Keras ValueError: Dimensions must be equal issue
Even after applying the suggestions in answer and comments, it looks like the dimension mismatch issue persists. This is exact code and data file to replicate as well: https://drive.google.com/drive/folders/1q67s0VhB-O7J8OtIhU2jmj7Kc4LxL3sf?usp=sharing
How can this be corrected!? Latest code, model summary, functions used and error I get is below
type_ae=='dcor'
#Wrappers for keras
def custom_loss1(y_true,y_pred):
dcor = -1*distance_correlation(y_true,encoded_layer)
return dcor
def custom_loss2(y_true,y_pred):
recon_loss = losses.categorical_crossentropy(y_true, y_pred)
return recon_loss
input_layer = Input(shape=(64,64,1))
encoded_layer = Conv2D(filters = 128, kernel_size = (5,5),padding = 'same',activation ='relu',
input_shape = (64,64,1))(input_layer)
encoded_layer = MaxPool2D(pool_size=(2,2))(encoded_layer)
encoded_layer = Dropout(0.25)(encoded_layer)
encoded_layer = (Conv2D(filters = 64, kernel_size = (3,3),padding = 'same',activation ='relu'))(encoded_layer)
encoded_layer = (MaxPool2D(pool_size=(2,2)))(encoded_layer)
encoded_layer = (Dropout(0.25))(encoded_layer)
encoded_layer = (Conv2D(filters = 64, kernel_size = (3,3),padding = 'same',activation ='relu'))(encoded_layer)
encoded_layer = (MaxPool2D(pool_size=(2,2)))(encoded_layer)
encoded_layer = (Dropout(0.25))(encoded_layer)
encoded_layer = Conv2D(filters = 1, kernel_size = (3,3),padding = 'same',activation ='relu',
input_shape = (64,64,1),strides=1)(encoded_layer)
encoded_layer = ZeroPadding2D(padding=(28, 28), data_format=None)(encoded_layer)
decoded_imag = Conv2D(8, (2, 2), activation='relu', padding='same')(encoded_layer)
decoded_imag = UpSampling2D((2, 2))(decoded_imag)
decoded_imag = Conv2D(8, (3, 3), activation='relu', padding='same')(decoded_imag)
decoded_imag = UpSampling2D((2, 2))(decoded_imag)
decoded_imag = Conv2D(16, (3, 3), activation='relu', padding='same')(decoded_imag)
decoded_imag = UpSampling2D((2, 2))(decoded_imag)
decoded_imag = Conv2D(1, (3, 3), activation='sigmoid', padding='same')(decoded_imag)
flat_layer = Flatten()(decoded_imag)
dense_layer = Dense(256,activation = "relu")(flat_layer)
dense_layer = Dense(64,activation = "relu")(dense_layer)
dense_layer = Dense(32,activation = "relu")(dense_layer)
output_layer = Dense(9, activation = "softmax")(dense_layer)
autoencoder = Model(input_layer, [encoded_layer,output_layer])
autoencoder.summary()
autoencoder.compile(optimizer='adadelta', loss=[custom_loss1,custom_loss2])
autoencoder.fit(x_train,[x_train, y_train],batch_size=32,epochs=3,shuffle=True,
validation_data=(x_val, [x_val,y_val]))
The data is of dimensions:
x_train.shape: (4000, 64, 64, 1)
x_val.shape: (1000, 64, 64, 1)
y_train.shape: (4000, 9)
y_val.shape: (1000, 9)
losses look like:
def custom_loss1(y_true,y_pred):
dcor = -1*distance_correlation(y_true,encoded_layer)
return dcor
def custom_loss2(y_true,y_pred):
recon_loss = losses.categorical_crossentropy(y_true, y_pred)
return recon_loss
The correlation function is based on tensors as follows:
def distance_correlation(y_true,y_pred):
pred_r = tf.reduce_sum(y_pred*y_pred,1)
pred_r = tf.reshape(pred_r,[-1,1])
pred_d = pred_r - 2*tf.matmul(y_pred,tf.transpose(y_pred))+tf.transpose(pred_r)
true_r = tf.reduce_sum(y_true*y_true,1)
true_r = tf.reshape(true_r,[-1,1])
true_d = true_r - 2*tf.matmul(y_true,tf.transpose(y_true))+tf.transpose(true_r)
concord = 1-tf.matmul(y_true,tf.transpose(y_true))
#print(pred_d)
#print(tf.reshape(tf.reduce_mean(pred_d,1),[-1,1]))
#print(tf.reshape(tf.reduce_mean(pred_d,0),[1,-1]))
#print(tf.reduce_mean(pred_d))
tf.check_numerics(pred_d,'pred_d has NaN')
tf.check_numerics(true_d,'true_d has NaN')
A = pred_d - tf.reshape(tf.reduce_mean(pred_d,1),[-1,1]) - tf.reshape(tf.reduce_mean(pred_d,0),[1,-1]) + tf.reduce_mean(pred_d)
