def calculate_accuracy(x, y, w, b):
"""
Given x, y and parameters (w, b), calculate accuracy and cross entropy loss.
"""
pred_y = predict(x, w, b)
pred_labels = nd.argmax(pred_y, axis=1)
true_labels = nd.argmax(y, axis=1)
return nd.mean(pred_labels == true_labels).asscalar(), cross_entropy_loss(pred_y, y).asscalar()
def plot_prediction(batch_x, batch_y, i, w, b):
"""
Given a batch (of x, y) and an index i, plot image and predictions.
"""
xi = batch_x[i:(i+1)]
yi = batch_y[i:(i+1)]
img = xi.asnumpy().reshape((28, 28))
plt.imshow(img, cmap='gray')
pred_label = nd.argmax(predict(xi, w, b), axis=1).asscalar()
true_label = nd.argmax(yi, axis=1).asscalar()
plt.title("Prediction: {}, True: {}".format(pred_label, true_label))
plt.show()
pass
def test_model_save_load(gluon_model, model_data, model_path):
_, _, test_data = model_data
expected = nd.argmax(gluon_model(test_data), axis=1)
mlflow.gluon.save_model(gluon_model, model_path)
# Loading Gluon model
model_loaded = mlflow.gluon.load_model(model_path, ctx.cpu())
actual = nd.argmax(model_loaded(test_data), axis=1)
assert all(expected == actual)
# Loading pyfunc model
pyfunc_loaded = mlflow.pyfunc.load_model(model_path)
test_pyfunc_data = pd.DataFrame(test_data.asnumpy())
pyfunc_preds = pyfunc_loaded.predict(test_pyfunc_data)
assert all(np.argmax(pyfunc_preds.values, axis=1) == expected.asnumpy())
def yolo2_decoder(x, num_class, anchor_scales):
"""
yolo2_decoder 会把卷积的通道分开，转换，最后转成我们需要的检测框
out: (index,score,xmin,ymin,xmax,ymax)
"""
stride = num_class + 5
x = x.transpose((0, 2, 3, 1)) # (Batch,H,W,Stride*Anchor)
x = x.reshape((0, 0, 0, -1, stride)) # (Batch,H,W,Anchor,Stride)
xy_pred = x.slice_axis(begin=0, end=2, axis=-1)
wh = x.slice_axis(begin=2, end=4, axis=-1)
score_pred = x.slice_axis(begin=4, end=5, axis=-1)
cls_pred = x.slice_axis(begin=5, end=stride, axis=-1)
xy = nd.sigmoid(xy_pred)
x, y = transform_center(xy)
w, h = transform_size(wh, anchor_scales)
score = nd.sigmoid(score_pred)
cid = nd.argmax(cls_pred, axis=-1, keepdims=True)
left = nd.clip(x - w / 2, 0, 1)
top = nd.clip(y - h / 2, 0, 1)
right = nd.clip(x + w / 2, 0, 1)
bottom = nd.clip(y + h / 2, 0, 1)
output = nd.concat(*[cid, score, left, top, right, bottom], dim=4)
return output, cls_pred, score, nd.concat(*[xy, wh], dim=4)
def yolo2_forward(x, num_class, anchor_scales):
"""Transpose/reshape/organize convolution outputs."""
stride = num_class + 5
# transpose and reshape, 4th dim is the number of anchors
x = x.transpose((0, 2, 3, 1))
x = x.reshape((0, 0, 0, -1, stride))
# now x is (batch, m, n, stride), stride = num_class + 1(object score) + 4(coordinates)
# class probs
cls_pred = x.slice_axis(begin=0, end=num_class, axis=-1)
# object score
score_pred = x.slice_axis(begin=num_class, end=num_class + 1, axis=-1)
score = nd.sigmoid(score_pred)
# center prediction, in range(0, 1) for each grid
xy_pred = x.slice_axis(begin=num_class + 1, end=num_class + 3, axis=-1)
xy = nd.sigmoid(xy_pred)
# width/height prediction
wh = x.slice_axis(begin=num_class + 3, end=num_class + 5, axis=-1)
# convert x, y to positions relative to image
x, y = transform_center(xy)
# convert w, h to width/height relative to image
w, h = transform_size(wh, anchor_scales)
# cid is the argmax channel
cid = nd.argmax(cls_pred, axis=-1, keepdims=True)
# convert to corner format boxes
half_w = w / 2
half_h = h / 2
left = nd.clip(x - half_w, 0, 1)
top = nd.clip(y - half_h, 0, 1)
right = nd.clip(x + half_w, 0, 1)
bottom = nd.clip(y + half_h, 0, 1)
output = nd.concat(*[cid, score, left, top, right, bottom], dim=4)
return output, cls_pred, score, nd.concat(*[xy, wh], dim=4)
def update(self, labels, preds):
"""Updates the internal evaluation result.
