def balance_class(self, prediction, target_idx, k = 1):
predict_idx = prediction.argmax(-1)
bg_predict = predict_idx == 0 # correct
fg_predict = predict_idx != 0 # error type or correction char idx
target_idx = target_idx.expand_dims(-1)
bg_target = target_idx == 0
fg_target = target_idx != 0
num_bg_predict = int(bg_predict.sum().asnumpy())
num_bg_target = int(bg_target.sum().asnumpy())
num_fg_predict = int(fg_predict.sum().asnumpy())
num_fg_target = int(fg_target.sum().asnumpy())
topk_target = nd.topk(bg_target.reshape(-1) + nd.random.uniform_like(bg_target, 0.0, 0.1).reshape(-1),
k = min(num_fg_target * k, int(np.prod(bg_target.shape))), ret_typ = 'mask').reshape_like(bg_target)
topk_predict = nd.topk(bg_predict.reshape(-1) + nd.random.uniform_like(bg_predict, 0.0, 0.1).reshape(-1),
k = min(num_fg_predict * k, int(np.prod(bg_predict.shape))), ret_typ = 'mask').reshape_like(bg_predict)
if len(topk_predict.shape) == 3:
topk_predict = topk_predict[:, :, 0]
if len(topk_target.shape) == 3:
topk_target = topk_target[:, :, 0]
if len(fg_target.shape) == 3:
fg_target = fg_target[:, :, 0]
mask = (topk_target + fg_target + fg_predict + topk_predict).clip(0, 1)
return mask
def topk(hm, k=100):
ctx = hm.context
batch_size, cat, height, width = hm.shape
hm = nms(hm)
hm = nd.reshape(hm, (0, 0, -1))
topk_scores, topk_idx = nd.topk(hm, k=k, ret_typ='both')
topk_x_idx = nd.floor(topk_idx/width)
topk_x_idx = nd.reshape(topk_x_idx, (0, -1))
topk_y_idx = (topk_idx%height)
topk_y_idx = nd.reshape(topk_y_idx, (0, -1))
topk_scores = nd.reshape(topk_scores, (0, -1))
topk_cat_scores, topk_cat_idx = nd.topk(topk_scores, k=k, ret_typ='both')
cls_id = nd.floor(topk_cat_idx/k)
batch_idx = nd.repeat(nd.arange(batch_size), repeats=k).reshape((1, -1))
batch_idx = batch_idx.as_in_context(ctx)
topk_cat_idx = nd.reshape(topk_cat_idx, (1, -1))
topk_cat_idices = nd.concat(batch_idx, topk_cat_idx, dim=0)
topk_cat_x_idx = nd.gather_nd(topk_x_idx, topk_cat_idices)
topk_cat_x_idx = nd.reshape(topk_cat_x_idx, (batch_size, k))
topk_cat_y_idx = nd.gather_nd(topk_y_idx, topk_cat_idices)
topk_cat_y_idx = nd.reshape(topk_cat_y_idx, (batch_size, k))
return topk_cat_x_idx, topk_cat_y_idx, cls_id
def _topk(scores, K=40):
batch, cat, height, width = scores.shape
[topk_scores, topk_inds] = nd.topk(nd.reshape(scores, (batch, cat, -1)),
ret_typ='both',
k=K) # return both value and indices
topk_inds = topk_inds % (height * width)
topk_ys = (topk_inds / width).astype('int32').astype('float32')
topk_xs = (topk_inds % width).astype('int32').astype('float32')
[topk_score, topk_ind] = nd.topk(nd.reshape(topk_scores, (batch, -1)),
ret_typ='both',
k=K)
topk_clses = (topk_ind / K).astype('int32')
topk_inds = _gather_feat(nd.reshape(topk_inds, (batch, -1, 1)), topk_ind)
topk_inds = nd.reshape(topk_inds, (batch, K))
