Develop Reference

How to make predictions on new dataset with tensorflow's gradient tape

While I'm able to understand how to use model.fit(x_train, y_train), I can't figure out how to make predictions on new data using tensorflow's gradient tape. My github repository with runnable code (up to an error) can be found here. What is currently working is that I get the trained model "network_output", however it appears that with gradient tape, argmax is being used on the model itself, where I'm used to model.fit() taking the test data as an input:
network_output = trained_network(input_images,input_number)
preds = np.argmax(network_output, axis=1)
Where "input_images" is an ndarray: (20,3,3,1) and "input_number" is an ndarray: (20,5).
Now I'm taking network_output as the trained model and would like to use it to predict similarly typed data of test_images, and test_number respectively.
The error 'tensorflow.python.framework.ops.EagerTensor' object has no attribute 'predict' here:
predicted_number = network_output.predict(test_images)
Which is because I don't know how to use the tape to make predictions. However once the prediction works I would guess I can compare the resulting "predicted_number" against the "test_number" as would usually be done using the model.fit method.
acc = 0
for i in range(len(test_images)):
if (predicted_number[i] == test_number[i]):
acc += 1
print("Accuracy: ", acc / len(input_images) * 100, "%")

In order to obtain prediction I usually iterate through batches manually like this:
predictions = []
for batch in range(num_batch):
logits = trained_network(x_test[batch * batch_size: (batch + 1) * batch_size], training=False)
# first obtain probabilities
# (if the last layer of the network has no activation, otherwise skip the softmax here)
prob = tf.nn.softmax(logits)
# putting back together predictions for all batches
predictions.extend(tf.argmax(input=prob, axis=1))
If you don't have a lot of data you can skip the loop, this is faster than using predict because you directly invoke the __call__ method of the model:
logits = trained_network(x_test, training=False)
prob = tf.nn.softmax(logits)
predictions = tf.argmax(input=prob, axis=1)
Finally you could also use predict. In this case the batches are handled automatically. It is easier to use when you have lots of data since you don't have to create a loop to interate through batches. The result is a numpy array of predictions. In can be used like this:
predictions = trained_network.predict(x_test) # you can set a batch_size if you want
What you're doing wrong is this part:
network_output = trained_network(input_images,input_number)
predicted_number = network_output.predict(test_images)
You have to call predict directly on your model trained_network.

