You can not select more than 25 topics
Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
166 lines
5.4 KiB
166 lines
5.4 KiB
"""
|
|
Simple feed-forward neural network in PyTorch for baseline results on Scicite data.
|
|
Date: July 5th, 2020
|
|
"""
|
|
|
|
import torch
|
|
#import os
|
|
#os.chdir('/Users/iriley/code/citation-analysis')
|
|
import sys
|
|
sys.path.append('/Users/iriley/code/citation-analysis')
|
|
from utils.nn_reader import read_csv_nn
|
|
from utils.nn_reader2 import read_csv_nn_dev
|
|
from sklearn.metrics import confusion_matrix
|
|
import pandas as pd
|
|
import numpy as np
|
|
|
|
class Feedforward(torch.nn.Module):
|
|
"""
|
|
Creates and trains a basic feedforward neural network.
|
|
"""
|
|
#
|
|
def __init__(self, input_size, hidden_size, output_size):
|
|
""" Sets up all basic elements of NN. """
|
|
super(Feedforward, self).__init__()
|
|
self.input_size = input_size
|
|
self.hidden_size = hidden_size
|
|
self.output_size = output_size
|
|
self.fc1 = torch.nn.Linear(self.input_size, self.hidden_size)
|
|
#self.relu = torch.nn.ReLU()
|
|
self.tanh = torch.nn.Tanh()
|
|
self.fc2 = torch.nn.Linear(self.hidden_size, self.output_size)
|
|
self.sigmoid = torch.nn.Sigmoid()
|
|
self.softmax = torch.nn.Softmax(dim=1)
|
|
|
|
def forward(self, x):
|
|
""" Computes output from a given input x. """
|
|
hidden = self.fc1(x)
|
|
#relu = self.relu(hidden)
|
|
tanh = self.tanh(hidden)
|
|
#output = self.fc2(relu)
|
|
output = self.fc2(tanh)
|
|
output = self.softmax(output)
|
|
return output
|
|
|
|
|
|
|
|
if __name__=='__main__':
|
|
""" Reads in the data, then trains and evaluates the neural network. """
|
|
|
|
X_train, y_train, X_test = read_csv_nn()
|
|
|
|
# balance classes
|
|
yclass = np.array([(x[1]==1)+2*(x[2]==1) for x in y_train])
|
|
is0 = yclass==0
|
|
is1 = yclass==1
|
|
is2 = yclass==2
|
|
X0 = torch.FloatTensor(X_train[is0])
|
|
X1 = torch.FloatTensor(X_train[is1])
|
|
X2 = torch.FloatTensor(X_train[is2])
|
|
y0 = torch.LongTensor(np.zeros((sum(is0),)))
|
|
y1 = torch.LongTensor(np.ones((sum(is1),)))
|
|
y2 = torch.LongTensor(2*np.ones((sum(is2),)))
|
|
l0 = sum(is0)
|
|
l1 = sum(is1)
|
|
l2 = sum(is2)
|
|
p0 = torch.randperm(l0)
|
|
p1 = torch.randperm(l1)
|
|
p2 = torch.randperm(l2)
|
|
p = torch.randperm(3000)
|
|
X_train = torch.cat((X0[p0][:1000], X1[p1][:1000], X2[p2][:1000]))[p]
|
|
y_train = torch.cat((y0[:1000], y1[:1000], y2[:1000]))[p]
|
|
|
|
#X_train = torch.FloatTensor(X_train)
|
|
X_test = torch.FloatTensor(X_test)
|
|
#y_train_ = torch.FloatTensor(y_train)
|
|
#y_train = torch.max(torch.FloatTensor(y_train_),1)[1]
|
|
|
|
model = Feedforward(28, 9, 3)
|
|
criterion = torch.nn.CrossEntropyLoss()
|
|
