def _async_copy(inputs, device_ids):
nr_devs = len(device_ids)
assert type(inputs) in (tuple, list)
assert len(inputs) == nr_devs
outputs = []
for i, dev in zip(inputs, device_ids):
with cuda.device(dev):
outputs.append(async_copy_to(i, dev))
return tuple(outputs)
def _async_copy_stream(inputs, device_ids):
nr_devs = len(device_ids)
assert type(inputs) in (tuple, list)
assert len(inputs) == nr_devs
outputs = []
streams = [_get_stream(d) for d in device_ids]
for i, dev, stream in zip(inputs, device_ids, streams):
with cuda.device(dev):
main_stream = cuda.current_stream()
with cuda.stream(stream):
outputs.append(async_copy_to(i, dev, main_stream=main_stream))
main_stream.wait_stream(stream)
return outputs
def _generic_fmm(proc_idx, queue, device_id):
# Unpack the function arguments
a: ArgsFmm = queue.get()
X1: torch.Tensor = a.X1
X2: torch.Tensor = a.X2
cuda_inputs = X1.is_cuda
out = a.out
kernel, gpu_dtype = a.kernel, a.gpu_dtype
max_mem = a.max_mem
num_streams = a.num_streams
# flags and local variables
change_dtype = gpu_dtype != X1.dtype
X1_equal_X2 = _gpu_tns_same_memory(X1, X2)
use_gpu_bufs = change_dtype or not cuda_inputs
stride = "F" if is_f_contig(out, strict=True) else "C"
j_iter = 0
dts = sizeof_dtype(gpu_dtype)
tc_device = torch.device('cuda:%d' % (int(device_id)))
avail_mem = max_mem / dts
# Choose block sizes n, m such that we won't run out of GPU memory
ntot, d = X1.shape
mtot = X2.shape[0]
extra_mem = kernel.extra_mem()
if cuda_inputs and not change_dtype:
# No allocation will be performed by us. Only in-kernel stuff.
n, m = select_dim_over_nm(max_n=ntot,
max_m=mtot,
d=d,
coef_nd=extra_mem.get('nd', 0),
coef_md=extra_mem.get('md', 0),
coef_nm=extra_mem.get('nm', 0),
coef_n=extra_mem.get('n', 0),
coef_m=extra_mem.get('m', 0),
rest=extra_mem.get('d', 0),
max_mem=avail_mem)
else:
n, m = select_dim_over_nm(
max_n=ntot,
max_m=mtot,
d=d,
coef_nd=num_streams * (extra_mem.get('nd', 0) + 1),
coef_md=num_streams * (extra_mem.get('md', 0) + 1),
coef_nm=num_streams * (extra_mem.get('nm', 0) + 1),
coef_n=extra_mem.get('n', 0),
coef_m=extra_mem.get('m', 0),
rest=extra_mem.get('d', 0),
max_mem=avail_mem)
# Create streams
streams = [tcd.Stream(device=tc_device) for _ in range(num_streams)]
# Create buffers
if use_gpu_bufs:
gX1 = create_same_stride((n, d), X1, gpu_dtype, tc_device)
gX2_list = [
create_same_stride((m, d), X2, gpu_dtype, tc_device)
for _ in range(num_streams)
]
gout_list = [
create_same_stride((n, m), out, gpu_dtype, tc_device)
for _ in range(num_streams)
]
if not cuda_inputs:
cpu_buf_list = [
create_same_stride((n, m), out, gpu_dtype, 'cpu', pin_memory=True)
for _ in range(num_streams)
]
# Define helpers for the copy-back operations (from cpu_buf to output)
copy_ops = [None] * num_streams
def wrap_copy_op(stream_idx):
if copy_ops[stream_idx] is not None:
copy_ops[stream_idx]()
copy_ops[stream_idx] = None
def do_copy_op(output, buf, i_, ic_, j_, jc_):
# This function will also do the type conversion
output[i_:i_ + ic_, j_:j_ + jc_].copy_(buf[:ic_, :jc_])
# Kernel computation begin
with tcd.device(tc_device):
for i in range(0, ntot, n):
ic = min(n, ntot - i)
with tcd.stream(streams[j_iter % len(streams)]):
X1_chunk = X1.narrow(0, i, ic)
if use_gpu_bufs:
cur_gX1 = gX1.narrow(0, 0, ic)
cur_gX1.copy_(X1_chunk, non_blocking=True)
else:
cur_gX1 = X1_chunk
for j in range(0, mtot, m):
jc = min(m, mtot - j)
# Choose the buffers for this inner iteration
stream_id = j_iter % len(streams)
stream = streams[stream_id]
if use_gpu_bufs:
gX2 = gX2_list[stream_id]
gout = gout_list[stream_id]
if not cuda_inputs:
cpu_buf = cpu_buf_list[stream_id]
# Sync for buffers we must use now (e.g. 2 previous iters)
with tcd.stream(stream): # Inner-loop
stream.synchronize()
wrap_copy_op(stream_id)
if X1_equal_X2 and j < i: # Shortcut for symmetric kernels
jc = min(m, mtot - j)
out[i:i + ic, j:j + jc].copy_(out[j:j + jc,
i:i + ic].T,
non_blocking=True)
j_iter += 1
continue
# Copy (CPU->GPU)
X2_chunk = X2.narrow(0, j, jc)
if use_gpu_bufs:
cur_gX2 = gX2.narrow(0, 0, jc)
cur_gX2.copy_(X2_chunk, non_blocking=True)
else:
cur_gX2 = X2_chunk
if use_gpu_bufs:
cur_gout = gout[:ic, :jc]
else:
cur_gout = out[i:i + ic, j:j + jc]
cur_gout.fill_(0.0)
# Compute
ddd = kernel._prepare(cur_gX1, cur_gX2)
kernel._apply(cur_gX1, cur_gX2.T, cur_gout)
cur_gout = kernel._finalize(cur_gout, ddd)
# Copy Back (GPU->CPU)
if not cuda_inputs:
# copy_ does not care about the contiguity of copies, as long as it's consistent
# however, in case of C-contiguous inputs it will create an intermediate array
# which is undesired. We use cuda_memcpy2d_async which works well with C-contiguous
# arrays.
if stride == "F":
copy_to_host(ic,
jc,
cur_gout,
0,
0,
cpu_buf,
0,
0,
s=stream)
else:
cuda_memcpy2d_async(dst=cpu_buf.data_ptr(),
dpitch=cpu_buf.stride(0) * dts,
src=cur_gout.data_ptr(),
spitch=cur_gout.stride(0) *
dts,
width=jc * dts,
height=ic,
stream=stream._as_parameter_)
copy_ops[stream_id] = partial(do_copy_op, out, cpu_buf,
i, ic, j, jc)
elif change_dtype:
out.narrow(0, i,
ic).narrow(1, j,
jc).copy_(cur_gout,
non_blocking=True)
j_iter += 1
for i in range(num_streams):
streams[i].synchronize()
wrap_copy_op(i)
return out
def prune_and_eval(rank, size, orig_fit, acc_constraint, valid, es, ref_model,
num_runs, final_results):
_valid = valid
gpu_id = GPU_ID
total_iterations = es.Tmax / es.popsize
individual_iter_count = 0
#ref_model = masked_models[rank]
X = torch.Tensor(copy.deepcopy(es.pop))
communicate_size = es.n + 4 # the size of tensors transfer accross computers
communicate_tensor = torch.FloatTensor(communicate_size * [0.])
fitness_list = []
itr_best_remain = 0
if rank == 0: # rank 0 is the main process to collect finesses
X.share_memory_()
#fitness_list = [torch.FloatTensor([0.0,0.1,0.2,0.3]).share_memory_() for i in range(size)]
fitness_list = [
torch.FloatTensor(communicate_size * [0.]).share_memory_()
for i in range(size)
]
if rank >= 1 and rank < size: # split tasks to different GPUs
gpu_id = other_GPU_IDs[rank - 1]
with cuda.device(gpu_id):
local_fields = onmt.IO.load_fields(torch.load(TRAIN_DATA +
'.vocab.pt'))
_valid.fields = local_fields # fields can not be packed, so reconstruct it in each threahds
while (individual_iter_count < total_iterations):
if rank == 0: # master node
itr_X = torch.Tensor(es.ask())
# broadcast the fathers
X.copy_(itr_X)
dist.broadcast(itr_X, 0)
else:
# recieve fathers from the source process
dist.broadcast(X, 0)
# apply MP on model
x = X.numpy()[rank]
ref_model.change_mask(x, apply_MP_on_mask)
ref_model.apply_mask()
# evaluate pruned network
fitness = evaluate(ref_model, _valid, local_fields)
communicate_tensor[0] = fitness[0]
communicate_tensor[1] = fitness[1]
communicate_tensor[2] = rank
communicate_tensor[3] = ref_model.get_sparsity()
for i in range(x.size):
communicate_tensor[i + 4] = X[rank, i] #x[i]
# sync fitness
if rank == 0: # collect fitness across processes
dist.gather(communicate_tensor, gather_list=fitness_list)
else:
dist.gather(communicate_tensor, dst=0)
# judge new solutions
if rank == 0: # negatively correlated search in master node
fit = []
X_ = []
for i in range(es.popsize):
the_fitness = 100
for j in range(len(
fitness_list)): # results of fitness evaluation
if int(fitness_list[j]
[2]) == i: # 0:ppl, 1:acc, 2:rank of individual
X_.append(fitness_list[j].numpy()[4:])
if orig_fit[1] - fitness_list[j][
1] <= acc_constraint:
the_fitness = -fitness_list[j][3]
else:
the_fitness = (orig_fit[1] - fitness_list[j][1]
) / acc_constraint
continue
fit.append(the_fitness)
es.tell(X_, fit)
itr_best_remain = min(fit)
final_results['result_NCS'].copy_(torch.Tensor(es.result()[0]))
individual_iter_count += 1
if rank == 0: # record status
logger.scalar_summary(
'ncs_%s_fitness' % num_runs,
es.result()[1],
num_runs * total_iterations + individual_iter_count)
logger.scalar_summary(
'ncs_%s_best_itr_remain' % num_runs, itr_best_remain,
num_runs * total_iterations + individual_iter_count)
logger.histo_summary(
'ncs_%s_pop' % num_runs,
es.result()[0],
num_runs * total_iterations + individual_iter_count)
logger.histo_summary(
'pop of 1', X_[0],
num_runs * total_iterations + individual_iter_count)
logger.scalar_summary(
'sp of 1', -fitness_list[0][3],
num_runs * total_iterations + individual_iter_count)
logger.scalar_summary(
'rank of 1', fitness_list[0][2],
num_runs * total_iterations + individual_iter_count)
logger.histo_summary(
'pop of 2', X_[1],
num_runs * total_iterations + individual_iter_count)
logger.scalar_summary(
'sp of 2', -fitness_list[1][3],
num_runs * total_iterations + individual_iter_count)
logger.scalar_summary(
'rank of 2', fitness_list[1][2],
num_runs * total_iterations + individual_iter_count)
#logger.histo_summary('pop of 3', X_[2], num_runs*total_iterations + individual_iter_count)
#logger.scalar_summary('sp of 3', -fitness_list[2][3], num_runs*total_iterations + individual_iter_count)
#logger.scalar_summary('rank of 3', fitness_list[2][2], num_runs*total_iterations + individual_iter_count)
ref_model.clear_cache()
def main():
data_path = "{}/data/penn".format(cfg.PROJECT_ROOT)
model_path = "{}/model/original_model/language_model/{}".format(
cfg.PROJECT_ROOT, 'lm_model_orignal.pt')
total_times = 20
run_times = 0
orginal_acc = 0
init_threshold = ...
start_t = time.time()
# get data
corpus = data.Corpus(data_path)
ntokens = len(corpus.dictionary)
eval_batch_size = TEST_BATCH_SIZE
train_data = batchify(corpus.train, TRAIN_BATCH_SIZE)
val_data = batchify(corpus.valid, TEST_BATCH_SIZE)
valid_data = val_data
test_data = batchify(corpus.test, TEST_BATCH_SIZE)
ref_model = None
# Load the best saved model.
with cuda.device(GPU_ID):
ff = open(model_path, 'rb')
ref_model = torch.load(ff)
ref_model.eval()
masked_model = MaskedModel(
ref_model, group_dict, cuda.current_device(), cuda.current_device(
)) # ref_model is at current_device, no copy will happen
#pdb.set_trace()
ff.close()
if GPU_ID:
cuda.set_device(GPU_ID)
print(time_now(), "get accuray of no pruning model")
masked_model.make_evaluable()
tmp_crate = len(masked_model.group_name_list) * [0]
masked_model.change_mask(tmp_crate, apply_MP_on_mask)
masked_model.apply_mask()
tmp_fit = evaluate_lm(masked_model.masked_model, valid_data, corpus,
TEST_BATCH_SIZE)
# 只需要原始的accuracy
acc_of_no_prune = tmp_fit[1]
fit_of_no_prune = tmp_fit
original_acc = acc_of_no_prune
pruning_arr = []
ppl_arr = []
#acc_of_no_prune = int(acc_of_no_prune*10)/10
print("init accuracy of model:", acc_of_no_prune)
print("accuracy constraint:", acc_percent_prune)
init_threshold = [0]
while run_times < total_times:
print("-----------------------------------------")
print("-----------------------------------------")
print("-----------------------------------------")
print("start Iteration ", run_times)
print("test model---------------")
LR = LR_INIT
previous_pr = None
previous_fit = None
best_pr = None
best_fit = None
for prune_rate in range(1, 100):
tmp_crate = len(masked_model.group_name_list) * [0.01 * prune_rate]
masked_model.change_mask(tmp_crate, apply_MP_on_mask)
masked_model.apply_mask()
tmp_fit = evaluate_lm(masked_model.masked_model, valid_data,
corpus, TEST_BATCH_SIZE)
print(
"each layer {} \% | {} % in total => validation acc {}\%, validation ppl {}"
.format(prune_rate,
masked_model.get_sparsity() * 100, tmp_fit[1] * 100.,
tmp_fit[0]))
if (not best_pr) and (tmp_fit[1] +
acc_percent_prune) < original_acc:
best_pr = previous_pr
best_fit = previous_fit
previous_pr = tmp_crate
previous_fit = tmp_fit
print('==============================')
print("The best pruning rates are: {}".format(best_pr))
if (not best_pr) or (best_pr[0] == init_threshold[0]):
print(
"Not better than last iteration of pruning, stop the process.")