Siamese Network for binary classification with pre-encoded inputs

I want to train a Siamese Network to compare vectors for similarity.
My dataset consist of pairs of vectors and a target column with "1" if they are the same and "0" otherwise (binary classification):
import pandas as pd
# Define train and test sets.
X_train_val = pd.read_csv("train.csv")
print(X_train_val.head())
y_train_val = X_train_val.pop("class")
print(y_train_val.value_counts())
# Keep 50% of X_train_val in validation set.
X_train, X_val = X_train_val[:991], X_train_val[991:]
y_train, y_val = y_train_val[:991], y_train_val[991:]
del X_train_val, y_train_val
# Split our data to 'left' and 'right' inputs (one for each side Siamese).
X_left_train, X_right_train = X_train.iloc[:, :200], X_train.iloc[:, 200:]
X_left_val, X_right_val = X_val.iloc[:, :200], X_val.iloc[:, 200:]
assert X_left_train.shape == X_right_train.shape
# Repeat for test set.
X_test = pd.read_csv("test.csv")
y_test = X_test.pop("class")
print(y_test.value_counts())
X_left_test, X_right_test = X_test.iloc[:, :200], X_test.iloc[:, 200:]
returns
v0 v1 v2 ... v397 v398 v399 class
0 0.003615 0.013794 0.030388 ... -0.093931 0.106202 0.034870 0.0
1 0.018988 0.056302 0.002915 ... -0.007905 0.100859 -0.043529 0.0
2 0.072516 0.125697 0.111230 ... -0.010007 0.064125 -0.085632 0.0
3 0.051016 0.066028 0.082519 ... 0.012677 0.043831 -0.073935 1.0
4 0.020367 0.026446 0.015681 ... 0.062367 -0.022781 -0.032091 0.0
1.0 1060
0.0 923
Name: class, dtype: int64
1.0 354
0.0 308
Name: class, dtype: int64
The rest of my script is as follows:
import keras
import keras.backend as K
from keras.layers import Dense, Dropout, Input, Lambda
from keras.models import Model
def euclidean_distance(vectors):
"""
Find the Euclidean distance between two vectors.
"""
x, y = vectors
sum_square = K.sum(K.square(x - y), axis=1, keepdims=True)
# Epsilon is small value that makes very little difference to the value of the denominator, but ensures that it isn't equal to exactly zero.
return K.sqrt(K.maximum(sum_square, K.epsilon()))
def contrastive_loss(y_true, y_pred):
"""
Distance-based loss function that tries to ensure that data samples that are semantically similar are embedded closer together.
See:
* https://gombru.github.io/2019/04/03/ranking_loss/
"""
margin = 1
return K.mean(y_true * K.square(y_pred) + (1 - y_true) * K.square(K.maximum(margin - y_pred, 0)))
def accuracy(y_true, y_pred):
"""
Compute classification accuracy with a fixed threshold on distances.
"""
return K.mean(K.equal(y_true, K.cast(y_pred < 0.5, y_true.dtype)))
def create_base_network(input_dim: int, dense_units: int, dropout_rate: float):
input1 = Input(input_dim, name="encoder")
x = input1
x = Dense(dense_units, activation="relu")(x)
x = Dropout(dropout_rate)(x)
x = Dense(dense_units, activation="relu")(x)
x = Dropout(dropout_rate)(x)
x = Dense(dense_units, activation="relu", name="Embeddings")(x)
return Model(input1, x)
def build_siamese_model(input_dim: int):
shared_network = create_base_network(input_dim, dense_units=128, dropout_rate=0.1)
left_input = Input(input_dim)
right_input = Input(input_dim)
# Since this is a siamese nn, both sides share the same network.
encoded_l = shared_network(left_input)
encoded_r = shared_network(right_input)
# The euclidean distance layer outputs close to 0 value when two inputs are similar and 1 otherwise.
distance = Lambda(euclidean_distance, name="Euclidean-Distance")([encoded_l, encoded_r])
siamese_net = Model(inputs=[left_input, right_input], outputs=distance)
siamese_net.compile(loss=contrastive_loss, optimizer="RMSprop", metrics=[accuracy])
return siamese_net
model = build_siamese_model(X_left_train.shape[1])
es_callback = keras.callbacks.EarlyStopping(monitor="val_loss", patience=3, verbose=0)
history = model.fit(
[X_left_train, X_right_train],
y_train,
validation_data=([X_left_val, X_right_val], y_val),
epochs=100,
callbacks=[es_callback],
verbose=1,
)
I have plotted the contrastive loss vs epoch and model accuracy vs epoch:
The validation line is almost flat, which seems odd to me (overfitted?).
After changing the dropout of the shared network from 0.1 to 0.5, I get the following results:
Somehow it looks better, but yields bad predictions as well.
My questions are:
Most examples of Siamese Networks I've seen so far involves embedding layers (text pairs) and/or Convolution layers (image pairs). My input pairs are the actual vector representation of some text, which is why I used Dense layers for the shared network. Is this the proper approach?
The output layer of my Siamese Network is as follows:
distance = Lambda(euclidean_distance, name="Euclidean-Distance")([encoded_l, encoded_r])
siamese_net = Model(inputs=[left_input, right_input], outputs=distance)
siamese_net.compile(loss=contrastive_loss, optimizer="RMSprop", metrics=[accuracy])
but someone over the internet suggested this instead:
distance = Lambda(lambda tensors: K.abs(tensors[0] - tensors[1]), name="L1-Distance")([encoded_l, encoded_r])
output = Dense(1, activation="sigmoid")(distance) # returns the class probability
siamese_net = Model(inputs=[left_input, right_input], outputs=output)
siamese_net.compile(loss="binary_crossentropy", optimizer="adam", metrics=["accuracy"])
I'm not sure which one to trust nor the difference between them (except that the former returns the distance and the latter returns the class probability). In my experiments, I get poor results with binary_crossentropy.
EDIT:
After following #PlzBePython suggestions, I come up with the following base network:
distance = Lambda(lambda tensors: K.abs(tensors[0] - tensors[1]), name="L1-Distance")([encoded_l, encoded_r])
output = Dense(1, activation="linear")(distance)
siamese_net = Model(inputs=[left_input, right_input], outputs=output)
siamese_net.compile(loss=contrastive_loss, optimizer="RMSprop", metrics=[accuracy])
Thank you for your help!

