labels.append(label)
return {"data": np.array(data).astype(np.float32), "label":np.array(labels)}
if __name__ == '__main__':
parser = argparse.ArgumentParser()
os.system("rm -r tbdata/")
tb = TB("tbdata/")
iswarmup = True
with TrainingEnv(name = "lyy.{}.test".format(net_name), part_count = 2, custom_parser = parser) as env:
args = parser.parse_args()
num_GPU = len(args.devices.split(','))
minibatch_size *= num_GPU
net = make_network(minibatch_size = minibatch_size)
preloss = net.loss_var
net.loss_var = WeightDecay(net.loss_var, {"*conv*": 1e-4, "*fc*": 1e-4})
train_func = env.make_func_from_loss_var(net.loss_var, "train", train_state = True)
lr = 0.1 * num_GPU
if iswarmup:
lr /= 10
optimizer = Momentum(lr, 0.9)
optimizer(train_func)
#train_func.comp_graph.share_device_memory_with(valid_func.comp_graph)
dic = {
"loss": net.loss_var,
"pre_loss": preloss,
"outputs": net.outputs[0]
(img, label) = p.get()
data.append(img)
labels.append(label)
return {
"data": np.array(data).astype(np.float32),
"label": np.array(labels)
}
if __name__ == '__main__':
with TrainingEnv(name="lyy.{}.test".format(net_name), part_count=2) as env:
net = make_network(minibatch_size=minibatch_size)
preloss = net.loss_var
net.loss_var = WeightDecay(net.loss_var, {
"*conv*:W": 1e-4,
"*fc*:W": 1e-4,
"*bnaff*:k": 1e-4,
"*offset*": 1e-4
})
train_func = env.make_func_from_loss_var(net.loss_var,
"train",
train_state=True)
valid_func = env.make_func_from_loss_var(net.loss_var,
"val",
train_state=False)
lr = 0.1
optimizer = MyMomentum(lr, 0.9)
#optimizer.learning_rate = 0.01
optimizer(train_func)
return {
"data": np.array(data).astype(np.float32),
"label": np.array(labels)
}
if __name__ == '__main__':
parser = argparse.ArgumentParser()
with TrainingEnv(name="lyy.{}.test".format(net_name),
part_count=2,
custom_parser=parser) as env:
print("A")
net = make_network(minibatch_size=minibatch_size)
preloss = net.loss_var
net.loss_var = WeightDecay(net.loss_var, {
"*encoder*": 1e-4,
"*outputs*": 1e-4
})
"""
print(isinstance(net.loss_var.owner_opr, WeightDecay))
print(net.loss_var.owner_opr._params)
print(type(net.loss_var.owner_opr._param_weights))
exit()
"""
train_func = env.make_func_from_loss_var(net.loss_var,
"train",
train_state=True)
valid_func = env.make_func_from_loss_var(net.loss_var,
"val",
train_state=False)
