def rmsprop(nn, X_train, y_train, val_set=None, alpha=1e-3, mb_size=256, n_iter=2000, print_after=100):
cache = {k: np.zeros_like(v) for k, v in nn.model.items()}
gamma = .9
minibatches = get_minibatch(X_train, y_train, mb_size)
if val_set:
X_val, y_val = val_set
for iter in range(1, n_iter + 1):
idx = np.random.randint(0, len(minibatches))
X_mini, y_mini = minibatches[idx]
grad, loss = nn.train_step(X_mini, y_mini)
if iter % print_after == 0:
if val_set:
val_acc = util.accuracy(y_val, nn.predict(X_val))
print('Iter-{} loss: {:.4f} validation: {:4f}'.format(iter, loss, val_acc))
print('grad:',grad)
else:
print('Iter-{} loss: {:.4f}'.format(iter, loss))
print('grad:',grad)
for k in grad:
cache[k] = util.exp_running_avg(cache[k], grad[k]**2, gamma)
nn.model[k] -= alpha * grad[k] / (np.sqrt(cache[k]) + c.eps)
return nn
def sgd(nn,
X_train,
y_train,
val_set=None,
alpha=1e-3,
mb_size=256,
n_iter=2000,
print_after=100):
minibatches = get_minibatch(X_train, y_train, mb_size)
if val_set:
X_val, y_val = val_set
for iter in range(1, n_iter + 1):
idx = np.random.randint(0, len(minibatches))
X_mini, y_mini = minibatches[idx]
grad, loss = nn.train_step(X_mini, y_mini)
if iter % print_after == 0:
if val_set:
val_acc = util.accuracy(y_val, nn.predict(X_val))
print('Iter-{} loss: {:.4f} validation: {:4f}'.format(
iter, loss, val_acc))
else:
print('Iter-{} loss: {:.4f}'.format(iter, loss))
for layer in grad:
nn.model[layer] -= alpha * grad[layer]
return nn
def nesterov(nn, X_train, y_train, val_set=None, alpha=1e-3, mb_size=256, n_iter=2000, print_after=100):
velocity = {k: np.zeros_like(v) for k, v in nn.model.items()}
gamma = .9
minibatches = get_minibatch(X_train, y_train, mb_size)
if val_set:
X_val, y_val = val_set
for iter in range(1, n_iter + 1):
idx = np.random.randint(0, len(minibatches))
X_mini, y_mini = minibatches[idx]
nn_ahead = copy.deepcopy(nn)
nn_ahead.model.update({k: v + gamma * velocity[k] for k, v in nn.model.items()})
grad, loss = nn_ahead.train_step(X_mini, y_mini)
if iter % print_after == 0:
if val_set:
val_acc = util.accuracy(y_val, nn.predict(X_val))
print('Iter-{} loss: {:.4f} validation: {:4f}'.format(iter, loss, val_acc))
print('grad:',grad)
else:
print('Iter-{} loss: {:.4f}'.format(iter, loss))
print('grad:',grad)
for layer in grad:
velocity[layer] = gamma * velocity[layer] + alpha * grad[layer]
nn.model[layer] -= velocity[layer]
return nn
def nesterov(nn, X_train, y_train, val_set=None, alpha=1e-3, mb_size=256, n_iter=2000, print_after=100):
velocity = {k: np.zeros_like(v) for k, v in nn.model.items()}
gamma = .9
minibatches = get_minibatch(X_train, y_train, mb_size)
if val_set:
X_val, y_val = val_set
for iter in range(1, n_iter + 1):
idx = np.random.randint(0, len(minibatches))
X_mini, y_mini = minibatches[idx]
nn_ahead = copy.deepcopy(nn)
nn_ahead.model.update({k: v + gamma * velocity[k] for k, v in nn.model.items()})
grad, loss = nn_ahead.train_step(X_mini, y_mini)
if iter % print_after == 0:
if val_set:
val_acc = util.accuracy(y_val, nn.predict(X_val))
print('Iter-{} loss: {:.4f} validation: {:4f}'.format(iter, loss, val_acc))
else:
print('Iter-{} loss: {:.4f}'.format(iter, loss))
for layer in grad:
velocity[layer] = gamma * velocity[layer] + alpha * grad[layer]
