i was approaching NLP sequence classification problem (3 classes) using huggingface transformers (funnel-transformer/large) and tensorflow.
first i created laserembedding like this :
from laserembeddings import Laser
laser = Laser()
df = pd.read_csv("mycsv.csv")
embeds = laser.embed_sentences(df['text'].values, lang='en')
write_pickle_to_file('train.pkl', embeds )
part 1 : Tensorflow version
for data preparation i use code like below :
df['text']=temp['column1']+tokenizer.sep_token+temp['column2']+tokenizer.sep_token+temp['column3']
def encode_text(texts):
enc_di = tokenizer.batch_encode_plus(
texts,
padding='max_length',
truncation=True,
return_token_type_ids=True,
pad_to_max_length=True,
max_length=cfg.max_len
)
return [np.asarray(enc_di['input_ids'], dtype=np.int64),
np.asarray(enc_di['attention_mask'], dtype=np.int64),
np.asarray(enc_di['token_type_ids'], dtype=np.int64)]
then inside training function :
x_train = encode_text(df.text.to_list())
train_ds = (
tf.data.Dataset
.from_tensor_slices((
{
"input_ids": x_train[0],
"input_masks": x_train[1],
"input_segments": x_train[2],
"lasers": np.array( train[laser_columns].values, dtype=np.float32 ) #laser_columns contains all the laser embedded columns
},
tf.one_hot(df["label"].to_list(), 3) #3 class
))
.repeat()
.shuffle(2048)
.batch(cfg.batch_size)
.prefetch(AUTO)
)
i add laser embedding in my model like this :
def create_model():
transformer = transformers.TFAutoModel.from_pretrained(cfg.pretrained,config=config,from_pt=True)
max_len=512
# transformer
input_ids = Input(shape=(max_len,), dtype="int32", name="input_ids")
input_masks = Input(shape=(max_len,), dtype="int32", name="input_masks")
input_segments = Input(shape=(max_len,), dtype="int32", name="input_segments")
sequence_output = transformer(input_ids, attention_mask=input_masks, token_type_ids=input_segments)[0]
cls_token = sequence_output[:, 0, :]
# lasers
lasers = Input(shape=(n_lasers,), dtype=tf.float32, name="lasers") #n_lasers = 1024
lasers_output = tf.keras.layers.Dense(n_lasers, activation='tanh')(lasers)
x = tf.keras.layers.Concatenate()([cls_token, lasers_output])
x = tf.keras.layers.Dropout(0.1)(x)
x = tf.keras.layers.Dense(2048, activation='tanh')(x)
x = tf.keras.layers.Dropout(0.1)(x)
out = tf.keras.layers.Dense(3, activation='softmax')(x)
model = Model(inputs=[input_ids, input_masks, input_segments, lasers], outputs=out)
model.compile(Adam(lr=1e-5), loss=losses.CategoricalCrossentropy(), metrics=["acc", metrics.CategoricalCrossentropy(name='xentropy')])
return model
now my question is, how do we do the same with pytorch for exact same problem and same dataset?
part 2 : pytorch version
df = pd.read_csv("mytrain.csv")
class myDataset(Dataset):
def __init__(self,df, max_length, tokenizer, training=True):
self.df = df
self.max_len = max_length
self.tokenizer = tokenizer
self.column1 = self.df['column1'].values
self.column2 = self.df['column2'].values
self.column3= self.df['column3'].values
self.column4= self.df['column4'].values
self.training = training
if self.training:
self.targets = self.df['label'].values
def __len__(self):
return len(self.df)
def __getitem__(self, index):
column1 = self.column1[index]
column2= self.column2[index]
column3= self.column3[index]
text0 = self.column4[index]
text1 = column1 + ' ' + column2+ ' ' + column3
inputs = self.tokenizer.encode_plus(
text1 ,
text0 ,
truncation = True,
add_special_tokens = True,
return_token_type_ids = True,
is_split_into_words=False,
max_length = self.max_len
)
samples = {
'input_ids': inputs['input_ids'],
'attention_mask': inputs['attention_mask'],
}
if 'token_type_ids' in inputs:
samples['token_type_ids'] = inputs['token_type_ids']
if self.training:
samples['target'] = self.targets[index]
return samples
collate_fn = DataCollatorWithPadding(tokenizer=CONFIG['tokenizer'])
class myModel(nn.Module):
def __init__(self, model_name):
super(myModel, self).__init__()
self.model = AutoModel.from_pretrained(model_name)
if(True):
print("using gradient_checkpoint...")
self.model.gradient_checkpointing_enable()
self.config = AutoConfig.from_pretrained(model_name)
self.config.update(
{
"output_hidden_states": True,
"hidden_dropout_prob": 0.0,
"layer_norm_eps": 1e-7,
"add_pooling_layer": False,
"attention_probs_dropout_prob":0.0,
}
)
self.fc = nn.Linear(self.config.hidden_size, 3)
def forward(self, ids, mask):
out = self.model(input_ids=ids,attention_mask=mask,output_hidden_states=False)
out = out[0][:, 0, :]
outputs = self.fc(out)
return outputs
and in train and validation loop i have code like this :
bar = tqdm(enumerate(dataloader), total=len(dataloader))
for step, data in bar:
ids = data['input_ids'].to(device, dtype = torch.long)
mask = data['attention_mask'].to(device, dtype = torch.long)
targets = data['target'].to(device, dtype=torch.long)
batch_size = ids.size(0)
optimizer.zero_grad()
# forward pass with `autocast` context manager
with autocast(enabled=True):
outputs = model(ids, mask)
loss = loss_fct(outputs, targets)
i would like to know where and how in my huggingface pytorch pipeline i can use the laserembedding that i created earlier and used in tensorflow huggingface model?
i would like to concat laserembeddings with funnel transformer’s simple CLS token output and train the transformers model with laser embed as extra feature in pytorch implementation exactly like i did in tensorflow example,do you know how to modify my pytorch code to make it working in pytorch? the tensorflow implementation with laserembedding concatenated above that i have posted here works good,i just wanted to do the same in pytorch implementation,your help is highly appreciated,thanks in advance