Parameters
----------
labels : list of `NDArray`
The labels of the data with class indices as values, one per sample.
preds : list of `NDArray`
Prediction values for samples. Each prediction value can either be the class index,
or a vector of likelihoods for all classes.
"""
labels, preds = check_label_shapes(labels, preds, True)
for label, pred_label in zip(labels, preds):
if pred_label.shape != label.shape:
pred_label = ndarray.argmax(pred_label, axis=self.axis)
pred_label = pred_label.asnumpy().astype('int32')
label = label.asnumpy().astype('int32')
# flatten before checking shapes to avoid shape miss match
label = label.flat
pred_label = pred_label.flat
check_label_shapes(label, pred_label)
num_correct = (pred_label == label).sum()
self.sum_metric += num_correct
self.global_sum_metric += num_correct
self.num_inst += len(pred_label)
self.global_num_inst += len(pred_label)
def main():
parser = argparse.ArgumentParser()
parser.add_argument(
"dir",
help=
"Directory containing the image files (.png) we'll run inference on. \
Relative to the root of the project (tulip-fields/)")
args = parser.parse_args()
ctx = mx.gpu(0)
batch_size = 8
img_size = 256
root = os.path.dirname(__file__)
imgdir = os.path.join(root, os.pardir, args.dir)
checkpoint_dir = os.path.join(root, 'checkpoints', 'unet')
# Instantiate a U-Net and train it
net = unet.Unet()
net.load_params(os.path.join(checkpoint_dir, 'best_unet.params'), ctx)
print("Scanning dir {}".format(imgdir))
files = glob.glob(os.path.join(imgdir, '*wms*.png'))
print("Found {} images".format(len(files)))
nbatches = int(math.ceil(len(files) / batch_size))
reader = ImageReader(img_size, ctx)
for n in range(nbatches):
files_batch = files[n * batch_size:(n + 1) * batch_size]
batch = reader.load_batch(files_batch)
batch = batch.as_in_context(ctx)
preds = nd.argmax(net(batch), axis=1)
save_batch(files_batch, preds)
def viterbi_decode(self, feats):
backpointers = []
init_vvars = nd.full((1, self.tagset_size), -10000.)
init_vvars[0][self.tag_dictionary.get_idx_for_item(START_TAG)] = 0
forward_var = init_vvars
for feat in feats:
next_tag_var = forward_var.reshape((1, -1)).tile((self.tagset_size, 1)) + self.transitions.data()
bptrs_t = nd.argmax(next_tag_var, axis=1)
viterbivars_t = next_tag_var[list(range(len(bptrs_t))), bptrs_t]
forward_var = viterbivars_t + feat
backpointers.append(bptrs_t)
terminal_var = forward_var + self.transitions.data()[self.tag_dictionary.get_idx_for_item(STOP_TAG)]
terminal_var[self.tag_dictionary.get_idx_for_item(STOP_TAG)] = -10000.
terminal_var[self.tag_dictionary.get_idx_for_item(START_TAG)] = -10000.
best_tag_id = int(terminal_var.argmax(axis=0).asscalar())
path_score = terminal_var[best_tag_id]
best_path = [best_tag_id]
for bptrs_t in reversed(backpointers):
best_tag_id = bptrs_t[best_tag_id]
best_path.append(int(best_tag_id.asscalar()))
start = best_path.pop()
assert start == self.tag_dictionary.get_idx_for_item(START_TAG)
best_path.reverse()
return path_score, best_path
def update(self, labels, preds):
"""Updates the internal evaluation result.
