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 plot_img(losses_log):
sw.add_image(tag='A', image=nd.clip(nd.concatenate([losses_log['real_A'][0][0:1],
losses_log['fake_B'][0][0:1],
losses_log['rec_A'][0][0:1],
losses_log['idt_A'][0][0:1]]) * 0.5 + 0.5, 0, 1))
sw.add_image(tag='B', image=nd.clip(nd.concatenate([losses_log['real_B'][0][0:1],
losses_log['fake_A'][0][0:1],
losses_log['rec_B'][0][0:1],
losses_log['idt_B'][0][0:1]]) * 0.5 + 0.5, 0, 1))
def plot_img(losses_log):
sw.add_image(tag='lr_img',
image=nd.clip(
nd.concatenate(losses_log['lr_img'])[0:4], 0, 1))
sw.add_image(tag='hr_img',
image=nd.clip(
nd.concatenate(losses_log['hr_img'])[0:4], 0, 1))
sw.add_image(tag='hr_img_fake',
image=nd.clip(
nd.concatenate(losses_log['hr_img_fake'])[0:4], 0, 1))
def get_rmse_log(net, X_train, y_train):
"""Gets root mse between the logarithms of the prediction and the truth."""
num_train = X_train.shape[0]
clipped_preds = nd.clip(net(X_train), 1, float('inf'))
return np.sqrt(2 * nd.sum(
square_loss(nd.log(clipped_preds), nd.log(y_train))).asscalar() /
num_train)
def update(self, index, weight, grad, state):
assert (isinstance(weight, NDArray))
assert (isinstance(grad, NDArray))
self._update_count(index)
lr = self._get_lr(index)
wd = self._get_wd(index)
is_sparse = grad.stype == 'row_sparse'
history = state
if is_sparse:
kwargs = {
'epsilon': self.float_stable_eps,
'rescale_grad': self.rescale_grad
}
if self.clip_gradient:
kwargs['clip_gradient'] = self.clip_gradient
sparse.adaalter_update(weight,
grad,
history,
out=weight,
lr=lr,
wd=wd,
**kwargs)
# raise NotImplementedError('AdaAlter has not been implemented for sparse nd')
else:
grad = grad * self.rescale_grad
if self.clip_gradient is not None:
grad = clip(grad, -self.clip_gradient, self.clip_gradient)
div = grad / sqrt(history + self.float_stable_eps)
weight[:] += (div + weight * wd) * -lr
history[:] += square(grad)
def get_srme_log(net, X_train, Y_train):
num_train = X_train.shape[0]
clipped_preds = nd.clip(net(X_train), 1,
float('inf')) #对net(X)结果进行截断[1,正无穷]
return np.sqrt(2 * nd.sum(
square_loss(nd.log(clipped_preds), nd.log(Y_train))).asscalar() /
num_train)
def update(self, index, weight, grad, state):
assert(isinstance(weight, NDArray))
assert(isinstance(grad, NDArray))
self._update_count(index)
lr = self._get_lr(index)
wd = self._get_wd(index)
t = self._index_update_count[index]
# preprocess grad
#grad = grad * self.rescale_grad + wd * weight
grad *= self.rescale_grad + wd * weight
if self.clip_gradient is not None:
grad = clip(grad, -self.clip_gradient, self.clip_gradient)
# warming momentum schedule
momentum_t = self.beta1 * (1. - 0.5 * (pow(0.96, t * self.schedule_decay)))
momentum_t_1 = self.beta1 * (1. - 0.5 * (pow(0.96, (t + 1) * self.schedule_decay)))
self.m_schedule = self.m_schedule * momentum_t
m_schedule_next = self.m_schedule * momentum_t_1
# update m_t and v_t
m_t, v_t = state
m_t[:] = self.beta1 * m_t + (1. - self.beta1) * grad
v_t[:] = self.beta2 * v_t + (1. - self.beta2) * grad * grad
grad_prime = grad / (1. - self.m_schedule)
m_t_prime = m_t / (1. - m_schedule_next)
v_t_prime = v_t / (1. - pow(self.beta2, t))
m_t_bar = (1. - momentum_t) * grad_prime + momentum_t_1 * m_t_prime
# update weight
weight[:] -= lr * m_t_bar / (sqrt(v_t_prime) + self.epsilon)
def get_rmse_log(net, X_train, y_train):
num_train = X_train.shape[0]
clipped_preds = nd.clip(net(X_train), 1, float('inf'))
return np.sqrt(2 * nd.sum(
square_loss(nd.log(clipped_preds), nd.log(y_train))).asscalar() /
num_train)
def update(self, index, weight, grad, state):
"""Update the parameters.
