How to model Convolutional recurrent network ( CRNN ) in Keras - keras

I was trying to port CRNN model to Keras.
But, I got stuck while connecting output of Conv2D layer to LSTM layer.
Output from CNN layer will have a shape of ( batch_size, 512, 1, width_dash) where first one depends on batch_size, and last one depends on input width of input ( this model can accept variable width input )
For eg: an input with shape [2, 1, 32, 829] was resulting output with shape of (2, 512, 1, 208)
Now, as per Pytorch model, we have to do squeeze(2) followed by permute(2, 0, 1)
it will result a tensor with shape [208, 2, 512 ]
I was trying to implement this is Keras, but I was not able to do that because, in Keras we can not alter batch_size dimension in a keras.models.Sequential model
Can someone please guide me how to port above part of this model to Keras?
Current state of ported CNN layer

You don't need to permute the batch axis in Keras. In a pytorch model you need to do it because a pytorch LSTM expects an input shape (seq_len, batch, input_size). However in Keras, the LSTM layer expects (batch, seq_len, input_size).
So after defining the CNN and squeezing out axis 2, you just need to permute the last two axes. As a simple example (in 'channels_first' Keras image format),
model = Sequential()
model.add(Conv2D(512, 3, strides=(32, 4), padding='same', input_shape=(1, 32, None)))
model.add(Reshape((512, -1)))
model.add(Permute((2, 1)))
model.add(LSTM(32))
You can verify the shapes with model.summary():
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d_4 (Conv2D) (None, 512, 1, None) 5120
_________________________________________________________________
reshape_3 (Reshape) (None, 512, None) 0
_________________________________________________________________
permute_4 (Permute) (None, None, 512) 0
_________________________________________________________________
lstm_3 (LSTM) (None, 32) 69760
=================================================================
Total params: 74,880
Trainable params: 74,880
Non-trainable params: 0
_________________________________________________________________

Related

BATCHNORM toward an axis in pytorch [duplicate]

I've a sample tiny CNN implemented in both Keras and PyTorch. When I print summary of both the networks, the total number of trainable parameters are same but total number of parameters and number of parameters for Batch Normalization don't match.
Here is the CNN implementation in Keras:
inputs = Input(shape = (64, 64, 1)). # Channel Last: (NHWC)
model = Conv2D(filters=32, kernel_size=(3, 3), padding='SAME', activation='relu', input_shape=(IMG_SIZE, IMG_SIZE, 1))(inputs)
model = BatchNormalization(momentum=0.15, axis=-1)(model)
model = Flatten()(model)
dense = Dense(100, activation = "relu")(model)
head_root = Dense(10, activation = 'softmax')(dense)
And the summary printed for above model is:
Model: "model_8"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input_9 (InputLayer) (None, 64, 64, 1) 0
_________________________________________________________________
conv2d_10 (Conv2D) (None, 64, 64, 32) 320
_________________________________________________________________
batch_normalization_2 (Batch (None, 64, 64, 32) 128
_________________________________________________________________
flatten_3 (Flatten) (None, 131072) 0
_________________________________________________________________
dense_11 (Dense) (None, 100) 13107300
_________________________________________________________________
dense_12 (Dense) (None, 10) 1010
=================================================================
Total params: 13,108,758
Trainable params: 13,108,694
Non-trainable params: 64
_________________________________________________________________
Here's the implementation of the same model architecture in PyTorch:
# Image format: Channel first (NCHW) in PyTorch
class CustomModel(nn.Module):
def __init__(self):
super(CustomModel, self).__init__()
self.layer1 = nn.Sequential(
nn.Conv2d(in_channels=1, out_channels=32, kernel_size=(3, 3), padding=1),
nn.ReLU(True),
nn.BatchNorm2d(num_features=32),
)
self.flatten = nn.Flatten()
self.fc1 = nn.Linear(in_features=131072, out_features=100)
self.fc2 = nn.Linear(in_features=100, out_features=10)
def forward(self, x):
output = self.layer1(x)
output = self.flatten(output)
output = self.fc1(output)
output = self.fc2(output)
return output
And following is the output of summary of the above model:
----------------------------------------------------------------
Layer (type) Output Shape Param #
================================================================
Conv2d-1 [-1, 32, 64, 64] 320
ReLU-2 [-1, 32, 64, 64] 0
BatchNorm2d-3 [-1, 32, 64, 64] 64
Flatten-4 [-1, 131072] 0
Linear-5 [-1, 100] 13,107,300
Linear-6 [-1, 10] 1,010
================================================================
Total params: 13,108,694
Trainable params: 13,108,694
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.02
Forward/backward pass size (MB): 4.00
Params size (MB): 50.01
Estimated Total Size (MB): 54.02
----------------------------------------------------------------
As you can see in above results, Batch Normalization in Keras has more number of parameters than PyTorch (2x to be exact). So what's the difference in above CNN architectures? If they are equivalent, then what am I missing here?
Keras treats as parameters (weights) many things that will be "saved/loaded" in the layer.
While both implementations naturally have the accumulated "mean" and "variance" of the batches, these values are not trainable with backpropagation.
Nevertheless, these values are updated every batch, and Keras treats them as non-trainable weights, while PyTorch simply hides them. The term "non-trainable" here means "not trainable by backpropagation", but doesn't mean the values are frozen.
In total they are 4 groups of "weights" for a BatchNormalization layer. Considering the selected axis (default = -1, size=32 for your layer)
scale (32) - trainable
offset (32) - trainable
accumulated means (32) - non-trainable, but updated every batch
accumulated std (32) - non-trainable, but updated every batch
The advantage of having it like this in Keras is that when you save the layer, you also save the mean and variance values the same way you save all other weights in the layer automatically. And when you load the layer, these weights are loaded together.