Parameters
----------
labels : list of `NDArray`
The labels of the data with class indices as values, one per sample.
preds : list of `NDArray`
Prediction values for samples. Each prediction value can either be the class index,
or a vector of likelihoods for all classes.
"""
labels, preds = check_label_shapes(labels, preds, True)
for label, pred_label in zip(labels, preds):
if pred_label.shape != label.shape:
pred_label = ndarray.argmax(pred_label, axis=self.axis)
pred_label = pred_label.asnumpy().astype('int32')
label = label.asnumpy().astype('int32')
labels, preds = check_label_shapes(label, pred_label)
valid = (labels.reshape(-1, 1) != self.ignore_labels).all(axis=-1)
self.sum_metric += np.logical_and(pred_label.flat == label.flat,
valid).sum()
self.num_inst += np.sum(valid)
def evaluate_accuracy(data_iterator, num_examples, batch_size, params, net,
pool_type, pool_size, pool_stride, act_type, dilate_size,
nf):
numerator = 0.
denominator = 0.
for batch_i, (data, label) in enumerate(data_iterator):
data = data.as_in_context(ctx).reshape((batch_size, 1, 1, -1))
label = label.as_in_context(ctx)
label_one_hot = nd.one_hot(label, 10)
output, _ = net(data,
params,
pool_type=pool_type,
pool_size=pool_size,
pool_stride=pool_stride,
act_type=act_type,
dilate_size=dilate_size,
nf=nf)
predictions = nd.argmax(output, axis=1)
numerator += nd.sum(predictions == label)
denominator += data.shape[0]
print('Evaluating accuracy. (complete percent: %.2f/100' %
(1.0 * batch_i / (num_examples // batch_size) * 100) + ')',
end='')
sys.stdout.write("\r")
return (numerator / denominator).asscalar()
def update(self, labels, preds):
"""Updates the internal evaluation result.
Parameters
----------
labels : list of `NDArray`
The labels of the data.
preds : list of `NDArray`
Predicted values.
"""
check_label_shapes(labels, preds)
for label, pred_label in zip(labels, preds):
if pred_label.shape != label.shape:
pred_label = ndarray.argmax(pred_label, axis=self.axis)
pred_label = pred_label.asnumpy().astype('int32')
label = label.asnumpy().astype('int32')
check_label_shapes(label, pred_label)
pred_label_ = pred_label.flat
label_ = label.flat
for i in range(len(pred_label_)):
if label_[i] == self.c:
if pred_label_[i] == self.c:
self.TP += 1
else:
self.FN += 1
elif pred_label_[i] == self.c:
self.FP += 1
def select_action(self, state, exploration_rate):
if np.random.rand() < exploration_rate:
return self.environment.action_space.sample()
else:
return int(
nd.argmax(self.model(nd.expand_dims(state, axis=0)),
0).asnumpy()[0])
def test(model, test_loader, ctx):
"""
Test the model on test dataset.
"""
print("Testing...")
start_time = time.time()
model.load_params(model_file, ctx=ctx) # restore the best parameters
y_pred, y_true = [], []
for data, label in test_loader:
data, label = data.as_in_context(ctx), label.as_in_context(ctx)
with autograd.record(
train_mode=False): # set the training_mode to False
output = model(data)
pred = nd.argmax(output, axis=1).asnumpy().tolist()
y_pred.extend(pred)
y_true.extend(label.asnumpy().tolist())
test_acc = metrics.accuracy_score(y_true, y_pred)
test_f1 = metrics.f1_score(y_true, y_pred, average='macro')
print("Test accuracy: {0:>7.2%}, F1-Score: {1:>7.2%}".format(
test_acc, test_f1))
print("Precision, Recall and F1-Score...")
print(
metrics.classification_report(y_true,
y_pred,
target_names=['POS', 'NEG']))
print('Confusion Matrix...')
cm = metrics.confusion_matrix(y_true, y_pred)
print(cm)
print("Time usage:", get_time_dif(start_time))
def update(self, labels, preds):
"""Updates the internal evaluation result.