Parameters
----------
index : int
An unique integer key used to index the parameters
weight : NDArray
weight ndarray
grad : NDArray
grad ndarray
state : NDArray or other objects returned by init_state
The auxiliary state used in optimization.
"""
assert (isinstance(weight, NDArray))
assert (isinstance(grad, NDArray))
(lr, momentum) = self._get_lr(index)
wd = self._get_wd(index)
self._update_count(index)
grad = grad * self.rescale_grad
if self.clip_gradient is not None:
grad = clip(grad, -self.clip_gradient, self.clip_gradient)
if state:
mom = state
mom[:] *= momentum
mom[:] += -lr * (1.0 - momentum) * (grad + wd * weight)
weight[:] += mom
else:
assert self.momentum == 0.0
weight[:] += -lr * (grad + self.wd * weight)
def update(self, index, weight, grad, state):
"""Update the parameters.
Parameters
----------
index : int
An unique integer key used to index the parameters
weight : NDArray
weight ndarray
grad : NDArray
grad ndarray
state : NDArray or other objects returned by init_state
The auxiliary state used in optimization.
"""
assert(isinstance(weight, NDArray))
assert(isinstance(grad, NDArray))
(lr, momentum) = self._get_lr(index)
wd = self._get_wd(index)
self._update_count(index)
grad = grad * self.rescale_grad
if self.clip_gradient is not None:
grad = clip(grad, -self.clip_gradient, self.clip_gradient)
if state:
mom = state
mom[:] *= momentum
mom[:] += -lr * (1.0 - momentum) * (grad + wd * weight)
weight[:] += mom
else:
assert self.momentum == 0.0
weight[:] += -lr * (grad + self.wd * weight)
def _realize_layer(sym, params, graph, inputs_ext, runtime):
logger = logging.getLogger('log.realize.layer')
name = sym.attr('name')
childs, attr = sym_iter(sym.get_children()), sym.list_attr()
if not _is_simulate_op(sym):
return sym, params
X, scale = childs[0], eval(attr['scale'])
in_prec, out_prec = eval(attr['in_prec']), eval(attr['out_prec'])
frac, sb = sim.extract_float(scale)
assert runtime in ['cvm', 'tvm']
_realize_func = _realize_cvm if runtime == 'cvm' else _realize_tvm
def cal_bit(A_bit, B_bit, sb):
max_bit = 32
total_bit = A_bit + B_bit
excess_bit = (total_bit - max_bit) // 2 if total_bit > max_bit else 0
A_target_bit = A_bit - excess_bit
B_target_bit = min(B_bit - excess_bit, 32 - A_target_bit)
A_sb, B_sb = A_bit - A_target_bit, B_bit - B_target_bit
Y_sb = (-sb) - A_sb - B_sb
return A_sb, A_target_bit, B_sb, B_target_bit, Y_sb
if scale == 1:
node = _realize_func(X, 0, out_prec, params, graph)
logger.debug("layer %-40s skip prec=%s", name, out_prec)
elif frac == 1:
node = _realize_func(X, -sb, out_prec, params, graph)
logger.debug("layer %-40s X(%s >> %s) prec=%s", name, in_prec, -sb,
out_prec)
else:
B_bit = math.ceil(math.log2(frac)) + 1
A_sb, A_tb, B_sb, B_tb, Y_sb = cal_bit(in_prec, B_bit, sb)
X = _realize_func(X, A_sb, A_tb, params, graph)
B_name = name + '_scale'
params[B_name] = nd.array([round(frac / (2**B_sb))])
B_range = 2**(B_tb - 1) - 1
params[B_name] = nd.clip(params[B_name], a_min=-B_range, a_max=B_range)
attr = {'precision': str(B_tb)}
B = graph[B_name] = mx.sym.var(B_name, shape=(1, ), attr=attr)
node = mx.sym.broadcast_mul(X, B)
node = _realize_func(node, Y_sb, out_prec, params, graph)
logger.debug(
"layer %-40s Y(INT%s >> %s) X(%s >> %s) B(%s vs. %s %s >> %s)",
name, out_prec, Y_sb, in_prec, A_sb, scale, frac, sb, B_sb)
# if childs[0].attr('name') in [
# 'yolov30_yolooutputv30_tile0',
# 'yolov30_yolooutputv31_tile0',
# 'yolov30_yolooutputv32_tile0',
# 'yolov30_yolooutputv30_expand_dims0',
# 'yolov30_yolooutputv31_expand_dims0',
# 'yolov30_yolooutputv32_expand_dims0',
# ]:
# sb = out_prec - 16
# node = _realize_func(node, sb, 16, params, graph)
return node, params
def get_a(self, item):
if isinstance(self.a, gluon.nn.Embedding):
a = squeeze(self.a(as_array(item)))
if self.a_range is not None:
return nd.clip(a, *self.a_range)
return a
else:
return self.a
def Prediction(test_data, network, ctx):
#Next Lotto Number!!!
