def issue_queries(self, query_samples):
for query_sample in query_samples:
waveform = self.qsl[query_sample.index]
assert waveform.ndim == 1
waveform_length = np.array(waveform.shape[0], dtype=np.int64)
waveform = np.expand_dims(waveform, 0)
waveform_length = np.expand_dims(waveform_length, 0)
with torch.no_grad():
waveform = torch.from_numpy(waveform)
waveform_length = torch.from_numpy(waveform_length)
feature, feature_length = self.audio_preprocessor.forward(
(waveform, waveform_length))
assert feature.ndim == 3
assert feature_length.ndim == 1
feature = feature.permute(2, 0, 1)
_, _, transcript = self.greedy_decoder.forward(
feature, feature_length)
assert len(transcript) == 1
response_array = array.array('q', transcript[0])
bi = response_array.buffer_info()
response = lg.QuerySampleResponse(query_sample.id, bi[0],
bi[1] * response_array.itemsize)
lg.QuerySamplesComplete([response])
def run_one_item(self, qitem):
# run the prediction
processed_results = []
try:
results = self.model.predict(qitem.batch_dense_X, qitem.batch_lS_o, qitem.batch_lS_i)
processed_results = self.post_process(results, qitem.batch_T, self.result_dict)
if self.take_accuracy:
self.post_process.add_results(processed_results)
self.result_timing.append(time.time() - qitem.start)
except Exception as ex: # pylint: disable=broad-except
log.error("thread: failed, %s", ex)
# since post_process will not run, fake empty responses
processed_results = [[]] * len(qitem.query_id)
finally:
response_array_refs = []
response = []
for idx, query_id in enumerate(qitem.query_id):
# NOTE: processed_results returned by DlrmPostProcess store both
# result = processed_results[idx][0] and target = processed_results[idx][1]
# also each idx might be a query of samples, rather than a single sample
# depending on the --samples-to-aggregate* arguments.
s_idx = qitem.idx_offsets[idx]
e_idx = qitem.idx_offsets[idx + 1]
# debug prints
# print("s,e:",s_idx,e_idx, len(processed_results))
response_array = array.array("B", np.array(processed_results[s_idx:e_idx], np.float32).tobytes())
response_array_refs.append(response_array)
bi = response_array.buffer_info()
response.append(lg.QuerySampleResponse(query_id, bi[0], bi[1]))
lg.QuerySamplesComplete(response)
def process_query_async(query_samples, i_slice):
time.sleep(LOOPBACK_LATENCY_S * (i_slice + 1))
responses = []
samples_to_complete = query_samples[i_slice:len(query_samples):NUM_AGENTS]
for j, s in enumerate(samples_to_complete):
responses.append(mlperf_loadgen.QuerySampleResponse(s.id, 0, 0))
mlperf_loadgen.QuerySamplesComplete(responses)
def run_one_item(self, qitem):
# run the prediction
processed_results = []
try:
results = self.model.predict({self.model.inputs[0]: qitem.img})
processed_results = self.post_process(results, qitem.content_id,
qitem.label,
self.result_dict)
if self.take_accuracy:
self.post_process.add_results(processed_results)
self.result_timing.append(time.time() - qitem.start)
except Exception as ex: # pylint: disable=broad-except
src = [self.ds.get_item_loc(i) for i in qitem.content_id]
log.error("thread: failed on contentid=%s, %s", src, ex)
# since post_process will not run, fake empty responses
processed_results = [[]] * len(qitem.query_id)
finally:
response_array_refs = []
response = []
for idx, query_id in enumerate(qitem.query_id):
response_array = array.array(
"B",
np.array(processed_results[idx], np.float32).tobytes())
response_array_refs.append(response_array)
bi = response_array.buffer_info()
response.append(lg.QuerySampleResponse(query_id, bi[0], bi[1]))
lg.QuerySamplesComplete(response)
def run_one_item(self, qitem, cloud_tpu_id):
# run the prediction
try:
for i in [1]:
data, _ = self.ds.get_indices(qitem.content_id)
results = self.models[cloud_tpu_id].predict(data)
except Exception as ex: # pylint: disable=broad-except
src = [self.ds.get_item_loc(i) for i in qitem.content_id]
log.error("thread: failed on contentid=%s, %s", src, ex)
results = [[]] * len(qitem.query_id)
# since post_process will not run, fake empty responses
finally:
response = []
response_arrays = []