B = true_d - tf.reshape(tf.reduce_mean(true_d,1),[-1,1]) - tf.reshape(tf.reduce_mean(true_d,0),[1,-1]) + tf.reduce_mean(true_d)
#dcor = -tf.reduce_sum(concord*pred_d)/tf.reduce_sum((1-concord)*pred_d)
dcor = -tf.log(tf.reduce_mean(A*B))+tf.log(tf.sqrt(tf.reduce_mean(A*A)*tf.reduce_mean(B*B)))#-tf.reduce_sum(concord*pred_d)/tf.reduce_sum((1-concord)*pred_d)
#print(dcor.shape)
#tf.Print(dcor,[dcor])
#dcor = tf.tile([dcor],batch_size)
return (dcor)
model summary looks like:
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input_5 (InputLayer) (None, 64, 64, 1) 0
_________________________________________________________________
conv2d_30 (Conv2D) (None, 64, 64, 128) 3328
_________________________________________________________________
max_pooling2d_13 (MaxPooling (None, 32, 32, 128) 0
_________________________________________________________________
dropout_13 (Dropout) (None, 32, 32, 128) 0
_________________________________________________________________
conv2d_31 (Conv2D) (None, 32, 32, 64) 73792
_________________________________________________________________
max_pooling2d_14 (MaxPooling (None, 16, 16, 64) 0
_________________________________________________________________
dropout_14 (Dropout) (None, 16, 16, 64) 0
_________________________________________________________________
conv2d_32 (Conv2D) (None, 16, 16, 64) 36928
_________________________________________________________________
max_pooling2d_15 (MaxPooling (None, 8, 8, 64) 0
_________________________________________________________________
dropout_15 (Dropout) (None, 8, 8, 64) 0
_________________________________________________________________
conv2d_33 (Conv2D) (None, 8, 8, 1) 577
_________________________________________________________________
zero_padding2d_5 (ZeroPaddin (None, 64, 64, 1) 0
_________________________________________________________________
conv2d_34 (Conv2D) (None, 64, 64, 8) 40
_________________________________________________________________
up_sampling2d_10 (UpSampling (None, 128, 128, 8) 0
_________________________________________________________________
conv2d_35 (Conv2D) (None, 128, 128, 8) 584
_________________________________________________________________
up_sampling2d_11 (UpSampling (None, 256, 256, 8) 0
_________________________________________________________________
conv2d_36 (Conv2D) (None, 256, 256, 16) 1168
_________________________________________________________________
up_sampling2d_12 (UpSampling (None, 512, 512, 16) 0
_________________________________________________________________
conv2d_37 (Conv2D) (None, 512, 512, 1) 145
_________________________________________________________________
flatten_4 (Flatten) (None, 262144) 0
_________________________________________________________________
dense_13 (Dense) (None, 256) 67109120
_________________________________________________________________
dense_14 (Dense) (None, 64) 16448
_________________________________________________________________
dense_15 (Dense) (None, 32) 2080
_________________________________________________________________
dense_16 (Dense) (None, 9) 297
=================================================================
Total params: 67,244,507
Trainable params: 67,244,507
Non-trainable params: 0
_________________________________________________________________
This is the error:
InvalidArgumentError Traceback (most recent call last)
~/anaconda3/lib/python3.6/site-packages/tensorflow/python/framework/ops.py in _create_c_op(graph, node_def, inputs, control_inputs)
1658 try:
-> 1659 c_op = c_api.TF_FinishOperation(op_desc)
1660 except errors.InvalidArgumentError as e:
InvalidArgumentError: Dimensions must be equal, but are 1 and 64 for 'loss_1/zero_padding2d_5_loss/MatMul' (op: 'BatchMatMul') with input shapes: [?,64,64,1], [1,64,64,?].
