code stringlengths 3 6.57k |
|---|
re.compile(r'reference\.md') |
re.compile(r'_episodes/.*\.md') |
re.compile(r'.*\.md') |
main() |
CustomDataset(torch.utils.data.Dataset) |
__init__(self, input_ids, token_type_ids, attention_masks) |
__len__(self) |
len(self.input_ids) |
__getitem__(self, index) |
weights_init(tensor: torch.Tensor) |
isinstance(tensor, nn.Linear) |
switch_init(tensor.weight.data) |
torch.nn.init.zeros_(tensor.bias.data) |
isinstance(tensor, nn.LayerNorm) |
torch.nn.init.zeros_(tensor.weight.data) |
torch.nn.init.zeros_(tensor.bias.data) |
switch_init(tensor: torch.Tensor, s: float = 0.1, mean: float=0) |
torch.nn.init._calculate_fan_in_and_fan_out(tensor) |
math.sqrt(s/fan_in) |
torch.nn.init.trunc_normal_(tensor=tensor, mean=mean, std=std) |
LaBSE_Switch(nn.Module) |
__init__(self, config, word_embeddings_module) |
super() |
__init__() |
MultiHeadAttention(config['heads'], config['d_model'], config['dropout']) |
FeedForward(config['d_model'], config['d_ff'], config['dropout_ffn']) |
int(768) |
self.switch_model.apply(weights_init) |
self.weight_init_from_teacher(teacher_model=teacher_model, int_matches=int_matches) |
weight_init_from_teacher(self, teacher_model, int_matches) |
load_teacher(device=torch.device('cuda') |
forward(self, input_ids, token_type_ids=None, attention_mask=None) |
self.word_embeddings(input_ids) |
print(_n_hid) |
self.switch_model(torch.reshape(input_embeddings, (_seq_len, _batch, _n_hid) |
load_student(name, student_config, device, teacher_model, int_matches, N_LAYERS) |
BertConfig.from_pretrained(name) |
BertModel.from_pretrained(name, config=student_config) |
student_model.set_input_embeddings(teacher_model.get_input_embeddings() |
student_model.float() |
student_model.to(device=device) |
deepcopy(teacher_model.get_input_embeddings() |
word_embedding_compression(word_embeddings, student_config['d_model']) |
LaBSE_Switch(config=student_config, word_embeddings_module=compressed_word_embeddings) |
student_model.switch_model.apply(weights_init) |
student_model.weight_init_from_teacher(teacher_model=teacher_model, int_matches=int_matches) |
student_model.float() |
student_model.to(device=device) |
load_teacher(device) |
AutoConfig.from_pretrained('sentence-transformers/LaBSE') |
AutoModel.from_pretrained('sentence-transformers/LaBSE', config=teacher_config) |
teacher_model.float() |
teacher_model.to(device=device) |
simple_adaptor(batch, model_outputs) |
values.append(i[1]) |
torch.stack(values) |
attentions.append(inv_softmax(j) |
torch.stack(attentions) |
inv_softmax(x,C=-50) |
torch.log(x) |
torch.where(result <= float('-inf') |
torch.full_like(result,C) |
teacher_adaptor(batch, model_outputs) |
values.append(i[1]) |
torch.stack(values) |
attentions.append(inv_softmax(j) |
torch.stack(attentions) |
print(model_outputs['pooler_output'].requires_grad) |
switch_student_adaptor(batch, model_outputs) |
reshape(layers, batch_size, heads, len, len) |
reshape(batch_size, len, d_model) |
print(model_outputs['pooler_output'].requires_grad) |
reshape(layers, batch_size, heads, len, embedding_per_head) |
predict(model, teacher_model, eval_dataset, step, device, STUDENT, BATCH_SIZE, eval_metric='cosine_similarity', feedback=True) |
model.eval() |
DataLoader(eval_dataset,batch_size=BATCH_SIZE) |
print('Running callback function on {} dev set samples...'.format(len(eval_dataset) |
to(device) |
to(device) |
torch.no_grad() |
model(input_ids=input_ids, attention_mask=attention_mask) |
teacher_model(input_ids=input_ids, attention_mask=attention_mask) |
logits_S.detach() |
cpu() |
logits_T.detach() |
cpu() |
detach() |
cpu() |
detach() |
cpu() |
range(len(cpu_logits_S) |
student_logits.append(cpu_logits_S[i].numpy() |
teacher_logits.append(cpu_logits_T[i].numpy() |
range(len(counts) |
batch_counts.append(counts[i].numpy() |
batch_n_dropped.append(n_dropped[i]) |
batch_route_prob.append(route_prob[i].numpy() |
model.train() |
np.array(student_logits) |
np.array(teacher_logits) |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.