for data, label in test_data:
data = data.as_in_context(ctx)
output = mx.nd.round(network(data))
output = nd.clip(data=output, a_min=0, a_max=45)
print(output.asnumpy()[0])
def get_c(self, item):
if isinstance(self.c, gluon.nn.Embedding):
c = squeeze(self.c(as_array(item)))
if self.c_range is not None:
return nd.clip(c, *self.c_range)
return c
else:
return self.c
def contrast_aug(self, src, x):
alpha = 1.0 + random.uniform(-x, x)
coef = np.array([[[0.299, 0.587, 0.114]]])
gray = src * coef
gray = (3.0 * (1.0 - alpha) / gray.size) * nd.sum(gray)
src *= alpha
src += gray
src = nd.clip(src, 0, 255)
return src
def save_image(data, file, normalize=True, img_range=None):
if img_range is None:
img_range = [min(data), max(data)]
norm_img = normalize_image(data, img_range[0], img_range[1])
img = nd.clip(norm_img * 255 + 0.5, 0, 255).asnumpy().astype(np.uint8)
img = Image.fromarray(np.transpose(img, (1, 2, 0)))
img.save(file)
def batched_l2_dist(a, b):
a_squared = nd.power(nd.norm(a, axis=-1), 2)
b_squared = nd.power(nd.norm(b, axis=-1), 2)
squared_res = nd.add(nd.linalg_gemm(
a, nd.transpose(b, axes=(0, 2, 1)), nd.broadcast_axes(nd.expand_dims(b_squared, axis=-2), axis=1, size=a.shape[1]), alpha=-2
), nd.expand_dims(a_squared, axis=-1))
res = nd.sqrt(nd.clip(squared_res, 1e-30, np.finfo(np.float32).max))
return res
def backward(self, out_grad, in_data, out_data, in_grad):
x = out_data[0]
action = in_data[1]
reward = in_data[2]
dx = in_grad[0]
dx[:] = 0
dx[:] = nd.fill_element_0index(dx,
nd.clip(nd.choose_element_0index(x, action) - reward, -1, 1),
action)
def W_decay(self, lr):
#go through weights
if self.L_coeff > 0: #L2
self.W -= self.L_coeff * lr * self.W
elif self.L_coeff < 0: #L1
self.W += self.L_coeff * lr * nd.sign(self.W)
else:
#fix boundary
self.W = nd.clip(self.W, -10.0, 10.0)
return
def hybrid_forward(self, F, x, weight):
x_norm = F.L2Normalization(x, mode='instance', name='x_n')
w_norm = F.L2Normalization(weight, mode='instance', name='w_n')
cos_theta = F.FullyConnected(x_norm,
w_norm,
no_bias=True,
num_hidden=self._units,
name='cos_theta')
cos_theta = F.clip(cos_theta, a_min=-1, a_max=1)
return cos_theta
def unsorted_1d_segment_mean(input, seg_id, n_segs, dim):
# TODO: support other dimensions
assert dim == 0, 'MXNet only supports segment mean on first dimension'
n_ones = nd.ones_like(seg_id).astype(input.dtype)
w = unsorted_1d_segment_sum(n_ones, seg_id, n_segs, 0)
w = nd.clip(w, a_min=1, a_max=np.inf)
y = unsorted_1d_segment_sum(input, seg_id, n_segs, dim)
y = y / w.reshape((-1, ) + (1, ) * (y.ndim - 1))
return y
def lab_to_rgb(lab, ctx=None):
if ctx is None:
raise ValueError("ctx can not be None")
if lab is None:
raise ValueError("lab can not be None")
with mx.Context(ctx):
lab = __check_image(lab)
lab_pixels = lab.reshape([-1, 3])
lab_to_fxfyfz = nd.array([
# fx fy fz
[1 / 116.0, 1 / 116.0, 1 / 116.0], # l
[1 / 500.0, 0.0, 0.0], # a
