def __init__(self,
config,
input_dim,
num_layers,
num_classes,
encoder_dim=None,
bert_pretrained=True,
bert_pretrained_model_name='bert-base-cased'):
super().__init__()
self.bert = get_bert(bert_pretrained, bert_pretrained_model_name)
self.bert_asr = get_bert(bert_pretrained, bert_pretrained_model_name)
self.aux_embedding = nn.Linear(config.enc_dim,
self.bert.config.hidden_size
) #bert_hidden_size = 768 enc_dim = 128
self.lugosch_model = lugosch.models.PretrainedModel(config)
pretrained_model_path = os.path.join(config.libri_folder,
"libri_pretraining",
"model_state.pth")
self.lugosch_model.load_state_dict(torch.load(pretrained_model_path))
self.config = config
# freeze phoneme and word layers
self.freeze_all_layers()
self.unfreezing_index = 1
self.maxpool = MaskedMaxPool()
self.classifier = nn.Linear(self.bert.config.hidden_size, num_classes)
def __init__(self, num_classes, pretrained=True):
super().__init__()
self.bert = get_bert(pretrained)
self.classifier = nn.Sequential(
nn.Dropout(0.3),
nn.Linear(self.bert.config.hidden_size, num_classes)
)
def __init__(self,
config,
input_dim,
num_layers,
num_classes,
encoder_dim=None,
bert_pretrained=True,
bert_pretrained_model_name='bert-base-cased'):
super().__init__()
self.bert = get_bert(bert_pretrained, bert_pretrained_model_name)
#model 3 is bert trained on GT and ASR
if config.bert_dir:
print(
f"loading model3 (bert pretrained on GT and ASR) from {config.bert_dir}"
)
chkpt_path = os.path.join(config.bert_dir, 'best_ckpt.pth')
model_dict = self.bert.state_dict()
pretrained_dict = torch.load(chkpt_path)
pretrained_dict = {
k.split(".", 1)[1]: v
for k, v in pretrained_dict.items()
if k.split(".", 1)[1] in model_dict
}
self.bert.load_state_dict(pretrained_dict)
### Comment out Alexa's encoder
self.aux_embedding = nn.Linear(config.enc_dim,
self.bert.config.hidden_size
) #bert_hidden_size = 768 enc_dim = 128
self.lugosch_model = lugosch.models.PretrainedModel(config)
pretrained_model_path = os.path.join(config.libri_folder,
"libri_pretraining",
"model_state.pth")
self.lugosch_model.load_state_dict(torch.load(pretrained_model_path))
self.config = config
# freeze phoneme and word layers
self.freeze_all_layers()
self.unfreezing_index = 1
self.maxpool = MaskedMaxPool()
self.classifier = nn.Linear(self.bert.config.hidden_size, num_classes)
def __init__(self,
input_dim,
num_layers,
num_classes,
encoder_dim=None,
bert_pretrained=True,
bert_pretrained_model_name='bert-base-cased'):
super().__init__()
self.bert = get_bert(bert_pretrained, bert_pretrained_model_name)
self.encoder_dim = encoder_dim
if encoder_dim is None:
self.speech_encoder = SubsampledBiLSTMEncoder(
input_dim=input_dim,
encoder_dim=self.bert.config.hidden_size // 2,
num_layers=num_layers)
else:
self.speech_encoder = SubsampledBiLSTMEncoder(
input_dim=input_dim,
encoder_dim=encoder_dim,
num_layers=num_layers)
self.aux_embedding = nn.Linear(2 * encoder_dim,
self.bert.config.hidden_size)
self.maxpool = MaskedMaxPool()
self.classifier = nn.Linear(self.bert.config.hidden_size, num_classes)
def __init__(self,
config,
input_dim,
num_layers,
num_classes,
encoder_dim=None,
bert_pretrained=True,
bert_pretrained_model_name='bert-base-cased'):
super().__init__()
self.bert = get_bert(bert_pretrained, bert_pretrained_model_name)
### Comment out Alexa's encoder
# self.encoder_dim = encoder_dim
# if encoder_dim is None:
# self.speech_encoder = SubsampledBiLSTMEncoder(input_dim=input_dim, encoder_dim=self.bert.config.hidden_size//2, num_layers=num_layers)
# else:
# self.speech_encoder = SubsampledBiLSTMEncoder(input_dim=input_dim, encoder_dim=encoder_dim, num_layers=num_layers)
self.aux_embedding = nn.Linear(config.enc_dim,
self.bert.config.hidden_size
) #bert_hidden_size = 768 enc_dim = 128
self.aux_reverse = nn.Linear(
self.bert.config.hidden_size,
config.enc_dim) #match bert hidden size with Lugosch's
self.lugosch_model = lugosch.models.PretrainedModel(config)
pretrained_model_path = os.path.join(config.libri_folder,
"libri_pretraining",
"model_state.pth")
self.lugosch_model.load_state_dict(torch.load(pretrained_model_path))
self.config = config
# freeze phoneme and word layers
self.freeze_all_layers()
self.unfreezing_index = 1
self.maxpool = MaskedMaxPool()
self.classifier = nn.Linear(self.bert.config.hidden_size, num_classes)
# Based on Lugosch's Intent Module (class Model in models.py)
self.intent_layers = []
self.num_values_total = num_classes #default for fluentai
self.num_rnn_layers = len(config.intent_rnn_num_hidden)
self.out_dim = config.word_rnn_num_hidden[-1]
if config.word_rnn_bidirectional:
self.out_dim *= 2
for idx in range(self.num_rnn_layers):
# recurrent
print("config.intent_rnn_bidirectional :",
config.intent_rnn_bidirectional)
layer = torch.nn.GRU(input_size=self.out_dim,
hidden_size=config.intent_rnn_num_hidden[idx],
batch_first=True,
bidirectional=config.intent_rnn_bidirectional)
layer.name = "intent_rnn%d" % idx
self.intent_layers.append(layer)
self.out_dim = config.intent_rnn_num_hidden[idx]
if config.intent_rnn_bidirectional:
self.out_dim *= 2
# grab hidden states of RNN for each timestep
layer = RNNSelect()
layer.name = "intent_rnn_select%d" % idx
self.intent_layers.append(layer)
# dropout
layer = torch.nn.Dropout(p=config.intent_rnn_drop[idx])
layer.name = "intent_dropout%d" % idx
self.intent_layers.append(layer)
# downsample
layer = Downsample(method=config.intent_downsample_type[idx],
factor=config.intent_downsample_len[idx],
axis=1)
layer.name = "intent_downsample%d" % idx
self.intent_layers.append(layer)
layer = torch.nn.Linear(self.out_dim, self.num_values_total)
layer.name = "final_classifier"
self.intent_layers.append(layer)
layer = FinalPool() #maxpool 3D - 2D
layer.name = "final_pool"
self.intent_layers.append(layer)
self.lugosch_classifier = torch.nn.ModuleList(self.intent_layers)