topk_ys = _gather_feat(nd.reshape(topk_ys, (batch, -1, 1)), topk_ind)
topk_ys = nd.reshape(topk_ys, (batch, K))
topk_xs = _gather_feat(nd.reshape(topk_xs, (batch, -1, 1)), topk_ind)
topk_xs = nd.reshape(topk_xs, (batch, K))
return topk_score, topk_inds, topk_clses, topk_ys, topk_xs
def check_topk():
a = create_vector(size=LARGE_X)
ind = nd.topk(a, k=10, axis=0, dtype=np.int64)
for i in range(10):
assert ind[i] == (LARGE_X - i - 1)
ind, val = mx.nd.topk(a, k=3, axis=0, dtype=np.int64, ret_typ="both", is_ascend=False)
assert np.all(ind == val)
val = nd.topk(a, k=1, axis=0, dtype=np.int64, ret_typ="value")
assert val == (LARGE_X - 1)
def test_topk():
b = create_2d_tensor(rows=LARGE_X, columns=SMALL_Y)
k = nd.topk(b, k=10, axis=0, dtype=np.int64)
assert np.sum(k.asnumpy() == (LARGE_X - 1)) == SMALL_Y
ind, val = mx.nd.topk(b, k=3, axis=0, dtype=np.int64, ret_typ="both", is_ascend=False)
assert np.all(ind == val)
b = create_2d_tensor(rows=SMALL_Y, columns=LARGE_X)
l = nd.topk(b, k=1, axis=-1, dtype=np.int64, ret_typ="value")
assert l.sum() == np.sum(np.arange(0, SMALL_Y))
def _predict(self, x, ctx_id=0, with_proba=False):
if with_proba:
return self._predict_proba(x, ctx_id=ctx_id)
x = self._predict_preprocess(x)
if isinstance(x, pd.DataFrame):
assert 'image' in x.columns, "Expect column `image` for input images"
df = self._predict(tuple(x['image']))
return df.reset_index(drop=True)
elif isinstance(x, (list, tuple)):
bs = self._cfg.valid.batch_size
self.net.hybridize()
results = []
topK = min(5, self.num_class)
loader = mx.gluon.data.DataLoader(ImageListDataset(
x, self._predict_preprocess),
batch_size=bs,
last_batch='keep')
idx = 0
for batch in loader:
batch = mx.gluon.utils.split_and_load(batch,
ctx_list=self.ctx,
even_split=False)
pred = [self.net(input) for input in batch]
for p in pred:
for ii in range(p.shape[0]):
ind = nd.topk(
p[ii], k=topK).astype('int').asnumpy().flatten()
probs = mx.nd.softmax(p[ii]).asnumpy().flatten()
for k in range(topK):
results.append({
'class': self.classes[ind[k]],
'score': probs[ind[k]],
'id': ind[k],
'image': x[idx]
})
idx += 1
return pd.DataFrame(results)
elif not isinstance(x, mx.nd.NDArray):
raise ValueError('Input is not supported: {}'.format(type(x)))
assert len(x.shape) == 4 and x.shape[
1] == 3, "Expect input to be (n, 3, h, w), given {}".format(
x.shape)
x = x.as_in_context(self.ctx[ctx_id])
pred = self.net(x)
topK = min(5, self.num_class)
ind = nd.topk(pred, k=topK)[0].astype('int').asnumpy().flatten()
probs = mx.nd.softmax(pred)[0].asnumpy().flatten()
df = pd.DataFrame([{
'class': self.classes[ind[i]],
'score': probs[ind[i]],
'id': ind[i]
} for i in range(topK)])
return df
def test_topk():
b = create_vector(size=LARGE_X)
ind = nd.topk(b, k=10, axis=0, dtype=np.int64)
assert np.sum(ind.asnumpy() == (LARGE_X - 1)) == 1
ind, val = mx.nd.topk(b,
k=3,
axis=0,
dtype=np.int64,
ret_typ="both",
