.xavier_initializer_conv2d()
conv1_weights = tf.get_variable(name="c1", initializer=xavier, shape=[3, 3, 1, 32])
conv1_biases = tf.Variable(tf.zeros([32]))
conv2_weights = tf.get_variable(name="c2", initializer=xavier, shape=[3, 3, 32, 32])
conv2_biases = tf.Variable(tf.zeros([32]))
conv3_weights = tf.get_variable(name="c3", initializer=xavier, shape=[3, 3, 32, 32])
conv3_biases = tf.Variable(tf.zeros([32]))
conv4_weights = tf.get_variable(name="c4", initializer=xavier, shape=[3, 3, 32, 32])
conv4_biases = tf.Variable(tf.zeros([32]))
conv5_weights = tf.get_variable(name="c5", initializer=xavier, shape=[3, 3, 32, 32])
conv5_biases = tf.Variable(tf.zeros([32]))

conv6_weights = tf.get_variable(name="c6", initializer=xavier, shape=[3, 3, 32, 64])
conv6_biases = tf.Variable(tf.zeros([64]))
conv7_weights = tf.get_variable(name="c7", initializer=xavier, shape=[3, 3, 64, 64])
conv7_biases = tf.Variable(tf.zeros([64]))
conv8_weights = tf.get_variable(name="c8", initializer=xavier, shape=[3, 3, 64, 64])
conv8_biases = tf.Variable(tf.zeros([64]))
conv9_weights = tf.get_variable(name="c9", initializer=xavier, shape=[3, 3, 64, 64])
conv9_biases = tf.Variable(tf.zeros([64]))

conv10_weights = tf.get_variable(name="c10", initializer=xavier, shape=[3, 3, 64, 128])
conv10_biases = tf.Variable(tf.zeros([128]))
conv11_weights = tf.get_variable(name="c11", initializer=xavier, shape=[3, 3, 128, 128])
conv11_biases = tf.Variable(tf.zeros([128]))
conv12_weights = tf.get_variable(name="c12", initializer=xavier, shape=[3, 3, 128, 128])
conv12_biases = tf.Variable(tf.zeros([128]))
conv13_weights = tf.get_variable(name="c13", initializer=xavier, shape=[3, 3, 128, 128])
conv13_biases = tf.Variable(tf.zeros([128]))

conv14_weights = tf.get_variable(name="c14", initializer=xavier, shape=[3, 3, 128, 256])
conv14_biases = tf.Variable(tf.zeros([256]))
conv15_weights = tf.get_variable(name="c15", initializer=xavier, shape=[3, 3, 256, 256])
conv15_biases = tf.Variable(tf.zeros([256]))
conv16_weights = tf.get_variable(name="c16", initializer=xavier, shape=[3, 3, 256, 256])
conv16_biases = tf.Variable(tf.zeros([256]))

fc1_weights = tf.Variable(tf.truncated_normal([1024, 200], stddev=0.1))
fc1_biases = tf.Variable(tf.zeros([200]))
fc2_weights = tf.Variable(tf.truncated_normal([200, 100], stddev=0.1))
fc2_biases = tf.Variable(tf.zeros([100]))

out_weights = tf.Variable(tf.truncated_normal([100, 48], stddev=0.1))
out_biases = tf.Variable(tf.zeros([48]))

Stack the Layers

reshaped_input = tf.reshape(x, [-1, 28, 28, 1], name="absolute_input")

Stack 1

conv1 = tf.nn.conv2d(reshaped_input, conv1_weights, strides=[1, 1, 1, 1], padding='SAME')
relu1 = tf.nn.relu(tf.nn.bias_add(conv1, conv1_biases))
conv2 = tf.nn.conv2d(relu1, conv2_weights, strides=[1, 1, 1, 1], padding='SAME')
relu2 = tf.nn.relu(tf.nn.bias_add(conv2, conv2_biases))
conv3 = tf.nn.conv2d(relu2, conv3_weights, strides=[1, 1, 1, 1], padding='SAME')
relu3 = tf.nn.relu(tf.nn.bias_add(conv3, conv3_biases))
conv4 = tf.nn.conv2d(relu3, conv4_weights, strides=[1, 1, 1, 1], padding='SAME')
relu4 = tf.nn.relu(tf.nn.bias_add(conv4, conv4_biases))
conv5 = tf.nn.conv2d(relu4, conv5_weights, strides=[1, 1, 1, 1], padding='SAME')
relu5 = tf.nn.relu(tf.nn.bias_add(conv5, conv5_biases))
pool1 = tf.nn.max_pool(relu5, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
print(pool1.shape)

Stack 2

conv6 = tf.nn.conv2d(pool1, conv6_weights, strides=[1, 1, 1, 1], padding='SAME')
relu6 = tf.nn.relu(tf.nn.bias_add(conv6, conv6_biases))
conv7 = tf.nn.conv2d(relu6, conv7_weights, strides=[1, 1, 1, 1], padding='SAME')
relu7 = tf.nn.relu(tf.nn.bias_add(conv7, conv7_biases))
conv8 = tf.nn.conv2d(relu7, conv8_weights, strides=[1, 1, 1, 1], padding='SAME')
relu8 = tf.nn.relu(tf.nn.bias_add(conv8, conv8_biases))
conv9 = tf.nn.conv2d(relu8, conv9_weights, strides=[1, 1, 1, 1], padding='SAME')
relu9 = tf.nn.relu(tf.nn.bias_add(conv9, conv9_biases))
pool2 = tf.nn.max_pool(relu9, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
print(pool2.shape)