nn.model[layer] -= velocity[layer]
return nn
def adam(nn, X_train, y_train, val_set=None, alpha=0.001, mb_size=256, n_iter=2000, print_after=100):
M = {k: np.zeros_like(v) for k, v in nn.model.items()}
R = {k: np.zeros_like(v) for k, v in nn.model.items()}
beta1 = .9
beta2 = .999
minibatches = get_minibatch(X_train, y_train, mb_size)
if val_set:
X_val, y_val = val_set
for iter in range(1, n_iter + 1):
t = iter
idx = np.random.randint(0, len(minibatches))
X_mini, y_mini = minibatches[idx]
grad, loss = nn.train_step(X_mini, y_mini)
if iter % print_after == 0:
if val_set:
val_acc = util.accuracy(y_val, nn.predict(X_val))
print('Iter-{} loss: {:.4f} validation: {:4f}'.format(iter, loss, val_acc))
else:
print('Iter-{} loss: {:.4f}'.format(iter, loss))
for k in grad:
M[k] = util.exp_running_avg(M[k], grad[k], beta1)
R[k] = util.exp_running_avg(R[k], grad[k]**2, beta2)
m_k_hat = M[k] / (1. - beta1**(t))
r_k_hat = R[k] / (1. - beta2**(t))
nn.model[k] -= alpha * m_k_hat / (np.sqrt(r_k_hat) + c.eps)
return nn
def rmsprop(nn, X_train, y_train, val_set=None, alpha=1e-3, mb_size=256, n_iter=2000, print_after=100):
cache = {k: np.zeros_like(v) for k, v in nn.model.items()}
gamma = .9
minibatches = get_minibatch(X_train, y_train, mb_size)
if val_set:
X_val, y_val = val_set
for iter in range(1, n_iter + 1):
idx = np.random.randint(0, len(minibatches))
X_mini, y_mini = minibatches[idx]
grad, loss = nn.train_step(X_mini, y_mini)
if iter % print_after == 0:
if val_set:
val_acc = util.accuracy(y_val, nn.predict(X_val))
print('Iter-{} loss: {:.4f} validation: {:4f}'.format(iter, loss, val_acc))
else:
print('Iter-{} loss: {:.4f}'.format(iter, loss))
for k in grad:
cache[k] = util.exp_running_avg(cache[k], grad[k]**2, gamma)
nn.model[k] -= alpha * grad[k] / (np.sqrt(cache[k]) + c.eps)
return nn
def sgd(nn,
X_train,
y_train,
f,
val_set=None,
alpha=1e-3,
mb_size=256,
n_iter=2000,
print_after=100):
minibatches = get_minibatch(X_train, y_train, mb_size)
accu = []
if val_set:
X_val, y_val = val_set
for iter in range(1, n_iter + 1):
idx = np.random.randint(0, len(minibatches))
X_mini, y_mini = minibatches[idx]
grad, loss = nn.train_step(X_mini, y_mini)
if iter % print_after == 0:
if val_set:
val_acc = util.accuracy(y_val, nn.predict(X_val))
accu.append(val_acc)
print('Iter-{} loss: {:.4f} validation: {:4f}'.format(
iter, loss, val_acc))
f.write('Iter-{} loss: {:.4f} validation: {:4f}'.format(
iter, loss, val_acc))
np.set_printoptions(threshold=np.NaN)
f.write('grad[W1] {}:{}'.format(iter, '\n'))
f.write('{} {}'.format(grad['W1'], '\n'))
f.write('grad[b1] {}:{}'.format(iter, '\n'))
f.write('{} {}'.format(grad['b1'], '\n'))
'''
f.write('grad[W2] {}:{}'.format(iter,'\n'))
f.write('{} {}'.format(grad['W2'],'\n'))
f.write('grad[b2] {}:{}'.format(iter,'\n'))
f.write('{} {}'.format(grad['b2'],'\n'))
f.write('grad[W3] {}:{}'.format(iter,'\n'))
f.write('{} {}'.format(grad['W3'],'\n'))
f.write('grad[b3] {}:{}'.format(iter,'\n'))
f.write('{} {}'.format(grad['b3'],'\n'))
'''
else:
print('Iter-{} loss: {:.4f}'.format(iter, loss))
print('grad:', grad)
for layer in grad:
nn.model[layer] -= alpha * grad[layer]
for content in accu:
f.write(str(content))
return nn
def sgd(nn,
X_train,
y_train,
val_set=None,
alpha=1e-3,
mb_size=256,
n_iter=2000,