is_ascend=False)
assert np.all(ind == val)
val = nd.topk(b, k=1, axis=0, dtype=np.int64, ret_typ="value")
assert val.sum() == (LARGE_X - 1)
def get_similar_tokens(query_token, k, embed):
W = embed.weight.data()
x = W(token_to_idx[query_token])
cos = nd.dot(W, x) / (nd.sum(W * W, axis=1) * nd.sum(x * x) + 1e-9).sqrt()
topk = nd.topk(cos, k=k + 1, ret_type='indices').asnumpy().astype('int32')
for i in topk[1:]:
print('cosine sim=%.3f: %s ' % (cos[i].asscalar(), (idx_to_token[i])))
def _predict(self, x):
resize = int(math.ceil(self.input_size / self._cfg.train.crop_ratio))
if isinstance(x, str):
x = transform_eval(mx.image.imread(x),
resize_short=resize,
crop_size=self.input_size)
elif isinstance(x, mx.nd.NDArray):
x = transform_eval(x,
resize_short=resize,
crop_size=self.input_size)
elif isinstance(x, pd.DataFrame):
assert 'image' in x.columns, "Expect column `image` for input images"
def _predict_merge(x):
y = self._predict(x)
y['image'] = x
return y
return pd.concat([_predict_merge(xx)
for xx in x['image']]).reset_index(drop=True)
else:
raise ValueError('Input is not supported: {}'.format(type(x)))
x = x.as_in_context(self.ctx[0])
pred = self.net(x)
topK = min(5, self.num_class)
ind = nd.topk(pred, k=topK)[0].astype('int').asnumpy().flatten()
probs = mx.nd.softmax(pred)[0].asnumpy().flatten()
df = pd.DataFrame([{
'class': self.classes[ind[i]],
'score': probs[ind[i]],
'id': ind[i]
} for i in range(topK)])
return df
def knn(W, x, k):
""" knn算法 """
# 添加的1e-9是为了数值稳定性
cos = nd.dot(W, x.reshape((-1, ))) / (
(nd.sum(W * W, axis=1) + 1e-9).sqrt() * nd.sum(x * x).sqrt())
topk = nd.topk(cos, k=k, ret_typ='indices').asnumpy().astype('int32')
return topk, [cos[i].asscalar() for i in topk]
def get_top_k_by_analogy(self, word1, word2, word3, k=1):
'''
Returns analogical word for the set of 3 words that are passed as arguments.
Analogy refers to:
king->queen ; man->woman
good->better ; bad->worse
do->did ; go->went
Eg.
emb= Embedder(dimensions=50)
print(emb.get_top_k_analogy('good','best','bad'))
Returns...
['worst']
Returns a list because you can have top k analogies as result
'''
word_vecs = self.__emb_mapper[word1, word2, word3]
word_diff = (word_vecs[1] - word_vecs[0] + word_vecs[2]).reshape(
(-1, 1))
vocab_vecs = self.__norm_vecs_by_row(self.__emb_mapper.idx_to_vec)
dot_product = nd.dot(vocab_vecs, word_diff)
indices = nd.topk(dot_product.reshape((len(self.__embedder), )),
k=k,
ret_typ='indices')
indices = [int(i.asscalar()) for i in indices]
return self.__embedder.to_tokens(indices)
def sample_neighbours(self, data, query_network):
num_stored_samples = self.key_memory.shape[0]
batch_size = data[0].shape[0]
query = query_network(*data).as_in_context(mx.cpu())
vec1 = nd.repeat(query, repeats=num_stored_samples, axis=0)
vec2 = nd.tile(self.key_memory, reps=(batch_size, 1))
diff = nd.subtract(vec1, vec2)
