lstm-textgen-pets / pipeline.py
miittnnss's picture
Update pipeline.py
6cd5136
raw
history blame contribute delete
No virus
2.51 kB
import torch
import torch.nn as nn
import random
from huggingface_hub import PyTorchModelHubMixin
class LSTMTextGenerator(nn.Module, PyTorchModelHubMixin):
def __init__(self, input_size=45, hidden_size=512, output_size=45, num_layers=2, dropout=0.5):
super(LSTMTextGenerator, self).__init__()
self.embedding = nn.Embedding(input_size, hidden_size)
self.lstm = nn.LSTM(hidden_size, hidden_size, num_layers, batch_first=True, dropout=dropout, bidirectional=False)
self.fc = nn.Linear(hidden_size, output_size)
self.num_layers = num_layers
self.hidden_size = hidden_size
def forward(self, x, hidden):
x = x.to(torch.long)
x = self.embedding(x)
x, hidden = self.lstm(x, hidden)
x = self.fc(x)
return x, hidden
def init_hidden(self, batch_size):
return (torch.zeros(self.num_layers, batch_size, self.hidden_size).to(device),
torch.zeros(self.num_layers, batch_size, self.hidden_size).to(device))
class PreTrainedPipeline():
def __init__(self, path=""):
self.model = LSTMTextGenerator.from_pretrained("miittnnss/lstm-textgen-pets")
self.chars = "!',.;ACDFGHIMORSTWabcdefghijklmnopqrstuvwxy"
self.char_to_index = {char: index for index, char in enumerate(self.chars)}
self.index_to_char = {index: char for char, index in self.char_to_index.items()}
self.output_size = len(self.chars)
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
def __call__(self, inputs: str):
seed_numerical_data = [self.char_to_index[char] for char in inputs]
with torch.no_grad():
input_sequence = torch.LongTensor([seed_numerical_data]).to(self.device)
hidden = self.model.init_hidden(1)
generated_text = inputs # Initialize generated text with seed text
temperature = 0.7 # Temperature for temperature sampling
for _ in range(500):
output, hidden = self.model(input_sequence, hidden)
probabilities = nn.functional.softmax(output[-1, 0] / temperature, dim=0).cpu().numpy()
predicted_index = random.choices(range(self.output_size), weights=probabilities, k=1)[0]
generated_text += self.index_to_char[predicted_index] # Append the generated character to the text
input_sequence = torch.LongTensor([[predicted_index]]).to(self.device)
return generated_text