optimizer = torch.optim.SGD(model.parameters(), lr = 0.01)
|
|
|
|
model.eval()
|
|
y_pred = model(X_train)
|
|
before_train = criterion(y_pred, y_train)
|
|
print('Test loss before training' , before_train.item())
|
|
|
|
l = 3000 # X_train.shape[0]
|
|
batch_indices = list(zip(list(range(0,l,16))[:-1], list(range(16,l,16))))# + [(l-l%16,l)]
|
|
batch_indices[-1] = (batch_indices[-1][0], l)
|
|
|
|
# train model
|
|
model.train()
|
|
epochs = 100
|
|
for epoch in range(epochs):
|
|
batch = 0
|
|
for a,b in batch_indices:
|
|
optimizer.zero_grad()
|
|
# forward pass
|
|
y_pred = model(X_train[a:b])
|
|
loss = criterion(y_pred, y_train[a:b])
|
|
print('Epoch {}, batch {}: train loss: {}'.format(epoch, batch, loss.item()))
|
|
# backward pass
|
|
loss.backward()
|
|
optimizer.step()
|
|
batch += 1
|
|
# get loss following epoch
|
|
y_pred = model.forward(X_train)
|
|
loss = criterion(y_pred, y_train)
|
|
print('Epoch {}: train loss: {}'.format(epoch, loss.item()))
|
|
# shuffle dataset
|
|
#p = torch.randperm(l)
|
|
#X_train = X_train[p]
|
|
#y_train = y_train[p]
|
|
p0 = torch.randperm(l0)
|
|
p1 = torch.randperm(l1)
|
|
p2 = torch.randperm(l2)
|
|
p = torch.randperm(3000)
|
|
X_train = torch.cat((X0[p0][:1000], X1[p1][:1000], X2[p2][:1000]))[p]
|
|
y_train = torch.cat((y0[:1000], y1[:1000], y2[:1000]))[p]
|
|
|
|
model.eval()
|
|
y_pred = model.forward(X_train)
|
|
after_train = criterion(y_pred, y_train)
|
|
print('Training loss after training' , after_train.item())
|
|
|
|
## reload the data to get original order
|
|
#X_train, y_train, X_test = read_csv_nn()
|
|
#X_train = torch.FloatTensor(X_train)
|
|
#X_test = torch.FloatTensor(X_test)
|
|
#y_train_ = torch.FloatTensor(y_train)
|
|
#y_train = torch.max(torch.FloatTensor(y_train_),1)[1]
|
|
|
|
# get dev set to make predictions
|
|
#X_dev, y_dev = read_csv_nn_dev()
|
|
#X_dev = torch.FloatTensor(X_dev)
|
|
#y_dev_pre = torch.FloatTensor(y_dev)
|
|
#y_dev = torch.max(torch.FloatTensor(y_dev_pre),1)[1]
|
|
|
|
# post-process to get predictions & get back to np format
|
|
y_pred = model.forward(X_test)
|
|
y_pred_np = y_pred.detach().numpy()
|
|
predmax = np.amax(y_pred_np, axis=1)
|
|
preds = 1*(y_pred_np[:,1]==predmax) + 2*(y_pred_np[:,2]==predmax)
|
|
#y_dev_ = y_dev.detach().numpy()
|
|
|
|
# create confusion matrix
|
|
#cm = confusion_matrix(y_dev_, preds)
|
|
#print(cm)
|
|
|
|
# save predictions
|
|
df = pd.DataFrame()
|
|
df['preds'] = preds
|
|
#df['true'] = y_dev_
|
|
probs = y_pred.detach().numpy()
|
|
df['pr0'] = probs[:,0]
|
|
df['pr1'] = probs[:,1]
|
|
df['pr2'] = probs[:,2]
|
|
#df['correct'] = df.preds==df.true
|
|
df.to_csv('/Users/iriley/code/machine_learning/lab2020/y_pred_model2.csv', index=False)
|
|
|
|
|
|
|
|
|
|
|
|
|