sq = nd.square(diff)
batch_sum = nd.sum(sq, exclude=1, axis=0)
sqrt = nd.sqrt(batch_sum)
dist = nd.reshape(sqrt, shape=(batch_size, num_stored_samples))
sample_ind = nd.topk(dist, k=self.k, axis=1, ret_typ="indices")
num_outputs = len(self.label_memory)
sample_labels = [
self.label_memory[i][sample_ind] for i in range(num_outputs)
]
sample_batches = [[
self.value_memory[j][sample_ind]
for j in range(len(self.value_memory))
], sample_labels]
return sample_batches
def beam_search_translate(encoder, decoder, input_seq, max_length, ctx,
beam_size, in_vocab, out_vocab):
in_tokens = input_seq.lower().split(' ')
in_tokens += [EOS] + [PAD] * (max_length - len(in_tokens) - 1)
enc_input = nd.array([in_vocab.to_indices(in_tokens)], ctx=ctx)
enc_state = encoder.begin_state(batch_size=1, ctx=ctx)
enc_output, enc_state = encoder(enc_input, enc_state)
dec_input = nd.array([out_vocab.token_to_idx[BOS]], ctx=ctx)
dec_state = decoder.begin_state(enc_state)
output_tokens = []
# the first character prediction
dec_output, dec_state = decoder(dec_input, dec_state, enc_output)
topk = nd.topk(dec_output, k=beam_size,
ret_typ='indices').asnumpy().astype('int32')
for idx in topk[0]:
score = nd.softmax(dec_output[0])[idx].asscalar()
sample_output = predict_rest(encoder, decoder, input_seq, max_length,
idx, dec_state, enc_output, score,
in_vocab, out_vocab, ctx)
output_tokens.append(sample_output)
for idx in range(len(output_tokens)):
output_tokens[idx][1] = math.log(output_tokens[idx][1]) / (len(
output_tokens[idx][0])**0.75)
return output_tokens
def get_similar_tokens(query_token, k, embed):
W = embed.weight.data()
x = W[token_to_idx[query_token]]
# Add 1e-9 for numerical stability
cos = nd.dot(W, x) / (nd.sum(W * W, axis=1) * nd.sum(x * x) + 1e-9).sqrt()
topk = nd.topk(cos, k=k + 1, ret_typ='indices').asnumpy().astype('int32')
for i in topk[1:]: # remove input word
print('cosine sim=%.3f: %s' % (cos[i].asscalar(), (idx_to_token[i])))
def predict(images, model_name, model_config):
""" Runs the model to infer on the given image
Arguments:
image {NDArray Image} -- image to do prediction on
model_name {str} -- name of the model to be used for prediction
Returns:
dict -- dictionary containing the predicted class
"""
outputs = {}
net = model_config.loaded_models[model_name]
# apply default data preprocessing
transformed_img = transform_eval(images)
if(model_config.models_config[model_name]== 'gpu'):
transformed_img = transformed_img.copyto(mx.gpu(0))
# run forward pass to obtain the predicted score for each class
pred = net(transformed_img)
# map predicted values to probability by softmax
prob = nd.softmax(pred)[0].asnumpy()
#prob = pred[0].asnumpy()
with open(os.path.join("/models",str(model_name),'config.json')) as config_file:
data = json.load(config_file)
max_number_of_predictions = data['max_number_of_predictions']