exit()
masked_model.change_mask(best_pr, apply_MP_on_mask)
masked_model.apply_mask()
test_fit = evaluate_lm(masked_model.masked_model, test_data, corpus,
TEST_BATCH_SIZE)
print("{} \% => validation acc {}\%, validation ppl {}".format(
best_pr[0], best_fit[1] * 100., best_fit[0]))
print("{} \% => test acc {}\%, test ppl {}".format(
best_pr[0], test_fit[1] * 100., test_fit[0]))
print('==============================')
init_threshold = best_pr
saved_model_name = 'ncs_pruned_model_%s_iteration%s_%s_%s_acc_cons_%s.pt' % (
name_mark, run_times, Model_type, layer_group_type,
str(acc_percent_prune))
torch.save(masked_model.masked_model,
cfg.LM_MODEL_TMP_FOLDER + saved_model_name)
#--------------- start retraining --------------
model_for_train = masked_model
with open(cfg.LM_MODEL_TMP_FOLDER + saved_model_name, 'rb') as f:
model_tmp_load = torch.load(f)
model_for_train.masked_model = model_tmp_load
model_for_train.change_mask(init_threshold, apply_MP_on_mask)
model_for_train.apply_mask()
model_for_train.make_trainable()
recovered = False
best_val_loss = None
try:
for epoch in range(1, RETRAIN_EPOCHS + 1):
epoch_start_time = time.time()
train(model_for_train, ntokens, train_data, TRAIN_BATCH_SIZE,
SEQ_LEN, corpus, GRAD_CLIP, TRAIN_LOG_INTERVAL, epoch)
val_eval = evaluate_lm(model_for_train.masked_model, val_data,
corpus, TEST_BATCH_SIZE)
print('-' * 89)
print(
'| end of epoch {:3d} | time: {:5.2f}s | valid acc {:5.2f} | '
'valid ppl {:8.2f}'.format(
epoch, (time.time() - epoch_start_time), val_eval[1],
val_eval[0]))
val_loss = val_eval[2]
print('-' * 89)
# Save the model if the validation loss is the best we've seen so far.
if not best_val_loss or val_loss < best_val_loss:
with open(
"{}/{}{}_iterative_retrain_model_runtime{}_epoch_{}.pt"
.format(cfg.LM_MODEL_PATH, name_mark,
acc_percent_prune, run_times, epoch),
'wb') as f:
torch.save(model_for_train, f)
best_val_loss = val_loss
else:
# Anneal the learning rate if no improvement has been seen in the validation dataset.
LR /= 4.0
if val_eval[1] >= original_acc:
recovered = True
break
except KeyboardInterrupt:
print('-' * 89)
print('Exiting from training early')
print(time_now(), "finish retraining ")
if not recovered:
exit()
else:
print("------------Accuracy recorverd!--------------------")
print("recovered accuracy (>= {})".format(acc_of_no_prune))
model_for_train.make_evaluable()
model_for_train.apply_mask()
ref_model = model_for_train.masked_model
print("validate acc of the model---------------")
tmp_fit = evaluate_lm(ref_model, valid_data, corpus, TEST_BATCH_SIZE)
print('ref_model', 'acc (%.4f), ppl (%.4f)' % (tmp_fit[1], tmp_fit[0]))
print("-------------print TEST evaluation info ---------------")
tmp_fit = evaluate_lm(ref_model, test_data, corpus, TEST_BATCH_SIZE)
print('percentage %s => acc (%.4f), ppl (%.4f)' %
(init_threshold[0] * 100, tmp_fit[1], tmp_fit[0]))
masked_model = model_for_train
run_times += 1
def _work(process_id, model, dataset, args):
databin = dataset[process_id]
n_gpus = torch.cuda.device_count()
data_loader = DataLoader(databin,
shuffle=False,
num_workers=args.num_workers // n_gpus,
pin_memory=False)
with torch.no_grad(), cuda.device(process_id):
model.cuda()
for iter, pack in enumerate(data_loader):
img_name = pack['name'][0]
label = pack['label'][0]
size = pack['size']
strided_size = imutils.get_strided_size(size, 4)
strided_up_size = imutils.get_strided_up_size(size, 16)
outputs = [
model(img[0].cuda(non_blocking=True)) for img in pack['img']
]
strided_cam = torch.sum(
torch.stack([
F.interpolate(torch.unsqueeze(o, 0),
strided_size,
mode='bilinear',
align_corners=False)[0] for o in outputs
]), 0)
highres_cam = [
F.interpolate(torch.unsqueeze(o, 1),
strided_up_size,
mode='bilinear',
align_corners=False) for o in outputs
]
highres_cam = torch.sum(torch.stack(highres_cam, 0),
0)[:, 0, :size[0], :size[1]]
valid_cat = torch.nonzero(label)[:, 0]
strided_cam = strided_cam[valid_cat]
strided_cam /= F.adaptive_max_pool2d(strided_cam, (1, 1)) + 1e-5
highres_cam = highres_cam[valid_cat]
highres_cam /= F.adaptive_max_pool2d(highres_cam, (1, 1)) + 1e-5
# save cams
np.save(
os.path.join(args.cam_out_dir, img_name + '.npy'), {
"keys": valid_cat,
"cam": strided_cam.cpu(),
"high_res": highres_cam.cpu().numpy()
})
if process_id == n_gpus - 1 and iter % (len(databin) // 20) == 0:
print("%d " % ((5 * iter + 1) // (len(databin) // 20)), end='')
def main():
data_path = "{}/data/penn".format(DATA_PATH)
model_path = "{}/deepModels/torch_models/language-model/{}".format(MODEL_PATH, 'model.pt')
#model_path = "{}/deepModels/torch_models/language-model/{}".format(MODEL_PATH, 'lstm_3layer.pt')
total_times = 1
run_times = 0
orginal_acc = 0
init_threshold = ...
start_t = time.time()
# get data
corpus = data.Corpus(data_path)
ntokens = len(corpus.dictionary)
eval_batch_size = TEST_BATCH_SIZE
train_data = batchify(corpus.train, TRAIN_BATCH_SIZE)
val_data = batchify(corpus.valid, TEST_BATCH_SIZE)
valid_data = val_data
test_data = batchify(corpus.test, TEST_BATCH_SIZE)
ref_model = None
# Load the best saved model.
masked_models = []
with cuda.device(GPU_ID):
ff = open(model_path, 'rb')
ref_model = torch.load(ff)
ref_model.eval()
masked_model = MaskedModel(ref_model, group_dict, cuda.current_device(), cuda.current_device()) # ref_model is at current_device, no copy will happen
#pdb.set_trace()
masked_models.append(masked_model)
ff.close()
if GPU_ID:
cuda.set_device(GPU_ID)
print(time_now(), "get accuray of no pruning model")
masked_model.make_evaluable()
tmp_crate = len(masked_model.group_name_list)*[0]
masked_model.change_mask(tmp_crate, apply_MP_on_mask)
masked_model.apply_mask()
tmp_fit = evaluate_lm(masked_model.masked_model, valid_data, corpus, TEST_BATCH_SIZE)
# 只需要原始的accuracy
acc_of_no_prune = tmp_fit[1]
fit_of_no_prune = tmp_fit
original_acc = acc_of_no_prune
#acc_of_no_prune = int(acc_of_no_prune*10)/10
print("=============TiPO start========================")
print("init accuracy of model:", acc_of_no_prune)
print("accuracy constraint:", acc_percent_prune)
previous_pr = None
best_pr = None
ncs_std = 0.05
while run_times < total_times:
print("-----------------------------------------")
print("-----------------------------------------")
print("-----------------------------------------")
print("start Iteration ", run_times)
print("test model---------------")
LR = LR_INIT
#ref_model.generator.eval()
print("test model---------------")
masked_models[0].make_evaluable()
tmp_fit = evaluate_lm(masked_models[0].masked_model, valid_data, corpus, TEST_BATCH_SIZE)
print('masked_models[0]','acc (%.4f), ppl (%.4f)' % (tmp_fit[1], tmp_fit[0]))
if run_times == 0:
init_threshold = len(masked_models[0].group_name_list) * [0.6]
itr_time = time.time()
for gpu_candidate in other_GPU_IDs:
with cuda.device(gpu_candidate):
masked_models.append(MaskedModel(ref_model, group_dict, GPU_ID, gpu_candidate)) # if the gpu_candidate is the same as ref_model, it will return the ref_model
#------------- Here -------------------------
# del ref_model
# do pruning
ncs_start = time.time()
print('Itration %d, model loading: %d sec' % (run_times, ncs_start - itr_time))
print("init threshold:", init_threshold)
best_found, saved_model, best_masked_model = NCS_MP(init_threshold, ncs_std, masked_models, valid_data, corpus, acc_percent_prune, fit_of_no_prune, run_times)
#best_found, saved_model, best_masked_model = init_threshold, '/raid/lab_tk/liguiying/deepModels/torch_models/language-model/prune_tmp/ncs_pruned_model_test_iteration0_LM_time_acc_cons_0.01.pt', masked_models[0]
init_threshold = best_found
#best_found, saved_model, best_masked_model = NCS_MP(init_threshold, 0.05, fields, masked_models, valid_data, 0.01, run_times, checkpoint)
end_t = time.time()
print('NCS Time: {} min'.format((end_t - itr_time)/60.))
print('Best found thresholds:')
for i in range(len(masked_models[0].group_name_list)):
print("layer {}: {}%".format(masked_models[0].group_name_list[i], 100*best_found[i]))
print("TEST PPL evaluation:")
tmp_fit = evaluate_lm(best_masked_model.masked_model, test_data, corpus, TEST_BATCH_SIZE)
print('Finsished => acc (%.4f percent), ppl (%.4f)' % (tmp_fit[1]*100, tmp_fit[0]))
# clear no used models
for gpu_model in masked_models:
del gpu_model
if not best_pr :
best_pr = best_masked_model.get_sparsity()
else:
tmp_pr = best_masked_model.get_sparsity()
if best_pr > tmp_pr:
print("No improvement! Stop the PROCESS.")
exit()
elif best_pr == tmp_pr:
if tmp_fit[1] <fit_of_no_prune[1]:
ncs_std /= 10
else:
ncs_std *= 10
else:
best_pr = tmp_pr
#if run_times % 5 == 0:
# ncs_std /= 10
#--------------- start retraining --------------
model_for_train = best_masked_model
#pretrained_leaf_dict = model_for_train.make_trainable()
#print(model_for_train.map_dict.keys())
#pdb.set_trace()
#fix_no_leaf(model_for_train, pretrained_leaf_dict)
#pdb.set_trace()
with open(SAVE_MODEL_TMP_FOLDER + saved_model, 'rb') as f:
model_tmp_load = torch.load(f)
model_for_train.masked_model = model_tmp_load.masked_model
model_for_train.change_mask(init_threshold, apply_MP_on_mask)
model_for_train.apply_mask()
model_for_train.make_trainable()
recovered = False
best_val_loss = None
try:
for epoch in range(1, RETRAIN_EPOCHS + 1):
epoch_start_time = time.time()
train(model_for_train, ntokens, train_data, TRAIN_BATCH_SIZE, SEQ_LEN, corpus, GRAD_CLIP, TRAIN_LOG_INTERVAL, epoch)
val_eval = evaluate_lm(model_for_train.masked_model, val_data, corpus, TEST_BATCH_SIZE)
print('-' * 89)
print('| end of epoch {:3d} | time: {:5.2f}s | valid acc {:5.2f} | '
'valid ppl {:8.2f}'.format(epoch, (time.time() - epoch_start_time),
val_eval[1], val_eval[0]))
val_loss = val_eval[2]
print('-' * 89)
# Save the model if the validation loss is the best we've seen so far.
if not best_val_loss or val_loss < best_val_loss:
with open("{}/{}{}_iterative_retrain_model_runtime{}_epoch_{}.pt".format(SAVE_MODEL_FOLDER, name_mark, acc_percent_prune, run_times, epoch), 'wb') as f:
torch.save(model_for_train, f)
best_val_loss = val_loss
else:
# Anneal the learning rate if no improvement has been seen in the validation dataset.
LR /= 4.0
if val_eval[1] >= original_acc:
recovered = True
except KeyboardInterrupt:
print('-' * 89)
print('Exiting from training early')
print(time_now(), "finish retraining ")
if not recovered:
print("NOT RECORVER!")
exit()
else:
print("------------Accuracy recorverd!--------------------")
print("recovered accuracy (>= {})".format(acc_of_no_prune))
model_for_train.make_evaluable()
model_for_train.apply_mask()
ref_model = model_for_train.masked_model
masked_models = [MaskedModel(ref_model, group_dict, cuda.current_device(), cuda.current_device())]
print("validate acc of the model---------------")
tmp_fit = evaluate_lm(ref_model, valid_data, corpus, TEST_BATCH_SIZE)
print('ref_model','acc (%.4f), ppl (%.4f)' % (tmp_fit[1], tmp_fit[0]))
tmp_fit = evaluate_lm(masked_models[0].masked_model, valid_data, corpus, TEST_BATCH_SIZE)
print('masked_models[0]','acc (%.4f), ppl (%.4f)' % (tmp_fit[1], tmp_fit[0]))
print('------------- save checkpoint ---------------')
saved_model = update_checkpoint(model_for_train, run_times, acc_percent_prune, t=True)
print(time_now(), ' saving model:', saved_model)
print("-------------print TEST evaluation info ---------------")
tmp_fit = evaluate_lm(model_for_train.masked_model, test_data, corpus, TEST_BATCH_SIZE)
print('percentage %s => acc (%.4f), ppl (%.4f)' % (model_for_train.get_sparsity()*100, tmp_fit[1], tmp_fit[0]))
run_times += 1
def distk_fdmmv(proc_idx, queue, device_id):
a: ArgsFdmmv = queue.get()
X1, X2, v, w, out = a.X1, a.X2, a.v, a.w, a.out
kernel: L2DistanceKernel = a.kernel
max_mem = a.max_mem
N, D = X1.size()
M = X2.size(0)
T = v.size(1) if v is not None else w.size(1)
dtype = X1.dtype
# Memory usage:
# v : M x T
# K : n x M
# X1ss : n x d
# X2s : M x d
# Kv : n x T
# out : M x T
# sq1 : n x 1
# sq2 : M x 1
# ------------
# total : n*d + M*d + n*(M + T + 1) + 2*M*T + M
avail_mem = max_mem / sizeof_dtype(dtype)
# FIXME: There seems to be a bug where if we let avail_mem like it is
# for 32-bit data-types some copy fails. In such case we need
# to free up some more memory and then everything runs fine.
rest_coef = 2 * M * T if v is not None else M * T
n, d = select_dim_over_d(maxD=D,
maxN=N,
coef_nd=1,
coef_n=M + T + 1,
coef_d=M,
rest=rest_coef + M,
tot=avail_mem)
ddev = torch.device('cuda:%d' % int(device_id))
s1 = tcd.Stream()
s2 = tcd.Stream()
with tcd.device(ddev), tcd.stream(s1):
if v is not None:
v_gpu = create_same_stride((M, T), v, dtype, ddev)
copy_to_device_noorder(M, T, v, 0, 0, v_gpu, 0, 0)
K_gpu = create_same_stride((n, M), X1, dtype, ddev)
X1ss_gpu = create_same_stride((n, d), X1, dtype, ddev)
X2s_gpu = create_same_stride((M, d), X2, dtype, ddev)
Kv_gpu = create_same_stride((n, T), X1, dtype, ddev)
if out.is_cuda:
out_gpu = out
else:
out_gpu = create_same_stride((M, T), out, dtype, ddev)
out_gpu.fill_(0.0)
sq1_gpu = create_same_stride((n, ), X1, dtype, ddev)
sq2_gpu = create_same_stride((M, ), X1, dtype, ddev)
#if (d == D):
# with torch.cuda.stream(s2):
# cur_X2s_gpu = copy_to_device_noorder(M, d, X2, 0, 0, X2s_gpu, 0, 0, s=s2)
# torch.norm(cur_X2s_gpu, p=2, dim=1, keepdim=True, out=sq2_gpu).pow_(2)
for i in range(0, N, n):
nb = min(N - i, n)
cur_K_gpu = K_gpu.narrow(0, 0, nb) # nb x M
cur_K_gpu.fill_(0.0)
for j in range(0, D, d):
db = min(D - j, d)
# Parallelize two matrix transfers (probably pointless)
#if d < D:
with torch.cuda.stream(s2):
cur_X2s_gpu = copy_to_device_noorder(M,
db,
X2,
0,
j,
X2s_gpu,
0,
0,
s=s2)
torch.norm(cur_X2s_gpu,
p=2,
dim=1,
keepdim=True,
out=sq2_gpu).pow_(2)
cur_X1ss_gpu = copy_to_device_noorder(nb,
db,
X1,
i,
j,
X1ss_gpu,
0,
0,
s=s1)
torch.norm(cur_X1ss_gpu, p=2, dim=1, keepdim=True,
out=sq1_gpu).pow_(2)
s2.synchronize()
s1.synchronize()
cur_K_gpu.addmm_(mat1=cur_X1ss_gpu,
mat2=cur_X2s_gpu.T,
alpha=-2.0)
cur_K_gpu.add_(sq1_gpu)
cur_K_gpu.add_(sq2_gpu.T)
cur_K_gpu.clamp_min_(0)
cur_K_gpu = kernel._transform(cur_K_gpu)
if w is not None:
# Copy split w to GPU into cur_Kv_gpu,
cur_Kv_gpu = copy_to_device_noorder(nb,
T,
w,
i,
0,
Kv_gpu,
0,
0,
s=s1) # n x T
if v is not None:
cur_Kv_gpu.addmm_(cur_K_gpu, v_gpu)
else:
# v cannot be None if w is None
cur_Kv_gpu = Kv_gpu.narrow(0, 0, nb) # n x T
torch.mm(cur_K_gpu, v_gpu, out=cur_Kv_gpu) # n x T
# Multiply transposed kernel with the Kv result.
out_gpu.addmm_(cur_K_gpu.T, cur_Kv_gpu) # M x T
s1.synchronize()
s1.synchronize()
if not out.is_cuda:
copy_to_host_noorder(M, T, out_gpu, 0, 0, out, 0, 0)
return out
def generic_fdmmv(proc_idx, queue, device_id):
a: ArgsFdmmv = queue.get()
X1, X2, v, w, out = a.X1, a.X2, a.v, a.w, a.out
kernel, max_mem = a.kernel, a.max_mem
dtype = X1.dtype
N, D = X1.size()
M = X2.size(0)
if v is None:
T = w.size(1)
else:
T = v.size(1)
# Memory usage:
# v : M x T
# K : n x M
# X1d : n x d
# X2d : M x d
# Kv : n x T
# out2 : M x T
# sq1 : n x 1
# sq2 : M x 1
# ------------
# total : n*d + M*d + n*(M + T) + 2*M*T + M
avail_mem = max_mem / sizeof_dtype(dtype)
# FIXME: There seems to be a bug where if we let avail_mem like it is
# for 32-bit data-types some copy fails. In such case we need
# to free up some more memory and then everything runs fine.
if sizeof_dtype(dtype) == 4:
avail_mem /= 2
rest_coef = 2 * M * T if v is not None else M * T
n, d = select_dim_over_d(maxD=D,
maxN=N,
coef_nd=1,
coef_n=M + T + 1,
coef_d=M,
rest=rest_coef + M,
tot=avail_mem)
ddev = torch.device('cuda:%d' % int(device_id))
with tcd.device(ddev):
# Initialize GPU data
ker_gpu = create_same_stride((n, M), out, dtype=dtype, device=ddev)
X1s_gpu = create_same_stride((n, d), X1, dtype, ddev)
X2s_gpu = create_same_stride((M, d), X2, dtype, ddev)
w_gpu = create_same_stride((n, T), ker_gpu, dtype, ddev)
if out.is_cuda:
out_gpu = out
else:
out_gpu = create_same_stride((M, T), out, dtype, ddev)
out_gpu.fill_(0.0)
if v is not None:
v_gpu = v.to(device=ddev) # M x T
for i in range(0, N, n):
ic = min(n, N - i)
ddd = kernel._prepare(X1.narrow(0, i, ic), X2)
c_g_ker = ker_gpu.narrow(0, 0, ic)
c_g_ker.fill_(0.0)
for k in range(0, D, d):
kc = min(d, D - k)
c_g_X1s = copy_to_device_noorder(ic, kc, X1, i, k, X1s_gpu, 0,
0)
c_g_X2s = copy_to_device_noorder(M, kc, X2, 0, k, X2s_gpu, 0,
0)
kernel._apply(c_g_X1s, c_g_X2s.T, c_g_ker)
kernel._finalize(c_g_ker, ddd)
if w is not None:
c_g_w = copy_to_device_noorder(ic, T, w, i, 0, w_gpu, 0, 0)
else:
c_g_w = w_gpu.narrow(0, 0, ic)
c_g_w.fill_(0.0)
if v is not None:
c_g_w.addmm_(c_g_ker, v_gpu)
out_gpu.addmm_(c_g_ker.T, c_g_w)
if not out.is_cuda:
copy_to_device_noorder(M, T, out_gpu, 0, 0, out, 0, 0)
return out
def main():
data_path = "{}/data/penn".format(DATA_PATH)
model_path = "{}/deepModels/SiPO/original_model/language_model/{}".format(MODEL_PATH, 'lm_model_orignal.pt')
total_times = 12
run_times = 0
orginal_acc = 0
init_threshold = ...
start_t = time.time()
# get data
corpus = data.Corpus(data_path)
ntokens = len(corpus.dictionary)
eval_batch_size = TEST_BATCH_SIZE
train_data = batchify(corpus.train, TRAIN_BATCH_SIZE)
val_data = batchify(corpus.valid, TEST_BATCH_SIZE)
valid_data = val_data
test_data = batchify(corpus.test, TEST_BATCH_SIZE)
ref_model = None
# Load the best saved model.
with cuda.device(GPU_ID):
ff = open(model_path, 'rb')
ref_model = torch.load(ff)
ref_model.eval()
masked_model = MaskedModel(ref_model, group_dict, cuda.current_device(), cuda.current_device()) # ref_model is at current_device, no copy will happen
#pdb.set_trace()
ff.close()
if GPU_ID:
cuda.set_device(GPU_ID)
print(time_now(), "get accuray of no pruning model")
masked_model.make_evaluable()
tmp_crate = len(masked_model.group_name_list)*[0]
masked_model.change_mask(tmp_crate, apply_MP_on_mask)
masked_model.apply_mask()
tmp_fit = evaluate_lm(masked_model.masked_model, valid_data, corpus, TEST_BATCH_SIZE)
# 只需要原始的accuracy
acc_of_no_prune = tmp_fit[1]
fit_of_no_prune = tmp_fit
original_acc = acc_of_no_prune
pruning_arr = []
ppl_arr = []
#acc_of_no_prune = int(acc_of_no_prune*10)/10
print("init accuracy of model:", acc_of_no_prune)
print("accuracy constraint:", acc_percent_prune)
init_threshold = [0]
print("-----------------------------------------")
print("-----------------------------------------")
print("-----------------------------------------")
SAVE_MODEL_FOLDER = '/fl/checkpoint/language-mode/prune/'
print("test model---------------")
LR = LR_INIT
previous_pr = [0.01] * 4
previous_fit = [0.01, 12.3]
best_pr = [0.01] * 4
best_fit = [0.01, 12.3]
for prune_rate in range(1, 100):
tmp_crate = len(masked_model.group_name_list)*[0.01 * prune_rate]
masked_model.change_mask(tmp_crate, apply_MP_on_mask)
masked_model.apply_mask()
tmp_fit = evaluate_lm(masked_model.masked_model, valid_data, corpus, TEST_BATCH_SIZE)
print("each layer {} \% | {} % in total => validation acc {}\%, validation ppl {}".format(prune_rate, masked_model.get_sparsity()*100, tmp_fit[1]*100., tmp_fit[0]))
if (tmp_fit[1] + acc_percent_prune) > original_acc:
best_pr = previous_pr
best_fit = previous_fit
previous_pr = tmp_crate
previous_fit = tmp_fit
test_fit = evaluate_lm(masked_model.masked_model, test_data, corpus, TEST_BATCH_SIZE)
print("{} \% => validation acc {}\%, validation ppl {}".format(best_pr[0], best_fit[1]*100., best_fit[0]))
print("{} \% => test acc {}\%, test ppl {}".format(best_pr[0], test_fit[1]*100., test_fit[0]))
print('==============================')
masked_model.apply_mask_init()
init_threshold = best_pr
saved_model_name = 'ncs_pruned_model_%s_iteration%s_%s_%s_acc_cons_%s.pt' % (name_mark, run_times, Model_type, layer_group_type, str(acc_percent_prune))
torch.save(masked_model.masked_model, SAVE_MODEL_FOLDER+saved_model_name)
#--------------- start retraining --------------
model_for_train = masked_model
with open(SAVE_MODEL_FOLDER + saved_model_name, 'rb') as f:
model_tmp_load = torch.load(f)
model_for_train.masked_model = model_tmp_load
model_for_train.change_mask(init_threshold, apply_MP_on_mask)
model_for_train.apply_mask()
model_for_train.make_trainable()
recovered = False
best_val_loss = None
train(model_for_train, ntokens, train_data, TRAIN_BATCH_SIZE, SEQ_LEN, corpus, GRAD_CLIP, TRAIN_LOG_INTERVAL, prune_rate)
print(time_now(), "finish retraining ")
print("------------Accuracy recorverd!--------------------")
print("recovered accuracy (>= {})".format(acc_of_no_prune))
model_for_train.make_evaluable()
model_for_train.apply_mask()
ref_model = model_for_train.masked_model
print("validate acc of the model---------------")
tmp_fit = evaluate_lm(ref_model, valid_data, corpus, TEST_BATCH_SIZE)
print('ref_model','acc (%.4f), ppl (%.4f)' % (tmp_fit[1], tmp_fit[0]))
print("-------------print TEST evaluation info ---------------")
tmp_fit = evaluate_lm(ref_model, test_data, corpus, TEST_BATCH_SIZE)
print('percentage %s => acc (%.4f), ppl (%.4f)' % (init_threshold[0]*100, tmp_fit[1], tmp_fit[0]))
masked_model = model_for_train
print('==============================')
print("The best pruning rates are: {}".format(best_pr))
masked_model.change_mask(best_pr, apply_MP_on_mask)
masked_model.apply_mask()
test_fit = evaluate_lm(masked_model.masked_model, test_data, corpus, TEST_BATCH_SIZE)
print("{} \% => validation acc {}\%, validation ppl {}".format(best_pr[0], best_fit[1]*100., best_fit[0]))
print("{} \% => test acc {}\%, test ppl {}".format(best_pr[0], test_fit[1]*100., test_fit[0]))
print('==============================')
def main():
data_path = "{}/data/penn".format(DATA_PATH)
model_path = "{}/deepModels/torch_models/language-model/{}".format(
MODEL_PATH, 'model.pt')
total_times = 12
run_times = 0
orginal_acc = 0
init_threshold = ...
start_t = time.time()
# get data
corpus = data.Corpus(data_path)
ntokens = len(corpus.dictionary)
eval_batch_size = TEST_BATCH_SIZE
train_data = batchify(corpus.train, TRAIN_BATCH_SIZE)
val_data = batchify(corpus.valid, TEST_BATCH_SIZE)
valid_data = val_data
test_data = batchify(corpus.test, TEST_BATCH_SIZE)
ref_model = None
masked_model = None
# Load the best saved model.
with cuda.device(GPU_ID):
ff = open(model_path, 'rb')
ref_model = torch.load(ff)
ref_model.eval()
masked_model = MaskedModel(
ref_model, group_dict, cuda.current_device(), cuda.current_device(
)) # ref_model is at current_device, no copy will happen
ff.close()
if GPU_ID:
cuda.set_device(GPU_ID)
if layer_group_type == 'simple':
print("MP start for LM")
elif layer_group_type == 'layer':
print("LMP start for LM")
print(time_now(), "get accuray of no pruning model")
masked_model.make_evaluable()
tmp_crate = len(masked_model.group_name_list) * [0]
masked_model.change_mask(tmp_crate, apply_MP_on_mask)
masked_model.apply_mask()
tmp_fit = evaluate_lm(masked_model.masked_model, valid_data, corpus,
TEST_BATCH_SIZE)
# 只需要原始的accuracy
acc_of_no_prune = tmp_fit[1]
fit_of_no_prune = tmp_fit
original_acc = acc_of_no_prune
#acc_of_no_prune = int(acc_of_no_prune*10)/10
print("init accuracy of model:", acc_of_no_prune)
print("accuracy constraint:", acc_percent_prune)
# the best pruning rate
previous_pr = None
previous_fit = None
best_pr = None
best_fit = None
for prune_rate in range(1, 100):
tmp_crate = len(masked_model.group_name_list) * [0.01 * prune_rate]
masked_model.change_mask(tmp_crate, apply_MP_on_mask)
masked_model.apply_mask()
tmp_fit = evaluate_lm(masked_model.masked_model, valid_data, corpus,
TEST_BATCH_SIZE)
print("{} \% => validation acc {}\%, validation ppl {}".format(
prune_rate, tmp_fit[1] * 100., tmp_fit[0]))
if (not best_pr) and (tmp_fit[1] + acc_percent_prune) < original_acc:
best_pr = previous_pr
best_fit = previous_fit
previous_pr = tmp_crate
previous_fit = tmp_fit
print('==============================')
print("The best pruning rates are: {}".format(best_pr))
masked_model.change_mask(best_pr, apply_MP_on_mask)
masked_model.apply_mask()
test_fit = evaluate_lm(masked_model.masked_model, test_data, corpus,
TEST_BATCH_SIZE)
print("{} \% => validation acc {}\%, validation ppl {}".format(
best_pr[0], best_fit[1] * 100., best_fit[0]))
print("{} \% => test acc {}\%, test ppl {}".format(best_pr[0],
test_fit[1] * 100.,
test_fit[0]))
print('==============================')
def main():
data_path = "{}/data/penn".format(cfg.PROJECT_ROOT)
model_path = "{}/model/original_model/language_model/{}".format(cfg.PROJECT_ROOT, 'lm_model_orignal.pt')
total_times = 2
run_times = 0
orginal_acc = 0
init_threshold = ...
start_t = time.time()
# get data
corpus = data.Corpus(data_path)
ntokens = len(corpus.dictionary)
eval_batch_size = TEST_BATCH_SIZE
train_data = batchify(corpus.train, TRAIN_BATCH_SIZE)
val_data = batchify(corpus.valid, TEST_BATCH_SIZE)
valid_data = val_data
test_data = batchify(corpus.test, TEST_BATCH_SIZE)
ref_model = None
# Load the best saved model.
masked_models = []
with cuda.device(GPU_ID):
ff = open(model_path, 'rb')
ref_model = torch.load(ff)
ref_model.eval()
masked_model = MaskedModel(ref_model, group_dict, cuda.current_device(), cuda.current_device()) # ref_model is at current_device, no copy will happen
#pdb.set_trace()
masked_models.append(masked_model)
ff.close()
if GPU_ID:
cuda.set_device(GPU_ID)
print(time_now(), "get accuray of no pruning model")
masked_model.make_evaluable()
tmp_crate = len(masked_model.group_name_list)*[0]
masked_model.change_mask(tmp_crate, apply_MP_on_mask)
masked_model.apply_mask()
tmp_fit = evaluate_lm(masked_model.masked_model, valid_data, corpus, TEST_BATCH_SIZE)
# 只需要原始的accuracy
acc_of_no_prune = tmp_fit[1]
fit_of_no_prune = tmp_fit
original_acc = acc_of_no_prune
pruning_arr = []
ppl_arr = []
#acc_of_no_prune = int(acc_of_no_prune*10)/10
print("init accuracy of model:", acc_of_no_prune)
print("accuracy constraint:", acc_percent_prune)
while run_times < total_times:
print("-----------------------------------------")
print("-----------------------------------------")
print("-----------------------------------------")
print("start Iteration ", run_times)
print("test model---------------")
LR = LR_INIT
ref_model.eval()
#ref_model.generator.eval()
tmp_fit = evaluate_lm(ref_model, valid_data, corpus, TEST_BATCH_SIZE)
print('ref_model','acc (%.4f), ppl (%.4f)' % (tmp_fit[1], tmp_fit[0]))
print("test model---------------")
masked_models[0].make_evaluable()
tmp_fit = evaluate_lm(masked_models[0].masked_model, valid_data, corpus, TEST_BATCH_SIZE)
print('masked_models[0]','acc (%.4f), ppl (%.4f)' % (tmp_fit[1], tmp_fit[0]))
itr_time = time.time()
for gpu_candidate in other_GPU_IDs:
with cuda.device(gpu_candidate):
masked_models.append(MaskedModel(ref_model, group_dict, GPU_ID, gpu_candidate)) # if the gpu_candidate is the same as ref_model, it will return the ref_model
#------------- Here -------------------------
# del ref_model
# do pruning
ncs_start = time.time()
print('Itration %d, model loading: %d sec' % (run_times, ncs_start - itr_time))
if run_times == 0:
init_threshold = len(masked_models[0].group_name_list)*[0.10]
print("init threshold:", init_threshold)
best_found, saved_model, best_masked_model = NCS_MP(init_threshold, 0.05, masked_models, valid_data, corpus, acc_percent_prune, fit_of_no_prune, run_times)
#best_found, saved_model, best_masked_model = init_threshold, '/raid/lab_tk/liguiying/deepModels/torch_models/language-model/prune_tmp/ncs_pruned_model_test_iteration0_LM_time_acc_cons_0.01.pt', masked_models[0]
#init_threshold = best_found
#best_found, saved_model, best_masked_model = NCS_MP(init_threshold, 0.05, fields, masked_models, valid_data, 0.01, run_times, checkpoint)
end_t = time.time()
print('NCS Time: {} min'.format((end_t - itr_time)/60.))