This is less of an answer and more writing my thoughts down and hoping they can help find an answer.
In general, everything you do seems pretty reasonable to me.
Regarding your Questions:
1:
Embedding or feature extraction layers are never a must, but almost always make it easier to learn the intended. You can think of them like providing your distance model with the comprehensive summary of a sentence instead of its raw words. This also makes your model not depend on the location of a word. In your case, creating the summary/important features of a sentence and embedding similar sentences close to each other is done by the same network. Of course, this can also work, and I don't even think it's a bad approach. However, I would maybe increase the network size.
2:
In my opinion, those two loss functions are not too different. Binary Crossentropy is defined as:
While Contrastive Loss (margin = 1) is:
So you basically swap a log function for a hinge function.
The only real difference comes from the distance calculation. You probably got suggested using some kind of L1 distance, since L2 distance is supposed to perform worse with higher dimensions (see for example here) and your dimensionality is 128. Personally, I would rather go with L1 in your case, but I don't think it's a dealbreaker.
What I would try is:
increase the margin parameter. "1" always results in a pretty low loss in the false positive case. This could slow down training in general
try out embedding into the [-inf, inf] space (change last layer embedding activation to "linear")
change "binary_crossentropy" loss into "keras.losses.BinaryCrossentropy(from_logits=True)" and last activation from "sigmoid" to "linear". This should actually not make a difference, but I've made some weird experiences with the keras binary crossentropy function and from_logits seems to help sometimes
increase parameters
Lastly, a validation accuracy of 90% actually looks pretty good to me. Keep in mind, that when the validation accuracy is calculated in the first epoch, the model already has done about 60 weight updates (batch_size = 32). That means, especially in the first episode, a validation accuracy that is higher than the training accuracy (which is calculated during training) is kind of to be expected. Also, this can sometimes cause the misbelief that training loss is increasing faster than validation loss.
EDIT
I recommended "linear" in the last layer, because tensorflow recommends it ("from_logits"=True which requires value in [-inf, inf]) for Binary Crossentropy. In my experience, it converges better.