minimum_confidence = data['minimum_confidence']
if(max_number_of_predictions>len(net.classes)):
max_number_of_predictions=len(net.classes)
if(max_number_of_predictions<1):
max_number_of_predictions=1
# find the 5 class indices with the highest score
ind = nd.topk(pred, k=max_number_of_predictions)[0].astype('int').asnumpy().tolist()
if(minimum_confidence>float(prob[ind[0]])):
minimum_confidence=float(prob[ind[-1]])
for i in range(max_number_of_predictions):
if(float(prob[ind[i]])>minimum_confidence):
outputs[net.classes[ind[i]]] = float(prob[ind[i]])
outputs_descending = OrderedDict(sorted(outputs.items(),
key=lambda kv: kv[1], reverse=True))
return outputs_descending
def beam_search(outputs, ctx, targets, max_seq_len, beam_width):
predicts = nd.topk(nd.softmax(outputs, axis=-1),
axis=-1, k=beam_width, ret_typ='both')
if not targets:
targets = {}
beam_result_idxs = []
beam_result_score = []
count = 0
for score, idx in zip(predicts[0][0], predicts[1][0]):
idx = [2] + [int(idx.asscalar())]
beam_result_idxs.append(idx)
beam_result_score.append(score)
targets.update(
{"beam_{}".format(count): {"idx": idx, "score": score}})
count += 1
result = []
for idx in beam_result_idxs:
idx = idx[:max_seq_len] + \
[0] * (max_seq_len - len(idx))
result.append(idx)
return nd.array(result, ctx=ctx), targets
else:
beam_idxs = []
beam_score = []
for scores, idxs, target in zip(predicts[0], predicts[1], targets.values()):
last_score = target["score"]
last_idxs = target["idx"]
max_score = 0
max_score_idx = []
for score, idx in zip(scores, idxs):
if last_score + score > max_score:
max_score = last_score + score
idx = int(idx.asscalar())
max_score_idx = last_idxs[:] + [idx]
beam_idxs.append(max_score_idx)
beam_score.append(max_score)
beam_score, beam_idxs = (list(t)
for t in zip(*sorted(zip(beam_score, beam_idxs), reverse=True)))
targets = {}
count = 0
for idx, score in zip(beam_idxs, beam_score):
targets.update(
{"beam_{}".format(count): {"idx": idx, "score": score}})
count += 1
result = []
for idx in beam_idxs:
idx = idx[:max_seq_len] + \
[0] * (max_seq_len - len(idx))
result.append(idx)
return nd.array(result, ctx=ctx), targets
def pick_top_n(preds, top_n=5):
top = nd.topk(preds, axis=1, k=top_n, ret_typ='both')
top_pred_prob = top[0]
top_pred_label = top[1].asnumpy()
top_pred_prob /= nd.sum(top_pred_prob, axis=1, keepdims=True)
top_pred_prob = top_pred_prob.asnumpy().reshape((-1, ))
top_pred_label = top_pred_label.reshape((-1, ))
c = np.random.choice(top_pred_label, size=1, p=top_pred_prob)
return c
def get_knn(word, k=2000):
word_vec = vocab.embedding[word].reshape((-1, 1))
vocab_vecs = norm_vecs_by_row(vocab.embedding.idx_to_vec)
dot_prod = nd.dot(vocab_vecs, word_vec)
indices = nd.topk(dot_prod.reshape((len(vocab), )),
k=k + 1,
ret_typ='indices')
indices = [int(i.asscalar()) for i in indices]