print('Best found thresholds:')
for i in range(len(masked_models[0].group_name_list)):
print("layer {}: {}%".format(masked_models[0].group_name_list[i], 100*best_found[i]))
print("TEST PPL evaluation:")
tmp_fit = evaluate_lm(best_masked_model.masked_model, test_data, corpus, TEST_BATCH_SIZE)
print('Finsished => acc (%.4f percent), ppl (%.4f)' % (tmp_fit[1]*100, tmp_fit[0]))
pruning_arr.append(masked_models[0].get_sparsity())
ppl_arr.append(tmp_fit[0])
# clear no used models
for gpu_model in masked_models:
del gpu_model
run_times += 1
pruning_arr = np.array(pruning_arr)
ppl_arr = np.array(ppl_arr)
print("Prunig rate : mean({}) std({})".format(pruning_arr.mean(), pruning_arr.std()))
print("PPL : mean({}) std({})".format(ppl_arr.mean(), ppl_arr.std()))
proc_pipe = transf.Compose(
[transf.ToPILImage(), img_resize,
transf.ToTensor(), img_norm])
train_dir = 'train_images'
val_dir = 'test_images'
train_loader = DataLoader(DataFeed(train_dir,
nat_sort=True,
transform=proc_pipe),
batch_size=batch_size,
shuffle=False)
val_loader = DataLoader(DataFeed(val_dir, nat_sort=True, transform=proc_pipe),
batch_size=batch_size,
shuffle=False)
# Network training:
with cuda.device(0):
top_1 = np.zeros((1, len(train_size)))
top_2 = np.zeros((1, len(train_size)))
top_3 = np.zeros((1, len(train_size)))
acc_loss = 0
itr = []
for idx, n in enumerate(train_size):
print('```````````````````````````````````````````````````````')
print('Training size is {}'.format(n))
# Build the network:
net = resnet18_mod(pretrained=True, progress=True, num_classes=64)
net = net.cuda()
layers = list(net.children())
# Optimization parameters:
criterion = nn.CrossEntropyLoss()
def _work(process_id, model, dataset, args):
n_gpus = torch.cuda.device_count()
databin = dataset[process_id]
# data_loader = DataLoader(databin,
# shuffle=False, num_workers=args.num_workers // n_gpus, pin_memory=False)
data_loader = DataLoader(databin, shuffle=False, pin_memory=False)
with torch.no_grad(), cuda.device(process_id):
model.cuda()
for iter, pack in enumerate(tqdm(data_loader)):
# img_name = voc12.dataloader.decode_int_filename(pack['name'][0])
img_name = pack['name'][0]
orig_img_size = np.asarray(pack['size'])
if args.dataset == 'voc12':
img_path = voc12.dataloader.get_img_path(
img_name, args.dev_root)
elif args.dataset in ['adp_morph', 'adp_func']:
img_path = adp.dataloader.get_img_path(
img_name, args.dev_root, args.split == 'evaluation')
elif args.dataset in ['deepglobe', 'deepglobe_balanced']:
img_path = deepglobe.dataloader.get_img_path(
img_name, args.dev_root)
else:
raise KeyError('Dataset %s not yet implemented' % args.dataset)
edge, dp = model(pack['img'][0].cuda(non_blocking=True))
# if img_name == '2007_001185':
# cv2.imwrite('edge.png', np.uint8(255 * cv2.resize(edge.cpu().numpy()[0], tuple(orig_img_size[::-1]))))
# D = dp.cpu().numpy()
# hsv = np.zeros((D.shape[1], D.shape[2], 3), dtype='uint8')
# hsv[..., 1] = 255
# mag, ang = cv2.cartToPolar(-D[0], -D[1])
# hsv[..., 0] = ang * 180 / np.pi / 2
# hsv[..., 2] = cv2.normalize(mag, None, 0, 255, cv2.NORM_MINMAX)
#
# rgb = cv2.cvtColor(hsv, cv2.COLOR_HSV2BGR)
# cv2.imwrite('dp.png', cv2.resize(rgb[:, :, ::-1], tuple(orig_img_size[::-1])))
# a=1
cam_dict = np.load(args.cam_out_dir + '/' + img_name + '.npy',
allow_pickle=True).item()
cam_downsized_values = torch.from_numpy(cam_dict['cam']).cuda()
if args.dataset == 'voc12':
if len(cam_dict['keys']) > 0:
keys = np.pad(cam_dict['keys'] + 1, (1, 0),
mode='constant')
if edge.shape[1:] != cam_downsized_values.shape[1:]:
edge = F.interpolate(
edge.unsqueeze(0),
size=(cam_downsized_values.shape[1:]),
mode='bilinear',
align_corners=False)
rw = indexing.propagate_to_edge(cam_downsized_values,
edge,
beta=args.beta,
exp_times=args.exp_times,
radius=5) # radius=5
# rw = indexing.propagate_to_edge(cam_downsized_values, edge, beta=args.beta, exp_times=3, radius=5)
rw_up = F.interpolate(
rw,
size=tuple(orig_img_size),
mode='bilinear',
align_corners=False)[
..., 0, :orig_img_size[0], :orig_img_size[1]]
rw_up = rw_up / torch.max(rw_up)
rw_up_bg = F.pad(rw_up, (0, 0, 0, 0, 1, 0),
value=args.sem_seg_bg_thres)
rw_pred = torch.argmax(rw_up_bg, dim=0).cpu().numpy()
rw_pred = keys[rw_pred]
else:
rw_pred = np.zeros(orig_img_size, dtype='uint8')
elif args.dataset in ['adp_morph', 'adp_func']:
keys = cam_dict['keys']
if edge.shape[1:] != cam_downsized_values.shape[1:]:
edge = F.interpolate(edge.unsqueeze(0),
size=(cam_downsized_values.shape[1:]),
mode='bilinear',
align_corners=False)
rw = indexing.propagate_to_edge(cam_downsized_values,
edge,
beta=args.beta,
exp_times=args.exp_times,
radius=5)
rw_up = F.interpolate(
rw,
size=tuple(orig_img_size),
mode='bilinear',
align_corners=False)[
..., 0, :orig_img_size[0], :orig_img_size[1]]
rw_up = rw_up / torch.max(rw_up)
rw_pred = torch.argmax(rw_up, dim=0).cpu().numpy()
rw_pred = keys[rw_pred]
elif args.dataset in ['deepglobe', 'deepglobe_balanced']:
if len(cam_dict['keys']) > 0:
keys = cam_dict['keys']
down_fac = 6
cam_downsized_values = F.interpolate(
cam_downsized_values.unsqueeze(0),
size=[
x // down_fac
for x in cam_downsized_values.shape[1:]
],
mode='bilinear',
align_corners=False)[0]
edge = F.interpolate(edge.unsqueeze(0),
size=(cam_downsized_values.shape[1:]),
mode='bilinear',
align_corners=False)
rw = indexing.propagate_to_edge(cam_downsized_values,
edge,
beta=args.beta,
exp_times=args.exp_times,
radius=5)
rw_up = F.interpolate(
rw,
size=tuple(orig_img_size // 4),
mode='bilinear',
align_corners=False)[..., 0, :orig_img_size[0] //
4, :orig_img_size[1] // 4]
rw_up = rw_up / torch.max(rw_up)
rw_pred = torch.argmax(rw_up, dim=0).cpu().numpy()
rw_pred = keys[rw_pred]
else:
rw_pred = 5 * np.ones(tuple(orig_img_size // 4))
else:
raise KeyError('Dataset %s not yet implemented' % args.dataset)
imageio.imsave(
os.path.join(args.sem_seg_out_dir, img_name + '.png'),
rw_pred.astype(np.uint8))
# Save with colour
rw_pred_clr = np.zeros(list(rw_pred.shape) + [3], dtype=np.uint8)
off = 0
for t in ['bg', 'fg']:
for i, c in enumerate(args.class_colours[t]):
for ch in range(3):
rw_pred_clr[:, :,
ch] += c[ch] * np.uint8(rw_pred == (i +
off))
off += len(args.class_colours[t])
imageio.imsave(
os.path.join(args.sem_seg_clr_out_dir, img_name + '.png'),
rw_pred_clr)
# Save with colour, overlaid on original image
if args.dataset not in ['deepglobe', 'deepglobe_balanced']:
orig_img = cv2.cvtColor(cv2.imread(img_path),
cv2.COLOR_BGR2RGB)
else:
orig_img = cv2.resize(
cv2.cvtColor(cv2.imread(img_path), cv2.COLOR_BGR2RGB),
rw_pred_clr.shape[:2])
if args.dataset in ['adp_morph', 'adp_func']:
rw_pred_clr = cv2.resize(rw_pred_clr, orig_img.shape[:2])
rw_pred_clr_over = np.uint8(
(1 - args.overlay_r) * np.float32(orig_img) +
args.overlay_r * np.float32(rw_pred_clr))
imageio.imsave(
os.path.join(args.sem_seg_clr_out_dir,
img_name + '_overlay.png'), rw_pred_clr_over)
def _work(process_id, model, dataset, args):
n_gpus = torch.cuda.device_count()
databin = dataset[process_id]
data_loader = DataLoader(databin,
shuffle=False, num_workers=args.num_workers // n_gpus, pin_memory=False)
with torch.no_grad(), cuda.device(process_id):
model.cuda()
for iter, pack in enumerate(data_loader):
img_name = voc12.data.decode_int_filename(pack['name'][0])
orig_img_size = np.asarray(pack['size'])
strided_size = imutils.get_strided_size(orig_img_size, 4)
out_setting = {
"upsize": strided_size,
"flip": True
}
img_o = pack['img'][0][0]
edge, dp = model(img_o.cuda(non_blocking=True), out_setting)
edge_avg = edge
edge_avg = torch.sigmoid(edge_avg)
cam_dict = np.load(args.cam_dir + '/' + img_name + '.npy', allow_pickle=True).item()
cams = cam_dict['cam']
keys = np.pad(cam_dict['keys'] + 1, (1, 0), mode='constant')
cam_downsized_values = cams.cuda()
edge_padded = F.pad(edge_avg, (5, 5, 0, 5), mode='constant', value=1.0)
path_index = adv_indexing.PathIndex(radius=5, default_size=(strided_size[0] + 5, strided_size[1] + 10))
sparse_aff = adv_indexing.edge_to_affinity(torch.unsqueeze(edge_padded, 0),
path_index.default_path_indices)
dense_aff = affinity_sparse2dense(sparse_aff, path_index.default_src_indices,
path_index.default_dst_indices, (strided_size[0] + 5) * (strided_size[1] + 10))
dense_aff = dense_aff.view(strided_size[0] + 5, strided_size[1] + 10, strided_size[0] + 5, -1)[:-5, 5:-5, :-5, 5:-5]
dense_aff = dense_aff.reshape(strided_size[0]*strided_size[1], -1)
trans_mat = to_transition_matrix(dense_aff, beta=args.beta, times=args.t)
cam_expanded = cam_downsized_values*(1 - edge_avg)
rw = torch.matmul(cam_expanded.view(cam_expanded.size(0), -1), trans_mat)
rw = rw.view(rw.size(0), 1, strided_size[0], strided_size[1])
rw_up = F.interpolate(rw, scale_factor=4, mode='bilinear', align_corners=False)[..., :orig_img_size[0], :orig_img_size[1]]
rw_up = rw_up[:, 0]
rw_up_norm = rw_up / torch.max(rw_up)
rw_up_norm_bg = F.pad(rw_up_norm, (0, 0, 0, 0, 1, 0), value=args.sem_seg_bg_thres)
rw_pred = torch.argmax(rw_up_norm_bg, dim=0).cpu().numpy()
rw_pred = keys[rw_pred]
imageio.imsave(os.path.join(args.segm_out_dir, img_name + '.png'), rw_pred.astype(np.uint8))
if process_id == n_gpus - 1 and iter % (len(databin) // 20) == 0:
print("%d " % ((5*iter+1)//(len(databin) // 20)), end='')
def main():
# load valid data
cuda.set_device(GPU_ID)
print('loading data...')
start_time = time.time()
valid_data = torch.load(os.path.join(data_path, 'len50_pywmt14.valid.pt'))
fields = onmt.IO.load_fields(
torch.load(os.path.join(data_path, 'len50_pywmt14.vocab.pt')))
valid_data.fields = fields
print('data loaded. time {} seconds.'.format(time.time() - start_time))
# load model
print('load pretrained model...')
start_time = time.time()
checkpoint = torch.load(model_path,
map_location=lambda storage, loc: storage)
model_opt = checkpoint['opt']
with cuda.device(GPU_ID):
ref_model = onmt.ModelConstructor.make_base_model(
model_opt, fields, True, checkpoint)
ref_model.eval()
ref_model.generator.eval()
lds_model = LDSMaskedNMT(ref_model, MODEL_TYPE)
lds_model.mask_pruned_model()
print('model loaded. time {} seconds.'.format(time.time() - start_time))
lds_model.cal_LDS_for_all_matrices()
lds_model.cal_dis_temp()
lds_model.cal_sorted_params()
tmp_fit = evaluate(lds_model.pruned_model, valid_data, fields)
print(tmp_fit)
test_metrics(lds_model.pruned_model, fields)
get_sparity(lds_model.pruned_model)
tea_model = copy.deepcopy(lds_model.pruned_model)
lds_model.prun_rnn(1.0, 1.0)
lds_model.prun_other(0.6)
lds_model.mask_pruned_model()
tmp_fit = evaluate(lds_model.pruned_model, valid_data, fields)
print(tmp_fit)
test_metrics(lds_model.pruned_model, fields)
stu_model = copy.deepcopy(lds_model.pruned_model)
# load train data
cuda.set_device(GPU_ID)
print('loading data...')