LSTM Autoencoder problems

TLDR:
Autoencoder underfits timeseries reconstruction and just predicts average value.
Question Set-up:
Here is a summary of my attempt at a sequence-to-sequence autoencoder. This image was taken from this paper: https://arxiv.org/pdf/1607.00148.pdf
Encoder: Standard LSTM layer. Input sequence is encoded in the final hidden state.
Decoder: LSTM Cell (I think!). Reconstruct the sequence one element at a time, starting with the last element x[N].
Decoder algorithm is as follows for a sequence of length N:
Get Decoder initial hidden state hs[N]: Just use encoder final hidden state.
Reconstruct last element in the sequence: x[N]= w.dot(hs[N]) + b.
Same pattern for other elements: x[i]= w.dot(hs[i]) + b
use x[i] and hs[i] as inputs to LSTMCell to get x[i-1] and hs[i-1]
Minimum Working Example:
Here is my implementation, starting with the encoder:
class SeqEncoderLSTM(nn.Module):
def __init__(self, n_features, latent_size):
super(SeqEncoderLSTM, self).__init__()
self.lstm = nn.LSTM(
n_features,
latent_size,
batch_first=True)
def forward(self, x):
_, hs = self.lstm(x)
return hs
Decoder class:
class SeqDecoderLSTM(nn.Module):
def __init__(self, emb_size, n_features):
super(SeqDecoderLSTM, self).__init__()
self.cell = nn.LSTMCell(n_features, emb_size)
self.dense = nn.Linear(emb_size, n_features)
def forward(self, hs_0, seq_len):
x = torch.tensor([])
# Final hidden and cell state from encoder
hs_i, cs_i = hs_0
# reconstruct first element with encoder output
x_i = self.dense(hs_i)
x = torch.cat([x, x_i])
# reconstruct remaining elements
for i in range(1, seq_len):
hs_i, cs_i = self.cell(x_i, (hs_i, cs_i))
x_i = self.dense(hs_i)
x = torch.cat([x, x_i])
return x
Bringing the two together:
class LSTMEncoderDecoder(nn.Module):
def __init__(self, n_features, emb_size):
super(LSTMEncoderDecoder, self).__init__()
self.n_features = n_features
self.hidden_size = emb_size
self.encoder = SeqEncoderLSTM(n_features, emb_size)
self.decoder = SeqDecoderLSTM(emb_size, n_features)
def forward(self, x):
seq_len = x.shape[1]
hs = self.encoder(x)
hs = tuple([h.squeeze(0) for h in hs])
out = self.decoder(hs, seq_len)
return out.unsqueeze(0)
And here's my training function:
def train_encoder(model, epochs, trainload, testload=None, criterion=nn.MSELoss(), optimizer=optim.Adam, lr=1e-6, reverse=False):
device = 'cuda' if torch.cuda.is_available() else 'cpu'
print(f'Training model on {device}')
model = model.to(device)
opt = optimizer(model.parameters(), lr)
train_loss = []
valid_loss = []
for e in tqdm(range(epochs)):
running_tl = 0
running_vl = 0
for x in trainload:
x = x.to(device).float()
opt.zero_grad()
x_hat = model(x)
if reverse:
x = torch.flip(x, [1])
loss = criterion(x_hat, x)
loss.backward()
opt.step()
running_tl += loss.item()
if testload is not None:
model.eval()
with torch.no_grad():
for x in testload:
x = x.to(device).float()
loss = criterion(model(x), x)
running_vl += loss.item()
valid_loss.append(running_vl / len(testload))
model.train()
train_loss.append(running_tl / len(trainload))
return train_loss, valid_loss
Data:
Large dataset of events scraped from the news (ICEWS). Various categories exist that describe each event. I initially one-hot encoded these variables, expanding the data to 274 dimensions. However, in order to debug the model, I've cut it down to a single sequence that is 14 timesteps long and only contains 5 variables. Here is the sequence I'm trying to overfit:
tensor([[0.5122, 0.0360, 0.7027, 0.0721, 0.1892],
[0.5177, 0.0833, 0.6574, 0.1204, 0.1389],
[0.4643, 0.0364, 0.6242, 0.1576, 0.1818],
[0.4375, 0.0133, 0.5733, 0.1867, 0.2267],
[0.4838, 0.0625, 0.6042, 0.1771, 0.1562],
[0.4804, 0.0175, 0.6798, 0.1053, 0.1974],
[0.5030, 0.0445, 0.6712, 0.1438, 0.1404],
[0.4987, 0.0490, 0.6699, 0.1536, 0.1275],
[0.4898, 0.0388, 0.6704, 0.1330, 0.1579],
[0.4711, 0.0390, 0.5877, 0.1532, 0.2201],
[0.4627, 0.0484, 0.5269, 0.1882, 0.2366],
[0.5043, 0.0807, 0.6646, 0.1429, 0.1118],
[0.4852, 0.0606, 0.6364, 0.1515, 0.1515],
[0.5279, 0.0629, 0.6886, 0.1514, 0.0971]], dtype=torch.float64)
And here is the custom Dataset class:
class TimeseriesDataSet(Dataset):
def __init__(self, data, window, n_features, overlap=0):
super().__init__()
if isinstance(data, (np.ndarray)):
data = torch.tensor(data)
elif isinstance(data, (pd.Series, pd.DataFrame)):
data = torch.tensor(data.copy().to_numpy())
else:
raise TypeError(f"Data should be ndarray, series or dataframe. Found {type(data)}.")
self.n_features = n_features
self.seqs = torch.split(data, window)
def __len__(self):
return len(self.seqs)
def __getitem__(self, idx):
try:
return self.seqs[idx].view(-1, self.n_features)
except TypeError:
raise TypeError("Dataset only accepts integer index/slices, not lists/arrays.")
Problem:
The model only learns the average, no matter how complex I make the model or now long I train it.
Predicted/Reconstruction:
Actual:
My research:
This problem is identical to the one discussed in this question: LSTM autoencoder always returns the average of the input sequence
The problem in that case ended up being that the objective function was averaging the target timeseries before calculating loss. This was due to some broadcasting errors because the author didn't have the right sized inputs to the objective function.
In my case, I do not see this being the issue. I have checked and double checked that all of my dimensions/sizes line up. I am at a loss.
Other Things I've Tried
I've tried this with varied sequence lengths from 7 timesteps to 100 time steps.
I've tried with varied number of variables in the time series. I've tried with univariate all the way to all 274 variables that the data contains.
I've tried with various reduction parameters on the nn.MSELoss module. The paper calls for sum, but I've tried both sum and mean. No difference.
The paper calls for reconstructing the sequence in reverse order (see graphic above). I have tried this method using the flipud on the original input (after training but before calculating loss). This makes no difference.
I tried making the model more complex by adding an extra LSTM layer in the encoder.
I've tried playing with the latent space. I've tried from 50% of the input number of features to 150%.
I've tried overfitting a single sequence (provided in the Data section above).
Question:
What is causing my model to predict the average and how do I fix it?

