I am learning about RNNs, and am trying to code one up using PyTorch.
I have some trouble understanding the output dimensions
Here is some code for a simple RNN architecture
class RNN(nn.Module):
def __init__(self, input_size, hidden_dim, n_layers):
super(RNN, self).__init__()
self.hidden_dim=hidden_dim
self.rnn = nn.RNN(input_size, hidden_dim, n_layers, batch_first=True)
def forward(self, x, hidden):
r_out, hidden = self.rnn(x, hidden)
return r_out, hidden
So, what I understand is the hidden_dim is the number blocks I will have in my hidden layer, and essentially the number of features in the output and in the hidden state.
I create some dummy data, to test it
test_rnn = RNN(input_size=1, hidden_dim=4, n_layers=1)
# generate evenly spaced, test data pts
time_steps = np.linspace(0, 6, 3)
data = np.sin(time_steps)
data.resize((3, 1))
test_input = torch.Tensor(data).unsqueeze(0) # give it a batch_size of 1 as first dimension
print('Input size: ', test_input.size())
# test out rnn sizes
test_out, test_h = test_rnn(test_input, None)
print('Hidden state size: ', test_h.size())
print('Output size: ', test_out.size())
What I get is
Input size: torch.Size([1, 3, 1])
Hidden state size: torch.Size([1, 1, 4])
Output size: torch.Size([1, 3, 4])
So I understand that the shape of x is determined like so
x = (batch_size, seq_length, input_size).. so 1 bath size, and input of 1 feature and 3 time steps(sequence length).
For hidden state, like so hidden = (n_layers, batch_size, hidden_dim).. so I had 1 layer, batch size 1, and 4 blocks in my hidden layer.
What I don't get is the RNN output. From the documentation, r_out = (batch_size, time_step, hidden_size).. Wasn't the output supposed to be the same as the hidden state that was output from the hidden units? That is, if I have 4 units in my hidden layer, I would expect it to output 4 numbers for the hidden state, and 4 numbers for the output. Why is the time step a dimension of the output? Because, each hidden unit, takes in some numbers, outputs a state S and output Y, and both of these are equal, yes? I tried a diagram, this is what I came up with. Help me understand what part of it I'm doing wrong.
So TL;DR
Why does output shape in a simple Elman RNN depend on the sequence length(while hidden state shape doesn't)? For in the diagram I drew, I see both of them being the same.
In the PyTorch API, the output is a sequence of hidden states during the RNN computation, i.e., there is one hidden state vector per input vector. The hidden state is the last hidden state, the state the RNN ends with after processing the input, so test_out[:, -1, :] = test_h.
Vector y in your diagrams is the same as a hidden state Ht, it indeed has 4 numbers, but the state is different for every time step, so you have 4 number for every time step.
The reason why PyTorch separates the sequence of outputs = hidden states (it's not the same in LSTMs, though) is that you can have a batch of sequences of different lengths. In that case, the final state is not simply test_out[:, -1, :], because you need to select final states based on the lengths of individual sequences.
Related
This seems to be one of the most common questions about LSTMs in PyTorch, but I am still unable to figure out what should be the input shape to PyTorch LSTM.
Even after following several posts (1, 2, 3) and trying out the solutions, it doesn't seem to work.
Background: I have encoded text sequences (variable length) in a batch of size 12 and the sequences are padded and packed using pad_packed_sequence functionality. MAX_LEN for each sequence is 384 and each token (or word) in the sequence has a dimension of 768. Hence my batch tensor could have one of the following shapes: [12, 384, 768] or [384, 12, 768].
The batch will be my input to the PyTorch rnn module (lstm here).
According to the PyTorch documentation for LSTMs, its input dimensions are (seq_len, batch, input_size) which I understand as following.
seq_len - the number of time steps in each input stream (feature vector length).
batch - the size of each batch of input sequences.
input_size - the dimension for each input token or time step.
lstm = nn.LSTM(input_size=?, hidden_size=?, batch_first=True)
What should be the exact input_size and hidden_size values here?
You have explained the structure of your input, but you haven't made the connection between your input dimensions and the LSTM's expected input dimensions.
Let's break down your input (assigning names to the dimensions):
batch_size: 12
seq_len: 384
input_size / num_features: 768
That means the input_size of the LSTM needs to be 768.
The hidden_size is not dependent on your input, but rather how many features the LSTM should create, which is then used for the hidden state as well as the output, since that is the last hidden state. You have to decide how many features you want to use for the LSTM.
