prakashpandey9 / text-classification-pytorch Goto Github PK
View Code? Open in Web Editor NEWText classification using deep learning models in Pytorch
License: MIT License
Text classification using deep learning models in Pytorch
License: MIT License
Hi! Your code is of great help to me and thanks to you so much.
When experimenting with the same test data of CNN in your repo, I compare the result with TF-IDF+LR I wrote and found out that the test data is not totally used in testing. And I write a version like
with torch.no_grad(): for idx, batch in enumerate(val_iter): text = batch.Text[0] pass_batch_size= None #Since the batch size would be passed into the model, I set the variance to keep the data size in this iteration if (text.size()[0] is not batch_size): pass_batch_size = len(text) # pass_batch_size is the batch size of the batch target = batch.Label target = torch.autograd.Variable(target).long() if torch.cuda.is_available(): # text = text.cuda() # target = target.cuda() text =text.to(device) target = target.to(device) prediction = model(text,pass_batch_size) $ pass into the model
In LSTM.py, the model initializes the h_0 and c_0. I think it's the reason you define a parameter 'batch_size=None' in 'forward()'.
Running the main.py
got the error:
C:\Python36\python.exe I:/github_repos/Text-Classification-Pytorch/main.py
Traceback (most recent call last):
File "I:/github_repos/Text-Classification-Pytorch/main.py", line 11, in <module>
TEXT, vocab_size, word_embeddings, train_iter, valid_iter, test_iter = load_data.load_dataset()
File "I:\github_repos\Text-Classification-Pytorch\load_data.py", line 31, in load_dataset
LABEL = data.LabelField(tensor_type=torch.FloatTensor)
File "C:\Python36\lib\site-packages\torchtext\data\field.py", line 699, in __init__
super(LabelField, self).__init__(**kwargs)
TypeError: __init__() got an unexpected keyword argument 'tensor_type'
Process finished with exit code 1
If it is because that it is a bi-RNN, then shouldn't it be 2 times the hidden_size only?
Thanks
I will complete a Chinese emotion analysis task. Can this project complete the task of Chinese emotion analysis?
Hello, If I want to load separate external CSV files for training and testing where the first column of the CSV contains sentences and 2nd column contains the labels then where should I change?
On lines 63 and 73 of Text-Classification-Pytorch/models/RCNN.py
, function permute
is used.
input = input.permute(1, 0, 2) # input.size() = (num_sequences, batch_size, embedding_length)
...
y = y.permute(0, 2, 1) # y.size() = (batch_size, hidden_size, num_sequences)
Could you please explain why it is necessary or useful to permute the dimensions of these tensors?
I find that "main.py" only support "LSTM". I want to know how to use other models to make text classification,such as RNN,CNN,RCNN,attention,self-attention.I find other models need parameter "weight" which is not found in "main.py".How to set the "weight" parameter to call the other models?
Thanks a lot!
can you make an example of how to use the cnn model, like you did for lstm?
i have trouble understanding the parameters required to get the cnn up and running.
thanks in advance!
Hi,
why are you using Conv2d module?
The author in the paper uses to 1D convolutions.
Thanks,
Pablo
Would you please add the reference for the implementation details of the attention layer?
Running torchtext.SST got the error:
/pytorch/aten/src/THCUNN/ClassNLLCriterion.cu:105: void cunn_ClassNLLCriterion_updateOutput_kernel(Dtype *, Dtype *, Dtype *, long *, Dtype *, int, int, int, int, long) [with Dtype = float, Acctype = float]: block: [0,0,0], thread: [0,0,0] Assertion
t >= 0 && t < n_classesfailed. /pytorch/aten/src/THCUNN/ClassNLLCriterion.cu:105: void cunn_ClassNLLCriterion_updateOutput_kernel(Dtype *, Dtype *, Dtype *, long *, Dtype *, int, int, int, int, long) [with Dtype = float, Acctype = float]: block: [0,0,0], thread: [11,0,0] Assertion
t >= 0 && t < n_classesfailed. /pytorch/aten/src/THCUNN/ClassNLLCriterion.cu:105: void cunn_ClassNLLCriterion_updateOutput_kernel(Dtype *, Dtype *, Dtype *, long *, Dtype *, int, int, int, int, long) [with Dtype = float, Acctype = float]: block: [0,0,0], thread: [19,0,0] Assertion
t >= 0 && t < n_classesfailed. /pytorch/aten/src/THCUNN/ClassNLLCriterion.cu:105: void cunn_ClassNLLCriterion_updateOutput_kernel(Dtype *, Dtype *, Dtype *, long *, Dtype *, int, int, int, int, long) [with Dtype = float, Acctype = float]: block: [0,0,0], thread: [20,0,0] Assertion
t >= 0 && t < n_classesfailed. /pytorch/aten/src/THCUNN/ClassNLLCriterion.cu:105: void cunn_ClassNLLCriterion_updateOutput_kernel(Dtype *, Dtype *, Dtype *, long *, Dtype *, int, int, int, int, long) [with Dtype = float, Acctype = float]: block: [0,0,0], thread: [26,0,0] Assertion
t >= 0 && t < n_classes failed. Traceback (most recent call last): File "./main.py", line 91, in <module> train_loss, train_acc = train_model(model, train_iter, epoch) File "./main.py", line 44, in train_model loss.backward() File "....", line 118, in backward torch.autograd.backward(self, gradient, retain_graph, create_graph) File "....", line 93, in backward allow_unreachable=True) # allow_unreachable flag RuntimeError: cuda runtime error (59) : device-side assert triggered at /pytorch/aten/src/THC/generic/THCTensorMath.cu:26
I checked the target values:
tensor([2, 1, 1, 0, 0, 1, 1, 0, 0, 1, 1, 2, 1, 0, 0, 0, 0, 0, 1, 2, 2, 0, 1, 1, 0, 1, 2, 1, 1, 0, 1, 1], device='cuda:0')
there is "2" here that I think makes the problem. Would you please guide me regarding this issue?