# Remove unknown and input tokens.
return vocab.to_tokens(indices[1:])
def eval(img):
# Transform
img = transform_eval(img)
pred = net(img)
ind = nd.topk(pred, k=1)[0].astype('int')
#print('The input picture is classified to be')
label = ind[0].asscalar()
prob = nd.softmax(pred)[0][label].asscalar()
return label, prob
def draw_line(self, s, e):
dir = 1
lu_l, rd_l, dir = (s[0], e[0], dir) if s[0] <= e[0] else (e[0], s[0],
-dir)
lu_c, rd_c, dir = (s[1], e[1], dir) if s[1] <= e[1] else (e[1], s[1],
-dir)
h, w = rd_l - lu_l + 1, rd_c - lu_c + 1
mimic_mat = nd.zeros((h, w))
if h == 1 and w == 1:
return s, e
if 1.0 * h / w <= 1:
mimic_w = nd.tile(nd.arange(1, w), (h, 1))
if dir == 1:
mimic_h = nd.tile(nd.arange(
0, h), (w - 1, 1)).T # notice the direction & vertical
else:
mimic_h = nd.tile(nd.arange(h - 1, -1, -1), (w - 1, 1)).T
start = nd.array([[h - 1]]) if dir == -1 else nd.array([[0]])
out = nd.abs(mimic_h / mimic_w - 1.0 * h / w)
loc_h = nd.topk(-out, axis=0)
loc_h = nd.concat(start, loc_h, dim=1)
a = nd.arange(w)
mimic_mat[loc_h, a] = 1
res_h = (loc_h + lu_l).asnumpy().astype(np.uint8).tolist()[0]
res_w = list(range(lu_c, lu_c + w))
return res_h, res_w
else:
mimic_h = nd.tile(nd.arange(1, h), (w, 1)).T
if dir == 1:
mimic_w = nd.tile(nd.arange(0, w), (h - 1, 1))
else:
mimic_w = nd.tile(nd.arange(w - 1, -1, -1), (h - 1, 1))
start = nd.array([[w - 1]]) if dir == -1 else nd.array([[0]])
out = nd.abs(mimic_h / mimic_w - 1.0 * h / w)
loc_h = nd.topk(-out, axis=1).T
loc_h = nd.concat(start, loc_h, dim=1)
a = nd.arange(h)
mimic_mat[a, loc_h] = 1
res_h = list(range(lu_l, lu_l + h))
res_w = (loc_h + lu_c).asnumpy().astype(np.uint8).tolist()[0]
return res_h, res_w
def predict(self, query_token, k, embed):
word = embed.weight.data()
x = word[self.token_to_idx[query_token]]
cos = nd.dot(word, x) / (nd.sum(word * word, axis=1) * nd.sum(x * x) +
1e-9).sqrt()
topk = nd.topk(cos, k=k + 1,
ret_typ='indices').asnumpy().astype('int32')
for i in topk[1:]:
print('cosine sim=%.3f: %s' % (cos[i].asscalar(),
(self.idx_to_token[i])))
def _topk_channel(scores, K=40):
batch, cat, height, width = scores.shape
[topk_scores, topk_inds] = nd.topk(scores.reshape((batch, cat, -1)), ret_typ = "both", k= K)
#[topk_score, topk_ind] = nd.topk(nd.reshape(topk_scores, (batch, -1)), ret_typ='both', k=K)
topk_inds = topk_inds % (height * width)
topk_ys = (topk_inds / width).astype('int32').astype('float32')
topk_xs = (topk_inds % width).astype('int32').astype('float32')
return topk_scores, topk_inds, topk_ys, topk_xs
def curvature_based_sample(nn_pts, k):
curvature = compute_curvature(nn_pts)
point_indices = nd.topk(curvature, axis=-1, k=k, ret_typ='indices')
pts_shape = nn_pts.shape
batch_size = pts_shape[0]
batch_indices = nd.tile(nd.reshape(nd.arange(batch_size), (-1, 1, 1)),
(1, k, 1))
indices = nd.concat(batch_indices,
nd.expand_dims(point_indices, axis=2),
dim=2)
return indices
def knn(W, x, k):
"""
使用余弦相似度来搜索近义词
:param W:
:param x:
:param k:
:return:
"""
cos = nd.dot(W, x.reshape(
(-1, ))) / (nd.sum(W * W, axis=1).sqrt() * nd.sum(x * x).sqrt())
top_k = nd.topk(cos, k=k, ret_typ="indices").asnumpy().astype("int32")
return top_k, [cos[i].asscalar() for i in top_k]
def get_attribute(self, image):
"""Face attribute predictor.
Parameters
----------
image: NDArray.
The NDArray data format for MXNet to process, such as (H, W, C).
Returns
-------
type: tuple
Results of Face Attribute Predict:
(str(gender), int(age), str(expression)).