# Ignore padded samples.
results = np.array(results).reshape(
(self.max_batchsize, -1)).tolist()
for idx, query_id in enumerate(qitem.query_id):
if query_id == -1:
continue
response_array = array.array(
"B",
np.array(results[idx], np.float32).tobytes())
response_arrays.append(response_array)
index = len(response_arrays) - 1
bi = response_arrays[index].buffer_info()
response.append(lg.QuerySampleResponse(query_id, bi[0], bi[1]))
lg.QuerySamplesComplete(response)
def handle_tasks(self, tpu_id):
while True:
# Block until an item becomes available
qitem = self.tasks.get(block=True)
# When a "None" item was added, it is a
# signal from the parent to indicate we should stop
# working (see finish)
if qitem is None:
break
results = self.process(qitem, tpu_id)
response = []
gc_hack = []
for res, q_id in zip(results, qitem.query_id_list):
result_arr = array.array("B", res)
gc_hack.append(result_arr)
r_info = result_arr.buffer_info()
response.append(
mlperf_loadgen.QuerySampleResponse(q_id, r_info[0],
r_info[1]))
# Tell loadgen that we're ready with this query
mlperf_loadgen.QuerySamplesComplete(response)
self.tasks.task_done()
def issue_queries(self, query_samples):
input_ids = np.zeros((len(query_samples), 1, 384), dtype=np.int32)
input_mask = np.zeros((len(query_samples), 1, 384), dtype=np.int32)
segment_ids = np.zeros((len(query_samples), 1, 384), dtype=np.int32)
for sample_idx in range(len(query_samples)):
eval_features = self.qsl.get_features(
query_samples[sample_idx].index)
input_ids[sample_idx, ...] = np.array(eval_features.input_ids)
input_mask[sample_idx, ...] = np.array(eval_features.input_mask)
segment_ids[sample_idx, ...] = np.array(eval_features.segment_ids)
def input_fn():
inputs = {
"input_ids": input_ids,
"input_mask": input_mask,
"segment_ids": segment_ids
}
return tf.data.Dataset.from_tensor_slices(inputs)
for i, result in enumerate(self.estimator.predict(input_fn)):
logits = [float(x) for x in result["logits"].flat]
response_array = array.array(
"B",
np.array(logits).astype(np.float32).tobytes())
bi = response_array.buffer_info()
response = lg.QuerySampleResponse(query_samples[i].id, bi[0],
bi[1])
lg.QuerySamplesComplete([response])
def issue_queries(query_samples):
global BATCH_SIZE
if VERBOSITY_LEVEL:
printable_query = [(qs.index, qs.id) for qs in query_samples]
print("issue_queries( {} )".format(printable_query))
tick('Q', len(query_samples))
for j in range(0, len(query_samples), BATCH_SIZE):
batch = query_samples[
j:j +
BATCH_SIZE] # NB: the last one may be shorter than BATCH_SIZE in length
batch_data = [preprocessed_image_buffer[qs.index] for qs in batch]
batch_predicted_labels = predict_labels_for_batch(batch_data)
tick('p', len(batch))
if VERBOSITY_LEVEL:
print(
"predicted_batch_results = {}".format(batch_predicted_labels))
response = []
response_array_refs = [
] # This is needed to guarantee that the individual buffers to which we keep extra-Pythonian references, do not get garbage-collected.
