from gpt2_client import GPT2Client
gpt2 = GPT2Client('117M')
gpt2.load_model()
#inputText = 'Trump is building a wall. A big wall. To stop CHINA from taking our jobs. Make America Great again!'
genText = gpt2.generate(interactive=True, n_samples=1, return_text=True)
#genText = gpt2.generate(return_text=True)
with open('test.txt', 'a') as filePtr:
#filePtr.write(inputText + '\n')
for line in genText:
filePtr.write(line + '\n')
filePtr.write('\n')
from gpt2_client import GPT2Client
gpt2 = GPT2Client('345M')
gpt2.load_model()
my_corpus = './_data/_data_en_es.txt'
genText = gpt2.finetune(my_corpus, return_text=True)
def __init__(self, **kwargs):
GPT2Client.__init__(self, **kwargs)
# note also you don't have to take the argmax from bert but can sample from it
# another primitive would be back-translating from an embedding
# if you can't balance the beam score, you can perhaps alternate faithfulness
# and maybe you can keep the "best of all time"
# would be good to do an autoencoder type of embed and decode primitive
# def get_score(sentence):
# tokenize_input = bert_tokeniser.tokenize(sentence)
# tensor_input = torch.tensor([bert_tokeniser.convert_tokens_to_ids(tokenize_input)])
# predictions=bert_model(tensor_input)
# loss_fct = torch.nn.CrossEntropyLoss()
# loss = loss_fct(predictions.squeeze(),tensor_input.squeeze()).data
# return math.exp(loss)
from gpt2_client import GPT2Client
gpt2 = GPT2Client('774M')
gpt2.load_model(force_download=False)
from transformers import GPT2Tokenizer, GPT2LMHeadModel
with torch.no_grad():
gpt2_model = GPT2LMHeadModel.from_pretrained('gpt2')
gpt2_model.eval()
gpt2_tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
def get_score(sentence):
tokenize_input = gpt2_tokenizer.encode(sentence)
tensor_input = torch.tensor([tokenize_input])
loss=gpt2_model(tensor_input, labels=tensor_input)[0]
return np.exp(loss.detach().numpy())
# syn_then_do("please cancel my restaurant reservations")
"""
The following snippet covers how to use GPT2Client to generate text randomly.
It will return an array containing the `n_samples` pieces of generated text.
"""
from gpt2_client import GPT2Client
gpt2 = GPT2Client('117M', save_dir="models")
# optional -> if you already have the assets, you can comment this out
gpt2.load_model(force_download=False)
# interative mode is False by default
# This will return 4 pieces of generated text
text = gpt2.generate(n_samples=4, return_text=True)
print(text)