start_time = time.time()
train_data = torch.load(os.path.join(data_path, 'len50_pywmt14.train.pt'))
train_fields = onmt.IO.load_fields(
torch.load(os.path.join(data_path, 'len50_pywmt14.vocab.pt')))
train_data.fields = train_fields
print('data loaded. time {} seconds.'.format(time.time() - start_time))
# retrain
train_ts(tea_model, stu_model, train_data, epoch=1)
tmp_fit = evaluate(stu_model, valid_data, fields)
print(tmp_fit)
test_metrics(stu_model, fields)
get_sparity(stu_model)
lds_model.mask_pruned_model()
tmp_fit = evaluate(stu_model, valid_data, fields)
print(tmp_fit)
test_metrics(stu_model, fields)
get_sparity(stu_model)
def sparse_fdmmv(proc_idx, queue, device_id):
a: ArgsFdmmv = queue.get()
X1: SparseTensor = a.X1
X2: SparseTensor = a.X2
v, w, out = a.v, a.w, a.out
kernel, max_mem = a.kernel, a.max_mem
dtype = X1.dtype
N, D = X1.shape
M = X2.size(0)
if v is None:
T = w.size(1)
else:
T = v.size(1)
# Memory needs:
# X1_chunk : ntot + 2 * D * ntot * density
# X2 : dtot + 2 * D * M * density (because is transposed)
# sparse_out : ntot + 2 * ntot * M * density (assume here density = 1)
# ker_gpu : M * ntot
# w_gpu : ntot * T
# v_gpu : M * T
# out_gpu : M * T
avail_mem = max_mem / sizeof_dtype(dtype)
den = 2 * D * X1.density + 2 + 3 * M + T
sub = D + 2 * D * M * X2.density + M * T
if v is not None:
sub += M * T
n = (avail_mem - sub) / den
n = min(int(n), N)
if n < 1:
raise MemoryError("Not enough memory to run sparse dfmmv")
ddev = torch.device('cuda:%d' % int(device_id))
with tcd.device(ddev):
# Initialize GPU data
w_gpu = create_same_stride((n, T), out, dtype, ddev)
if out.is_cuda:
out_gpu = out
else:
out_gpu = create_same_stride((M, T), out, dtype, ddev)
out_gpu.fill_(0.0)
ker_gpu = create_fortran((n, M), dtype, ddev)
if v is not None:
v_gpu = v.to(device=ddev) # M x T
X2_d = SparseTensor.from_scipy(
X2.transpose_csc().to_scipy().tocsr(copy=False)) \
.index_to_int() \
.to(device=ddev)
for i in range(0, N, n):
ic = min(n, N - i)
X1_chunk = X1.narrow_rows(i, ic)
X1_chunk_d = X1_chunk.index_to_int().to(device=ddev)
ker_chunk = ker_gpu[:ic]
ker_chunk.fill_(0.0)
# TODO: This is wasteful (X2 will be prepared many times over)
ddd = kernel._prepare_sparse(X1_chunk, X2)
ker_chunk = kernel._apply_sparse(X1_chunk_d, X2_d, ker_chunk)
ker_chunk = kernel._finalize(ker_chunk, ddd)
if w is not None:
c_g_w = copy_to_device_noorder(ic, T, w, i, 0, w_gpu, 0, 0)
else:
c_g_w = w_gpu.narrow(0, 0, ic)
c_g_w.fill_(0.0)
if v is not None:
c_g_w.addmm_(ker_chunk, v_gpu)
out_gpu.addmm_(ker_chunk.T, c_g_w)
del ddd, X1_chunk, X1_chunk_d
if not out.is_cuda:
copy_to_device_noorder(M, T, out_gpu, 0, 0, out, 0, 0)
return out
def main():
total_times = 100
run_times = 0
init_threshold = ...
start_t = time.time()
valid_data = torch.load(TRAIN_DATA + '.valid.pt')
fields = onmt.IO.load_fields(torch.load(TRAIN_DATA + '.vocab.pt'))
# fields = onmt.IO.load_fields_from_vocab(torch.load(TRAIN_DATA + '.vocab.pt'))
valid_data.fields = fields # we need to clear this assignment relationg if we want to transfere valid among threads
checkpoint = torch.load(weights, map_location=lambda storage, loc: storage)
model_opt = checkpoint['opt']
# masked_models = []
with cuda.device(GPU_ID):
ref_model = onmt.ModelConstructor.make_base_model(
model_opt, fields, True, checkpoint)
ref_model.eval()
ref_model.generator.eval()
masked_model = MaskedModel(
ref_model, group_dict, cuda.current_device(), cuda.current_device(
)) # ref_model is at current_device, no copy will happen
# masked_models.append(masked_model)
train_opt, _, _ = opt_initialize(checkpoint, 'opennmt_translate_opt.pt',
'opennmt_translate_dummy_opt.pt')
if GPU_ID:
cuda.set_device(GPU_ID)
# 只需要原始的accuracy
acc_of_no_prune = 0
get_acc_of_no_prune = False
print(time_now(), "start while")
while run_times < total_times:
print("-----------------------------------------")
print("-----------------------------------------")
print("start Iteration ", run_times)
# init threshold
best_threshold = 0
itr_time = time.time()
xxx = np.arange(0., 1, 0.01)
print(time_now(), "start testing pruning")
masked_model.make_evaluable()
for i in range(len(xxx)):
# best_threshold = 0.2
# break
tmp_crate = len(masked_model.group_name_list) * [xxx[i]]
masked_model.change_mask(tmp_crate, apply_MP_on_mask)
masked_model.apply_mask()
tmp_fit = evaluate(masked_model, valid_data, fields)
print('percentage %s => acc (%.4f), ppl (%.4f)' %
(xxx[i] * 100, tmp_fit[1], tmp_fit[0]))
if i == 0 and not get_acc_of_no_prune:
acc_of_no_prune = tmp_fit[1]
acc_of_no_prune = int(acc_of_no_prune * 10) / 10
get_acc_of_no_prune = True
elif acc_of_no_prune - tmp_fit[1] > acc_percent_prune:
best_threshold = xxx[i] - 0.01
break
# -------------------------------------------------
# Start writing
# prune again
print(time_now(), " init accuracy of model:", acc_of_no_prune)
print("accuracy constraint:", acc_percent_prune)
print("-------test------------:", get_acc_of_no_prune)
print(time_now(), " apply pruning with threshold:", best_threshold)
tmp_crate = len(masked_model.group_name_list) * [best_threshold]
masked_model.change_mask(tmp_crate, apply_MP_on_mask)
masked_model.apply_mask()
# print information
tmp_fit = evaluate(masked_model, valid_data, fields)
print('percentage %s => acc (%.4f), ppl (%.4f)' %
(best_threshold * 100, tmp_fit[1], tmp_fit[0]))
model_sparsity = masked_model.get_sparsity()
print('Sparsity: {}'.format(model_sparsity))
#--------------- start retraining --------------
# first store model
print(time_now(), "start saving model")
_, saved_model = update_checkpoint(checkpoint, masked_model, run_times,
acc_percent_prune)
print(time_now(), "finish saving model:", saved_model)
model_for_train = masked_model
pretrained_leaf_dict = model_for_train.make_trainable()
optim = build_optim(model_for_train.masked_model, checkpoint,
train_opt, pretrained_leaf_dict)
print("finish building optim")
print(time_now(), "start loading data for retraining")
train = torch.load(train_opt.data + '.train.pt')
valid = torch.load(train_opt.data + '.valid.pt')
train_fields = load_fields(train, valid, checkpoint, train_opt)
print(time_now(), "finish data loading")
recovered = train_model(model_for_train, train, valid, train_fields,
optim, train_opt, run_times, acc_of_no_prune)
print(time_now(), "finish retraining ")
if not recovered:
exit()
else:
print("------------Accuracy recorverd!--------------------")
print("recovered accuracy:", acc_of_no_prune)
#-------------------------------------------------
print('------------- save checkpoint ---------------')
_, saved_model = update_checkpoint(checkpoint,
model_for_train,
run_times,
t=True)
print(time_now(), ' saving model:', saved_model)
print("-------------print evaluation info ---------------")
model_for_train.make_evaluable()
tmp_fit = evaluate(model_for_train, valid_data, fields)
print('percentage %s => acc (%.4f), ppl (%.4f)' %
(best_found * 100, tmp_fit[1], tmp_fit[0]))
model_sparsity = model_for_train.get_sparsity()
print('Sparsity: {}'.format(model_sparsity))
#--------------------------------------------------
print("BLEU evaluation:")
translate_opt, translate_dummy_opt = translate_opt_initialize(
'opennmt_translate_opt.pt', 'opennmt_translate_dummy_opt.pt')
translator = init_translate_model(translate_opt, translate_dummy_opt)
del translator.model
translator.model = model_for_train.masked_model
tt = open(translate_opt.tgt, 'r')
references = [[t] for t in tt]
translate_data = onmt.IO.ONMTDataset(translate_opt.src,
translate_opt.tgt,
fields,
use_filter_pred=False)
prune_data = onmt.IO.OrderedIterator(dataset=translate_data,
device=GPU_ID,
batch_size=1,
train=False,
sort=False,
shuffle=False)
tmp_fit2 = evaluate_trans(translator, references, prune_data,
translate_data)
print('Finsished => bleu (%.4f), ppl (%.4f)' %
(tmp_fit2[1] * 100, tmp_fit2[0]))
#--------------------------------------------------
run_times += 1
def distk_fmmv(proc_idx, queue, device_id):
a: ArgsFmmv = queue.get()
X1, X2, v, out = a.X1, a.X2, a.v, a.out
kernel: L2DistanceKernel = a.kernel
max_mem = a.max_mem
N, D = X1.shape
M = X2.shape[0]
T = v.shape[1]
dtype = X1.dtype
# GPU memory usage:
# X1s : n x D
# X2s : m x D
# vs : m x T
# nm : n x m
# out : n x T
# -----------
# total: n*m + n * (D + T) + m * (D + T) = R
avail_mem = max_mem / sizeof_dtype(dtype)
#if sizeof_dtype(dtype) == 4:
# avail_mem /= 2
n, m = select_dim_over_m(maxM=M,
maxN=N,
coef_nm=1.0,
coef_n=D + T,
coef_m=D + T,
tot=avail_mem)
ddev = torch.device('cuda:%d' % int(device_id))
with tcd.device(ddev):
nm_gpu = create_same_stride((n, m), X1, dtype, ddev)
out_gpu = create_same_stride((n, T), out, dtype, ddev)
X1s_gpu = create_same_stride((n, D), X1, dtype, ddev)
X2s_gpu = create_same_stride((m, D), X2, dtype, ddev)
vs_gpu = create_same_stride((m, T), v, dtype, ddev)
for i in range(0, N, n):
nb = min(n, N - i)
cur_X1s_gpu = copy_to_device_noorder(nb, D, X1, i, 0, X1s_gpu, 0,
0)
sq1 = torch.norm(cur_X1s_gpu, p=2, dim=1, keepdim=True).pow_(2)
cur_out_gpu = out_gpu.narrow(0, 0, nb) # n x T
cur_out_gpu.fill_(0.0)
for j in range(0, M, m):
mb = min(m, M - j)
cur_X2s_gpu = copy_to_device_noorder(mb, D, X2, j, 0, X2s_gpu,
0, 0)
cur_vs_gpu = copy_to_device_noorder(mb, T, v, j, 0, vs_gpu, 0,
0) # m x T
cur_nm_gpu = nm_gpu[:nb, :mb] # n x m
sq2 = torch.norm(cur_X2s_gpu, p=2, dim=1, keepdim=True).pow_(2)
torch.mm(cur_X1s_gpu, cur_X2s_gpu.T, out=cur_nm_gpu)
cur_nm_gpu.mul_(-2.0)
cur_nm_gpu.add_(sq1)
cur_nm_gpu.add_(sq2.T)
cur_nm_gpu.clamp_min_(0)
kernel._transform(cur_nm_gpu)
# Multiply by the vector v
# FIXME: This is the cause of mapping errors in case of float32 calculations.
cur_out_gpu.addmm_(cur_nm_gpu, cur_vs_gpu) # n x T
# send result to CPU
copy_to_host_noorder(nb, T, out_gpu, 0, 0, out, i, 0)
return out
def main():
total_times = 10
run_times = 0
init_threshold = ...
start_t = time.time()
valid_data = torch.load(TRAIN_DATA + '.valid.pt')
fields = onmt.IO.load_fields(torch.load(TRAIN_DATA + '.vocab.pt'))
valid_data.fields = fields # we need to clear this assignment relationg if we want to transfere valid among threads
if GPU_ID:
cuda.set_device(GPU_ID)
while run_times < total_times:
itr_time = time.time()
# intit pruning
checkpoint = torch.load(weights,
map_location=lambda storage, loc: storage)
model_opt = checkpoint['opt']
masked_models = []
with cuda.device(GPU_ID):
ref_model = onmt.ModelConstructor.make_base_model(
model_opt, fields, True, checkpoint)
ref_model.eval()
ref_model.generator.eval()
masked_model = MaskedModel(
ref_model, group_dict, cuda.current_device(),
cuda.current_device(
)) # ref_model is at current_device, no copy will happen
masked_models.append(masked_model)
'''
display all the names of parameters
'''
'''
aa=ref_model.named_parameters
aa_namelist = [ak[0] for ak in aa]
'''
'''
test MP
'''
'''
translate_opt, translate_dummy_opt = translate_opt_initialize('opennmt_translate_opt.pt', 'opennmt_translate_dummy_opt.pt')
translator = init_translate_model(translate_opt, translate_dummy_opt)
del translator.model
translator.model = masked_model
tt=open(translate_opt.tgt, 'r')
references = [[t] for t in tt]
xxx=np.arange(0.,1, 0.01)
#for i in range(len(masked_model.group_name_list)):
# tmp_crate = len(masked_model.group_name_list)*[0.]
for i in range(len(xxx)):
translate_data = onmt.IO.ONMTDataset(
translate_opt.src, translate_opt.tgt, fields,
use_filter_pred=False)
prune_data = onmt.IO.OrderedIterator(
dataset=translate_data, device=GPU_ID,
batch_size=1, train=False, sort=False,
shuffle=False)
tmp_crate = len(masked_model.group_name_list)*[xxx[i]]
#tmp_crate[i] = 0.01
masked_model.change_mask(tmp_crate, apply_MP_on_mask)
masked_model.apply_mask()
tmp_fit = evaluate(masked_model, valid_data, fields)
#tmp_fit = evaluate_trans(translator, references, prune_data, translate_data)
#logger.scalar_summary('test_bleu', tmp_fit[1]*100, int(xxx[i]*100))
logger.scalar_summary('acc', tmp_fit[1], int(xxx[i]*100))
logger.scalar_summary('ppl', tmp_fit[0], int(xxx[i]*100))
#logger.scalar_summary('test_ppl', tmp_fit[0], int(xxx[i]*100))
#print('group %s => acc (%.4f), ppl (%.4f)' % (masked_model.group_name_list[i], tmp_fit[1], tmp_fit[0]))
#print('percentage %s => bleu (%.4f), ppl (%.4f)' % (xxx[i]*100, tmp_fit[1]*100, tmp_fit[0]))
print('percentage %s => acc (%.4f), ppl (%.4f)' % (xxx[i]*100, tmp_fit[1], tmp_fit[0]))
exit()
'''
with cuda.device(GPU_ID):
masked_models.append(
MaskedModel(ref_model, group_dict, GPU_ID, gpu_candidate)
) # if the gpu_candidate is the same as ref_model, it will return the ref_model
del ref_model
# do pruning
ncs_start = time.time()
print('Itration %d, model loading: %d sec' %
(run_times, ncs_start - itr_time))
init_threshold = len(masked_models[0].group_name_list) * [0.10]
best_found, saved_model, best_masked_model = NCS_MP(
init_threshold, 0.05, fields, masked_models, valid_data, 1.0,
run_times, checkpoint)
init_thresholds = best_found[0] # best pop found
end_t = time.time()
print('NCS Time: {} min'.format((end_t - itr_time) / 60.))