Keras LSTM Layer ValueError: Dimensions must be equal, but are 17 and 2

I'm working on a basic RNN model for a multiclass task and I'm facing some issues with output dimensions.
This is my input/output shapes:
input.shape = (50000, 2, 5) # (samples, features, feature_len)
output.shape = (50000, 17, 185) # (samples, features, feature_len) <-- one hot encoded
input[0].shape = (2, 5)
output[0].shape = (17, 185)
This is my model, using Keras functional API:
inp = tf.keras.Input(shape=(2, 5,))
x = tf.keras.layers.LSTM(128, input_shape=(2, 5,), return_sequences=True, activation='relu')(inp)
out = tf.keras.layers.Dense(185, activation='softmax')(x)
model = tf.keras.models.Model(inputs=inp, outputs=out)
This is my model.summary():
Layer (type) Output Shape Param #
=================================================================
input_1 (InputLayer) [(None, 2, 5)] 0
_________________________________________________________________
lstm (LSTM) (None, 2, 128) 68608
_________________________________________________________________
dense (Dense) (None, 2, 185) 23865
=================================================================
Total params: 92,473
Trainable params: 92,473
Non-trainable params: 0
_________________________________________________________________
Then I compile the model and run fit():
model.compile(optimizer='adam',
loss=tf.nn.softmax_cross_entropy_with_logits,
metrics='accuracy')
model.fit(x=input, y=output, epochs=5)
And I'm getting a dimension error:
ValueError: Dimensions must be equal, but are 17 and 2 for '{{node Equal}} = Equal[T=DT_INT64, incompatible_shape_error=true](ArgMax, ArgMax_1)' with input shapes: [?,17], [?,2].
The error is clear, the model output a dimension 2 and my output has dimension 17, although I understand the issue, I can't find a way of fixing it, any ideas?
I think your output shape is not "output[0].shape = (17, 185)" but "dense (Dense) (None, 2, 185) ".
You need to change your output shape or change your layer structure.
LSTM output is a list of encoder_outputs, when you specify return_sequences=True. hence; I suggest just using the last item of encoder_outputs as the input of your Dense layer. you can see the example section of this link to the documentation. It may help you.

ValueError: expected dense_22 to have shape (None, 37) but got array with shape (1000, 2)

