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.
128 lines
4.7 KiB
128 lines
4.7 KiB
"""
|
|
Simple feed-forward neural network in PyTorch for baseline results on Scicite data.
|
|
Created: July 5th, 2020
|
|
"""
|
|
|
|
import torch
|
|
from utils.nn_reader import read_csv_nn
|
|
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.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)
|
|
tanh = self.tanh(hidden)
|
|
output = self.fc2(tanh)
|
|
output = self.softmax(output)
|
|
return output
|
|
|
|
def read_data(self):
|
|
"""" Reads in training and test data and converts it to proper format. """
|
|
self.X_train_, self.y_train_, self.X_test_ = read_csv_nn()
|
|
yclass = np.array([(x[1] == 1) + 2 * (x[2] == 1) for x in self.y_train_])
|
|
is0 = yclass == 0
|
|
is1 = yclass == 1
|
|
is2 = yclass == 2
|
|
self.X0 = torch.FloatTensor(self.X_train_[is0])
|
|
self.X1 = torch.FloatTensor(self.X_train_[is1])
|
|
self.X2 = torch.FloatTensor(self.X_train_[is2])
|
|
self.y0 = torch.LongTensor(np.zeros((sum(is0),)))
|
|
self.y1 = torch.LongTensor(np.ones((sum(is1),)))
|
|
self.y2 = torch.LongTensor(2 * np.ones((sum(is2),)))
|
|
self.l0 = sum(is0)
|
|
self.l1 = sum(is1)
|
|
self.l2 = sum(is2)
|
|
|
|
def fit(self, epochs=100, batch_size=16, lr=0.01, samples0=1000, samples1=1000, samples2=1000):
|
|
""" Trains model, using cross entropy loss and SGD optimizer. """
|
|
self.criterion = torch.nn.CrossEntropyLoss()
|
|
self.optimizer = torch.optim.SGD(self.parameters(), lr)
|
|
self.samples0 = samples0
|
|
self.samples1 = samples1
|
|
self.samples2 = samples2
|
|
|
|
self.eval() # put into eval mode
|
|
|
|
# initialize training data
|
|
self.shuffle()
|
|
y_pred = self.forward(self.X_train)
|
|
before_train = self.criterion(y_pred, self.y_train)
|
|
print('Test loss before training', before_train.item())
|
|
|
|
# setup for batches
|
|
l = self.samples0 + self.samples1 + self.samples2 # total length
|
|
batch_indices = list(zip(list(range(0, l, batch_size))[:-1], list(range(16, l, batch_size))))
|
|
batch_indices[-1] = (batch_indices[-1][0], l)
|
|
|
|
# train model
|
|
self.train() # put into training mode
|
|
for epoch in range(epochs):
|
|
batch = 0
|
|
for a, b in batch_indices:
|
|
self.optimizer.zero_grad()
|
|
|
|
# forward pass
|
|
y_pred = model(X_train[a:b])
|
|
loss = self.criterion(y_pred, self.y_train[a:b])
|
|
|
|
# backward pass
|
|
loss.backward()
|
|
self.optimizer.step()
|
|
batch += 1
|
|
|
|
# get loss following epoch
|
|
y_pred = self.forward(self.X_train)
|
|
loss = self.criterion(y_pred, self.y_train)
|
|
print('Epoch {}: train loss: {}'.format(epoch, loss.item()))
|
|
|
|
# shuffle dataset
|
|
self.shuffle()
|
|
|
|
# display final loss
|
|
self.eval() # back to eval mode
|
|
y_pred = self.forward(self.X_train)
|
|
after_train = self.criterion(y_pred, self.y_train)
|
|
print('Training loss after training', after_train.item())
|
|
|
|
def predict(self):
|
|
""" Generates predictions from model, using test data. """
|
|
|
|
# post-process to get predictions & get back to np format
|
|
y_pred = self.forward(self.X_test)
|
|
y_pred_np = y_pred.detach().numpy()
|
|
predmax = np.amax(y_pred_np, axis=1)
|
|
self.preds = 1 * (y_pred_np[:, 1] == predmax) + 2 * (y_pred_np[:, 2] == predmax)
|
|
self.probs = y_pred.detach().numpy()
|
|
|
|
def shuffle(self):
|
|
""" Samples and shuffles training data. """
|
|
|
|
# create permutations for shuffling
|
|
p0 = torch.randperm(self.l0)
|
|
p1 = torch.randperm(self.l1)
|
|
p2 = torch.randperm(self.l2)
|
|
n = self.l0 + self.l1 + self.l2
|
|
p = torch.randperm(n)
|
|
|
|
# sample and shuffle data
|
|
self.X_train = \
|
|
torch.cat((self.X0[p0][:self.samples0], self.X1[p1][:self.samples1], self.X2[p2][:self.samples2]))[p]
|
|
self.y_train = torch.cat((self.y0[:self.samples0], self.y1[:self.samples1], self.y2[:self.samples2]))[p]
|