for qs, predicted_label in zip(batch, batch_predicted_labels):
response_array = array.array(
"B",
np.array(predicted_label, np.float32).tobytes())
response_array_refs.append(response_array)
bi = response_array.buffer_info()
response.append(lg.QuerySampleResponse(qs.id, bi[0], bi[1]))
lg.QuerySamplesComplete(response)
#tick('R', len(response))
sys.stdout.flush()
def response_loadgen(out_queue):
global out_queue_cnt
while True:
next_task = out_queue.get()
if next_task is None:
# None means shutdown
log.info('Exiting response thread')
break
query_id_list = next_task.query_id_list
result = next_task.result
batch_size = len(query_id_list)
result.reshape(batch_size, -1, 2)
out_list = np.split(result, batch_size, axis=0)
#responses = []
for i, o in enumerate(out_list):
response_array = array.array(
"B",
np.array(o).astype(np.float32).tobytes())
bi = response_array.buffer_info()
#responses.append(lg.QuerySampleResponse(query_id_list[i], bi[0], bi[1]))
responses = [
lg.QuerySampleResponse(query_id_list[i], bi[0], bi[1])
]
out_queue_cnt += 1
#print('Response loadgen ({}), query_id {}, out_queue_cnt {}'.format(os.getpid(), query_id_list[i], out_queue_cnt))
lg.QuerySamplesComplete(responses)
def handle_tasks(self):
while True:
# Block until an item becomes available
qitem = self.tasks.get(block=True)
# When a "None" item was added, it is a
# signal from the parent to indicate we should stop
# working (see finish)
if qitem is None:
break
result = self.process(qitem)
response = []
# TBD: do something when we are running accuracy mode
# We need to properly store the result. Perhaps through QuerySampleResponse, otherwise internally
# in this instance of Runner.
# QuerySampleResponse contains an ID, a size field and a data pointer field
for query_id in qitem.query_id:
response.append(mlperf_loadgen.QuerySampleResponse(query_id, 0, 0))
# Tell loadgen that we're ready with this query
mlperf_loadgen.QuerySamplesComplete(response)
self.tasks.task_done()
def handle_tasks(self, tasks_queue):
"""Worker thread."""
while True:
qitem = tasks_queue.get()
if qitem is None:
# None in the queue indicates the parent want us to exit
tasks_queue.task_done()
break
try:
# run the prediction
results = self.model.predict({self.model.inputs[0]: qitem.img})
if self.take_accuracy:
response = self.post_process(results, qitem.content_id,
qitem.label, self.result_dict)
except Exception as ex: # pylint: disable=broad-except
src = [self.ds.get_item_loc(i) for i in qitem.content_id]
log.error("thread: failed on contentid=%s, %s", src, ex)
finally:
response = []
for query_id in qitem.query_id:
# FIXME: unclear what to return here
response.append(lg.QuerySampleResponse(query_id, 0, 0))
lg.QuerySamplesComplete(response)
tasks_queue.task_done()
def issue_queries(query_samples):
printable_query = [(qs.index, qs.id) for qs in query_samples]
print("LG: issue_queries( {} )".format(printable_query))
predicted_results = {}
for qs in query_samples:
query_index, query_id = qs.index, qs.id
x_vector = dataset[query_index]
predicted_label = predict_label(x_vector)
predicted_results[query_index] = predicted_label
print("LG: predicted_results = {}".format(predicted_results))
response = []
for qs in query_samples:
query_index, query_id = qs.index, qs.id
response_array = array.array(
"B",
np.array(predicted_results[query_index], np.float32).tobytes())
bi = response_array.buffer_info()
response.append(lg.QuerySampleResponse(query_id, bi[0], bi[1]))
lg.QuerySamplesComplete(response)
def process_query_async(query_samples, i_slice):
time.sleep(.001 * (i_slice + 1))
responses = []
samples_to_complete = query_samples[i_slice:len(query_samples):3]
for s in samples_to_complete:
responses.append(mlperf_loadgen.QuerySampleResponse(s.id, 0, 0))
mlperf_loadgen.QuerySamplesComplete(responses)
def handle_tasks(self, tasks_queue):
"""Worker thread."""