"""
img = transform_eval(image,
resize_short=self._image_size,
crop_size=self._image_size)
img = img.as_in_context(self.ctx[0])
tic = time.time()
pred = self.net(img)
toc = time.time() - tic
print('Attribute inference time: %fms' % (toc * 1000))
topK = 1
topK_age = 6
topK_exp = 2
age = 0
ind_1 = nd.topk(pred[0], k=topK)[0].astype('int')
ind_2 = nd.topk(pred[1], k=topK_age)[0].astype('int')
ind_3 = nd.topk(pred[2], k=topK_exp)[0].astype('int')
for i in range(topK_age):
age += int(
nd.softmax(pred[1])[0][ind_2[i]].asscalar() *
self.attribute_map2[1][ind_2[i].asscalar()])
gender = self.attribute_map2[0][ind_1[0].asscalar()]
if nd.softmax(pred[2])[0][ind_3[0]].asscalar() < 0.45:
expression = self.attribute_map2[2][7]
else:
expression_1 = self.attribute_map2[2][ind_3[0].asscalar()]
expression_2 = self.attribute_map2[2][ind_3[1].asscalar()]
return (gender, age, (expression_1, expression_2))
def explore_seeds(self, x, proportion=0.6):
proportion_self = get_self_handle()
b, _, h, w = x.shape
k = int(h * w * proportion)
top_k_values = nd.topk(x.detach().flatten(), k=k, ret_typ='value')
idx = [random.choice(range(int(k * (proportion_self - 0.2)), int(k * proportion_self)))
for i in range(b)]
idx = nd.array(idx).as_in_context(x.context)
# line = list(range(b))
threshold = top_k_values.pick(idx)
threshold = threshold.reshape((-1, 1, 1, 1))
seeds = (x > threshold) * 5.0
return seeds
def test_accur(target, it, *input):
LambdaMin = 5.0
LambdaMax = 1500.0
lamb = 1500.0
theta, phi = input
batch_size = target.size
lamb = max(LambdaMin, LambdaMax / (1 + 0.1 * it))
# because indexing is not differentiable in mxnet, we must do this
output = theta - theta / (1 + lamb) + phi / (1 + lamb)
nd.softmax(output, out=output)
v, idx = nd.topk(output, ret_typ='both')
real = (idx == target.reshape(-1, 1).astype(idx.dtype))
return nd.sum(real) / batch_size, nd.sum(real * v) / batch_size
def _gather_topk_beams(list,
score_or_log_probs,
batch_size,
beam_size,
cache=None):
"""Gather top beams from nested structure."""
score_or_log_probs = nd.array(score_or_log_probs, ctx=ctx)
topk_indexes = nd.topk(score_or_log_probs, k=beam_size)
return _gather_beams(list,
topk_indexes,
batch_size,
beam_size,
cache=cache)
def forward(self, x):
root = next(iter(self._structure.items()))[0]
router, router_mat, weight, embedd = self._contextify(x)(root)
presence = nd.sum(router_mat, axis=2)
weight_adj = presence * weight
depth = len(self._weightlayer) - nd.topk(nd.reverse(presence,
axis=1)) - 1
depth = depth[:, 0]
remainder = 1 - nd.sum(weight_adj, axis=1)
remainder += nd.choose_element_0index(weight_adj, depth)
weight_adj = nd.fill_element_0index(weight_adj, remainder, depth)
return (router, weight, weight_adj, router_mat)
def get_action(self, net):
if len(self.frames) < self.SAMPLE_DURATION:
return None
clip_input = self.frames
transform_fn = video.VideoGroupValTransform(size=224,
mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
clip_input = transform_fn(clip_input)
print(f"INFO: action input shape:")
print([clip.shape for clip in clip_input])
clip_input = np.stack(clip_input, axis=0)
clip_input = clip_input.reshape((-1, ) + (32, 3, 224, 224))
clip_input = np.transpose(clip_input, (0, 2, 1, 3, 4))
pred = net(nd.array(clip_input))
classes = net.classes
topK = 1
ind = nd.topk(pred, k=topK)[0].astype('int')
return classes[ind[0].asscalar()]
