I have a problem with PyTorch, cause I am a beginner. I am trying to implement some encoding algorithm using vocabularies and torch.zero matrix. But I have a trouble, it did not worked properly:
For example having the same function with or without break
gives a different output matrixes. I checked it only for first matrix and I guess it the same with others. Function with break
gives correct matrix, with the 1
in 45 indexes (see example below), function without break in the for cycle contains one in the incorrect place (in 18 index). However I have changed nothing in code despite break
.
Here is a function without break:
def encode_dialog_states(df, symptoms_voc, pathos_voc, max_turn_num=30):
N = max_turn_num
M = len(symptoms_voc) + len(pathos_voc)
print("Dialog state matrix shape is:", N, 'x', M)
df['DIALOG_STATE_MATRIX'] = [None] * len(df)
X = torch.zeros(N,M)
for idx, row in df.iterrows():
X.zero_()
patho_idx = pathos_voc.get(row['PATHOLOGY'])
X[:, patho_idx] = 1
#print('patho_idx', patho_idx)
for t, evidence in enumerate(row['EVIDENCES']):
sym_idx = symptoms_voc.get(evidence)
X[t:N, len(pathos_voc) + sym_idx] = 1
df.at[idx, 'DIALOG_STATE_MATRIX'] = X
return df
Here is the same function with break in first cycle:
def encode_dialog_states(df, symptoms_voc, pathos_voc, max_turn_num=30):
N = max_turn_num
M = len(symptoms_voc) + len(pathos_voc)
print("Dialog state matrix shape is:", N, 'x', M)
df['DIALOG_STATE_MATRIX'] = [None] * len(df)
X = torch.zeros(N,M)
for idx, row in df.iterrows():
X.zero_()
patho_idx = pathos_voc.get(row['PATHOLOGY'])
X[:, patho_idx] = 1
#print('patho_idx', patho_idx)
for t, evidence in enumerate(row['EVIDENCES']):
sym_idx = symptoms_voc.get(evidence)
X[t:N, len(pathos_voc) + sym_idx] = 1
df.at[idx, 'DIALOG_STATE_MATRIX'] = X
break
return df
I checked both of the functions for following example and get different output:
For the pathology train_set.loc[0, 'PATHOLOGY']
== URTI
indexed in vocabulary by 45
index:
Here the first matrix's first row returned by without break
function with incorrectly located value 1
:
tensor([0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
1., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 1., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0.])
And the function with break
for the same matrix on the first row has correctly located value:
tensor([0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 1., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 1., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0.])
As you can see, the index in without break
is incorrect (the value 1
is located on 18 position) and the with break
version has correct value location (on 45 position). So, I want to have the correct matrixes as in the last example for all of the samples, without using break. What's the problem of the function and how to fix it?
P.S. I also have checked the size of matrixes for both cases and they are correct (1x561):
I have resolved the issues. The problem was the referencing to the same tensor for all of the samples in df
. To solve the issue I've just added .detach().clone() after X before assigning. However, in my case this would cause the memory leak, so before cloning I made the tensors sparse.
Here is modified code:
def encode_dialog_states(df, symptoms_voc, pathos_voc, max_turn_num=30):
N = max_turn_num
M = len(symptoms_voc) + len(pathos_voc)
print("Dialog state matrix shape is:", N, 'x', M)
df['DIALOG_STATE_MATRIX'] = [None] * len(df)
X = torch.zeros(N,M)
for idx, row in df.iterrows():
X.zero_()
patho_idx = pathos_voc.get(row['PATHOLOGY'])
X[:, patho_idx] = 1
#print('patho_idx', patho_idx)
for t, evidence in enumerate(row['EVIDENCES']):
sym_idx = symptoms_voc.get(evidence)
X[t:N, len(pathos_voc) + sym_idx] = 1
df.at[idx, 'DIALOG_STATE_MATRIX'] = X.to_sparse().clone().detach()
return df
```