Okay, after some debugging I think I know the reasons.
TLDR
You try to predict next timestep value instead of difference between current timestep and the previous one
Your hidden_features number is too small making the model unable to fit even a single sample
Analysis
Code used
Let's start with the code (model is the same):
import seaborn as sns
import matplotlib.pyplot as plt
def get_data(subtract: bool = False):
# (1, 14, 5)
input_tensor = torch.tensor(
[
[0.5122, 0.0360, 0.7027, 0.0721, 0.1892],
[0.5177, 0.0833, 0.6574, 0.1204, 0.1389],
[0.4643, 0.0364, 0.6242, 0.1576, 0.1818],
[0.4375, 0.0133, 0.5733, 0.1867, 0.2267],
[0.4838, 0.0625, 0.6042, 0.1771, 0.1562],
[0.4804, 0.0175, 0.6798, 0.1053, 0.1974],
[0.5030, 0.0445, 0.6712, 0.1438, 0.1404],
[0.4987, 0.0490, 0.6699, 0.1536, 0.1275],
[0.4898, 0.0388, 0.6704, 0.1330, 0.1579],
[0.4711, 0.0390, 0.5877, 0.1532, 0.2201],
[0.4627, 0.0484, 0.5269, 0.1882, 0.2366],
[0.5043, 0.0807, 0.6646, 0.1429, 0.1118],
[0.4852, 0.0606, 0.6364, 0.1515, 0.1515],
[0.5279, 0.0629, 0.6886, 0.1514, 0.0971],
]
).unsqueeze(0)
if subtract:
initial_values = input_tensor[:, 0, :]
input_tensor -= torch.roll(input_tensor, 1, 1)
input_tensor[:, 0, :] = initial_values
return input_tensor
if __name__ == "__main__":
torch.manual_seed(0)
HIDDEN_SIZE = 10
SUBTRACT = False
input_tensor = get_data(SUBTRACT)
model = LSTMEncoderDecoder(input_tensor.shape[-1], HIDDEN_SIZE)
optimizer = torch.optim.Adam(model.parameters())
criterion = torch.nn.MSELoss()
for i in range(1000):
outputs = model(input_tensor)
loss = criterion(outputs, input_tensor)
loss.backward()
optimizer.step()
optimizer.zero_grad()
print(f"{i}: {loss}")
if loss < 1e-4:
break
# Plotting
sns.lineplot(data=outputs.detach().numpy().squeeze())
sns.lineplot(data=input_tensor.detach().numpy().squeeze())
plt.show()
What it does:
get_data either works on the data your provided if subtract=False or (if subtract=True) it subtracts value of the previous timestep from the current timestep
Rest of the code optimizes the model until 1e-4 loss reached (so we can compare how model's capacity and it's increase helps and what happens when we use the difference of timesteps instead of timesteps)
We will only vary HIDDEN_SIZE and SUBTRACT parameters!
NO SUBTRACT, SMALL MODEL
HIDDEN_SIZE=5
SUBTRACT=False
In this case we get a straight line. Model is unable to fit and grasp the phenomena presented in the data (hence flat lines you mentioned).
1000 iterations limit reached
SUBTRACT, SMALL MODEL
HIDDEN_SIZE=5
SUBTRACT=True
Targets are now far from flat lines, but model is unable to fit due to too small capacity.
1000 iterations limit reached
NO SUBTRACT, LARGER MODEL
HIDDEN_SIZE=100
SUBTRACT=False
It got a lot better and our target was hit after 942 steps. No more flat lines, model capacity seems quite fine (for this single example!)
SUBTRACT, LARGER MODEL
HIDDEN_SIZE=100
SUBTRACT=True
Although the graph does not look that pretty, we got to desired loss after only 215 iterations.
Finally
Usually use difference of timesteps instead of timesteps (or some other transformation, see here for more info about that). In other cases, neural network will try to simply... copy output from the previous step (as that's the easiest thing to do). Some minima will be found this way and going out of it will require more capacity.
When you use the difference between timesteps there is no way to "extrapolate" the trend from previous timestep; neural network has to learn how the function actually varies
Use larger model (for the whole dataset you should try something like 300 I think), but you can simply tune that one.
Don't use flipud. Use bidirectional LSTMs, in this way you can get info from forward and backward pass of LSTM (not to confuse with backprop!). This also should boost your score
Questions
Okay, question 1: You are saying that for variable x in the time
series, I should train the model to learn x[i] - x[i-1] rather than
the value of x[i]? Am I correctly interpreting?
Yes, exactly. Difference removes the urge of the neural network to base it's predictions on the past timestep too much (by simply getting last value and maybe changing it a little)
Question 2: You said my calculations for zero bottleneck were
incorrect. But, for example, let's say I'm using a simple dense
network as an auto encoder. Getting the right bottleneck indeed
depends on the data. But if you make the bottleneck the same size as
the input, you get the identity function.
Yes, assuming that there is no non-linearity involved which makes the thing harder (see here for similar case). In case of LSTMs there are non-linearites, that's one point.
Another one is that we are accumulating timesteps into single encoder state. So essentially we would have to accumulate timesteps identities into a single hidden and cell states which is highly unlikely.
One last point, depending on the length of sequence, LSTMs are prone to forgetting some of the least relevant information (that's what they were designed to do, not only to remember everything), hence even more unlikely.
Is num_features * num_timesteps not a bottle neck of the same size as
the input, and therefore shouldn't it facilitate the model learning
the identity?
It is, but it assumes you have num_timesteps for each data point, which is rarely the case, might be here. About the identity and why it is hard to do with non-linearities for the network it was answered above.
One last point, about identity functions; if they were actually easy to learn, ResNets architectures would be unlikely to succeed. Network could converge to identity and make "small fixes" to the output without it, which is not the case.
I'm curious about the statement : "always use difference of timesteps
instead of timesteps" It seem to have some normalizing effect by
bringing all the features closer together but I don't understand why
this is key ? Having a larger model seemed to be the solution and the
substract is just helping.
Key here was, indeed, increasing model capacity. Subtraction trick depends on the data really. Let's imagine an extreme situation:
We have 100 timesteps, single feature
Initial timestep value is 10000
Other timestep values vary by 1 at most
What the neural network would do (what is the easiest here)? It would, probably, discard this 1 or smaller change as noise and just predict 1000 for all of them (especially if some regularization is in place), as being off by 1/1000 is not much.
What if we subtract? Whole neural network loss is in the [0, 1] margin for each timestep instead of [0, 1001], hence it is more severe to be wrong.
And yes, it is connected to normalization in some sense come to think about it.

