step += len(inputs)
inp = inputs[:, 6:7]
# output is (b, d1, d2, ..., dn, c) c is onehot logits
# inputs = random_mask(targets, ARCDataset.WordMap["start_symbol"], ARCDataset.WordMap['pad_symbol'])
answers = targets.ne(inp).to(torch.long)
# print(inputs.shape, inputs_mask.shape)
answers[inputs_mask[:, 6:7].eq(
False)] = TaskSpecificARCDataset.WordMap['pad_symbol']
outputs = train_forward(net, inputs, inputs_mask, task, task_mask)
# print(inputs.shape, outputs.shape, targets.shape, answers.shape)
loss = compute_balance_loss(
outputs, targets, answers,
TaskSpecificARCDataset.WordMap['pad_symbol'])
_, output_index = torch.max(outputs, dim=-1)
output_index = output_index.to(outputs.dtype)
element_accuracy = compute_element_accuracy(
output_index, targets,
TaskSpecificARCDataset.WordMap['pad_symbol'])
mask_accuracy = compute_mask_accuracy(
inp, output_index, targets,
TaskSpecificARCDataset.WordMap['pad_symbol'])
correct_accuracy = compute_corrects_accuracy(
output_index, targets,
TaskSpecificARCDataset.WordMap['pad_symbol'])
loss.backward()
optimizer.step()
inputs, ctx_input, targets, ctx_targets, task = to_same_size(
padding, *to_device(device, *batch))
start_time = time.time()
step += len(inputs)
answers = targets.ne(inputs).to(torch.long)
answers[inputs.eq(padding)] = padding
outputs, predict_task = train_forward(
net, inputs,
inputs.ne(padding).to(torch.float), ctx_input,
ctx_input.ne(padding).to(torch.float), ctx_targets,
ctx_targets.ne(padding).to(torch.float))
loss = compute_balance_loss(outputs, targets, answers,
padding) + compute_task_loss(
predict_task, task[:, 0, 0], padding)
_, predict_task = torch.max(predict_task, dim=-1)
_, output_index = torch.max(outputs, dim=-1)
output_index = output_index.to(outputs.dtype)
task_accuracy = compute_task_accuracy(predict_task, task[:, 0, 0],
padding)
element_accuracy = compute_element_accuracy(output_index, targets,
padding)
mask_accuracy = compute_mask_accuracy(inputs, output_index, targets,
padding)
correct_accuracy = compute_corrects_accuracy(output_index, targets,
padding)
loss.backward()
optimizer.step()