def scale_signed_dumping_worker(gpu, ngpus_per_node, config, tmp_path):
distributed_init_test_default(gpu, ngpus_per_node, config)
data_loader = create_rank_dataloader(config, gpu)
model = safe_thread_call(partial(squeezenet1_1, pretrained=True))
config.register_extra_structs([QuantizationRangeInitArgs(data_loader)])
quant_model, compression_ctrl = create_compressed_model_and_algo_for_test(
model, config)
compression_scheduler = compression_ctrl.scheduler
quant_model = post_compression_test_distr_init(compression_ctrl, config,
ngpus_per_node, quant_model)
criterion = torch.nn.MSELoss().cuda(config.gpu)
optimizer = torch.optim.Adam(quant_model.parameters(), lr=0.01)
torch.backends.cudnn.benchmark = True
# just to reproduce the same scale values without Dropout
quant_model.eval()
act_sum = 0
for layer in get_all_modules_by_type(quant_model,
"SymmetricQuantizer").values():
act_sum += layer.scale.sum()
ref_sum = 4447.291
assert act_sum.item() == approx(ref_sum, 0.01), \
'sum of scales is not expected {} vs {} rank {}'.format(act_sum.item(), ref_sum, config.rank)
out_file_path = get_path_after_broadcast(tmp_path, config.rank)
save_params(quant_model, out_file_path)
compression_scheduler.step()
for i, (input_, _) in enumerate(data_loader):
if i > 5:
break
output = quant_model(input_)
optimizer.zero_grad()
dummy_target = torch.randn(1000).cuda(config.gpu, non_blocking=True)
loss = criterion(output, dummy_target)
compression_scheduler.step()
loss.backward()
optimizer.step()
compression_scheduler.step()
out_file_path = get_path_path_after_train_iters(tmp_path, config.rank)
save_params(quant_model, out_file_path)
def hawq_dumping_worker(gpu, ngpus_per_node, config, tmp_path):
distributed_init_test_default(gpu, ngpus_per_node, config)
data_loader = create_rank_dataloader(config, gpu)
model = safe_thread_call(partial(mobilenet_v2, pretrained=True))
model.eval()
criterion = torch.nn.MSELoss().cuda(config.gpu)
config = register_default_init_args(config,
data_loader,
criterion,
autoq_eval_fn=lambda *x: 0,
autoq_eval_loader=data_loader)
quant_model, compression_ctrl = create_compressed_model_and_algo_for_test(
model, config)
quant_model = post_compression_test_distr_init(compression_ctrl, config,
ngpus_per_node, quant_model)
# just to reproduce the same scale values without Dropout
quant_model.eval()
act_bitwidth_per_scope = get_bitwidth_per_scope(quant_model.module)
out_file_path = get_path_to_bitwidth_dump(tmp_path, config.rank)
torch.save(act_bitwidth_per_scope, str(out_file_path))