# clear no used models
for gpu_model in masked_models:
del gpu_model
#-----------------------------------------------------------------
'''
masked_models = []
with cuda.device(GPU_ID):
ref_model = best_masked_model
masked_model = MaskedModel(ref_model, group_dict2, cuda.current_device(), cuda.current_device()) # ref_model is at current_device, no copy will happen
masked_models.append(masked_model)
for gpu_candidate in other_GPU_IDs:
with cuda.device(gpu_candidate):
masked_models.append(MaskedModel(ref_model, group_dict2, GPU_ID, gpu_candidate)) # if the gpu_candidate is the same as ref_model, it will return the ref_model
del ref_model
# do pruning
ncs_start = time.time()
print('*Itration %d, model loading: %d sec' % (run_times, ncs_start - itr_time))
init_threshold = len(masked_models[0].group_name_list)*[0.10]
best_found, saved_model, best_masked_model = NCS_MP(init_threshold, 0.05, fields, masked_models, valid_data, 0.5, run_times, checkpoint)
init_thresholds = best_found[0] # best pop found
end_t = time.time()
print('NCS Time: {} min'.format((end_t - itr_time)/60.))
# clear no used models
for gpu_model in masked_models:
del gpu_model
'''
exit()
# training
checkpoint = torch.load(saved_model,
map_location=lambda storage, loc: storage)
train_opt, _, _ = opt_initialize(checkpoint,
'opennmt_translate_opt.pt',
'opennmt_translate_dummy_opt.pt')
# train data loading
train = torch.load(train_opt.data + '.train.pt')
valid = torch.load(train_opt.data + '.valid.pt')
train_fields = load_fields(train, valid, checkpoint, train_opt)
model_for_train = init_train_model(checkpoint1, train_opt,
train_fields) # fields need data
optim = build_optim(model_for_train, checkpoint, train_opt)
model_sparsity = get_sparsity(model_for_train)
logger.scalar_summary('model_sparsity_%s' % num_runs, model_sparsity,
run_times)
print('Sparsity: {}'.format(model_sparsity))
train_model(model_for_train, train, valid, train_fields, optim,
train_opt, run_times)
# update global variabel weights
run_times += 1
def sparse_fmmv(proc_idx, queue, device_id):
a: ArgsFmmv = queue.get()
X1: SparseTensor = a.X1
X2: SparseTensor = a.X2
v, out = a.v, a.out
kernel, max_mem = a.kernel, a.max_mem
dtype = X1.dtype
ntot, dtot = X1.shape
mtot, T = v.size()
avail_mem = max_mem / sizeof_dtype(dtype)
# Memory needs:
# X1_chunk : N + 2*D*N*density
# X2_chunk : D + 2*D*M*density (because is transposed)
# sparse_out : N + 2*N*M*(density) (assume density = 1)
# ker_gpu : M*N
# mmv_gpu : N*T
# v_gpu : M*T
# Other: GPU buffer
n, m = select_dim_over_m(
maxM=mtot,
maxN=ntot,
tot=avail_mem,
coef_nm=3,
coef_n=2 + 2 * dtot * X1.density + T,
coef_m=2 * dtot * X2.density + T,
rest=dtot,
)
ddev = torch.device('cuda:%d' % int(device_id))
with tcd.device(ddev):
v_gpu = v.to(device=ddev) # M x T
mmv_gpu = create_same_stride((n, T), out, dtype, ddev)
# ker_gpu should be fortran-ordered due to cusparse csr2dense function
ker_gpu = create_fortran((n, m), dtype=dtype, device=ddev)
for i in range(0, ntot, n):
ic = min(n, ntot - i)
cur_mmv_gpu = mmv_gpu[:ic] # n x T
cur_mmv_gpu.fill_(0.0)
X1_chunk = X1.narrow_rows(i, ic)
X1_chunk_d = X1_chunk.index_to_int().to(device=ddev)
for j in range(0, mtot, m):
jc = min(m, mtot - j)
X2_chunk = X2.narrow_rows(j, jc)
# Prepare sparse on CPU
ddd = kernel._prepare_sparse(X1_chunk, X2_chunk)
# Transpose X2-chunk and convert it to CSR. This uses lots of RAM
X2_chunk_d = SparseTensor.from_scipy(
X2_chunk.transpose_csc().to_scipy().tocsr(copy=False)) \
.index_to_int() \
.to(device=ddev)
cur_ker_gpu = ker_gpu[:ic, :jc]
cur_ker_gpu.fill_(0.0)
# Run the matrix multiplication (kernel apply)
cur_ker_gpu = kernel._apply_sparse(X1_chunk_d, X2_chunk_d,
cur_ker_gpu)
cur_ker_gpu = kernel._finalize(cur_ker_gpu, ddd)
# Multiply by the vector v
cur_mmv_gpu.addmm_(cur_ker_gpu, v_gpu.narrow(0, j, jc))
del ddd, X2_chunk, X2_chunk_d
# send result to CPU
copy_to_host_noorder(ic, T, cur_mmv_gpu, 0, 0, out, i, 0)
del X1_chunk, X1_chunk_d
return out
def _align_nn(self,
model_path,
source_kb,
target_kb,
candidate_selector,
cuda_device,
batch_size=128):
"""
Align using neural network model
:param source_kb:
:param target_kb:
:param candidate_selector:
:param cuda_device: GPU device number
:return:
"""
# returns json representation of entity
def _form_json_entity(ent_to_json, kb):
all_rels = [
kb.relations[r_id] for r_id in ent_to_json.relation_ids
]
par_ents = [
kb.get_entity_by_research_entity_id(r.entity_ids[1])
for r in all_rels
if r.relation_type in constants.UMLS_PARENT_REL_LABELS
]
chd_ents = [
kb.get_entity_by_research_entity_id(r.entity_ids[1])
for r in all_rels
if r.relation_type in constants.UMLS_CHILD_REL_LABELS
]
return {
'research_entity_id': ent_to_json.research_entity_id,
'canonical_name': ent_to_json.canonical_name,
'aliases': ent_to_json.aliases,
'definition': ent_to_json.definition,
'other_contexts': ent_to_json.other_contexts,
'par_relations': [e.canonical_name for e in par_ents],
'chd_relations': [e.canonical_name for e in chd_ents]
}
from emma.allennlp_classes.ontoemma_dataset_reader import OntologyMatchingDatasetReader
from emma.allennlp_classes.ontoemma_model import OntoEmmaNN
from emma.allennlp_classes.ontoemma_predictor import OntoEmmaPredictor
alignment, s_ent_ids, t_ent_ids = self._align_string_equiv(
source_kb, target_kb)
sys.stdout.write("%i alignments with string equivalence\n" %
len(alignment))
if cuda_device > 0:
with device(cuda_device):
archive = load_archive(model_path, cuda_device=cuda_device)
else:
archive = load_archive(model_path, cuda_device=cuda_device)
predictor = Predictor.from_archive(archive, 'ontoemma-predictor')
sys.stdout.write("Making predictions...\n")
s_ent_tqdm = tqdm.tqdm(s_ent_ids, total=len(s_ent_ids))
temp_alignments = defaultdict(list)
if cuda_device > 0:
with device(cuda_device):
batch_json_data = []
for s_ent_id in s_ent_tqdm:
s_ent = source_kb.get_entity_by_research_entity_id(
s_ent_id)
for t_ent_id in candidate_selector.select_candidates(
s_ent_id)[:constants.KEEP_TOP_K_CANDIDATES]:
t_ent = target_kb.get_entity_by_research_entity_id(
t_ent_id)
json_data = {
'source_ent': _form_json_entity(s_ent, source_kb),
'target_ent': _form_json_entity(t_ent, target_kb),
'label': 0
}
batch_json_data.append(json_data)
if len(batch_json_data) == batch_size:
results = predictor.predict_batch_json(
batch_json_data, cuda_device)
for model_input, output in zip(
batch_json_data, results):
if output['predicted_label'] == [1.0]:
temp_alignments[model_input['source_ent'][
'research_entity_id']].append(
(model_input['target_ent']
['research_entity_id'],
output['score'][0]))
batch_json_data = []
else:
for s_ent_id in s_ent_tqdm:
s_ent = source_kb.get_entity_by_research_entity_id(s_ent_id)
for t_ent_id in candidate_selector.select_candidates(
s_ent_id)[:constants.KEEP_TOP_K_CANDIDATES]:
t_ent = target_kb.get_entity_by_research_entity_id(
t_ent_id)
json_data = {
'source_ent': _form_json_entity(s_ent, source_kb),
'target_ent': _form_json_entity(t_ent, target_kb),
'label': 0
}
output = predictor.predict_json(json_data, cuda_device)
if output['predicted_label'] == [1.0]:
temp_alignments[json_data['source_ent'][
'research_entity_id']].append(
(json_data['target_ent']['research_entity_id'],
output['score'][0]))
alignment = []
for s_ent_id, matches in temp_alignments.items():
if len(matches) > 0:
m_sort = sorted(matches, key=lambda p: p[1], reverse=True)
if m_sort[0][1] >= constants.NN_SCORE_THRESHOLD:
alignment.append((s_ent_id, m_sort[0][0], m_sort[0][1]))
return alignment
def _work(process_id, model, dataset, args):
databin = dataset[process_id]
n_gpus = torch.cuda.device_count()
data_loader = DataLoader(databin,
shuffle=False,
num_workers=args.num_workers // n_gpus,
pin_memory=False)
with torch.no_grad(), cuda.device(process_id):
model.cuda()
for iter, pack in enumerate(tqdm(data_loader)):
img_name = pack['name'][0]
label = pack['label'][0]
size = pack['size']
strided_size = imutils.get_strided_size(size, 4)
strided_up_size = imutils.get_strided_up_size(size, 16)
# Run through each scale of image
outputs = [
model(img[0].cuda(non_blocking=True)) for img in pack['img']
]
# Each output is resized to strided_size (lower than original) and summed
strided_cam = torch.sum(
torch.stack([
F.interpolate(torch.unsqueeze(o, 0),
strided_size,
mode='bilinear',
align_corners=False)[0] for o in outputs
]), 0)
# Each output is resized to strided_up_size (which should be orignal size?)
highres_cam = [
F.interpolate(torch.unsqueeze(o, 1),
strided_up_size,
mode='bilinear',
align_corners=False) for o in outputs
]
highres_cam = torch.sum(torch.stack(highres_cam, 0),
0)[:, 0, :size[0], :size[1]]
# Pick the cams corresponding to image-level labels
# Normalize by max value across H x W dimension for each channel
valid_cat = torch.nonzero(label, as_tuple=False)[:, 0]
strided_cam = strided_cam[valid_cat]
strided_cam /= F.adaptive_max_pool2d(strided_cam, (1, 1)) + 1e-5
highres_cam = highres_cam[valid_cat]
highres_cam /= F.adaptive_max_pool2d(highres_cam, (1, 1)) + 1e-5
# save cams
np.save(
os.path.join(args.cam_out_dir, img_name + '.npy'), {
"keys": valid_cat,
"cam": strided_cam.cpu(),
"high_res": highres_cam.cpu().numpy()
})
if process_id == n_gpus - 1 and iter % (len(databin) // 20) == 0:
print("%d " % ((5 * iter + 1) // (len(databin) // 20)), end='')
sys.stdout.flush()
def __init__(self, pretrained, group_dict, source_device_id,
target_device_id):
super(MaskedModel, self).__init__()
self.skip_mark = 'skip' #skip the layer while pruning
self.mask_dict = {}
#self.sort_tensors = {}
self.layer_element_num = []
self.layer_num = 0
self.layer_name_dict = {}
self.sparsity = 1.0
self.total_parameter_num = 0
# used for group parameters
self.group_name_list = []
self.map_dict = {}
self.group_num_dict = {}
self.group_parameter_dict = {}
self.group_threshold_list = {}
# gpu realted
self.sgpu_id = source_device_id
self.tgpu_id = target_device_id
# settings for retraining
self.pre_forward_fn = None
self.forward_fn = None
# init group dicts
self.group_name_list = [
k for k in group_dict.keys()
] # the indices of list will map the the thresholds list which is accepted at pruning
self.group_threshold_list = len(self.group_name_list) * [0.]
for key, layer_names in group_dict.items():
self.group_num_dict[key] = 0
# maping the layer name to group name
for layer_name in layer_names:
self.map_dict[layer_name] = key
pretrained_model_on_device = None
# for each retrieval, transfer the pretrained model to a dictionary
if source_device_id == target_device_id:
pretrained_model_on_device = pretrained
else:
pretrained_model_on_device = my_replicate(
pretrained, source_device_id,
target_device_id) #[source_device_id, target1, ...]
with cuda.device(target_device_id):
self.pretrained_model_dict = dict([
(n, v)
for n, v in pretrained_model_on_device.named_parameters()
]) # 20191023 lgy
#self.pretrained_model_dict = dict([(n,v) for n,v in pretrained_model_on_device.state_dict().items()])
for param_name, module_param in self.pretrained_model_dict.items():
if param_name in self.map_dict: # ignore no-grouped layers
self.group_num_dict[
self.map_dict[param_name]] += module_param.nelement()
for group_name in self.group_name_list:
self.group_parameter_dict[group_name] = torch.cuda.FloatTensor(
1, self.group_num_dict[group_name])
self.generate_mask(
self.pretrained_model_dict) # init self.mask_dict
self.masked_model = my_replicate(pretrained, source_device_id,
target_device_id)
#self.generator = self.masked_model.generator
self.encoder = self.masked_model.encoder
self.decoder = self.masked_model.decoder
def _work(process_id, model, dataset, args):
databin = dataset[process_id]
n_gpus = torch.cuda.device_count()
data_loader = DataLoader(databin, shuffle=False, pin_memory=False)
with torch.no_grad(), cuda.device(process_id):
model.cuda()
with tqdm(total=len(data_loader)) as pbar:
for iter, pack in enumerate(data_loader):
img_name = pack['name'][0]
size = pack['size']
strided_size = imutils.get_strided_size(size, 4)
strided_up_size = imutils.get_strided_up_size(size, 16)
if args.dataset in ['adp_morph', 'adp_func']:
outputs, labels = zip(*[
model(img.cuda(
non_blocking=True), orig_img.cuda(
non_blocking=True))
for img, orig_img in zip(pack['img'], pack['orig_img'])
])
else:
outputs, labels = zip(*[
model(img.cuda(non_blocking=True))
for img in pack['img']
])
if 'train' in args.split:
label = pack['label'][0]
else:
label = labels[0][args.use_cls]
valid_cat = torch.nonzero(label)[:, 0]
if args.dataset in ['adp_morph', 'adp_func']:
if torch.cuda.is_available():
valid_cat = torch.cat(
(torch.from_numpy(
np.array(range(len(args.class_names['bg'])),
dtype=np.int64)).cuda(),
valid_cat.cuda() + len(args.class_names['bg'])))
else:
valid_cat = torch.cat(
(torch.from_numpy(
np.array(range(len(args.class_names['bg'])),
dtype=np.int64)),
valid_cat + len(args.class_names['bg'])))
if len(valid_cat) > 0:
strided_cam = torch.sum(
torch.stack([
F.interpolate(torch.unsqueeze(o, 0),
strided_size,
mode='bilinear',
align_corners=False)[0]
for o in outputs
]), 0)
highres_cam = [
F.interpolate(torch.unsqueeze(o, 1),
strided_up_size,
mode='bilinear',
align_corners=False) for o in outputs
]
highres_cam = torch.sum(torch.stack(tuple(highres_cam), 0),
0)[:, 0, :size[0], :size[1]]
strided_cam = strided_cam[valid_cat]
strided_cam /= F.adaptive_max_pool2d(strided_cam,
(1, 1)) + 1e-5
highres_cam = highres_cam[valid_cat]
highres_cam /= F.adaptive_max_pool2d(highres_cam,
(1, 1)) + 1e-5
# save cams
if args.dataset not in ['deepglobe', 'deepglobe_balanced']:
np.save(
os.path.join(args.cam_out_dir, img_name + '.npy'),
{
"keys": valid_cat.cpu().numpy(),
"cam": strided_cam.cpu().numpy(),
"high_res": highres_cam.cpu().numpy()
})
else:
np.save(
os.path.join(args.cam_out_dir, img_name + '.npy'),
{
"keys": valid_cat.cpu().numpy(),
"cam": strided_cam.cpu().numpy()
})
else:
np.save(
os.path.join(args.cam_out_dir, img_name + '.npy'), {
"keys": np.empty(0),
"cam": np.empty(0),
"high_res": np.empty(0)
})
pbar.update(1)
def main():
total_times = 100
run_times = 0
init_threshold = ...