Finally, for the input shape, setting batch_first=True requires the input to have the shape [batch_size, seq_len, input_size], in your case that would be [12, 384, 768].
import torch
import torch.nn as nn
# Size: [batch_size, seq_len, input_size]
input = torch.randn(12, 384, 768)
lstm = nn.LSTM(input_size=768, hidden_size=512, batch_first=True)
output, _ = lstm(input)
output.size() # => torch.Size([12, 384, 512])
The image passed to CNN layer and lstm layer,the feature map shape changes like this
BCHW->BCHW(BxCx1xW),
the CNN's output shape should has the height 1.
then sqeeze the dim of height.
BCHW->BCW
in rnn ,shape name changes,[batch ,seqlen,input_size],in image,[batch,width,channel],
**BCW->BWC,**this is batch_first tensor for LSTM layer(like pytorch).
Finally:
BWC is [batch,seqlen,channel].
I came across several examples of classifying MNIST digit using a RNN, what it the reason to initialize the hidden state with a sequence_length=1? If you were doing 1 step ahead prediction of a video frame prediction, how would you initialize it?
def init_hidden(self, x, device=None): # input 4D tensor: (batch size, channels, width, height)
# initialize the hidden and cell state to zero
# vectors:(number of layer, sequence length, number of hidden nodes)
if (self.bidirectional):
h0 = torch.zeros(2*self.n_layers, 1, self.n_hidden)
else:
h0 = torch.zeros(self.n_layers, 1, self.n_hidden)
if device is not None:
h0 = h0.to(device)
self.hidden = h0
The input is usually represented as
inputs = inputs.view(batch_size*image_height, 1, image_width)
In this above example are the images passed columns-wise? Is there another way to represent inputs images in RNN? And how does it related to how one initialize the hidden state?
When initializing the hidden state, the second dimension is actually not the sequence-length, it is the batch size:
hidden = torch.zeros(layers, batch_size, hidden_nodes)
For the MNIST rnn I would say that the input shape is 28x1 (shape of one row) and the sequence-length is also 28 (there are 28 rows).
input_size = 28
hidden_nodes = 128 # for example
layers = 2 # for example
dropout = 0.35
rnn = nn.RNN(input_size=input_size, hidden_size=hidden_nodes, num_layers=layers, dropout=dropout, batch_first=True)
Now lets init the hidden-state:
hidden = torch.zeros(layers, batch_size, hidden_nodes)
You dont have to tell the hidden-state how long the sequence is and also not how long an element of the sequence is. Just how big the hidden-layer should be.
So as you can the the sequence length for mnist can`t be 1, it has to be 28 since there are 28 rows. An RNN with sequence-size 1 makes no sense, because a sequence is only a sequence if it has more than 1 element.
Edit to answer question in the comments:
It would be (batch_size, 28, 28). Just the way you would pass an image to a cnn just without the channel dimension. The first 28 stands for the sequence length. The second 28 for how long one sequence is.
Maybe another example makes it more clear: If you would have an RNN which takes (for what ever reason) 4 letters as input and every letter is one-hot encoded (so the letter a for example would be a vector of length 26, length of the alphabet, where every element is zero but the first one is 1) the input dimension would look like this: (batch_size, 4, 26), batch_size, sequence-length is 4 (4 letters) and every element/letter in the sequence has length of 28 (one-hot encoded alphabet).
I started working with Pytorch recently so my understanding of it isn't quite strong. I previously had a 1 layer CNN but wanted to extend it to 2 layers, but the input and output channels have been throwing errors I can seem to decipher. Why does it expect 192 channels? Can someone give me a pointer to help me understand this better? I have seen several related problems on here, but I don't understand those solutions either.
import numpy as np
import pandas as pd
import torch
import torch.nn as nn
from transformers import BertConfig, BertModel, BertTokenizer
import math
from transformers import AdamW, get_linear_schedule_with_warmup
def pad_sents(sents, pad_token): # Pad list of sentences according to the longest sentence in the batch.