Training a LSTM auto-encoder gets NaN / super high MSE loss

I'm trying to train a LSTM ae.
It's like a seq2seq model, you throw a signal in to get a reconstructed signal sequence. And the I'm using a sequence which should be quite easy. The loss function and metric is MSE. The first hundred epochs went well. However after some epochs I got MSE which is super high and it goes to NaN sometimes. I don't know what causes this.
Can you inspect the code and give me a hint?
The sequence gets normalization before, so it's in a [0,1] range, how can it produce such a high MSE error?
This is the input sequence I get from training set:
sequence1 = x_train[0][:128]
looks like this:
I get the data from a public signal dataset(128*1)
This is the code: (I modify it from keras blog)
# lstm autoencoder recreate sequence
from numpy import array
from keras.models import Sequential
from keras.layers import LSTM
from keras.layers import Dense
from keras.layers import RepeatVector
from keras.layers import TimeDistributed
from keras.utils import plot_model
from keras import regularizers
# define input sequence. sequence1 is only a one dimensional list
# reshape sequence1 input into [samples, timesteps, features]
n_in = len(sequence1)
sequence = sequence1.reshape((1, n_in, 1))
# define model
model = Sequential()
model.add(LSTM(1024, activation='relu', input_shape=(n_in,1)))
model.add(RepeatVector(n_in))
model.add(LSTM(1024, activation='relu', return_sequences=True))
model.add(TimeDistributed(Dense(1)))
model.compile(optimizer='adam', loss='mse')
for epo in [50,100,1000,2000]:
model.fit(sequence, sequence, epochs=epo)
The first few epochs went all well. all the losses are about 0.003X or so. Then it became big suddenly, to some very big number, the goes to NaN all the way up.

You might have a problem with exploding gradient values when doing the backpropagation.
Try using the clipnorm and clipvalue parameters to control gradient clipping: https://keras.io/optimizers/
Alternatively, what is the learning rate you are using? I would also try to reduce the learning rate by 10,100,1000 to check if you observe the same behavior.