During handling of the above exception, another exception occurred:
ValueError Traceback (most recent call last)
<ipython-input-11-0e924885fc6b> in <module>
40 autoencoder = Model(input_layer, [encoded_layer,output_layer])
41 autoencoder.summary()
---> 42 autoencoder.compile(optimizer='adadelta', loss=[custom_loss1,custom_loss2])
43 autoencoder.fit(x_train,[x_train, y_train],batch_size=32,epochs=3,shuffle=True,
44 validation_data=(x_val, [x_val,y_val]))
~/anaconda3/lib/python3.6/site-packages/keras/engine/training.py in compile(self, optimizer, loss, metrics, loss_weights, sample_weight_mode, weighted_metrics, target_tensors, **kwargs)
340 with K.name_scope(self.output_names[i] + '_loss'):
341 output_loss = weighted_loss(y_true, y_pred,
--> 342 sample_weight, mask)
343 if len(self.outputs) > 1:
344 self.metrics_tensors.append(output_loss)
~/anaconda3/lib/python3.6/site-packages/keras/engine/training_utils.py in weighted(y_true, y_pred, weights, mask)
402 """
403 # score_array has ndim >= 2
--> 404 score_array = fn(y_true, y_pred)
405 if mask is not None:
406 # Cast the mask to floatX to avoid float64 upcasting in Theano
<ipython-input-11-0e924885fc6b> in custom_loss1(y_true, y_pred)
2 #Wrappers for keras
3 def custom_loss1(y_true,y_pred):
----> 4 dcor = -1*distance_correlation(y_true,encoded_layer)
5 return dcor
6
<ipython-input-6-f282528532cc> in distance_correlation(y_true, y_pred)
2 pred_r = tf.reduce_sum(y_pred*y_pred,1)
3 pred_r = tf.reshape(pred_r,[-1,1])
----> 4 pred_d = pred_r - 2*tf.matmul(y_pred,tf.transpose(y_pred))+tf.transpose(pred_r)
5 true_r = tf.reduce_sum(y_true*y_true,1)
6 true_r = tf.reshape(true_r,[-1,1])
~/anaconda3/lib/python3.6/site-packages/tensorflow/python/ops/math_ops.py in matmul(a, b, transpose_a, transpose_b, adjoint_a, adjoint_b, a_is_sparse, b_is_sparse, name)
2415 adjoint_b = True
2416 return gen_math_ops.batch_mat_mul(
-> 2417 a, b, adj_x=adjoint_a, adj_y=adjoint_b, name=name)
2418
2419 # Neither matmul nor sparse_matmul support adjoint, so we conjugate
~/anaconda3/lib/python3.6/site-packages/tensorflow/python/ops/gen_math_ops.py in batch_mat_mul(x, y, adj_x, adj_y, name)
1421 adj_y = _execute.make_bool(adj_y, "adj_y")
1422 _, _, _op = _op_def_lib._apply_op_helper(
-> 1423 "BatchMatMul", x=x, y=y, adj_x=adj_x, adj_y=adj_y, name=name)
1424 _result = _op.outputs[:]
1425 _inputs_flat = _op.inputs
~/anaconda3/lib/python3.6/site-packages/tensorflow/python/framework/op_def_library.py in _apply_op_helper(self, op_type_name, name, **keywords)
786 op = g.create_op(op_type_name, inputs, output_types, name=scope,
787 input_types=input_types, attrs=attr_protos,
--> 788 op_def=op_def)
789 return output_structure, op_def.is_stateful, op
790
~/anaconda3/lib/python3.6/site-packages/tensorflow/python/util/deprecation.py in new_func(*args, **kwargs)
505 'in a future version' if date is None else ('after %s' % date),
506 instructions)
--> 507 return func(*args, **kwargs)
508
509 doc = _add_deprecated_arg_notice_to_docstring(
~/anaconda3/lib/python3.6/site-packages/tensorflow/python/framework/ops.py in create_op(***failed resolving arguments***)
3298 input_types=input_types,
3299 original_op=self._default_original_op,
-> 3300 op_def=op_def)
3301 self._create_op_helper(ret, compute_device=compute_device)
3302 return ret
~/anaconda3/lib/python3.6/site-packages/tensorflow/python/framework/ops.py in __init__(self, node_def, g, inputs, output_types, control_inputs, input_types, original_op, op_def)
1821 op_def, inputs, node_def.attr)
1822 self._c_op = _create_c_op(self._graph, node_def, grouped_inputs,
-> 1823 control_input_ops)
1824
1825 # Initialize self._outputs.