sents_padded = []
max_len = max(len(s) for s in sents)
for s in sents:
padded = [pad_token] * max_len
padded[:len(s)] = s
sents_padded.append(padded)
return sents_padded
def sents_to_tensor(tokenizer, sents, device):
tokens_list = [tokenizer.tokenize(str(sent)) for sent in sents]
sents_lengths = [len(tokens) for tokens in tokens_list]
tokens_list_padded = pad_sents(tokens_list, '[PAD]')
sents_lengths = torch.tensor(sents_lengths, device=device)
masks = []
for tokens in tokens_list_padded:
mask = [0 if token == '[PAD]' else 1 for token in tokens]
masks.append(mask)
masks_tensor = torch.tensor(masks, dtype=torch.long, device=device)
tokens_id_list = [tokenizer.convert_tokens_to_ids(tokens) for tokens in tokens_list_padded]
sents_tensor = torch.tensor(tokens_id_list, dtype=torch.long, device=device)
return sents_tensor, masks_tensor, sents_lengths
class ConvModel(nn.Module):
def __init__(self, device, dropout_rate, n_class, out_channel=16):
super(ConvModel, self).__init__()
self.bert_config = BertConfig.from_pretrained('bert-base-uncased', output_hidden_states=True)
self.dropout_rate = dropout_rate
self.n_class = n_class
self.out_channel = out_channel
self.bert = BertModel.from_pretrained('bert-base-uncased', config=self.bert_config)
self.out_channels = self.bert.config.num_hidden_layers * self.out_channel
self.tokenizer = BertTokenizer.from_pretrained('bert-base-uncased', config=self.bert_config)
self.conv = nn.Conv2d(in_channels=self.bert.config.num_hidden_layers,
out_channels=self.out_channels,
kernel_size=(3, self.bert.config.hidden_size),
groups=self.bert.config.num_hidden_layers)
self.conv1 = nn.Conv2d(in_channels=self.out_channels,
out_channels=48,
kernel_size=(3, self.bert.config.hidden_size),
groups=self.bert.config.num_hidden_layers)
self.hidden_to_softmax = nn.Linear(self.out_channels, self.n_class, bias=True)
self.dropout = nn.Dropout(p=self.dropout_rate)
self.device = device
def forward(self, sents):
sents_tensor, masks_tensor, sents_lengths = sents_to_tensor(self.tokenizer, sents, self.device)
encoded_layers = self.bert(input_ids=sents_tensor, attention_mask=masks_tensor)
hidden_encoded_layer = encoded_layers[2]
hidden_encoded_layer = hidden_encoded_layer[0]
hidden_encoded_layer = torch.unsqueeze(hidden_encoded_layer, dim=1)
hidden_encoded_layer = hidden_encoded_layer.repeat(1, 12, 1, 1)
conv_out = self.conv(hidden_encoded_layer) # (batch_size, channel_out, some_length, 1)
conv_out = self.conv1(conv_out)
conv_out = torch.squeeze(conv_out, dim=3) # (batch_size, channel_out, some_length)
conv_out, _ = torch.max(conv_out, dim=2) # (batch_size, channel_out)
pre_softmax = self.hidden_to_softmax(conv_out)
return pre_softmax
def batch_iter(data, batch_size, shuffle=False, bert=None):
batch_num = math.ceil(data.shape[0] / batch_size)
index_array = list(range(data.shape[0]))
if shuffle:
data = data.sample(frac=1)
for i in range(batch_num):
indices = index_array[i * batch_size: (i + 1) * batch_size]
examples = data.iloc[indices]
sents = list(examples.train_BERT_tweet)
targets = list(examples.train_label.values)
yield sents, targets # list[list[str]] if not bert else list[str], list[int]
def train():
label_name = ['Yes', 'Maybe', 'No']
device = torch.device("cpu")
df_train = pd.read_csv('trainn.csv') # , index_col=0)
train_label = dict(df_train.train_label.value_counts())
label_max = float(max(train_label.values()))
train_label_weight = torch.tensor([label_max / train_label[i] for i in range(len(train_label))], device=device)
model = ConvModel(device=device, dropout_rate=0.2, n_class=len(label_name))
optimizer = AdamW(model.parameters(), lr=1e-3, correct_bias=False)
scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=100, num_training_steps=1000) # changed the last 2 arguments to old ones
model = model.to(device)
model.train()
cn_loss = torch.nn.CrossEntropyLoss(weight=train_label_weight, reduction='mean')
train_batch_size = 16
for epoch in range(1):
for sents, targets in batch_iter(df_train, batch_size=train_batch_size, shuffle=True): # for each epoch
optimizer.zero_grad()
pre_softmax = model(sents)
loss = cn_loss(pre_softmax, torch.tensor(targets, dtype=torch.long, device=device))
loss.backward()
optimizer.step()
scheduler.step()
TrainingModel = train()
Here's a snippet of data https://github.com/Kosisochi/DataSnippet
It seems that the original version of the code you had in this question behaved differently. The final version of the code you have here gives me a different error from what you posted, more specifically - this:
RuntimeError: Calculated padded input size per channel: (20 x 1). Kernel size: (3 x 768). Kernel size can't be greater than actual input size
I apologize if I misunderstood the situation, but it seems to me that your understanding of what exactly nn.Conv2d layer does is not 100% clear and that is the main source of your struggle. I interpret the part "detailed explanation on 2 layer CNN in Pytorch" you requested as an ask to explain in detail on how that layer works and I hope that after this is done there will be no problem applying it 1 time, 2 times or more.