[0.0, 0.0, -1 / 200.0], # b
], ctx=ctx)
fxfyfz_pixels = nd.dot(lab_pixels + nd.array([16.0, 0.0, 0.0], ctx=ctx), lab_to_fxfyfz)
# convert to xyz
epsilon = 6 / 29
linear_mask = fxfyfz_pixels <= epsilon
exponential_mask = fxfyfz_pixels > epsilon
xyz_pixels = (3 * epsilon ** 2 * (fxfyfz_pixels - 4 / 29)) * linear_mask + (fxfyfz_pixels ** 3) * exponential_mask
xyz_pixels = nd.multiply(xyz_pixels, nd.array([0.950456, 1.0, 1.088754]))
xyz_to_rgb =nd.array([
# r g b
[3.2404542, -0.9692660, 0.0556434], # x
[-1.5371385, 1.8760108, -0.2040259], # y
[-0.4985314, 0.0415560, 1.0572252], # z
])
rgb_pixels = nd.dot(xyz_pixels, xyz_to_rgb)
nd.clip(rgb_pixels, 0.0, 1.0, out=rgb_pixels)
linear_mask = rgb_pixels <= 0.0031308
exponential_mask = rgb_pixels > 0.0031308
step1 = nd.multiply(nd.multiply(rgb_pixels, 12.92), linear_mask)
step2 = nd.multiply(nd.multiply(nd.power(rgb_pixels, (1 / 2.4)), 1.055) - 0.055, exponential_mask)
srgb_pixels = step1 + step2
return srgb_pixels.reshape(lab.shape)
def run_demo(eval_args):
## load parameters
#content_image = 'images/content/xjtlu.jpg'
#style_image = 'images/styles/starry_night.jpg'
#eval_args = program_args(content_image,content_image,style_image,128,128,0)
#eval_args = camero_args(style_image)
if eval_args.cuda == 0:
ctx = mx.cpu()
else:
ctx = mx.gpu()
## Change the content and style image using Style Loader
#content_image = utils.tensor_load_rgbimage(eval_args.contentImage, ctx, size=eval_args.size, keep_asp=True)
style_image = utils.tensor_load_rgbimage(eval_args.styleImage,
ctx,
size=eval_args.size)
style_image = utils.preprocess_batch(style_image)
style_model = net.Net(ngf=eval_args.ngf)
style_model.load_parameters(eval_args.model, ctx=ctx)
style_model.set_target(style_image)
cam = cv2.VideoCapture(0)
while True:
## read frame
ret, frame = cam.read()
# read content image (cimg)
#cimg = img.copy()
#img = np.array(img).transpose(2, 0, 1)
content_img = load_image(frame, ctx, eval_args.size)
output = style_model(content_img)
tensor = output[0]
#(b, g, r) = F.split(tensor, num_outputs=3, axis=0)
#tensor = F.concat(r, g, b, dim=0)
img = F.clip(tensor, 0, 255).asnumpy()
img = img.transpose(1, 2, 0).astype('uint8')
img = Image.fromarray(img)
image = np.array(
img.resize((frame.shape[1], frame.shape[0]), Image.ANTIALIAS))
#print(frame.shape,image.shape)
numpy_horizontal = np.hstack((frame, image))
#cv2.imshow("Content Window",frame)
#cv2.imshow("Style Window",grey)
cv2.imshow("Test Window Shape", numpy_horizontal)
if cv2.waitKey(1) & 0xFF == ord('q'):
break
cam.release()
cv2.destroyAllWindows()
def generate(self,
v_q: nd.NDArray,
x_context: nd.NDArray,
v_context: nd.NDArray,
include_intermediate: bool = False,
**kwargs) -> Union[nd.NDArray, Tuple[nd.NDArray, nd.NDArray]]:
"""
Generate a batch of samples from model. See Algorithm S3 in paper.
:param v_q: Query view camera info.
:param x_context: Context frames.
