def load_data(self, prepared_data_file=None):
prepared_data_file = prepared_data_file or self.prepared_data_file
print(f"Loading prepared data from {prepared_data_file} ...")
data = torch.load(prepared_data_file)
self.data = {"train": Dataset(data['train']),
"valid": Dataset(data["valid"]),
"test": Dataset(data["test"])}
print("Number of examples:",
" ".join(f"{k.upper()}-{len(v)}" for k, v in self.data.items()))
def transform(self, data_file, batch_size,
data_type="test", shuffle=False, device=None):
"""
Transform raw text from data_file to Dataset and create data loader.
"""
raw_data = self.read_data(data_file, data_type=data_type)
examples = self.build_examples(raw_data)
data = Dataset(examples)
data_loader = data.create_batches(batch_size, shuffle, device)
return data_loader
def load_data(self, prepared_data_file=None):
"""
load_data
"""
prepared_data_file = prepared_data_file or self.prepared_data_file
print("Loading prepared data from {} ...".format(prepared_data_file))
data = torch.load(prepared_data_file)
self.data = {"train": Dataset(data['train']),
"valid": Dataset(data["valid"]),
"test": Dataset(data["test"])}
print("Number of examples:",
" ".join("{}-{}".format(k.upper(), len(v)) for k, v in self.data.items()))
def reload(self, data_type='test'):
data_file = os.path.join(self.data_dir, self.data_prefix + "." + data_type)
data_raw = self.read_data(data_file, data_type="test")
data_examples = self.build_examples(data_raw)
self.data[data_type] = Dataset(data_examples)
print("Number of examples:",
" ".join(f"{k.upper()}-{len(v)}" for k, v in self.data.items()))
def reload(self, data_type='test'):
"""
reload
训练时把demo.test覆盖了，测试时不用删除，reload会重新搞的
"""
data_file = os.path.join(self.data_dir, self.data_prefix + "." + data_type)
data_raw = self.read_data(data_file, data_type="test")
data_examples = self.build_examples(data_raw)
self.data[data_type] = Dataset(data_examples)
print("Number of examples:",
" ".join("{}-{}".format(k.upper(), len(v)) for k, v in self.data.items()))
def reload(self, data_type='test', data_file=None):
"""
reload
"""
if data_file is not None:
data_raw = self.read_data(data_file, data_type="test")
data_examples = self.build_examples(data_raw)
self.data[data_type] = Dataset(data_examples)
print(
"Number of examples:", " ".join("{}-{}".format(k.upper(), len(v))
for k, v in self.data.items()))
def gen_response(self, contexts):
"""
Return a list of responses to each context.
:param contexts: list, a list of context, each context is a dict that contains the dialogue history and personal
profile of each speaker
this dict contains following keys:
context['dialog']: a list of string, dialogue histories (tokens in each utterances
are separated using spaces).
context['uid']: a list of int, indices to the profile of each speaker
context['profile']: a list of dict, personal profiles for each speaker
context['responder_profile']: dict, the personal profile of the responder
:return: list, responses for each context, each response is a list of tokens.
e.g.
contexts:
[{ "dialog": [ ["How are you ?"], ["I am fine , thank you . And you ?"] ],
"uid": [0, 1],
"profile":[ { "loc":"Beijing", "gender":"male", "tag":"" },
{ "loc":"Shanghai", "gender":"female", "tag":"" } ],
"responder_profile":{ "loc":"Beijing", "gender":"male", "tag":"" }
}]
==> [['I', 'am', 'fine', 'too', '!']]
"""
test_raw = self.read_data(contexts[0])
test_data = self.corpus.build_examples(test_raw, data_type='test')
dataset = Dataset(test_data)
data_iter = dataset.create_batches(batch_size=1,
shuffle=False,
device=self.config.gpu)
results = self.generator.generate(batch_iter=data_iter)
res = [result.preds[0].split(" ") for result in results]
return res
def reload(self, data_type='test'):
"""
reload
"""
data_file1 = os.path.join(self.data_dir,
self.data_prefix + "." + 'test1')
data_file2 = os.path.join(self.data_dir,
self.data_prefix + "." + 'test2')
data_raw1, data_raw2 = self.read_data_multitask(data_file1,
data_file2,
data_type="test")
data_examples1 = self.build_examples(data_raw1)
data_examples2 = self.build_examples(data_raw2)
self.data[data_type] = Dataset((data_examples1, data_examples2))
print(
"Number of examples:", " ".join("{}-{}".format(k.upper(), len(v))
for k, v in self.data.items()))
def next_word_probability(self, context, partial_out):
"""
Return probability distribution over next words given a partial true output.
This is used to calculate the per-word perplexity.
:param context: dict, contexts containing the dialogue history and personal
profile of each speaker
this dict contains following keys:
context['dialog']: a list of string, dialogue histories (tokens in each utterances
are separated using spaces).
context['uid']: a list of int, indices to the profile of each speaker
context['profile']: a list of dict, personal profiles for each speaker
context['responder_profile']: dict, the personal profile of the responder
:param partial_out: list, previous "true" words
:return: a list, the first element is a dict, where each key is a word and each value is a probability
score for that word. Unset keys assume a probability of zero.
the second element is the probability for the EOS token
e.g.
context:
{ "dialog": [ ["How are you ?"], ["I am fine , thank you . And you ?"] ],
"uid": [0, 1],
"profile":[ { "loc":"Beijing", "gender":"male", "tag":"" },
{ "loc":"Shanghai", "gender":"female", "tag":"" } ],
"responder_profile":{ "loc":"Beijing", "gender":"male", "tag":"" }
}
partial_out:
['I', 'am']
==> {'fine': 0.9}, 0.1
"""
test_raw = self.read_data(context)
test_data = self.corpus.build_examples(test_raw, data_type='test')
dataset = Dataset(test_data)
data_iter = dataset.create_batches(batch_size=1,
shuffle=False,
device=self.config.gpu)
inputs = None
for batch in data_iter:
inputs = batch
break
partial_out_idx = [
self.stoi[s] if s in self.stoi.keys() else self.stoi['<unk>']
for s in partial_out
]
# switch the model to evaluate mode
self.model.eval()
with torch.no_grad():
enc_outputs, dec_init_state = self.model.encode(inputs)
long_tensor_type = torch.cuda.LongTensor if self.config.use_gpu else torch.LongTensor
# Initialize the input vector
input_var = long_tensor_type([self.BOS] * 1)
# Inflate the initial hidden states to be of size: (1, H)
dec_state = dec_init_state.inflate(1)
for t in range(len(partial_out_idx)):
# Run the RNN one step forward
output, dec_state, attn = self.model.decode(
input_var, dec_state)
input_var = long_tensor_type([partial_out_idx[t]])
output, dec_state, attn = self.model.decode(input_var, dec_state)
log_softmax_output = output.squeeze(1)
log_softmax_output = log_softmax_output.cpu().numpy()
prob_output = [math.exp(i) for i in log_softmax_output[0]]
# The first 4 tokens are: '<pad>' '<unk>' '<bos>' '<eos>'
freq_dict = {}
for i in range(4, len(self.itos)):
freq_dict[self.itos[i]] = prob_output[i]
eos_prob = prob_output[3]
return freq_dict, eos_prob