while True:
qitem = tasks_queue.get()
if qitem is None:
# None in the queue indicates the parent want us to exit
tasks_queue.task_done()
break
try:
# run the prediction
results = self.model.predict({self.model.inputs[0]: qitem.img})
# and keep track of how long it took
took = time.time() - qitem.start
response = []
for idx, result in enumerate(results[0]):
result = self.post_process(result)
if qitem.label[idx] == result:
self.result_dict["good"] += 1
self.result_dict["total"] += 1
# FIXME: unclear what to return here
response.append(lg.QuerySampleResponse(
qitem.id[idx], 0, 0))
self.result_list.append(took)
lg.QuerySamplesComplete(response)
except Exception as ex: # pylint: disable=broad-except
log.error("execute_parallel thread: %s", ex)
tasks_queue.task_done()
def post_process(self, query_ids, results):
response = []
for res, q_id in zip(results, query_ids):
response.append(mlperf_loadgen.QuerySampleResponse(q_id, 0, 0))
# Tell loadgen that we're ready with this query
mlperf_loadgen.QuerySamplesComplete(response)
def issue_queries(self, query_samples):
with torch.no_grad():
for i in range(len(query_samples)):
data = self.qsl.get_features(query_samples[i].index)
print("Processing sample id {:d} with shape = {:}".format(
query_samples[i].index, data.shape))
image = torch.from_numpy(data[np.newaxis,
...]).float().to(self.device)
prediction = self.trainer.network(image)
softmax = F.softmax(prediction[0],
dim=1).cpu().numpy().astype(np.float16)
transpose_forward = self.trainer.plans.get("transpose_forward")
transpose_backward = self.trainer.plans.get(
"transpose_backward")
assert transpose_forward == [
0, 1, 2
], "Unexpected transpose_forward {:}".format(transpose_forward)
assert transpose_backward == [
0, 1, 2
], "Unexpected transpose_backward {:}".format(
transpose_backward)
response_array = array.array("B", softmax.tobytes())
bi = response_array.buffer_info()
response = lg.QuerySampleResponse(query_samples[i].id, bi[0],
bi[1])
lg.QuerySamplesComplete([response])
def issue_queries(query_samples):
global BATCH_SIZE
global model
if VERBOSITY_LEVEL > 2:
printable_query = [(qs.index, qs.id) for qs in query_samples]
print("issue_queries( {} )".format(printable_query))
tick('Q', len(query_samples))
for j in range(0, len(query_samples), BATCH_SIZE):
batch = query_samples[
j:j +
BATCH_SIZE] # NB: the last one may be shorter than BATCH_SIZE in length
batch_data = preprocessed_image_buffer[preprocessed_image_map[[
qs.index for qs in batch
]]]
torch_batch = torch.from_numpy(batch_data)
begin_time = time.time()
# move the input to GPU for speed if available
if USE_CUDA:
torch_batch = torch_batch.to('cuda')
with torch.no_grad():
trimmed_batch_results = model(torch_batch)
inference_time_s = time.time() - begin_time
actual_batch_size = len(trimmed_batch_results)
if VERBOSITY_LEVEL > 1:
print("[batch of {}] inference={:.2f} ms".format(
actual_batch_size, inference_time_s * 1000))
batch_predicted_labels = torch.argmax(trimmed_batch_results,
dim=1).tolist()
tick('p', len(batch))
if VERBOSITY_LEVEL > 2:
print(
"predicted_batch_results = {}".format(batch_predicted_labels))
response = []
response_array_refs = [
] # This is needed to guarantee that the individual buffers to which we keep extra-Pythonian references, do not get garbage-collected.