You can find all the documentation about the layer here, but let me give you a recap which hopefully will help to understand more the errors you're getting.
First of all nn.Conv2d inputs are 4-d tensors of the shape (BatchSize, ChannelsIn, Height, Width) and outputs are 4-d tensors of the shape (BatchSize, ChannelsOut, HeightOut, WidthOut). The simplest way to think about nn.Conv2d is of something applied to 2d images with pixel grid of size Height x Width and having ChannelsIn different colors or features per pixel. Even if your inputs have nothing to do with actual images the behavior of the layer is still the same. Simplest situation is when the nn.Conv2d is not using padding (as in your code). In that case the kernel_size=(kernel_height, kernel_width) argument specifies the rectangle which you can imagine sweeping through Height x Width rectangle of your inputs and producing one pixel for each valid position. Without padding the coordinate of the rectangle's point can be any pair of indicies (x, y) with x between 0 and Height - kernel_height and y between 0 and Width - kernel_width. Thus the output will look like a 2d image of size (Height - kernel_height + 1) x (Width - kernel_width + 1) and will have as many output channels as specified to nn.Conv2d constructor, so the output tensor will be of shape (BatchSize, ChannelsOut, Height - kernel_height + 1, Width - kernel_width + 1).
The parameter groups is not affecting how shapes are changed by the layer - it is only controlling which input channels are used as inputs for the output channels (groups=1 means that every input channel is used as input for every output channel, otherwise input and output channels are divided into corresponding number of groups and only input channels from group i are used as inputs for the output channels from group i).
Now in your current version of the code you have BatchSize = 16 and the output of pre-trained model is (BatchSize, DynamicSize, 768) with DynamicSize depending on the input, e.g. 22. You then introduce additional dimension as axis 1 with unsqueeze and repeat the values along that dimension transforming the tensor of shape (16, 22, 768) into (16, 12, 22, 768). Effectively you are using the output of the pre-trained model as 12-channel (with each channel having same values as others) 2-d images here of size (22, 768), where 22 is not fixed (depends on the batch). Then you apply a nn.Conv2d with kernel size (3, 768) - which means that there is no "wiggle room" for width and output 2-d images will be of size (20, 1) and since your layer has 192 channels final size of the output of first convolution layer has shape (16, 192, 20, 1). Then you try to apply second layer of convolution on top of that with kernel size (3, 768) again, but since your 2-d "image" is now just (20 x 1) there is no valid position to fit (3, 768) kernel rectangle inside a rectangle (20 x 1) which leads to the error message Kernel size can't be greater than actual input size.
Hope this explanation helps. Now to the choices you have to avoid the issue:
(a) is to add padding in such a way that the size of the output is not changing comparing to input (I won't go into details here,
because I don't think this is what you need)
(b) Use smaller kernel on both first and/or second convolutions (e.g. if you don't change first convolution the only valid width for
the second kernel would be 1).
(c) Looking at what you're trying to do my guess is that you actually don't want to use 2d convolution, you want 1d convolution (on the sequence) with every position described by 768 values. When you're using one convolution layer with 768 width kernel (and same 768 width input) you're effectively doing exactly same thing as 1d convolution with 768 input channels, but then if you try to apply second one you have a problem. You can specify kernel width as 1 for the next layer(s) and that will work for you, but a more correct way would be to transpose pre-trained model's output tensor by switching the last dimensions - getting shape (16, 768, DynamicSize) from (16, DynamicSize, 768) and then apply nn.Conv1d layer with 768 input channels and arbitrary ChannelsOut as output channels and 1d kernel_size=3 (meaning you look at 3 consecutive elements of the sequence for convolution). If you do that than without padding input shape of (16, 768, DynamicSize) will become (16, ChannelsOut, DynamicSize-2), and after you apply second Conv1d with e.g. the same settings as first one you'll get a tensor of shape (16, ChannelsOut, DynamicSize-4), etc. (each time the 1d length will shrink by kernel_size-1). You can always change number of channels/kernel_size for each subsequent convolution layer too.