'relu' is the main culprit - see here. Possible solutions:
Initialize weights to smaller values, e.g. keras.initializers.TruncatedNormal(mean=0.0, stddev=0.01)
Clip weights (at initialization, or via kernel_constraint, recurrent_constraint, ...)
Increase weight decay
Use a warmup learning rate scheme (start low, gradually increase)
Use 'selu' activation, which is more stable, is ReLU-like, and works better than ReLU on some tasks
Since your training went stable for many epochs, 3 sounds the most promising, as it seems that eventually your weights norm gets too large and gradients explode. Generally, I suggest keeping the weight norms around 1 for 'relu'; you can monitor the l2 norms using the function below. I also recommend See RNN for inspecting layer activations & gradients.
def inspect_weights_l2(model, names='lstm', axis=-1):
def _get_l2(w, axis=-1):
axis = axis if axis != -1 else len(w.shape) - 1
reduction_axes = tuple([ax for ax in range(len(w.shape)) if ax != axis])
return np.sqrt(np.sum(np.square(w), axis=reduction_axes))
def _print_layer_l2(layer, idx, axis=-1):
W = layer.get_weights()
l2_all = []
txt = "{} "
for w in W:
txt += "{:.4f}, {:.4f} -- "
l2 = _get_l2(w, axis)
l2_all.extend([l2.max(), l2.mean()])
txt = txt.rstrip(" -- ")
print(txt.format(idx, *l2_all))
names = [names] if isinstance(names, str) else names
for idx, layer in enumerate(model.layers):
if any([name in layer.name.lower() for name in names]):
_print_layer_l2(layer, idx, axis=axis)

How to handle variable sized input in CNN with Keras?

I am trying to perform the usual classification on the MNIST database but with randomly cropped digits.
Images are cropped the following way : removed randomly first/last and/or row/column.
I would like to use a Convolutional Neural Network using Keras (and Tensorflow backend) to perform convolution and then the usual classification.
Inputs are of variable size and i can't manage to get it to work.
Here is how I cropped digits
import numpy as np
from keras.utils import to_categorical
from sklearn.datasets import load_digits
digits = load_digits()
X = digits.images
X = np.expand_dims(X, axis=3)
X_crop = list()
for index in range(len(X)):
X_crop.append(X[index, np.random.randint(0,2):np.random.randint(7,9), np.random.randint(0,2):np.random.randint(7,9), :])
X_crop = np.array(X_crop)
y = to_categorical(digits.target)
from sklearn.model_selection import train_test_split
X_train, X_test, y_train, y_test = train_test_split(X_crop, y, train_size=0.8, test_size=0.2)
And here is the architecture of the model I want to use
from keras.layers import Dense, Dropout
from keras.layers.convolutional import Conv2D
from keras.models import Sequential
model = Sequential()
model.add(Conv2D(filters=10,
kernel_size=(3,3),
input_shape=(None, None, 1),
data_format='channels_last'))
model.add(Dense(128, activation='relu'))
model.add(Dropout(0.2))
model.add(Dense(10, activation='softmax'))
model.compile(loss='categorical_crossentropy', optimizer='sgd', metrics=['accuracy'])
model.summary()
model.fit(X_train, y_train, epochs=100, batch_size=16, validation_data=(X_test, y_test))
Does someone have an idea on how to handle variable sized input in my neural network?
And how to perform classification?