Parameters
----------
labels : list of `NDArray`
The labels of the data.
preds : list of `NDArray`
Predicted values.
"""
check_label_shapes(labels, preds)
for label, pred_label in zip(labels, preds):
if pred_label.shape != label.shape:
pred_label = ndarray.argmax(pred_label, axis=self.axis)
pred_label = pred_label.asnumpy().astype('int32')
label = label.asnumpy().astype('int32')
print 'pred: ',pred_label
print 'gt: ',label
check_label_shapes(label, pred_label)
keep_inds = np.where(label != 0)
pred_label = pred_label[keep_inds]
label = label[keep_inds]
self.sum_metric += (pred_label.flat == label.flat).sum()
self.num_inst += len(pred_label.flat)
def verify_loaded_model(net):
"""Run inference using ten random images.
Print both input and output of the model"""
def transform(data, label):
return data.astype(np.float32)/255, label.astype(np.float32)
# Load ten random images from the test dataset
sample_data = mx.gluon.data.DataLoader(mx.gluon.data.vision.MNIST(train=False, transform=transform),
10, shuffle=True)
for data, label in sample_data:
# Display the images
img = nd.transpose(data, (1,0,2,3))
img = nd.reshape(img, (28,10*28,1))
imtiles = nd.tile(img, (1,1,3))
plt.imshow(imtiles.asnumpy())
plt.show()
# Display the predictions
data = nd.transpose(data, (0, 3, 1, 2))
out = net(data.as_in_context(ctx))
predictions = nd.argmax(out, axis=1)
print('Model predictions: ', predictions.asnumpy())
break
def evaluate(self, eval_data):
acc = mx.metric.Accuracy()
for eval_x, eval_y in eval_data:
output = self.net(eval_x)
prediction = ndarray.argmax(output, axis=1)
acc.update(labels=eval_y, preds=prediction)
return acc.get()[1]
def train_block(net, train_data, eval_data=None):
net.initialize()
accuracy = mx.metric.Accuracy()
loss = gluon.loss.SoftmaxCrossEntropyLoss()
trainer = gluon.Trainer(net.collect_params(), 'sgd', optimizer_params={'learning_rate':0.1})
epoch = 10
for e in range(epoch):
for i, (train_x, train_y) in enumerate(train_data):
train_x = train_x.as_in_context(mx.cpu())#.reshape((-1, 784))
train_y = train_y.as_in_context(mx.cpu())
with autograd.record():
output = net(train_x)
one_loss = loss(output, train_y)
one_loss.backward()
prediction = ndarray.argmax(output, axis=1)
accuracy.update(labels=train_y, preds=prediction)
trainer.step(train_x.shape[0])
curr_loss = ndarray.mean(one_loss).asscalar()
if i%200 == 0:
print('epoch:{}, step:{}, loss:{:.4f}, accuracy:{:.4f}'.format(e, i, curr_loss, accuracy.get()[1]))
eval_accu = evaluate(net, eval_data)
print("epoch:{}, evaluate accuracy:{:.2f}".format(e, eval_accu))
def batch_intersection_union(output, target, nclass, ignore_bg=False):
"""mIoU"""
# inputs are NDarray, output 4D, target 3D
# ignore_bg=True, ignoring class 0; ignore_bg = False, use class 0
predict = F.argmax(output, 1)
target = target.astype(predict.dtype)
mini = 0
maxi = nclass - 1
nbins = nclass
predict = predict.asnumpy()
target = target.asnumpy()
if ignore_bg:
mini = 1
nbins -= 1
predict = predict * (target > 0).astype(predict.dtype)
else:
predict = predict * (target >= 0).astype(predict.dtype)
#intersection = predict * (F.equal(predict, target)).astype(predict.dtype)
intersection = predict * (predict == target)
# areas of intersection and union
area_inter, _ = np.histogram(intersection, bins=nbins, range=(mini, maxi))