start_t = time.time()
valid_data = torch.load(TRAIN_DATA + '.valid.pt')
fields = onmt.IO.load_fields(torch.load(TRAIN_DATA + '.vocab.pt'))
# fields = onmt.IO.load_fields_from_vocab(torch.load(TRAIN_DATA + '.vocab.pt'))
valid_data.fields = fields # we need to clear this assignment relationg if we want to transfere valid among threads
checkpoint = torch.load(weights, map_location=lambda storage, loc: storage)
model_opt = checkpoint['opt']
masked_models = []
with cuda.device(GPU_ID):
ref_model = onmt.ModelConstructor.make_base_model(model_opt, fields, True, checkpoint)
ref_model.eval()
ref_model.generator.eval()
masked_model = MaskedModel(ref_model, group_dict, cuda.current_device(), cuda.current_device()) # ref_model is at current_device, no copy will happen
masked_models.append(masked_model)
train_opt, _, _ = opt_initialize(checkpoint, 'opennmt_translate_opt.pt', 'opennmt_translate_dummy_opt.pt')
if GPU_ID:
cuda.set_device(GPU_ID)
print("BLEU evaluation:")
translate_opt, translate_dummy_opt = translate_opt_initialize('opennmt_translate_opt.pt', 'opennmt_translate_dummy_opt.pt')
translator = init_translate_model(translate_opt, translate_dummy_opt)
del translator.model
translator.model = masked_model.masked_model
tt=open(translate_opt.tgt, 'r')
references = [[t] for t in tt]
translate_data = onmt.IO.ONMTDataset(translate_opt.src, translate_opt.tgt,fields,use_filter_pred=False)
prune_data = onmt.IO.OrderedIterator(dataset=translate_data, device=GPU_ID,batch_size=1, train=False, sort=False,shuffle=False)
tmp_fit2 = evaluate_trans(translator, references, prune_data, translate_data)
print('Finsished => bleu (%.4f), ppl (%.4f)' % (tmp_fit2[1]*100, tmp_fit2[0]))
exit()
# print(time_now(), "get accuray of no pruning model")
# masked_model.make_evaluable()
# tmp_crate = len(masked_model.group_name_list)*[0]
# masked_model.change_mask(tmp_crate, apply_MP_on_mask)
# masked_model.apply_mask()
# tmp_fit = evaluate(masked_model, valid_data, fields)
# # 只需要原始的accuracy
# acc_of_no_prune = tmp_fit[1]
# acc_of_no_prune = int(acc_of_no_prune*10)/10
print("init accuracy of model:", acc_of_no_prune)
print("accuracy constraint:", acc_percent_prune)
while run_times < total_times:
print("-----------------------------------------")
print("-----------------------------------------")
print("-----------------------------------------")
print(time_now(), "start Iteration ", run_times)
print("test model---------------")
ref_model.eval()
ref_model.generator.eval()
tmp_fit = evaluate(ref_model, valid_data, fields)
print('ref_model','acc (%.4f), ppl (%.4f)' % (tmp_fit[1], tmp_fit[0]))
print("test model---------------")
masked_models[0].make_evaluable()
tmp_fit = evaluate(masked_models[0], valid_data, fields)
print('masked_models[0]','acc (%.4f), ppl (%.4f)' % (tmp_fit[1], tmp_fit[0]))
model_sparsity = masked_models[0].get_sparsity()
print('masked_models[0] Sparsity: {}'.format(model_sparsity))
itr_time = time.time()
for gpu_candidate in other_GPU_IDs:
with cuda.device(gpu_candidate):
masked_models.append(MaskedModel(ref_model, group_dict, GPU_ID, gpu_candidate)) # if the gpu_candidate is the same as ref_model, it will return the ref_model
#------------- Here -------------------------
# del ref_model
# do pruning
ncs_start = time.time()
print('Itration %d, model loading: %d sec' % (run_times, ncs_start - itr_time))
if run_times == 0:
if START_THRESHOLD is not None:
init_threshold = START_THRESHOLD
else:
init_threshold = len(masked_models[0].group_name_list)*[0.25]
# if run_times == 0:
# init_threshold = len(masked_models[0].group_name_list)*[0.25]
print("init threshold:", init_threshold)
prune_acc_now = acc_percent_prune+tmp_fit[1]-acc_of_no_prune
print('pruning acc now:', prune_acc_now)
best_found, saved_model, best_masked_model = NCS_MP(init_threshold, 0.05, fields, masked_models, valid_data, prune_acc_now, run_times, checkpoint)
init_threshold = best_found
#best_found, saved_model, best_masked_model = NCS_MP(init_threshold, 0.05, fields, masked_models, valid_data, 0.01, run_times, checkpoint)
end_t = time.time()
print('NCS Time: {} min'.format((end_t - itr_time)/60.))
print('Best found thresholds:')
for i in range(len(masked_models[0].group_name_list)):
print("layer {}: {}%".format(masked_models[0].group_name_list[i], 100*best_found[i]))
print("BLEU evaluation:")
translate_opt, translate_dummy_opt = translate_opt_initialize('opennmt_translate_opt.pt', 'opennmt_translate_dummy_opt.pt')
translator = init_translate_model(translate_opt, translate_dummy_opt)
del translator.model
translator.model = best_masked_model
tt=open(translate_opt.tgt, 'r')
references = [[t] for t in tt]
translate_data = onmt.IO.ONMTDataset(translate_opt.src, translate_opt.tgt,fields,use_filter_pred=False)
prune_data = onmt.IO.OrderedIterator(dataset=translate_data, device=GPU_ID,batch_size=1, train=False, sort=False,shuffle=False)
tmp_fit = evaluate_trans(translator, references, prune_data, translate_data)
print('Finsished => bleu (%.4f), ppl (%.4f)' % (tmp_fit[1]*100, tmp_fit[0]))
# clear no used models
for gpu_model in masked_models:
del gpu_model
#--------------- start retraining --------------
model_for_train = best_masked_model
pretrained_leaf_dict = model_for_train.make_trainable()
optim = build_optim(model_for_train.masked_model, checkpoint, train_opt, pretrained_leaf_dict)
print(time_now(), "start loading data for retraining")
train = torch.load(train_opt.data+ '.train.pt')
valid = torch.load(train_opt.data + '.valid.pt')
train_fields = load_fields(train, valid, checkpoint, train_opt)
print(time_now(), "finish data loading")
model_for_train.change_mask(init_threshold, apply_MP_on_mask)
model_for_train.apply_mask()
model_for_train.make_trainable()
recovered = train_model(model_for_train, train, valid, train_fields, optim, train_opt, run_times, acc_of_no_prune)
print(time_now(), "finish retraining ")
if not recovered:
exit()
else:
print("------------Accuracy recorverd!--------------------")
print("recovered accuracy:", acc_of_no_prune)
model_for_train.make_evaluable()
ref_model = model_for_train.masked_model
masked_models = [MaskedModel(ref_model, group_dict, cuda.current_device(), cuda.current_device())]
print("test model---------------")
tmp_fit = evaluate(ref_model, valid_data, fields)
print('ref_model','acc (%.4f), ppl (%.4f)' % (tmp_fit[1], tmp_fit[0]))
print("test model---------------")
tmp_fit = evaluate(masked_models[0], valid_data, fields)
print('masked_models[0]','acc (%.4f), ppl (%.4f)' % (tmp_fit[1], tmp_fit[0]))
model_sparsity = masked_models[0].get_sparsity()
print('masked_models[0] Sparsity: {}'.format(model_sparsity))
print('------------- save checkpoint ---------------')
_, saved_model = update_checkpoint(checkpoint, model_for_train, run_times, acc_percent_prune, t=True)
print(time_now(), ' saving model:', saved_model)
print("-------------print evaluation info ---------------")
tmp_fit = evaluate(model_for_train, valid_data, fields)
print('percentage %s => acc (%.4f), ppl (%.4f)' % (best_found*100, tmp_fit[1], tmp_fit[0]))
#--------------------------------------------------
print("BLEU evaluation:")
translate_opt, translate_dummy_opt = translate_opt_initialize('opennmt_translate_opt.pt', 'opennmt_translate_dummy_opt.pt')
translator = init_translate_model(translate_opt, translate_dummy_opt)
del translator.model
translator.model = model_for_train.masked_model
tt=open(translate_opt.tgt, 'r')
references = [[t] for t in tt]
translate_data = onmt.IO.ONMTDataset(translate_opt.src, translate_opt.tgt,fields,use_filter_pred=False)
prune_data = onmt.IO.OrderedIterator(dataset=translate_data, device=GPU_ID,batch_size=1, train=False, sort=False,shuffle=False)
tmp_fit2 = evaluate_trans(translator, references, prune_data, translate_data)
print('Finsished => bleu (%.4f), ppl (%.4f)' % (tmp_fit2[1]*100, tmp_fit2[0]))
#--------------------------------------------------
run_times += 1
def _work(process_id, model, dataset, args):
databin = dataset[process_id]
n_gpus = torch.cuda.device_count()
data_loader = DataLoader(databin,
shuffle=False,
num_workers=args.num_workers // n_gpus,
pin_memory=True)
print("dcpu", args.num_workers // n_gpus)
cam_sizes = [[], [], [], []] # scale 0,1,2,3
with cuda.device(process_id):
model.cuda()
gcam = GradCAM(model=model, candidate_layers=[args.target_layer])
for iter, pack in enumerate(data_loader):
img_name = pack['name'][0]
if os.path.exists(os.path.join(args.cam_out_dir,
img_name + '.npy')):
continue
size = pack['size']
strided_size = imutils.get_strided_size(size, 4)
strided_up_size = imutils.get_strided_up_size(size, 16)
outputs_cam = []
n_classes = len(list(torch.nonzero(pack['label'][0])[:, 0]))
for s_count, size_idx in enumerate([1, 0, 2, 3]):
orig_img = pack['img'][size_idx].clone()
for c_idx, c in enumerate(
list(torch.nonzero(pack['label'][0])[:, 0])):
pack['img'][size_idx] = orig_img
img_single = pack['img'][size_idx].detach()[
0] # [:, 1]: flip
if size_idx != 1:
total_adv_iter = args.adv_iter
else: # size_idx == 0
if args.adv_iter > 10:
total_adv_iter = args.adv_iter // 2
mul_for_scale = 2
elif args.adv_iter < 6:
total_adv_iter = args.adv_iter
mul_for_scale = 1
else:
total_adv_iter = 5
mul_for_scale = float(total_adv_iter) / 5
for it in range(total_adv_iter):
img_single.requires_grad = True
outputs = gcam.forward(
img_single.cuda(non_blocking=True))
if c_idx == 0 and it == 0:
cam_all_classes = torch.zeros([
n_classes, outputs.shape[2], outputs.shape[3]
])
gcam.backward(ids=c)
regions = gcam.generate(target_layer=args.target_layer)
regions = regions[0] + regions[1].flip(-1)
if it == 0:
init_cam = regions.detach()
cam_all_classes[c_idx] += regions[0].data.cpu(
) * mul_for_scale
logit = outputs
logit = F.relu(logit)
logit = torchutils.gap2d(logit, keepdims=True)[:, :, 0,
0]
valid_cat = torch.nonzero(pack['label'][0])[:, 0]
logit_loss = -2 * (logit[:,
c]).sum() + torch.sum(logit)
expanded_mask = torch.zeros(regions.shape)
expanded_mask = add_discriminative(
expanded_mask, regions, score_th=args.score_th)
L_AD = torch.sum((torch.abs(regions - init_cam)) *
expanded_mask.cuda())
loss = -logit_loss - L_AD * args.AD_coeff
model.zero_grad()
img_single.grad.zero_()
loss.backward()
data_grad = img_single.grad.data
perturbed_data = adv_climb(img_single,
args.AD_stepsize, data_grad)
img_single = perturbed_data.detach()
outputs_cam.append(cam_all_classes)
strided_cam = torch.sum(
torch.stack([
F.interpolate(torch.unsqueeze(o, 0),
strided_size,
mode='bilinear',
align_corners=False)[0] for o in outputs_cam
]), 0)
highres_cam = [
F.interpolate(torch.unsqueeze(o, 1),
strided_up_size,
mode='bilinear',
align_corners=False) for o in outputs_cam
]
highres_cam = torch.sum(torch.stack(highres_cam, 0),
0)[:, 0, :size[0], :size[1]]
strided_cam /= F.adaptive_max_pool2d(strided_cam, (1, 1)) + 1e-5
highres_cam /= F.adaptive_max_pool2d(highres_cam, (1, 1)) + 1e-5
np.save(
os.path.join(args.cam_out_dir, img_name + '.npy'), {
"keys": valid_cat,
"cam": strided_cam.cpu(),
"high_res": highres_cam.cpu().numpy()
})
def _align_nn(self,
model_path,
source_kb,
target_kb,
candidate_selector,
cuda_device,
batch_size=128):
"""
Align using neural network model
:param source_kb:
:param target_kb:
:param candidate_selector:
:param cuda_device: GPU device number
:return:
"""
from emma.allennlp_classes.ontoemma_dataset_reader import OntologyMatchingDatasetReader
from emma.allennlp_classes.ontoemma_model import OntoEmmaNN
from emma.allennlp_classes.ontoemma_predictor import OntoEmmaPredictor
alignment, s_ent_ids, t_ent_ids = self._align_string_equiv(
source_kb, target_kb, candidate_selector)
sys.stdout.write("%i alignments with string equivalence\n" %
len(alignment))
if cuda_device > 0:
with device(cuda_device):
archive = load_archive(model_path, cuda_device=cuda_device)
else:
archive = load_archive(model_path, cuda_device=cuda_device)
predictor = Predictor.from_archive(archive, 'ontoemma-predictor')
sys.stdout.write("Making predictions...\n")
s_ent_tqdm = tqdm.tqdm(s_ent_ids, total=len(s_ent_ids))
sim_scores = dict()
if cuda_device > 0:
with device(cuda_device):
batch_json_data = []
for s_ent_id in s_ent_tqdm:
s_ent = source_kb.get_entity_by_research_entity_id(
s_ent_id)
for t_ent_id in candidate_selector.select_candidates(
s_ent_id)[:constants.KEEP_TOP_K_CANDIDATES]:
t_ent = target_kb.get_entity_by_research_entity_id(
t_ent_id)
json_data = {
'source_ent': s_ent.form_json(),
'target_ent': t_ent.form_json(),
'label': 0
}
batch_json_data.append(json_data)
if len(batch_json_data) == batch_size:
results = predictor.predict_batch_json(
batch_json_data, cuda_device)
for model_input, output in zip(
batch_json_data, results):
sim_scores[(model_input['source_ent']
['research_entity_id'],
model_input['target_ent']
['research_entity_id']
)] = output['score'][0]
batch_json_data = []
# finish last batch
if batch_json_data:
results = predictor.predict_batch_json(
batch_json_data, cuda_device)
for model_input, output in zip(batch_json_data, results):
sim_scores[(
model_input['source_ent']['research_entity_id'],
model_input['target_ent']['research_entity_id']
)] = output['score'][0]
else:
for s_ent_id in s_ent_tqdm:
s_ent = source_kb.get_entity_by_research_entity_id(s_ent_id)
for t_ent_id in candidate_selector.select_candidates(
s_ent_id)[:constants.KEEP_TOP_K_CANDIDATES]:
t_ent = target_kb.get_entity_by_research_entity_id(
t_ent_id)
json_data = {
'source_ent': s_ent.form_json(),
'target_ent': t_ent.form_json(),
'label': 0
}
output = predictor.predict_json(json_data, cuda_device)
sim_scores[(json_data['source_ent']['research_entity_id'],
json_data['target_ent']['research_entity_id']
)] = output['score'][0]
return sim_scores
def main():
total_times = 100
run_times = 0
init_threshold = ...