Stack 3

conv10 = tf.nn.conv2d(pool2, conv10_weights, strides=[1, 1, 1, 1], padding='SAME')
relu10 = tf.nn.relu(tf.nn.bias_add(conv10, conv10_biases))
conv11 = tf.nn.conv2d(relu10, conv11_weights, strides=[1, 1, 1, 1], padding='SAME')
relu11 = tf.nn.relu(tf.nn.bias_add(conv11, conv11_biases))
conv12 = tf.nn.conv2d(relu11, conv12_weights, strides=[1, 1, 1, 1], padding='SAME')
relu12 = tf.nn.relu(tf.nn.bias_add(conv12, conv12_biases))
conv13 = tf.nn.conv2d(relu12, conv13_weights, strides=[1, 1, 1, 1], padding='SAME')
relu13 = tf.nn.relu(tf.nn.bias_add(conv13, conv13_biases))
pool3 = tf.nn.max_pool(relu13, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
print(pool3.shape)

Stack 4

conv14 = tf.nn.conv2d(pool3, conv14_weights, strides=[1, 1, 1, 1], padding='SAME')
relu14 = tf.nn.relu(tf.nn.bias_add(conv14, conv14_biases))
conv15 = tf.nn.conv2d(relu14, conv15_weights, strides=[1, 1, 1, 1], padding='SAME')
relu15 = tf.nn.relu(tf.nn.bias_add(conv15, conv15_biases))
conv16 = tf.nn.conv2d(relu15, conv16_weights, strides=[1, 1, 1, 1], padding='SAME')
relu16 = tf.nn.relu(tf.nn.bias_add(conv16, conv16_biases))
pool4 = tf.nn.max_pool(relu16, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
print(pool4.shape)

layer 4

pool_shape = pool4.get_shape().as_list()
reshaped = tf.reshape(pool3, [-1, pool_shape[1] * pool_shape[2] * pool_shape[3]])
fc1 = tf.nn.relu(tf.add(tf.matmul(reshaped, fc1_weights), fc1_biases))
fc2 = tf.nn.relu(tf.add(tf.matmul(fc1, fc2_weights), fc2_biases))
print(fc2.shape)
y = tf.add(tf.matmul(fc2, out_weights), out_biases, name="absolute_output")
print(y.shape,y_.shape)

Define loss and optimizer

cross_entropy = tf.reduce_mean(tf.losses.softmax_cross_entropy(y_,y))

train_step = tf.train.AdamOptimizer(learning_rate=0.001).minimize(cross_entropy)

Train the model

sess = tf.InteractiveSession()
tf.initialize_all_variables().run()
for i in tqdm(range(num_epochs)):
for i in range(total_batch):
batch_x, batch_y =train_batches[i]
print("batch",batch_x.shape,batch_y.shape)
sess.run(train_step, feed_dict={x: batch_x, y_: batch_y})

Test trained model

correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(y_, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
test_acc = []
train_acc = []
for i in tqdm(range(total_batch)):
batch_x, batch_y = train_batches[i]
test_acc.append(sess.run(accuracy, feed_dict={x: batch_x, y_: batch_y}))
for i in tqdm(range(len(test_batches))):
batch_x, batch_y = test_batches[i]
train_acc.append(sess.run(accuracy, feed_dict={x: batch_x, y_: batch_y}))
print(np.mean(train_acc), np.mean(test_acc))

I searched on Google but I didn't find any solution. It's saying logits have size [64,48] and labels have size [32,48]

My batch size is 32 and number of classes is 48

[ramadevsaiteja]

Sorry! I found the error. The problem is at flattening final pool layer. By mistake I flattened the last before pool layer and passed it through fully connected layer.

[prabaldeshar]

Can u elaborate where u found the error. I have the same problem

[ghost]

@ramadevsaiteja will you elaborate more?? I also have the same error.

[OneManArmy93]

@ramadevsaiteja Im encountering the same issue. Can you please elaborate what is the error and how you fixed it? Thank you.

[MagnusFelinto]

Estou com mesmo alguém saber soluciona lo ?

[dinis-rodrigues]

@MagnusFelinto @OneManArmy93
This happened to me as well, but in regards to the shape of the labels, instead of the batch. Basically, my final classification was defined with 4 outputs.
But then I changed my dataset, to have only 3 labels, and tried to run the script again.

I just had to change my last layer of the model, from:

tf.keras.layers.Dense(4, activation="softmax")

to:

tf.keras.layers.Dense(3, activation="softmax")

If you are using, Keras Sequence Generator, don't forget to take a look into it as well when defining the returned labels etc..

[monjaraz85]

@MagnusFelinto @OneManArmy93
This happened to me as well, but in regards to the shape of the labels, instead of the batch. Basically, my final classification was defined with 4 outputs.
But then I changed my dataset, to have only 3 labels, and tried to run the script again.

I just had to change my last layer of the model, from:

tf.keras.layers.Dense(4, activation="softmax")

to:

tf.keras.layers.Dense(3, activation="softmax")

If you are using, Keras Sequence Generator, don't forget to take a look into it as well when defining the returned labels etc..

That does it!!

[MonkeyDLy]

@MagnusFelinto @OneManArmy93
This happened to me as well, but in regards to the shape of the labels, instead of the batch. Basically, my final classification was defined with 4 outputs.
But then I changed my dataset, to have only 3 labels, and tried to run the script again.

I just had to change my last layer of the model, from:

tf.keras.layers.Dense(4, activation="softmax")

to:

tf.keras.layers.Dense(3, activation="softmax")

If you are using, Keras Sequence Generator, don't forget to take a look into it as well when defining the returned labels etc..

Thank you.

[radprad]

it got resolved when the number of classes in train and test were different while using image generator. It got fixed when i matched the train and test number of class

[Herrytheeagle]

Why was the final classification defined with four outputs initially ?