for qs, predicted_label in zip(batch, batch_predicted_labels):
response_array = array.array(
"B",
np.array(predicted_label, np.float32).tobytes())
response_array_refs.append(response_array)
bi = response_array.buffer_info()
response.append(lg.QuerySampleResponse(qs.id, bi[0], bi[1]))
lg.QuerySamplesComplete(response)
#tick('R', len(response))
sys.stdout.flush()
def process_query_async(query_samples):
time.sleep(.001)
responses = []
response_data = array.array(
'B', [0, 1, 7, 8, 15, 16, 31, 32, 63, 64, 127, 128, 254, 255])
rdbi = response_data.buffer_info()
for s in query_samples:
responses.append(
mlperf_loadgen.QuerySampleResponse(s.id, rdbi[0], rdbi[1]))
mlperf_loadgen.QuerySamplesComplete(responses)
def issue_queries(query_samples):
global BATCH_SIZE
global model_output_volume
global num_classes
if VERBOSITY_LEVEL > 2:
printable_query = [(qs.index, qs.id) for qs in query_samples]
print("issue_queries( {} )".format(printable_query))
tick('Q', len(query_samples))
for j in range(0, len(query_samples), BATCH_SIZE):
batch = query_samples[
j:j +
BATCH_SIZE] # NB: the last one may be shorter than BATCH_SIZE in length
batch_data = preprocessed_image_buffer[preprocessed_image_map[[
qs.index for qs in batch
]]]
trimmed_batch_results, inference_time_s = inference_for_given_batch(
batch_data)
actual_batch_size = len(trimmed_batch_results)
if VERBOSITY_LEVEL > 1:
print("[batch of {}] inference={:.2f} ms".format(
actual_batch_size, inference_time_s * 1000))
if model_output_volume == 1:
batch_predicted_labels = trimmed_batch_results
else:
batch_predicted_labels = [
np.argmax(trimmed_batch_results[k][-num_classes:])
for k in range(actual_batch_size)
]
tick('p', len(batch))
if VERBOSITY_LEVEL > 2:
print(
"predicted_batch_results = {}".format(batch_predicted_labels))
response = []
response_array_refs = [
] # This is needed to guarantee that the individual buffers to which we keep extra-Pythonian references, do not get garbage-collected.
for qs, predicted_label in zip(batch, batch_predicted_labels):
response_array = array.array(
"B",
np.array(predicted_label, np.float32).tobytes())
response_array_refs.append(response_array)
bi = response_array.buffer_info()
response.append(lg.QuerySampleResponse(qs.id, bi[0], bi[1]))
lg.QuerySamplesComplete(response)
#tick('R', len(response))
sys.stdout.flush()
def issue_queries(query_samples):
global BATCH_SIZE
if VERBOSITY_LEVEL:
printable_query = [(qs.index, qs.id) for qs in query_samples]
print("issue_queries( {} )".format(printable_query))
tick('Q', len(query_samples))
for j in range(0, len(query_samples), BATCH_SIZE):
batch = query_samples[
j:j +
BATCH_SIZE] # NB: the last one may be shorter than BATCH_SIZE in length
batch_data = [preprocessed_image_buffer[qs.index] for qs in batch]
predictions_for_a_batch = predict_labels_for_batch(batch_data)
tick('p', len(batch))
if VERBOSITY_LEVEL:
print(
"predicted_batch_results = {}".format(predictions_for_a_batch))
response = []
response_array_refs = [
] # This is needed to guarantee that the individual buffers to which we keep extra-Pythonian references, do not get garbage-collected.