start_t = time.time()
valid_data = torch.load(TRAIN_DATA + '.valid.pt')
fields = onmt.IO.load_fields(torch.load(TRAIN_DATA + '.vocab.pt'))
# fields = onmt.IO.load_fields_from_vocab(torch.load(TRAIN_DATA + '.vocab.pt'))
valid_data.fields = fields # we need to clear this assignment relationg if we want to transfere valid among threads
checkpoint = torch.load(weights, map_location=lambda storage, loc: storage)
model_opt = checkpoint['opt']
masked_models = []
with cuda.device(GPU_ID):
ref_model = onmt.ModelConstructor.make_base_model(
model_opt, fields, True, checkpoint)
ref_model.eval()
ref_model.generator.eval()
masked_model = MaskedModel(
ref_model, group_dict, cuda.current_device(), cuda.current_device(
)) # ref_model is at current_device, no copy will happen
masked_models.append(masked_model)
if GPU_ID:
cuda.set_device(GPU_ID)
# 1 means 1% acc
acc_percent_prune = 1
# 只需要原始的accuracy
acc_of_no_prune = 0
get_acc_of_no_prune = False
print(time_now(), "start while")
while run_times < total_times:
print("-----------------------------------------")
print("start Iteration ", run_times)
# init threshold
best_threshold = 0
itr_time = time.time()
'''
display all the names of parameters
'''
'''
aa=ref_model.named_parameters
aa_namelist = [ak[0] for ak in aa]
'''
'''
test MP
'''
translate_opt, translate_dummy_opt = translate_opt_initialize(
'opennmt_translate_opt.pt', 'opennmt_translate_dummy_opt.pt')
translator = init_translate_model(translate_opt, translate_dummy_opt)
del translator.model
translator.model = masked_model
tt = open(translate_opt.tgt, 'r')
references = [[t] for t in tt]
xxx = np.arange(0., 1, 0.01)
#for i in range(len(masked_model.group_name_list)):
# tmp_crate = len(masked_model.group_name_list)*[0.]
print(time_now(), "start testing pruning")
masked_model.make_evaluable()
for i in range(len(xxx)):
# best_threshold = 0.55
# break
translate_data = onmt.IO.ONMTDataset(translate_opt.src,
translate_opt.tgt,
fields,
use_filter_pred=False)
prune_data = onmt.IO.OrderedIterator(dataset=translate_data,
device=GPU_ID,
batch_size=1,
train=False,
sort=False,
shuffle=False)
tmp_crate = len(masked_model.group_name_list) * [xxx[i]]
#tmp_crate[i] = 0.01
masked_model.change_mask(tmp_crate, apply_MP_on_mask)
masked_model.apply_mask()
tmp_fit = evaluate(masked_model, valid_data, fields)
#tmp_fit = evaluate_trans(translator, references, prune_data, translate_data)
#logger.scalar_summary('test_bleu', tmp_fit[1]*100, int(xxx[i]*100))
#logger.scalar_summary('acc', tmp_fit[1], int(xxx[i]*100))
#logger.scalar_summary('ppl', tmp_fit[0], int(xxx[i]*100))
#logger.scalar_summary('test_ppl', tmp_fit[0], int(xxx[i]*100))
#print('group %s => acc (%.4f), ppl (%.4f)' % (masked_model.group_name_list[i], tmp_fit[1], tmp_fit[0]))
#print('percentage %s => bleu (%.4f), ppl (%.4f)' % (xxx[i]*100, tmp_fit[1]*100, tmp_fit[0]))
# print('percentage %s => acc (%.4f), ppl (%.4f)' % (xxx[i]*100, tmp_fit[1], tmp_fit[0]))
if i == 0 and not get_acc_of_no_prune:
acc_of_no_prune = tmp_fit[1]
acc_of_no_prune = int(acc_of_no_prune * 100) / 100
get_acc_of_no_prune = True
elif acc_of_no_prune - tmp_fit[1] > acc_percent_prune:
best_threshold = xxx[i] - 0.01
break
# -------------------------------------------------
# Start writing
# prune again
print(time_now(), " start accuracy:", acc_of_no_prune)
print("-------test------------:", get_acc_of_no_prune)
print(time_now(), " apply pruning with threshold:", best_threshold)
tmp_crate = len(masked_model.group_name_list) * [best_threshold]
masked_model.change_mask(tmp_crate, apply_MP_on_mask)
masked_model.apply_mask()
# print information
tmp_fit = evaluate(masked_model, valid_data, fields)
print('percentage %s => acc (%.4f), ppl (%.4f)' %
(best_threshold * 100, tmp_fit[1], tmp_fit[0]))
model_sparsity = masked_model.get_sparsity()
print('Sparsity: {}'.format(model_sparsity))
#--------------- start retraining --------------
# first store model
print(time_now(), "start saving model")
_, saved_model = update_checkpoint(checkpoint, masked_model, run_times)
print(time_now(), "finish saving model")
print(time_now(), "start loading model")
checkpoint = torch.load(SAVE_MODEL_TMP_FOLDER + saved_model,
map_location=lambda storage, loc: storage)
train_opt, _, _ = opt_initialize(checkpoint,
'opennmt_translate_opt.pt',
'opennmt_translate_dummy_opt.pt')
# train data loading
print(time_now(), "start loading data for retraining")
train = torch.load(train_opt.data + '.train.pt')
valid = torch.load(train_opt.data + '.valid.pt')
print(time_now(), "finish data loading")
train_fields = load_fields(train, valid, checkpoint, train_opt)
model_for_train = init_train_model(checkpoint, train_opt, train_fields)
masked_model = MaskedModel(model_for_train, group_dict,
cuda.current_device(),
cuda.current_device())
masked_model.make_trainable()
print(time_now(), "building optm")
optim = build_optim(model_for_train, checkpoint, train_opt)
print(time_now(), "start restraining")
recovered = train_model(model_for_train, train, valid, train_fields,
optim, train_opt, run_times, acc_of_no_prune)
print(time_now(), "finish retraining ")
if not recovered:
exit()
else:
print("------------Accuracy recorverd!--------------------")
print("recovered accuracy:", acc_of_no_prune)
run_times += 1
masked_model.make_evaluable()
tmp_fit = evaluate(masked_model, valid_data, fields)
print("------------------for test-------------------")
print('percentage %s => acc (%.4f), ppl (%.4f)' %
(best_threshold * 100, tmp_fit[1], tmp_fit[0]))
def _work_gpu(process_id, model, dataset, args):
n_gpus = torch.cuda.device_count()
databin = dataset[process_id]
data_loader = DataLoader(databin,
shuffle=False,
num_workers=args.num_workers // n_gpus,
pin_memory=False)
with torch.no_grad(), cuda.device(process_id):
model.cuda()
for iter, pack in tqdm(enumerate(data_loader), total=len(databin)):
img_name = pack['name'][0]
size = np.asarray(pack['size'])
edge, dp = model(pack['img'][0].cuda(non_blocking=True))
dp = dp.cpu().numpy()
cam_dict = np.load(args.cam_out_dir + '/' + img_name + '.npy',
allow_pickle=True).item()
cams = cam_dict['cam'].cuda()
keys = cam_dict['keys']
centroids = find_centroids_with_refinement(dp)
instance_map = cluster_centroids(centroids, dp)
instance_cam = separte_score_by_mask(cams, instance_map)
rw = indexing.propagate_to_edge(instance_cam,
edge,
beta=args.beta,
exp_times=args.exp_times,
radius=5)
rw_up = F.interpolate(rw,
scale_factor=4,
mode='bilinear',
align_corners=False)[:,
0, :size[0], :size[1]]
rw_up = rw_up / torch.max(rw_up)
rw_up_bg = F.pad(rw_up, (0, 0, 0, 0, 1, 0),
value=args.ins_seg_bg_thres)
num_classes = len(keys)
num_instances = instance_map.shape[0]
instance_shape = torch.argmax(rw_up_bg, 0).cpu().numpy()
instance_shape = pyutils.to_one_hot(
instance_shape,
maximum_val=num_instances * num_classes + 1)[1:]
instance_class_id = np.repeat(keys, num_instances)
detected = detect_instance(rw_up.cpu().numpy(),
instance_shape,
instance_class_id,
max_fragment_size=size[0] * size[1] *
0.01)
np.save(os.path.join(args.ins_seg_out_dir, img_name + '.npy'),
detected)
if process_id == n_gpus - 1 and iter % (len(databin) // 4) == 0:
print("%d " % ((5 * iter + 1) // (len(databin) // 4)), end='')
def clear_cache(self):
with cuda.device(self.tgpu_id):
cuda.empty_cache()
def distk_fdmmv(proc_idx, queue, device_id):
a: ArgsFdmmv = queue.get()
X1, X2, v, w, out = a.X1, a.X2, a.v, a.w, a.out
kernel: L2DistanceKernel = a.kernel
max_mem = a.max_mem
N, D = X1.size()
M = X2.size(0)
T = v.shape[1] if v is not None else w.shape[1]
dtype = X1.dtype
cuda_inputs = X1.is_cuda
# Memory usage:
# v : M x T
# K : n x M
# X1ss : n x d
# X2s : M x d
# Kv : n x T
# out : M x T
# sq1 : n x 1
# sq2 : M x 1
# ------------
# total : n*d + M*d + n*(M + T + 1) + 2*M*T + M
avail_mem = max_mem / sizeof_dtype(dtype)
rest_coef = 2 * M * T if v is not None else M * T
n, d = select_dim_over_nd(max_n=N,
max_d=D,
coef_nd=1,
coef_n=M + T + 1,
coef_d=M,
rest=rest_coef + M,
max_mem=avail_mem)
ddev = torch.device('cuda:%d' % int(device_id))
s1 = tcd.Stream(ddev)
s2 = tcd.Stream(ddev)
with tcd.device(ddev), tcd.stream(s1):
# First collect necessary memory
mem_needed = n * M + n * T + n + M
if not cuda_inputs:
mem_needed += n * d + M * d
if v is not None:
mem_needed += M * T
if not out.is_cuda:
mem_needed += M * T
# Create flat tensor
flat_gpu_tn = torch.empty(size=(mem_needed, ),
dtype=dtype,
device=ddev)
# Extract the sub-tensors
flat_offset = 0
if v is not None:
if not cuda_inputs:
v_gpu = extract_same_stride(flat_gpu_tn,
size=(M, T),
other=v,
offset=flat_offset)
flat_offset += np.prod(v_gpu.shape)
copy_to_device_noorder(M, T, v, 0, 0, v_gpu, 0, 0)
else:
v_gpu = v
K_gpu = extract_same_stride(flat_gpu_tn,
size=(n, M),
other=X1,
offset=flat_offset)
flat_offset += np.prod(K_gpu.shape)
Kv_gpu = extract_same_stride(flat_gpu_tn,
size=(n, T),
other=X1,
offset=flat_offset)
flat_offset += np.prod(Kv_gpu.shape)
if out.is_cuda:
out_gpu = out
else:
out_gpu = extract_same_stride(flat_gpu_tn,
size=(M, T),
other=out,
offset=flat_offset)
flat_offset += np.prod(out_gpu.shape)
out_gpu.fill_(0.0)
if not cuda_inputs:
X1ss_gpu = extract_same_stride(flat_gpu_tn,
size=(n, d),
other=X1,
offset=flat_offset)
flat_offset += np.prod(X1ss_gpu.shape)
X2s_gpu = extract_same_stride(flat_gpu_tn,
size=(M, d),
other=X2,
offset=flat_offset)
flat_offset += np.prod(X2s_gpu.shape)
sq1_gpu = extract_same_stride(flat_gpu_tn,
size=(n, ),
other=X1,
offset=flat_offset)
flat_offset += np.prod(sq1_gpu.shape)
sq2_gpu = extract_same_stride(flat_gpu_tn,
size=(M, ),
other=X1,
offset=flat_offset)
for i in range(0, N, n):
nb = min(N - i, n)
cur_K_gpu = K_gpu[:nb] # nb x M
cur_K_gpu.fill_(0.0)
for j in range(0, D, d):
db = min(D - j, d)
s1.synchronize(
) # need that the add_(sq2_gpu.T) op is complete to avoid overwrite
# Parallelize two matrix transfers
with tcd.stream(s2):
if cuda_inputs:
cur_X2s_gpu = X2[:, j:j + db]
else:
cur_X2s_gpu = copy_to_device_noorder(M,
db,
X2,
0,
j,
X2s_gpu,
0,
0,
s=s2)
torch.norm(cur_X2s_gpu,
p=2,
dim=1,
keepdim=True,
out=sq2_gpu).pow_(2)
if cuda_inputs:
cur_X1ss_gpu = X1[i:i + nb, j:j + db]
else:
cur_X1ss_gpu = copy_to_device_noorder(nb,
db,
X1,
i,
j,
X1ss_gpu,
0,
0,
s=s1)
torch.norm(cur_X1ss_gpu, p=2, dim=1, keepdim=True,
out=sq1_gpu).pow_(2)
s2.synchronize(
) # need that cur_X2s_gpu and sq2_gpu are available.
cur_K_gpu.addmm_(mat1=cur_X1ss_gpu,
mat2=cur_X2s_gpu.T,
alpha=-2.0)
cur_K_gpu.add_(sq1_gpu)
cur_K_gpu.add_(sq2_gpu.T)
cur_K_gpu.clamp_min_(0)
cur_K_gpu = kernel._transform(cur_K_gpu)
if w is not None:
cur_Kv_gpu = copy_to_device_noorder(nb,
T,
w,
i,
0,
Kv_gpu,
0,
0,
s=s1) # n x T
if v is not None:
cur_Kv_gpu.addmm_(cur_K_gpu, v_gpu)
else:
# v cannot be None if w is None
cur_Kv_gpu = Kv_gpu.narrow(0, 0, nb) # n x T
torch.mm(cur_K_gpu, v_gpu, out=cur_Kv_gpu) # n x T
# Multiply transposed kernel with the Kv result.
out_gpu.addmm_(cur_K_gpu.T, cur_Kv_gpu) # M x T
if not out.is_cuda:
copy_to_host_noorder(M, T, out_gpu, 0, 0, out, 0, 0)
s1.synchronize()
return out