area_pred, _ = np.histogram(predict, bins=nbins, range=(mini, maxi))
area_lab, _ = np.histogram(target, bins=nbins, range=(mini, maxi))
area_union = area_pred + area_lab - area_inter
return area_inter, area_union
def accuracy(output, label):
"""
output : [1,2,3,4,5], label : 2
"""
labelHat = nd.argmax(output, axis=1)
return nd.sum(labelHat == label).asscalar()
def batch_pix_accuracy(output, target):
"""PixAcc"""
# inputs are NDarray, output 4D, target 3D
predict = F.argmax(output, 1) + 1
target = target.astype(predict.dtype) + 1
pixel_labeled = (target > 0).sum().asscalar()
pixel_correct = (F.equal(predict, target)*(target > 0)).sum().asscalar()
return pixel_correct, pixel_labeled
def predict_one_image(self,X):
prdY = self.net(X)
prdCls, prdObj, prdXYXY = self.cvt_output_for_predict(prdY)
cid = nd.argmax(prdCls, axis=-1, keepdims=True)
output = nd.concat(cid, prdObj, prdXYXY,dim=-1)
output = output.reshape((0,-1,6)) #cid, objectness x0,y0,x1,y1
output = nd.contrib.box_nms(output) #cid may be changed
return output
def predict2ndimg(predict):
result = ndarray.argmax(predict, axis=1)
colormap = ndarray.array(MICCAI_colormap, ctx=mx.gpu(),
dtype='uint8') # voc_colormap
ndimg = colormap[result[:, :, :]]
ndimg = ndarray.transpose(ndimg, (0, 3, 1, 2))
ndimg = ndimg.astype(('float32'))
return ndimg
def batch_pix_accuracy(output, target):
"""PixAcc"""
# inputs are NDarray, output 4D, target 3D
predict = F.argmax(output, 1)
predict = predict.asnumpy() + 1
target = target.asnumpy().astype(predict.dtype) + 1
pixel_labeled = np.sum(target > 0)
pixel_correct = np.sum((predict == target) * (target > 0))
assert pixel_correct <= pixel_labeled, "Correct area should be smaller than Labeled"
return pixel_correct, pixel_labeled
def process(self, element):
"""
Returns clear images after filtering the cloudy ones
:param element:
:return:
"""
batch = self.reader.load_batch(element)
batch = batch.as_in_context(self.ctx)
preds = nd.argmax(self.net(batch), axis=1)
self.save_batch(element, preds)
def test_model_log_load(gluon_model, model_data, model_path):
_, _, test_data = model_data
expected = nd.argmax(gluon_model(test_data), axis=1)
artifact_path = "model"
with kiwi.start_run():
kiwi.gluon.log_model(gluon_model, artifact_path=artifact_path)
model_uri = "runs:/{run_id}/{artifact_path}".format(
run_id=kiwi.active_run().info.run_id, artifact_path=artifact_path)
# Loading Gluon model
model_loaded = kiwi.gluon.load_model(model_uri, ctx.cpu())
actual = nd.argmax(model_loaded(test_data), axis=1)
assert all(expected == actual)
# Loading pyfunc model
pyfunc_loaded = kiwi.pyfunc.load_model(model_uri)
test_pyfunc_data = pd.DataFrame(test_data.asnumpy())
pyfunc_preds = pyfunc_loaded.predict(test_pyfunc_data)
assert all(np.argmax(pyfunc_preds.values, axis=1) == expected.asnumpy())
def character_recognize(pic, text_area):
img = cv2.imread(pic)
mser = cv2.MSER_create()
import mxnet.ndarray as nd
import mxnet as mx
with open('chinese.txt') as to_read:
chinese = [m_line.strip() for m_line in to_read]
net = get_net(len(chinese))
ctx = mx.gpu()
net.load_params(
'./single_character_recognition/train_model/chinese_2.para', ctx)
for m_text_area in text_area:
point1_x, point1_y, point2_x, point2_y = m_text_area[0], m_text_area[
1], m_text_area[2], m_text_area[3]
# roi of text area
pic_text_area = img[int(point1_y):int(point2_y),
int(point1_x):int(point2_x), :]
gray = cv2.cvtColor(pic_text_area, cv2.COLOR_RGB2GRAY)
binaray = cv2.adaptiveThreshold(gray, 255, cv2.ADAPTIVE_THRESH_MEAN_C,
cv2.THRESH_BINARY_INV, 5, 5)
regions, boxes = mser.detectRegions(gray)
to_predict_pics = []
boxes = sorted(boxes, key=lambda x: x[0])
if len(boxes) == 0:
continue
# remove the similar boxes from redundant boxes
refined_boxes = [boxes[0]]
superposition_ratio = .5
for i in range(1, len(boxes)):
last_x, last_width = boxes[i - 1][0], boxes[i - 1][2]
cur_x, cur_width = boxes[i][0], boxes[i][2]
total_len = cur_x + cur_width - last_x
share_len = last_width - cur_x + last_x
if share_len / total_len < superposition_ratio:
refined_boxes.append(boxes[i])
for box in refined_boxes:
x, y, w, h = box
if w / pic_text_area.shape[1] <= 0.9 and w / h < 1.5:
char_pic = cv2.resize(binaray[y:y + h, x:x + w], (28, 28))
to_predict_pics.append(char_pic / 255)
if len(to_predict_pics) == 0:
continue
output = net(
nd.array(to_predict_pics).reshape(
(-1, 1, 28, 28)).as_in_context(ctx))
predictions = nd.argmax(output, axis=1).asnumpy()
predictions_char = [
chinese[int(m_prediction)] for m_prediction in predictions
]
print(' '.join(predictions_char))
cv2.imshow('to_recognize', pic_text_area)
cv2.waitKey(0)
def verifyLoadedModel(net, data):
data = nd.transpose(nd.array(data), (0, 3, 1, 2))
out = net(data)
predictions = nd.argmax(out, axis=1)
text_labels = [
't-shirt', 'trouser', 'pullover', 'dress', 'coat', 'sandal', 'shirt',
'sneaker', 'bag', 'ankle boot'
]
return [(int(p), text_labels[int(p)]) for p in predictions.asnumpy()]
def batch_pix_accuracy(output, target):
"""PixAcc"""
# inputs are NDarray, output 4D, target 3D
# the category -1 is ignored class, typically for background / boundary
predict = F.argmax(output, 1)
predict = predict.asnumpy() + 1
target = target.asnumpy().astype(predict.dtype) + 1
pixel_labeled = np.sum(target > 0)
pixel_correct = np.sum((predict == target)*(target > 0))
assert pixel_correct <= pixel_labeled, "Correct area should be smaller than Labeled"
return pixel_correct, pixel_labeled
def accuracy(data):
good = 0
total = 0
for (X, Y) in data:
features = X.as_in_context(model_ctx)
label = Y.as_in_context(model_ctx).reshape(Y.size, -1)
prediction = nd.argmax(net(features),
axis=1).reshape(Y.size, -1)
good += nd.sum(prediction == label).asscalar()
total += len(X)
return good / total
def check_rgKL(self, v):
ndx = nd.argmax(v.reshape([-1, self.n_val]), axis=1)
pv = np.ones(shape=self.prob_RGs.shape)
for intcoor in ndx.asnumpy().reshape([-1, self.n_vis]):
rg2 = cal_rg2(int2xy(intcoor))
i = int(rg2 * 2)
if i >= len(pv): i = len(pv) - 1
pv[i] += 1
pv /= np.sum(pv)
prg = self.prob_RGs.asnumpy()
KL = np.sum(prg * np.log(prg / pv))
return KL
def update(self, labels, preds):
"""Updates the internal evaluation result.
Parameters
----------
labels : list of `NDArray`
The labels of the data with class indices as values, one per sample.
preds : list of `NDArray`
Prediction values for samples. Each prediction value can either be the class index,
or a vector of likelihoods for all classes.