for qs, all_boxes_for_this_sample in zip(batch,
predictions_for_a_batch):
num_active_boxes_for_this_sample = all_boxes_for_this_sample[
MODEL_MAX_PREDICTIONS * 7].view('int32')
global_image_index = qs.index
width_orig, height_orig = original_w_h[global_image_index]
reformed_active_boxes_for_this_sample = []
for i in range(num_active_boxes_for_this_sample):
(image_id, ymin, xmin, ymax, xmax, confidence_score,
class_number) = all_boxes_for_this_sample[i * 7:(i + 1) * 7]
if class_map:
class_number = float(class_map[int(class_number)])
reformed_active_boxes_for_this_sample += [
float(global_image_index), ymin, xmin, ymax, xmax,
confidence_score, class_number
]
response_array = array.array(
"B",
np.array(reformed_active_boxes_for_this_sample,
np.float32).tobytes())
response_array_refs.append(response_array)
bi = response_array.buffer_info()
response.append(lg.QuerySampleResponse(qs.id, bi[0], bi[1]))
lg.QuerySamplesComplete(response)
#tick('R', len(response))
sys.stdout.flush()
def worker_code():
while True:
## Grab a new batch:
#
grabbed_count = 0
deadline_ts = None
batch_jobs = []
batch_inputs = []
for grabbed_count in range(
BATCH_CAPACITY
): # may not run to the end due to the cumulative timeout
try:
grab_timeout = None if deadline_ts == None else max(
0, deadline_ts -
time.time()) # no waiting limit on the first job
job = task_queue.get(timeout=grab_timeout)
batch_jobs.append(job)
batch_inputs.append(job['inputs'])
if grabbed_count == 0:
deadline_ts = job['ts_submitted'] + TOPUP_TIME_S
except queue.Empty:
break # we ran out of TOPUP_TIME_S
print(f"LG: worker grabbed and submitted {len(batch_jobs)} jobs")
## Predict the whole batch:
#
predicted_labels = predict_labels(
batch_inputs) # takes LATENCY_S of time
ts_predicted = time.time()
## Report batch results:
#
for index_in_batch, job in enumerate(batch_jobs):
predicted_label = predicted_labels[index_in_batch]
one_input = job['inputs']
query_sample = job['query_sample']
ts_submitted = job['ts_submitted']
print(
f"LG: worker predicted: for input={one_input} label={predicted_label} in {(ts_predicted-ts_submitted)*1000} ms"
)
response_array = array.array(
"B",
np.array(predicted_label, np.float32).tobytes())
bi = response_array.buffer_info()
response = lg.QuerySampleResponse(query_sample.id, bi[0], bi[1])
lg.QuerySamplesComplete([response])
task_queue.task_done()
def process_query_async(query_samples):
time.sleep(.001)
responses = []
response_array = array.array(
'f', [0,1,7,8,15,16,31,32,63,64,127,128,254,255])
response_info = response_array.buffer_info()
response_data = response_info[0]
response_size = response_info[1] * response_array.itemsize
for s in query_samples:
responses.append(
mlperf_loadgen.QuerySampleResponse(
s.id, response_data, response_size))
mlperf_loadgen.QuerySamplesComplete(responses)
def issue_queries(self, query_samples):
with torch.no_grad():
for i in range(len(query_samples)):
eval_features = self.qsl.get_features(query_samples[i].index)
start_scores, end_scores = self.model.forward(input_ids=torch.LongTensor(eval_features.input_ids).unsqueeze(0).cuda(),
attention_mask=torch.LongTensor(eval_features.input_mask).unsqueeze(0).cuda(),
token_type_ids=torch.LongTensor(eval_features.segment_ids).unsqueeze(0).cuda())
output = torch.stack([start_scores, end_scores], axis=-1).squeeze(0).cpu().numpy()
response_array = array.array("B", output.tobytes())
bi = response_array.buffer_info()
response = lg.QuerySampleResponse(query_samples[i].id, bi[0], bi[1])
lg.QuerySamplesComplete([response])
def run_one_item(self, qitem):
# run the prediction
processed_results = []
query_id, content_id, feed = qitem
results = self.session.run(self.onnx_output_names, feed)
processed_results = [[]] * len(query_id)
response_array_refs = []
response = []
for idx, qid in enumerate(query_id):
response_array = array.array("B", np.array(processed_results[idx], np.float32).tobytes())
response_array_refs.append(response_array)
bi = response_array.buffer_info()
response.append(lg.QuerySampleResponse(qid, bi[0], bi[1]))
lg.QuerySamplesComplete(response)
def issue_queries(self, query_samples):
for i in range(len(query_samples)):
data = self.qsl.get_features(query_samples[i].index)
print("Processing sample id {:d} with shape = {:}".format(
query_samples[i].index, data.shape))
output = self.sess.run(self.output, feed_dict={self.input: data[np.newaxis, ...]})[0].astype(np.float16)
response_array = array.array("B", output.tobytes())
bi = response_array.buffer_info()
response = lg.QuerySampleResponse(query_samples[i].id, bi[0],
bi[1])
lg.QuerySamplesComplete([response])
def issue_queries(self, query_samples):
for i in range(len(query_samples)):
data = self.qsl.get_features(query_samples[i].index)
print("Processing sample id {:d} with shape = {:}".format(query_samples[i].index, data.shape))