:param v_context: Context camera info.
:param include_intermediate: If True, samples from all timesteps (not only the last timestep) are returned.
:return: n x *image_shape array of generated samples. If include_intermediate is True,
then steps x n x *image_shape.
"""
u = nd.zeros((self._batch_size, *self._upsample_output_shape),
ctx=self._ctx) # canvas (reconstruction)
h_dec = nd.zeros((self._batch_size, *self._rnn_hidden_shape),
ctx=self._ctx)
c_dec = nd.zeros((self._batch_size, *self._rnn_hidden_shape),
ctx=self._ctx)
# reshape camera information so we can concat it to image data
v_q = nd.broadcast_to(
nd.expand_dims(nd.expand_dims(v_q, axis=-1), axis=-1),
(0, 0, *self._downsample_output_shape[1:]))
outs = [] # sample(s) over time
r = self._representation_nn(x_context, v_context)
for i in range(self._num_steps):
outs.append(self._out_layer(u))
# Eq. S11
p = self._p_layer(h_dec)
p = nd.reshape(p, (self._batch_size, -1))
z = self._latent_layer(p)
gen_z = nd.reshape(z, (self._batch_size, self._num_latent_maps,
*self._downsample_output_shape[1:]))
_, (h_dec, c_dec) = self._gen_rnn(nd.concat(gen_z, v_q, r, dim=1),
[h_dec, c_dec])
u = u + self._upsample_nn(h_dec)
outs.append(self._out_layer(u))
if include_intermediate:
samples = nd.stack(*outs, axis=0)
else:
samples = outs[-1]
return nd.clip(samples, a_min=0.0, a_max=1.0)
def update(self, index, weight, grad, state):
assert (isinstance(weight, NDArray))
assert (isinstance(grad, NDArray))
self._update_count(index)
lr = self._get_lr(index)
wd = self._get_wd(index)
t = self._index_update_count[index]
with bulk(self._bulk):
# preprocess grad
grad *= self.rescale_grad
if self.clip_gradient is not None:
grad = clip(grad, -self.clip_gradient, self.clip_gradient)
mean, var = state
mean *= self.beta1
mean += (1. - self.beta1) * grad
var *= self.beta2
var += (1. - self.beta2) * square(grad)
r1 = weight.norm()
if not self.bias_correction:
r1 = minimum(maximum(r1, self.lower_bound), self.upper_bound)
sqrt_var = sqrt(var)
sqrt_var += self.epsilon
g = mean / sqrt_var
g += wd * weight
else:
# apply bias correction
mean_hat = mean / (1. - power(self.beta1, t))
var_hat = var / (1. - power(self.beta2, t))
if self._eps_after_sqrt:
sqrt(var_hat, out=var_hat)
var_hat += self.epsilon
else:
var_hat += self.epsilon
sqrt(var_hat, out=var_hat)
mean_hat /= var_hat
mean_hat += wd * weight
g = mean_hat
r2 = g.norm()
# calculate lamb_trust_ratio
ratio = r1 / r2
# becomes NaN if ratio == NaN or 0, otherwise 0
nan_or_zero = 1 - ratio / ratio
r = where(nan_or_zero, ones_like(ratio), ratio)
lr *= r
# update weight
g *= lr
weight[:] -= g
def update(self, index, weight, grad, state):
assert (isinstance(weight, NDArray))
assert (isinstance(grad, NDArray))
lr = self._get_lr(index)
self._update_count(index)
grad = grad * self.rescale_grad
if self.clip_gradient is not None:
grad = clip(grad, -self.clip_gradient, self.clip_gradient)
history = state
history[:] += (grad * grad)
weight[:] += -lr * (grad / sqrt(history + self.float_stable_eps) +
self.wd * weight)
def parse_net_output(Y,numClass, box_per_cell):
pred = nd.transpose(Y,(0,2,3,1))
pred = pred.reshape((0,0,0,box_per_cell,numClass + 5)) #add one dim for boxes
predCls = nd.slice_axis(pred, begin = 0, end = numClass,axis=-1)
predObject = nd.slice_axis(pred,begin=numClass,end=numClass+1,axis=-1)
#predObject = nd.sigmoid(predObject)
predXY = nd.slice_axis(pred, begin = numClass + 1, end = numClass + 3, axis=-1)
predWH = nd.slice_axis(pred, begin = numClass + 3, end = numClass + 5, axis=-1)
#predXY = nd.sigmoid(predXY)
x,y = convert_xy(predXY)
w,h = convert_wh(predWH)
w = nd.clip(w,0,1)
h = nd.clip(h,0,1)
x0 = nd.clip(x, 0, 1)
y0 = nd.clip(y,0,1)
x1 = nd.clip(x0 + w,0,1)
y1 = np.clip(y0 + h, 0,1)
x = x0
y = y0
w = x1 - x0
h = y1 - y0
XYWH = nd.concat(x,y,w,h,dim=-1)
return predCls, predObject, XYWH
def plot_images(images, path, ncols, nrows):
fig = plt.figure(figsize=(25, 25))
grid = ImageGrid(
fig,
111, # similar to subplot(111)
nrows_ncols=(nrows, ncols), # creates 2x2 grid of axes
axes_pad=0.1, # pad between axes in inch.