Please help
Hi
I am trying to use the models u implemented with bert embedding for Arabic language but I am getting very low accuracy. So I am wondering if I am doing things wrong especially that I al newby to deep learning
here is my modification to the attention model
class SelfAttention(nn.Module):
def __init__(self, batch_size, output_size, hidden_size, bert):
super(SelfAttention, self).__init__()
"""
Arguments
---------
batch_size : Size of the batch which is same as the batch_size of the data returned by the TorchText BucketIterator
output_size : 2 = (pos, neg)
hidden_sie : Size of the hidden_state of the LSTM
vocab_size : Size of the vocabulary containing unique words
embedding_length : Embeddding dimension of GloVe word embeddings
weights : Pre-trained GloVe word_embeddings which we will use to create our word_embedding look-up table
--------
"""
self.batch_size = batch_size
self.output_size = output_size
self.hidden_size = hidden_size
#self.vocab_size = vocab_size
self.bert = bert
embedding_length=bert.config.to_dict()['hidden_size']
print(embedding_length)
#self.word_embeddings = nn.Embedding(vocab_size, embedding_dim)
#self.word_embeddings.weights = nn.Parameter(weights, requires_grad=False)
self.dropout = 0.8
self.bilstm = nn.LSTM(embedding_length, hidden_size, dropout=self.dropout, bidirectional=True)
# We will use da = 350, r = 30 & penalization_coeff = 1 as per given in the self-attention original ICLR paper
self.W_s1 = nn.Linear(2*hidden_size, 350)
self.W_s2 = nn.Linear(350, 30)
self.fc_layer = nn.Linear(30*2*hidden_size, 2000)
self.label = nn.Linear(2000, output_size)
def attention_net(self, lstm_output):
"""
Now we will use self attention mechanism to produce a matrix embedding of the input sentence in which every row represents an
encoding of the inout sentence but giving an attention to a specific part of the sentence. We will use 30 such embedding of
the input sentence and then finally we will concatenate all the 30 sentence embedding vectors and connect it to a fully
connected layer of size 2000 which will be connected to the output layer of size 2 returning logits for our two classes i.e.,
pos & neg.
Arguments
---------
lstm_output = A tensor containing hidden states corresponding to each time step of the LSTM network.
---------
Returns : Final Attention weight matrix for all the 30 different sentence embedding in which each of 30 embeddings give
attention to different parts of the input sentence.
Tensor size : lstm_output.size() = (batch_size, num_seq, 2*hidden_size)
attn_weight_matrix.size() = (batch_size, 30, num_seq)
"""
attn_weight_matrix = self.W_s2(F.tanh(self.W_s1(lstm_output)))
attn_weight_matrix = attn_weight_matrix.permute(0, 2, 1)
attn_weight_matrix = F.softmax(attn_weight_matrix, dim=2)
return attn_weight_matrix
def forward(self, input_sentences, batch_size=None):
"""
Parameters
----------
input_sentence: input_sentence of shape = (batch_size, num_sequences)
batch_size : default = None. Used only for prediction on a single sentence after training (batch_size = 1)
Returns
-------
Output of the linear layer containing logits for pos & neg class.
"""
with torch.no_grad():
input = self.bert(input_sentences)[0]
input = input.permute(1, 0, 2)
if batch_size is None:
h_0 = Variable(torch.zeros(2, self.batch_size, self.hidden_size).cuda())
c_0 = Variable(torch.zeros(2, self.batch_size, self.hidden_size).cuda())
else:
h_0 = Variable(torch.zeros(2, batch_size, self.hidden_size).cuda())
c_0 = Variable(torch.zeros(2, batch_size, self.hidden_size).cuda())
output, (h_n, c_n) = self.bilstm(input, (h_0, c_0))
output = output.permute(1, 0, 2)
# output.size() = (batch_size, num_seq, 2*hidden_size)
# h_n.size() = (1, batch_size, hidden_size)
# c_n.size() = (1, batch_size, hidden_size)
attn_weight_matrix = self.attention_net(output)
# attn_weight_matrix.size() = (batch_size, r, num_seq)
# output.size() = (batch_size, num_seq, 2*hidden_size)
hidden_matrix = torch.bmm(attn_weight_matrix, output)
# hidden_matrix.size() = (batch_size, r, 2*hidden_size)
# Let's now concatenate the hidden_matrix and connect it to the fully connected layer.
fc_out = self.fc_layer(hidden_matrix.view(-1, hidden_matrix.size()[1]*hidden_matrix.size()[2]))
logits = self.label(fc_out)
# logits.size() = (batch_size, output_size)
return logits
could u help please
Thanks
Hello i am new to this application. can you please help me to run this application on CPU configuration
A declarative, efficient, and flexible JavaScript library for building user interfaces.
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
An Open Source Machine Learning Framework for Everyone
The Web framework for perfectionists with deadlines.
A PHP framework for web artisans
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
Some thing interesting about web. New door for the world.
A server is a program made to process requests and deliver data to clients.
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
Some thing interesting about visualization, use data art
Some thing interesting about game, make everyone happy.
We are working to build community through open source technology. NB: members must have two-factor auth.
Open source projects and samples from Microsoft.
Google ❤️ Open Source for everyone.
Alibaba Open Source for everyone
Data-Driven Documents codes.
China tencent open source team.