I am currently working on a question answering system. I create a synthetic dataset that contains multiple words in the answers. But, the answers are not a span of the given context.
Initially, I am planning to test it using a deep learning-based model. But I have some problems building the model.
This is how I vectorized data.
def vectorize(data, word2idx, story_maxlen, question_maxlen, answer_maxlen):
""" Create the story and question vectors and the label """
Xs, Xq, Y = [], [], []
for story, question, answer in data:
xs = [word2idx[word] for word in story]
xq = [word2idx[word] for word in question]
y = [word2idx[word] for word in answer]
#y = np.zeros(len(word2idx) + 1)
#y[word2idx[answer]] = 1
Xs.append(xs)
Xq.append(xq)
Y.append(y)
return (pad_sequences(Xs, maxlen=story_maxlen),
pad_sequences(Xq, maxlen=question_maxlen),
pad_sequences(Y, maxlen=answer_maxlen))
#np.array(Y))
below is how I create the model.
# story encoder. Output dim: (None, story_maxlen, EMBED_HIDDEN_SIZE)
story_encoder = Sequential()
story_encoder.add(Embedding(input_dim=vocab_size,
output_dim=EMBED_HIDDEN_SIZE,
input_length=story_maxlen))
story_encoder.add(Dropout(0.3))
# question encoder. Output dim: (None, question_maxlen, EMBED_HIDDEN_SIZE)
question_encoder = Sequential()
question_encoder.add(Embedding(input_dim=vocab_size,
output_dim=EMBED_HIDDEN_SIZE,
input_length=question_maxlen))
question_encoder.add(Dropout(0.3))
# episodic memory (facts): story * question
# Output dim: (None, question_maxlen, story_maxlen)
facts_encoder = Sequential()
facts_encoder.add(Merge([story_encoder, question_encoder],
mode="dot", dot_axes=[2, 2]))
facts_encoder.add(Permute((2, 1)))
## combine response and question vectors and do logistic regression
answer = Sequential()
answer.add(Merge([facts_encoder, question_encoder],
mode="concat", concat_axis=-1))
answer.add(LSTM(LSTM_OUTPUT_SIZE, return_sequences=True))
answer.add(Dropout(0.3))
answer.add(Flatten())
answer.add(Dense(vocab_size,activation= "softmax"))
answer.compile(optimizer="rmsprop", loss="categorical_crossentropy",
metrics=["accuracy"])
answer.fit([Xs_train, Xq_train], Y_train,
batch_size=BATCH_SIZE, nb_epoch=NBR_EPOCHS,
validation_data=([Xs_test, Xq_test], Y_test))
and this is the summary of the model
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
merge_46 (Merge) (None, 5, 616) 0
_________________________________________________________________
lstm_23 (LSTM) (None, 5, 32) 83072
_________________________________________________________________
dropout_69 (Dropout) (None, 5, 32) 0
_________________________________________________________________
flatten_9 (Flatten) (None, 160) 0
_________________________________________________________________
dense_22 (Dense) (None, 37) 5957
=================================================================
Total params: 93,765.0
Trainable params: 93,765.0
Non-trainable params: 0.0
_________________________________________________________________
It gives the following error.
ValueError: Error when checking model target: expected dense_22 to have shape (None, 37) but got array with shape (1000, 2)
I think the error is related to Y_train, Y_test. I should encode them to categorical values and the answers are not spans of text, but sequential. I don't know what/how to do it.
how can I fix it? any ideas?
EDIT:
When I use sparse_categorical_crossentropy in the loss, and Reshape(2,-1);
answer.summary()
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
merge_94 (Merge) (None, 5, 616) 0
_________________________________________________________________
lstm_65 (LSTM) (None, 5, 32) 83072
_________________________________________________________________
dropout_139 (Dropout) (None, 5, 32) 0
_________________________________________________________________
reshape_22 (Reshape) (None, 2, 80) 0
_________________________________________________________________
dense_44 (Dense) (None, 2, 37) 2997
=================================================================
Total params: 90,805.0
Trainable params: 90,805.0
Non-trainable params: 0.0
_________________________________________________________________
EDIT2:
The model after modifications
# story encoder. Output dim: (None, story_maxlen, EMBED_HIDDEN_SIZE)
story_encoder = Sequential()
story_encoder.add(Embedding(input_dim=vocab_size,
output_dim=EMBED_HIDDEN_SIZE,
input_length=story_maxlen))
story_encoder.add(Dropout(0.3))
# question encoder. Output dim: (None, question_maxlen, EMBED_HIDDEN_SIZE)
question_encoder = Sequential()
question_encoder.add(Embedding(input_dim=vocab_size,
output_dim=EMBED_HIDDEN_SIZE,
input_length=question_maxlen))
question_encoder.add(Dropout(0.3))
# episodic memory (facts): story * question
# Output dim: (None, question_maxlen, story_maxlen)
facts_encoder = Sequential()
facts_encoder.add(Merge([story_encoder, question_encoder],
mode="dot", dot_axes=[2, 2]))
facts_encoder.add(Permute((2, 1)))
## combine response and question vectors and do logistic regression
## combine response and question vectors and do logistic regression
answer = Sequential()
answer.add(Merge([facts_encoder, question_encoder],
mode="concat", concat_axis=-1))
answer.add(LSTM(LSTM_OUTPUT_SIZE, return_sequences=True))
answer.add(Dropout(0.3))
#answer.add(Flatten())
answer.add(keras.layers.Reshape((2, -1)))
answer.add(Dense(vocab_size,activation= "softmax"))
answer.compile(optimizer="rmsprop", loss="sparse_categorical_crossentropy",
metrics=["accuracy"])
answer.fit([Xs_train, Xq_train], Y_train,
batch_size=BATCH_SIZE, nb_epoch=NBR_EPOCHS,
validation_data=([Xs_test, Xq_test], Y_test))
It still gives
ValueError: Error when checking model target: expected dense_46 to have 3 dimensions, but got array with shape (1000, 2)
As far as I understand - Y_train, Y_test comprise of indexes (not one-hot vectors). If so - change loss to sparse_categorical_entropy:
answer.compile(optimizer="rmsprop", loss="sparse_categorical_crossentropy",
metrics=["accuracy"])
As far as I understand - Y_train, Y_test have a sequence dimension. And the length of questions (5) doesn't equal to the length of the answers (2). This dimension is removed by Flatten(). Try to replace Flatten() by Reshape():
# answer.add(Flatten())
answer.add(tf.keras.layers.Reshape((2, -1)))

logits and labels must have the same first dimension, got logits shape [3662,5] and labels shape [18310]