)
for ax, img in zip(grid, images):
norm_img = normalize_image(img, -1, 1)
img = nd.clip(norm_img * 255 + 0.5, 0, 255).asnumpy().astype(np.uint8)
img = np.transpose(img, (1, 2, 0))
ax.imshow(img)
ax.axis('off')
plt.savefig(path, bbox_inches='tight')
def step(self, indices, weights, grads, states):
"""Perform an optimization step using gradients and states.
Parameters
----------
indices : list of int
List of unique indices of the parameters into the individual learning rates
and weight decays. Learning rates and weight decay may be set via `set_lr_mult()`
and `set_wd_mult()`, respectively.
weights : list of NDArray
List of parameters to be updated.
grads : list of NDArray
List of gradients of the objective with respect to this parameter.
states : List of any obj
List of state returned by `create_state()`.
"""
for index, weight, grad, state in zip(indices, weights, grads, states):
self._update_count(index)
lr = self._get_lr(index)
wd = self._get_wd(index)
t = self._index_update_count[index]
if len(list(grad.shape_array())) > 3:
grad -= grad.mean(axis=0, exclude=True, keepdims=True)
# preprocess grad
grad *= self.rescale_grad
if self.clip_gradient is not None:
grad = clip(grad, -self.clip_gradient, self.clip_gradient)
grad += wd * weight
coef1 = 1. - self.beta1**t
coef2 = 1. - self.beta2**t
lr *= math.sqrt(coef2) / coef1
# update mean and var
mean, var = state
mean[:] *= self.beta1
mean[:] += (1. - self.beta1) * grad
var[:] *= self.beta2
var[:] += (1. - self.beta2) * square(grad)
# update weight
d = mean / (sqrt(var) + self.epsilon)
weight[:] -= lr * d
def step(self, indices, weights, grads, states):
"""Perform an optimization step using gradients and states.
Parameters
----------
indices : list of int
List of unique indices of the parameters into the individual learning rates
and weight decays. Learning rates and weight decay may be set via `set_lr_mult()`
and `set_wd_mult()`, respectively.
weights : list of NDArray
List of parameters to be updated.
grads : list of NDArray
List of gradients of the objective with respect to this parameter.
states : List of any obj
List of state returned by `create_state()`.
"""
for index, weight, grad, state in zip(indices, weights, grads, states):
if len(list(grad.shape_array())) > 3:
grad -= grad.mean(axis=0, exclude=True, keepdims=True)
self._update_count(index)
lr = self._get_lr(index)
wd = self._get_wd(index)
# preprocess grad
grad *= self.rescale_grad
if self.clip_gradient is not None:
grad = clip(grad, -self.clip_gradient, self.clip_gradient)
grad += wd * weight
# update mom
mom = state
if mom is not None:
mom[:] *= self.momentum
mom[:] -= lr * grad
d = self.momentum * mom - lr * grad
else:
d = -lr * grad
# update weight
weight[:] += d
def imagenet_clamp_batch(batch, low, high):
""" Not necessary in practice """
F.clip(batch[:,0,:,:],low-123.680, high-123.680)
F.clip(batch[:,1,:,:],low-116.779, high-116.779)
F.clip(batch[:,2,:,:],low-103.939, high-103.939)
def main():
parser = argparse.ArgumentParser(description='Script to test the trained network on a game.')
parser.add_argument('-r', '--rom', required=False, type=str,
default=os.path.join('roms', 'breakout.bin'),
help='Path of the ROM File.')