"""
labels, preds = check_label_shapes(labels, preds, True)
for label, pred_label in zip(labels, preds):
if pred_label.shape != label.shape:
pred_label = ndarray.argmax(pred_label, axis=self.axis)
pred_label = pred_label.asnumpy().astype('int32')
label = label.asnumpy().astype('int32')
labels, preds = check_label_shapes(label, pred_label)
valid = (labels.reshape(-1, 1) != self.ignore_labels).all(axis=-1)
self.sum_metric += np.logical_and(pred_label.flat == label.flat, valid).sum()
self.num_inst += np.sum(valid)
for i in range(total_batch):
num_valid = batch_size if (i + 1) * batch_size <= X_test.shape[0]\
else X_test.shape[0] - i * batch_size
data_npy = np.take(X_test,
indices=np.arange(i * batch_size, (i + 1) * batch_size),
axis=0,
mode="clip")
label_npy = np.take(y_test,
indices=np.arange(i * batch_size, (i + 1) * batch_size),
axis=0,
mode="clip")
test_net.forward(data_batch=mx.io.DataBatch(data=[nd.array(data_npy)],
label=None),
is_train=False)
logits_nd = test_net.get_outputs()[0]
pred_cls = nd.argmax(logits_nd, axis=-1).asnumpy()
correct_count += (pred_cls[:num_valid] == label_npy[:num_valid]).sum()
acc = correct_count / float(X_test.shape[0])
print('Accuracy:', acc)
# 6. Get one and predict
test_net.reshape(data_shapes=[mx.io.DataDesc(name='data', shape=(1, 1, 28, 28), layout='NCHW')],
label_shapes=None)
r = np.random.randint(0, X_test.shape[0])
test_net.forward(data_batch=mx.io.DataBatch(data=[nd.array(X_test[r:r + 1])],
label=None))
logits_nd = test_net.get_outputs()[0]
print("Label: ", int(y_test[r]))
print("Prediction: ", int(nd.argmax(logits_nd, axis=1).asnumpy()[0]))
'''
Epoch: 0001 cost = 0.222577997
def argmax(vec):
# return the argmax as a python int
idx = nd.argmax(vec, axis=1)
return to_scalar(idx)
correct_counts = [0 for i in range(num_models)]
ensemble_correct_count = 0
total_num = 0
for i in range(total_batch):
num_valid = batch_size if (i + 1) * batch_size <= X_test.shape[0]\
else X_test.shape[0] - i * batch_size
data_npy, label_npy, num_valid = get_batch(i, batch_size, X_test, y_test)
prob_ensemble = nd.zeros(shape=(label_npy.shape[0], 10), ctx=mx.gpu())
for i, test_net in enumerate(test_nets):
test_net.forward(data_batch=mx.io.DataBatch(data=[nd.array(data_npy)],
label=None),
is_train=False)
logits_nd = test_net.get_outputs()[0]
prob_nd = nd.softmax(logits_nd)
prob_ensemble += prob_nd
pred_cls = nd.argmax(prob_nd, axis=-1).asnumpy()
correct_counts[i] += (pred_cls[:num_valid] == label_npy[:num_valid]).sum()
prob_ensemble /= num_models
ensemble_pred_cls = nd.argmax(prob_ensemble, axis=-1).asnumpy()
ensemble_correct_count += (ensemble_pred_cls[:num_valid] == label_npy[:num_valid]).sum()
for i in range(num_models):
print(i, 'Accuracy:', correct_counts[i] / float(X_test.shape[0]))
print('Ensemble accuracy:', ensemble_correct_count / float(X_test.shape[0]))
'''
Learning Started!
Epoch: 0001 cost = [ 0.23813407 0.23717315]
Epoch: 0002 cost = [ 0.07455271 0.07434764]
Epoch: 0003 cost = [ 0.05925059 0.06024327]
Epoch: 0004 cost = [ 0.05032205 0.04895757]
Epoch: 0005 cost = [ 0.04573197 0.0439943 ]
Epoch: 0006 cost = [ 0.04143022 0.0416003 ]