# Follow the PyTorch implementation.
# The ONNX file has five outputs, but we only care about the one named "output".
output = self.sess.run(["output"], {"input": data[np.newaxis, ...]})[0].squeeze(0).astype(np.float16)
response_array = array.array("B", output.tobytes())
bi = response_array.buffer_info()
response = lg.QuerySampleResponse(query_samples[i].id, bi[0], bi[1])
lg.QuerySamplesComplete([response])
def issue_queries(query_samples):
global so
global last_timeing
global result_timeing
idx = np.array([q.index for q in query_samples]).astype(np.int32)
query_id = [q.id for q in query_samples]
if args.dataset == 'brats2019':
start = time.time()
response_array_refs = []
response = []
for i, qid in enumerate(query_id):
processed_results = so.IssueQuery(1, idx[i][np.newaxis])
processed_results = json.loads(
processed_results.decode('utf-8'))
response_array = array.array(
"B",
np.array(processed_results[0], np.float16).tobytes())
response_array_refs.append(response_array)
bi = response_array.buffer_info()
response.append(lg.QuerySampleResponse(qid, bi[0], bi[1]))
result_timeing.append(time.time() - start)
lg.QuerySamplesComplete(response)
else:
start = time.time()
processed_results = so.IssueQuery(len(idx), idx)
result_timeing.append(time.time() - start)
processed_results = json.loads(processed_results.decode('utf-8'))
response_array_refs = []
response = []
for idx, qid in enumerate(query_id):
response_array = array.array(
"B",
np.array(processed_results[idx], np.float32).tobytes())
response_array_refs.append(response_array)
bi = response_array.buffer_info()
response.append(lg.QuerySampleResponse(qid, bi[0], bi[1]))
lg.QuerySamplesComplete(response)
def issue_queries(self, query_samples):
for i in range(len(query_samples)):
data = self.qsl.get_features(query_samples[i].index)
print("Processing sample id {:d} with shape = {:}".format(
query_samples[i].index, data.shape))
before_softmax = self.exec_net.infer(
inputs={self.input_name: data[np.newaxis,
...]})[self.output_name]
after_softmax = softmax(before_softmax, axis=1).astype(np.float16)
response_array = array.array("B", after_softmax.tobytes())
bi = response_array.buffer_info()
response = lg.QuerySampleResponse(query_samples[i].id, bi[0],
bi[1])
lg.QuerySamplesComplete([response])
def postprocess(self, qitem):
qitem.readyEvent.synchronize()
with torch.cuda.stream(self.ppStream):
qitem.results = qitem.results.cpu()
qitem.results = qitem.results.numpy()
processed_results = self.post_proc(qitem.results, qitem.label,
self.result_dict)
response_array_refs = []
response = []
for result, query_id in zip(processed_results, qitem.query_id):
response_array = array.array("B", np.array(result, np.float32).tobytes())
response_array_refs.append(response_array)
bi = response_array.buffer_info()
response.append(lg.QuerySampleResponse(query_id, bi[0], bi[1]))
lg.QuerySamplesComplete(response)
if qitem.index == self.batches - 1:
self.finishTime = time.time()
def issue_queries(self, query_samples):
for i in range(len(query_samples)):
eval_features = self.qsl.get_features(query_samples[i].index)
input_ids = np.array([eval_features.input_ids])
input_mask = np.array([eval_features.input_mask])
segment_ids = np.array([eval_features.segment_ids])
feeds = {
'input_ids:0': input_ids,
'input_mask:0': input_mask,
'segment_ids:0': segment_ids
}
result = self.sess.run(["logits:0"], feed_dict=feeds)
logits = [float(x) for x in result[0].flat]
response_array = array.array("B", np.array(logits).astype(np.float32).tobytes())
bi = response_array.buffer_info()
response = lg.QuerySampleResponse(query_samples[i].id, bi[0], bi[1])
lg.QuerySamplesComplete([response])