~/anaconda3/lib/python3.6/site-packages/tensorflow/python/framework/ops.py in _create_c_op(graph, node_def, inputs, control_inputs)
1660 except errors.InvalidArgumentError as e:
1661 # Convert to ValueError for backwards compatibility.
-> 1662 raise ValueError(str(e))
1663
1664 return c_op
ValueError: Dimensions must be equal, but are 1 and 64 for 'loss_1/zero_padding2d_5_loss/MatMul' (op: 'BatchMatMul') with input shapes: [?,64,64,1], [1,64,64,?].
Answer
You are having two loss functions and so you have to pass two y (ground truths) for evaluating the loss with respect to the predictions.
Your first prediction is the output of layer encoded_layer which has a size of (None, 8, 8, 128) as observed from the model.summary for conv2d_59 (Conv2D)
But what you are passing in the fit for y is [x_train, y_train]. The loss_1 is expecting input of size (None, 8, 8, 128) but you are passing x_train which has a different size.
If you want the loss_1 to find the correlation of input image with the encoded image then stack the convolutions such that the output of the convolutions will result in the shape which is the same as your x_train image shape. Use model.summary to see the output shape of convolutions.
No use the padding, strides and kernel size of the convolution layer to get the desired output size of convolutions. use formula W2=(W1−F+2P)/S+1 and H2=(H1−F+2P)/S+1 to find the output width and height of convolutions. Check this reference
There are two major issues with your approach.
- Your loss function is checking the correlation between the encoded image and the actual image. The correct way to do it is to decode the image back from the encoded image and then check the correlation between the decoded image and the actual image (in lines of Autoencoder)
- Your loss 1 is using numpy arrays. For a loss function to be part of a computation graph it should use tensor operations, not numy operations.
Below is the working code. However, for loss 1 I am using l2 norm of the two images. If you want to use correlation then you have to somehow convert it into tensor operations (which is a different issue from this question)
def image_loss(y_true,y_pred):
return tf.norm(y_true - y_pred)
def label_loss(y_true,y_pred):
return categorical_crossentropy(y_true, y_pred)
input_img = Input(shape=(64, 64, 1))
enocded_imag = Conv2D(16, (3, 3), activation='relu', padding='same')(input_img)
enocded_imag = MaxPooling2D((2, 2), padding='same')(enocded_imag)
enocded_imag = Conv2D(8, (3, 3), activation='relu', padding='same')(enocded_imag)
enocded_imag = MaxPooling2D((2, 2), padding='same')(enocded_imag)
enocded_imag = Conv2D(8, (3, 3), activation='relu', padding='same')(enocded_imag)
enocded_imag = MaxPooling2D((2, 2), padding='same')(enocded_imag)
decoded_imag = Conv2D(8, (2, 2), activation='relu', padding='same')(enocded_imag)
decoded_imag = UpSampling2D((2, 2))(decoded_imag)
decoded_imag = Conv2D(8, (3, 3), activation='relu', padding='same')(decoded_imag)
decoded_imag = UpSampling2D((2, 2))(decoded_imag)
decoded_imag = Conv2D(16, (3, 3), activation='relu', padding='same')(decoded_imag)
decoded_imag = UpSampling2D((2, 2))(decoded_imag)
decoded_imag = Conv2D(1, (3, 3), activation='sigmoid', padding='same')(decoded_imag)
flat_layer = Flatten()(enocded_imag)
dense_layer = Dense(32,activation = "relu")(flat_layer)
output_layer = Dense(9, activation = "softmax")(dense_layer)
model = Model(input_img, [decoded_imag, output_layer])
model.compile(optimizer='adadelta', loss=[image_loss, label_loss])
images = np.random.randn(10,64,64,1)
model.fit(images, [images, np.random.randn(10,9)])
The loss function distance_correlation you have coded assumes that each row in y_true and y_pred represent an image. When you use Dense layers it will work because Dense layer outputs a batch of (row) vectors, where each vector represents an individual image. However, 2D convolutions operate on a batch of 2d tensors with multiple channels ( you have only 1 channel). So to use the distance_correlation loss function you have to reshape your tensor such that each row corresponds to an image. Add below two lines to reshape your tensors.
def distance_correlation(y_true,y_pred):
y_true = tf.reshape(tf.squeeze(y_true), [-1,64*64])
y_pred = tf.reshape(tf.squeeze(y_pred), [-1,64*64])
.... REST OF THE CODE ....