TL/DR - go to point 4
So - before we get to the point - let's fix some problems with your network:
Your network will not work because of activation: with categorical_crossentropy you need to have a softmax activation:
model.add(Dense(10, activation='softmax'))
Vectorize spatial tensors: as Daniel mentioned - you need to, at some stage, switch your vectors from spatial (images) to vectorized (vectors). Currently - applying Dense to output from a Conv2D is equivalent to (1, 1) convolution. So basically - output from your network is spatial - not vectorized what causes dimensionality mismatch (you can check that by running your network or checking the model.summary(). In order to change that you need to use either GlobalMaxPooling2D or GlobalAveragePooling2D. E.g.:
model.add(Conv2D(filters=10,
kernel_size=(3, 3),
input_shape=(None, None, 1),
padding="same",
data_format='channels_last'))
model.add(GlobalMaxPooling2D())
model.add(Dense(128, activation='relu'))
model.add(Dropout(0.2))
model.add(Dense(10, activation='softmax'))
Concatenated numpy arrays need to have the same shape: if you check the shape of X_crop you'll see that it's not a spatial matrix. It's because you concatenated matrices with different shapes. Sadly - it's impossible to overcome this issue as numpy.array need to have a fixed shape.
How to make your network train on examples of different shape: The most important thing in doing this is to understand two things. First - is that in a single batch every image should have the same size. Second - is that calling fit multiple times is a bad idea - as you reset inner model states. So here is what needs to be done:
a. Write a function which crops a single batch - e.g. a get_cropped_batches_generator which given a matrix cuts a batch out of it and crops it randomly.
b. Use train_on_batch method. Here is an example code:
from six import next
batches_generator = get_cropped_batches_generator(X, batch_size=16)
losses = list()
for epoch_nb in range(nb_of_epochs):
epoch_losses = list()
for batch_nb in range(nb_of_batches):
# cropped_x has a different shape for different batches (in general)
cropped_x, cropped_y = next(batches_generator)
current_loss = model.train_on_batch(cropped_x, cropped_y)
epoch_losses.append(current_loss)
losses.append(epoch_losses.sum() / (1.0 * len(epoch_losses))
final_loss = losses.sum() / (1.0 * len(losses))
So - a few comments to code above: First, train_on_batch doesn't use nice keras progress bar. It returns a single loss value (for a given batch) - that's why I added logic to compute loss. You could use Progbar callback for that also. Second - you need to implement get_cropped_batches_generator - I haven't written a code to keep my answer a little bit more clear. You could ask another question on how to implement it. Last thing - I use six to keep compatibility between Python 2 and Python 3.

Usually, a model containing Dense layers cannot have variable size inputs, unless the outputs are also variable. But see the workaround and also the other answer using GlobalMaxPooling2D - The workaround is equivalent to GlobalAveragePooling2D. These are layers that can eliminiate the variable size before a Dense layer and suppress the spatial dimensions.
For an image classification case, you may want to resize the images outside the model.
When my images are in numpy format, I resize them like this:
from PIL import Image
im = Image.fromarray(imgNumpy)
im = im.resize(newSize,Image.LANCZOS) #you can use options other than LANCZOS as well
imgNumpy = np.asarray(im)
Why?
A convolutional layer has its weights as filters. There is a static filter size, and the same filter is applied to the image over and over.
But a dense layer has its weights based on the input. If there is 1 input, there is a set of weights. If there are 2 inputs, you've got twice as much weights. But weights must be trained, and changing the amount of weights will definitely change the result of the model.
As #Marcin commented, what I've said is true when your input shape for Dense layers has two dimensions: (batchSize,inputFeatures).
But actually keras dense layers can accept inputs with more dimensions. These additional dimensions (which come out of the convolutional layers) can vary in size. But this would make the output of these dense layers also variable in size.
Nonetheless, at the end you will need a fixed size for classification: 10 classes and that's it. For reducing the dimensions, people often use Flatten layers, and the error will appear here.
A possible fishy workaround (not tested):
At the end of the convolutional part of the model, use a lambda layer to condense all the values in a fixed size tensor, probably taking a mean of the side dimensions and keeping the channels (channels are not variable)
Suppose the last convolutional layer is:
model.add(Conv2D(filters,kernel_size,...))
#so its output shape is (None,None,None,filters) = (batchSize,side1,side2,filters)
Let's add a lambda layer to condense the spatial dimensions and keep only the filters dimension:
import keras.backend as K
def collapseSides(x):
axis=1 #if you're using the channels_last format (default)
axis=-1 #if you're using the channels_first format
#x has shape (batchSize, side1, side2, filters)
step1 = K.mean(x,axis=axis) #mean of side1
return K.mean(step1,axis=axis) #mean of side2
#this will result in a tensor shape of (batchSize,filters)
Since the amount of filters is fixed (you have kicked out the None dimensions), the dense layers should probably work:
model.add(Lambda(collapseSides,output_shape=(filters,)))
model.add(Dense.......)
.....
In order for this to possibly work, I suggest that the number of filters in the last convolutional layer be at least 10.
With this, you can make input_shape=(None,None,1)
If you're doing this, remember that you can only pass input data with a fixed size per batch. So you have to separate your entire data in smaller batches, each batch having images all of the same size. See here: Keras misinterprets training data shape