I am trying to create a CNN with tensorflow, my images are 64x64x1 images and I have an array of 3662 images which I am using for training. I have total 5 labels which I have one-hot encoded. I am getting this error everytime:
InvalidArgumentError: logits and labels must have the same first dimension, got logits shape [3662,5] and labels shape [18310]
[[{{node loss_2/dense_5_loss/SparseSoftmaxCrossEntropyWithLogits/SparseSoftmaxCrossEntropyWithLogits}}]]
my neural network structure is this:
def cnn_model():
model = models.Sequential()
# model.add(layers.Dense(128, activation='relu', ))
model.add(layers.Conv2D(128, (3, 3), activation='relu',input_shape=(64, 64, 1)))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Conv2D(64, (3, 3), activation='relu',padding = 'same'))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Conv2D(64, (3, 3), activation='relu', padding='same'))
model.add(layers.Dense(64, activation='relu'))
model.add(layers.Flatten())
model.add(layers.Dense(5, activation='softmax'))
model.compile(optimizer='adam',
loss=tf.keras.losses.SparseCategoricalCrossentropy(),
metrics=['accuracy'])
print(model.summary())
return model
My model summary is this:
Model: "sequential_3"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d_9 (Conv2D) (None, 62, 62, 128) 1280
_________________________________________________________________
max_pooling2d_6 (MaxPooling2 (None, 31, 31, 128) 0
_________________________________________________________________
conv2d_10 (Conv2D) (None, 31, 31, 64) 73792
_________________________________________________________________
max_pooling2d_7 (MaxPooling2 (None, 15, 15, 64) 0
_________________________________________________________________
conv2d_11 (Conv2D) (None, 15, 15, 64) 36928
_________________________________________________________________
dense_4 (Dense) (None, 15, 15, 64) 4160
_________________________________________________________________
flatten_2 (Flatten) (None, 14400) 0
_________________________________________________________________
dense_5 (Dense) (None, 5) 72005
=================================================================
Total params: 188,165
Trainable params: 188,165
Non-trainable params: 0
my output array is of the shape (3662,5,1). I have seen other answers to same questions but I can't figure out the problem with mine. Where am I wrong?
Edit: My labels are stored in one hot encoded form using these:
df = pd.get_dummies(df)
diag = np.array(df)
diag = np.reshape(diag,(3662,5,1))
I have tried as numpy array and after converting to tensor(same for input as per documentation)
The problem lines within the choice of the loss function tf.keras.losses.SparseCategoricalCrossentropy(). According to what you are trying to achieve you should use tf.keras.losses.CategoricalCrossentropy(). Namely, the documentation of tf.keras.losses.SparseCategoricalCrossentropy() states:
Use this crossentropy loss function when there are two or more label classes. We expect labels to be provided as integers.
On the other hand, the documentation of tf.keras.losses.CategoricalCrossentropy() states:
We expect labels to be provided in a one_hot representation.
And because your labels are encoded as one-hot, you should use tf.keras.losses.CategoricalCrossentropy().

How to have a 3D filter for Conv2D in Keras?

My input images have 8 channels and my output (label) has 1 channel and my CNN in keras is like below:
def set_model(ks1=5, ks2=5, nf1=64, nf2=1):
model = Sequential()
model.add(Conv2D(nf1, padding="same", kernel_size=(ks1, ks1),
activation='relu', input_shape=(62, 62, 8)))
model.add(Conv2D(nf2, padding="same", kernel_size=(ks2, ks2),
activation='relu'))
model.compile(loss=keras.losses.binary_crossentropy,
optimizer=keras.optimizers.Adadelta())
return model
The filter I have here is the same for all 8 channels. What I would like to have is a 3D filter, something like (8, 5, 5) such that every channel has a separate filter because these channels have not the same importance.
Below is the summary of the model implemented above:
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d_1 (Conv2D) (None, 62, 62, 64) 12864
_________________________________________________________________
conv2d_2 (Conv2D) (None, 62, 62, 1) 1601
=================================================================
Total params: 14,465
Trainable params: 14,465
Non-trainable params: 0
_________________________________________________________________
And when I get the shape of weights for the first layer I have the following results:
for layer in model.layers:
weights = layer.get_weights()
len(weights)
2
a = np.array(weights[0])
a.shape
(5, 5, 64, 1)
And I am wondering where is 8 in the shape of weights of the first layer?

Resources