What is the difference between LSTM and LSTMCell in Pytorch (currently version 1.1)? It seems that LSTMCell is a special case of LSTM (i.e. with only one layer, unidirectional, no dropout).
Then, what's the purpose of having both implementations? Unless I'm missing something, it's trivial to use an LSTM object as an LSTMCell (or alternatively, it's pretty easy to use multiple LSTMCells to create the LSTM object)
Yes, you can emulate one by another, the reason for having them separate is efficiency.
LSTMCell is a cell that takes arguments:
Input of shape batch × input dimension;
A tuple of LSTM hidden states of shape batch x hidden dimensions.
It is a straightforward implementation of the equations.
LSTM is a layer applying an LSTM cell (or multiple LSTM cells) in a "for loop", but the loop is heavily optimized using cuDNN. Its input is
A three-dimensional tensor of inputs of shape batch × input length × input dimension;
Optionally, an initial state of the LSTM, i.e., a tuple of hidden states of shape batch × hidden dim (or tuple of such tuples if the LSTM is bidirectional)
You often might want to use the LSTM cell in a different context than apply it over a sequence, i.e. make an LSTM that operates over a tree-like structure. When you write a decoder in sequence-to-sequence models, you also call the cell in a loop and stop the loop when the end-of-sequence symbol is decoded.
Let me show some specific examples:
# LSTM example:
>>> rnn = nn.LSTM(10, 20, 2)
>>> input = torch.randn(5, 3, 10)
>>> h0 = torch.randn(2, 3, 20)
>>> c0 = torch.randn(2, 3, 20)
>>> output, (hn, cn) = rnn(input, (h0, c0))
# LSTMCell example:
>>> rnn = nn.LSTMCell(10, 20)
>>> input = torch.randn(3, 10)
>>> hx = torch.randn(3, 20)
>>> cx = torch.randn(3, 20)
>>> output = []
>>> for i in range(6):
hx, cx = rnn(input[i], (hx, cx))
output.append(hx)
The key difference:
LSTM: the argument 2, stands num_layers, number of recurrent layers. There are seq_len * num_layers=5 * 2 cells. No loop but more cells.
LSTMCell: in for loop (seq_len=5 times), each output of ith instance will be input of (i+1)th instance. There is only one cell, Truly Recurrent
If we set num_layers=1 in LSTM or add one more LSTMCell, the codes above will be the same.
Obviously, It is easier to apply parallel computing in LSTM.
I was just modifying some an LSTM network I had written to print out the test error. The issues, I realized, is that the model I had defined depends on the batch size.
Specifically, the input is a tensor of shape [batch_size, time_steps, features]. The input enters the LSTM cell and the output, which I turn into a list of time_steps 2D tensors, with each 2D tensor having shape [batch_size, hidden_units]. Each 2D tensor is then multiplied by a weight vector of shape [hidden_units] to yield a vector of shape [batch_size] which has added to it a bias vector of shape [batch_size].
In words, I give the model N sequences, and I expect it to output a scalar for each time step for each sequence. That is, the output is a list of N vectors, one for each time step.
For training, I give the model batches of size 13. For the test data, I feed the entire data set, which consists of over 400 examples. Thus, an error is raised, since the bias has fixed shape batch_size.
I haven't found a way to make it's shape variable without raising an error.
I can add complete code if requested. Added code anyways.
Thanks.
def basic_lstm(inputs, number_steps, number_features, number_hidden_units, batch_size):
weights = {
'out': tf.Variable(tf.random_normal([number_hidden_units, 1]))
}
biases = {
'out': tf.Variable(tf.constant(0.1, shape=[batch_size, 1]))
}
lstm_cell = rnn.BasicLSTMCell(number_hidden_units)
init_state = lstm_cell.zero_state(batch_size, dtype=tf.float32)
hidden_layer_outputs, states = tf.nn.dynamic_rnn(lstm_cell, inputs,
initial_state=init_state, dtype=tf.float32)
results = tf.squeeze(tf.stack([tf.matmul(output, weights['out'])
+ biases['out'] for output
in tf.unstack(tf.transpose(hidden_layer_outputs, (1, 0, 2)))], axis=1))
return results
You want the biases to be a shape of (batch_size, )
For example (using zeros instead of tf.constant but similar problem), I was able to specify the shape as a single integer:
biases = tf.Variable(tf.zeros(10,dtype=tf.float32))
print(biases.shape)
prints:
(10,)