parser.add_argument('-v', '--visualization', action='store_true',
help='Visualize the runs.')
parser.add_argument('--lr', required=False, type=float, default=0.01,
help='Learning rate of the AdaGrad optimizer')
parser.add_argument('--eps', required=False, type=float, default=0.01,
help='Eps of the AdaGrad optimizer')
parser.add_argument('--clip-gradient', required=False, type=float, default=None,
help='Clip threshold of the AdaGrad optimizer')
parser.add_argument('--double-q', action='store_true',
help='Use Double DQN only if specified')
parser.add_argument('--wd', required=False, type=float, default=0.0,
help='Weight of the L2 Regularizer')
parser.add_argument('-c', '--ctx', required=False, type=str, default='gpu',
help='Running Context. E.g `-c gpu` or `-c gpu1` or `-c cpu`')
parser.add_argument('-d', '--dir-path', required=False, type=str, default='',
help='Saving directory of model files.')
parser.add_argument('--start-eps', required=False, type=float, default=1.0,
help='Eps of the epsilon-greedy policy at the beginning')
parser.add_argument('--replay-start-size', required=False, type=int, default=50000,
help='The step that the training starts')
parser.add_argument('--kvstore-update-period', required=False, type=int, default=1,
help='The period that the worker updates the parameters from the sever')
parser.add_argument('--kv-type', required=False, type=str, default=None,
help='type of kvstore, default will not use kvstore, could also be dist_async')
parser.add_argument('--optimizer', required=False, type=str, default="adagrad",
help='type of optimizer')
args = parser.parse_args()
if args.dir_path == '':
rom_name = os.path.splitext(os.path.basename(args.rom))[0]
args.dir_path = 'dqn-%s-lr%g' % (rom_name, args.lr)
replay_start_size = args.replay_start_size
max_start_nullops = 30
replay_memory_size = 1000000
history_length = 4
rows = 84
cols = 84
ctx = parse_ctx(args.ctx)
q_ctx = mx.Context(*ctx[0])
game = AtariGame(rom_path=args.rom, resize_mode='scale', replay_start_size=replay_start_size,
resized_rows=rows, resized_cols=cols, max_null_op=max_start_nullops,
replay_memory_size=replay_memory_size, display_screen=args.visualization,
history_length=history_length)
##RUN NATURE
freeze_interval = 10000
epoch_num = 200
steps_per_epoch = 250000
update_interval = 4
discount = 0.99
eps_start = args.start_eps
eps_min = 0.1
eps_decay = (eps_start - eps_min) / 1000000
eps_curr = eps_start
freeze_interval /= update_interval
minibatch_size = 32
action_num = len(game.action_set)
data_shapes = {'data': (minibatch_size, history_length) + (rows, cols),
'dqn_action': (minibatch_size,), 'dqn_reward': (minibatch_size,)}
dqn_sym = dqn_sym_nature(action_num)
qnet = Base(data_shapes=data_shapes, sym_gen=dqn_sym, name='QNet',
initializer=DQNInitializer(factor_type="in"),
ctx=q_ctx)
target_qnet = qnet.copy(name="TargetQNet", ctx=q_ctx)
use_easgd = False
optimizer = mx.optimizer.create(name=args.optimizer, learning_rate=args.lr, eps=args.eps,
clip_gradient=args.clip_gradient,
rescale_grad=1.0, wd=args.wd)
updater = mx.optimizer.get_updater(optimizer)
qnet.print_stat()
target_qnet.print_stat()
# Begin Playing Game
training_steps = 0
total_steps = 0
for epoch in range(epoch_num):
# Run Epoch
steps_left = steps_per_epoch
episode = 0
epoch_reward = 0
start = time.time()
game.start()
while steps_left > 0:
# Running New Episode
episode += 1
episode_loss = 0.0
episode_q_value = 0.0
episode_update_step = 0
episode_action_step = 0
time_episode_start = time.time()
game.begin_episode(steps_left)
while not game.episode_terminate:
# 1. We need to choose a new action based on the current game status
if game.state_enabled and game.replay_memory.sample_enabled:
do_exploration = (npy_rng.rand() < eps_curr)
eps_curr = max(eps_curr - eps_decay, eps_min)
if do_exploration:
action = npy_rng.randint(action_num)
else:
# TODO Here we can in fact play multiple gaming instances simultaneously and make actions for each
# We can simply stack the current_state() of gaming instances and give prediction for all of them
# We need to wait after calling calc_score(.), which makes the program slow
# TODO Profiling the speed of this part!
current_state = game.current_state()
state = nd.array(current_state.reshape((1,) + current_state.shape),
ctx=q_ctx) / float(255.0)
qval_npy = qnet.forward(is_train=False, data=state)[0].asnumpy()
action = numpy.argmax(qval_npy)
episode_q_value += qval_npy[0, action]
episode_action_step += 1
else:
action = npy_rng.randint(action_num)
# 2. Play the game for a single mega-step (Inside the game, the action may be repeated for several times)
game.play(action)
total_steps += 1
# 3. Update our Q network if we can start sampling from the replay memory
# Also, we update every `update_interval`
if total_steps % update_interval == 0 and game.replay_memory.sample_enabled:
# 3.1 Draw sample from the replay_memory
training_steps += 1
episode_update_step += 1
states, actions, rewards, next_states, terminate_flags \
= game.replay_memory.sample(batch_size=minibatch_size)
states = nd.array(states, ctx=q_ctx) / float(255.0)
next_states = nd.array(next_states, ctx=q_ctx) / float(255.0)
actions = nd.array(actions, ctx=q_ctx)
rewards = nd.array(rewards, ctx=q_ctx)
terminate_flags = nd.array(terminate_flags, ctx=q_ctx)
# 3.2 Use the target network to compute the scores and
# get the corresponding target rewards
if not args.double_q:
target_qval = target_qnet.forward(is_train=False, data=next_states)[0]
target_rewards = rewards + nd.choose_element_0index(target_qval,
nd.argmax_channel(target_qval))\
* (1.0 - terminate_flags) * discount
else:
target_qval = target_qnet.forward(is_train=False, data=next_states)[0]
qval = qnet.forward(is_train=False, data=next_states)[0]
target_rewards = rewards + nd.choose_element_0index(target_qval,
nd.argmax_channel(qval))\
* (1.0 - terminate_flags) * discount
outputs = qnet.forward(is_train=True,
data=states,
dqn_action=actions,
dqn_reward=target_rewards)
qnet.backward()
qnet.update(updater=updater)
# 3.3 Calculate Loss
diff = nd.abs(nd.choose_element_0index(outputs[0], actions) - target_rewards)
quadratic_part = nd.clip(diff, -1, 1)
loss = 0.5 * nd.sum(nd.square(quadratic_part)).asnumpy()[0] +\
nd.sum(diff - quadratic_part).asnumpy()[0]
episode_loss += loss
# 3.3 Update the target network every freeze_interval
if training_steps % freeze_interval == 0:
qnet.copy_params_to(target_qnet)
steps_left -= game.episode_step
time_episode_end = time.time()
# Update the statistics
epoch_reward += game.episode_reward
info_str = "Epoch:%d, Episode:%d, Steps Left:%d/%d, Reward:%f, fps:%f, Exploration:%f" \
% (epoch, episode, steps_left, steps_per_epoch, game.episode_reward,
game.episode_step / (time_episode_end - time_episode_start), eps_curr)
if episode_update_step > 0:
info_str += ", Avg Loss:%f/%d" % (episode_loss / episode_update_step,
episode_update_step)
if episode_action_step > 0:
info_str += ", Avg Q Value:%f/%d" % (episode_q_value / episode_action_step,
episode_action_step)
if episode % 100 == 0:
logging.info(info_str)
end = time.time()
fps = steps_per_epoch / (end - start)
qnet.save_params(dir_path=args.dir_path, epoch=epoch)
logging.info("Epoch:%d, FPS:%f, Avg Reward: %f/%d"
% (epoch, fps, epoch_reward / float(episode), episode))
def tensor_save_rgbimage(img, filename, cuda=False):
img = F.clip(img, 0, 255).asnumpy()
img = img.transpose(1, 2, 0).astype('uint8')
img = Image.fromarray(img)
img.save(filename)
def get_rmse_log(net, X_train, y_train):
"""Gets root mse between the logarithms of the prediction and the truth."""
num_train = X_train.shape[0]
clipped_preds = nd.clip(net(X_train), 1, float('inf'))
return np.sqrt(2 * nd.sum(square_loss(
nd.log(clipped_preds), nd.log(y_train))).asscalar() / num_train)