print_after=100):
minibatches = get_minibatch(X_train, y_train, mb_size)
if val_set:
X_val, y_val = val_set
start = time.time()
for iter in range(1, n_iter + 1):
if iter != 0 and iter % 20000 == 0:
print('Learning rate halved')
alpha /= 2
idx = np.random.randint(0, len(minibatches))
X_mini, y_mini = minibatches[idx]
grad, loss = nn.train_step(X_mini, y_mini, iter)
if iter % print_after == 0:
if val_set:
end = time.time()
val_acc = util.accuracy(y_val, nn.predict(X_val))
test_acc = util.accuracy(y_mini, nn.predict(X_mini))
print(
'Iter-{} loss: {:.4f} test: {:4f} time: {:4f} validation: {:4f}'
.format(iter, loss, test_acc, end - start, val_acc))
else:
print('Iter-{} loss: {:.4f}'.format(iter, loss))
for layer in grad:
nn.model[layer] -= alpha * grad[layer]
return nn
def momentum1(nn, X_train, y_train,worker_num ,val_set=None, alpha=1e-3, mb_size=256, n_iter=2000, print_after=100):
gamma = .9
velocity = [[]for i in range(worker_num)]
minibatches =[[]for i in range(worker_num)]
X_mini,y_mini=[[] for i in range(worker_num)],[[] for i in range(worker_num)]
X_val,y_val =[[] for i in range(worker_num)],[[] for i in range(worker_num)]
grad,loss = [[] for i in range(worker_num)],[[] for i in range(worker_num)]
val_acc =[[] for i in range(worker_num)]
index = ['W1', 'W2', 'W4', 'W5', 'b1', 'b2', 'b4', 'b5', 'gamma4', 'gamma5', 'beta4', 'beta5']
except_index=[]
average_grad = dict()
accu = [[] for i in range(worker_num)]
for k in range(worker_num):
minibatches[k] = get_minibatch(X_train[k], y_train[k], mb_size)
velocity[k] = {k: np.zeros_like(v) for k, v in nn[k].model.items()}
if val_set:
X_val, y_val = val_set
for iter in range(1, n_iter + 1):
for k in range(worker_num):
idx = np.random.randint(0, len(minibatches[k]))
X_mini[k], y_mini[k] = minibatches[k][idx]
grad[k], loss[k] = nn[k].train_step(X_mini[k], y_mini[k])
if iter % print_after == 0:
if val_set:
val_acc[k] = util.accuracy(y_val, nn[k].predict(X_val))
print('Iter-{} loss: {:.4f} validation: {:4f}'.format(iter, loss[k], val_acc[k]))
#print('grad:',grad)
else:
print('Iter-{} loss: {:.4f}'.format(iter, loss[k]))
#print('grad:',grad)
for k in range(worker_num):
for layer in grad[0]:
if iter%15 ==0:
velocity[k][layer] = gamma * (velocity[k][layer])+ alpha * (grad[0][layer]+grad[1][layer]+grad[2][layer]+grad[3][layer])/worker_num
else:
velocity[k][layer] = gamma*(velocity[k][layer])+alpha*grad[k][layer]
nn[k].model[layer] -= velocity[k][layer]
return nn
def momentum(nn,
X_train,
y_train,
val_set=None,
alpha=1e-3,
mb_size=256,
n_iter=2000,
print_after=100,
max_norm=None):
velocity = {k: np.zeros_like(v) for k, v in nn.model.items()}
gamma = .9
minibatches = get_minibatch(X_train, y_train, mb_size)
if val_set:
X_val, y_val = val_set
for iter in range(1, n_iter + 1):
idx = np.random.randint(0, len(minibatches))
X_mini, y_mini = minibatches[idx]
grad, loss = nn.train_step(X_mini, y_mini)
if iter % print_after == 0:
if val_set:
val_acc = util.accuracy(y_val, nn.predict(X_val))
print('Iter-{} loss: {:.4f} validation: {:4f}'.format(
iter, loss, val_acc))
else:
print('Iter-{} loss: {:.4f}'.format(iter, loss))
for layer in grad:
velocity[layer] = gamma * velocity[layer] + alpha * grad[layer]
nn.model[layer] -= velocity[layer]
if max_norm != None:
nn.model[layer] = reg.limit_norm(nn.model[layer],
max_val=max_norm)
return nn
def sgd(nn, X_train, y_train, val_set=None, alpha=1e-3, mb_size=256, n_iter=2000, print_after=100):
minibatches = get_minibatch(X_train, y_train, mb_size)
if val_set:
X_val, y_val = val_set
for iter in range(1, n_iter + 1):
idx = np.random.randint(0, len(minibatches))
X_mini, y_mini = minibatches[idx]
grad, loss = nn.train_step(X_mini, y_mini)
if iter % print_after == 0:
if val_set:
val_acc = util.accuracy(y_val, nn.predict(X_val))
print('Iter-{} loss: {:.4f} validation: {:4f}'.format(iter, loss, val_acc))
else:
print('Iter-{} loss: {:.4f}'.format(iter, loss))
for layer in grad:
nn.model[layer] -= alpha * grad[layer]
return nn
def adagrad(nn,
X_train,
y_train,
val_set=None,
alpha=1e-3,
mb_size=256,
n_iter=2000,
print_after=100,
max_norm=None):
cache = {k: np.zeros_like(v) for k, v in nn.model.items()}
minibatches = get_minibatch(X_train, y_train, mb_size)
if val_set:
X_val, y_val = val_set
for iter in range(1, n_iter + 1):
idx = np.random.randint(0, len(minibatches))
X_mini, y_mini = minibatches[idx]
grad, loss = nn.train_step(X_mini, y_mini)
if iter % print_after == 0:
if val_set:
val_acc = util.accuracy(y_val, nn.predict(X_val))
print('Iter-{} loss: {:.4f} validation: {:4f}'.format(
iter, loss, val_acc))
else:
print('Iter-{} loss: {:.4f}'.format(iter, loss))
for k in grad:
cache[k] += grad[k]**2
nn.model[k] -= alpha * grad[k] / (np.sqrt(cache[k]) + c.eps)
if max_norm != None:
nn.model[k] = reg.limit_norm(nn.model[k], max_val=max_norm)
return nn
def sgd3(nn,
X_train,
y_train,
worker_num,
val_set=None,
alpha=1e-3,
mb_size=256,
n_iter=2000,
print_after=100):
minibatches = [[] for i in range(worker_num)]
X_mini, y_mini = [[] for i in range(worker_num)
], [[] for i in range(worker_num)]
X_val, y_val = [[] for i in range(worker_num)
], [[] for i in range(worker_num)]
grad, loss = [[] for i in range(worker_num)], [[]
for i in range(worker_num)]
val_acc = [[] for i in range(worker_num)]
share_time = 15
start_time = [[[] for j in range(share_time)] for i in range(worker_num)]
totoal_time = [[] for i in range(worker_num)]
for k in range(worker_num):
minibatches[k] = get_minibatch(X_train[k], y_train[k], mb_size)
if val_set:
#X_val[k], y_val[k] = val_set[k]
X_val, y_val = val_set
def f(x, y):
return x + a[y]
for iter in range(1, n_iter + 1):
for k in range(worker_num):
start_time[k][iter % share_time] = time.time()
idx = np.random.randint(0, len(minibatches[k]))
X_mini[k], y_mini[k] = minibatches[k][idx]
grad[k], loss[k] = nn[k].train_step(X_mini[k], y_mini[k])
if iter % print_after == 0:
if val_set:
#val_acc[k] = util.accuracy(y_val[k], nn[k].predict(X_val[k]))
val_acc[k] = util.accuracy(y_val, nn[k].predict(X_val))
print(
'Iter-{} worker {}, loss: {:.4f} validation: {:4f} {}'.
format(iter, k + 1, loss[k], val_acc[k], '\n'))
#f[k].write('Iter-{} worker {}, loss: {:.4f} validation: {:4f} {}'.format(iter,k+1 ,loss[k], val_acc[k],'\n'))
#np.set_printoptions(threshold=np.NaN)
#np.set_printoptions(precision=8)
#f[k].write('grad[{}][W1]{}:{}'.format(k+1,iter,'\n'))
#f[k].write('{}{}'.format(grad[k]['W1'],'\n'))
#f[k].write('grad[{}][b1]{}:{}'.format(k+1,iter,'\n'))
#f[k].write('{}{}'.format(grad[k]['b1'],'\n'))
#f[k].write('grad[{}][W2]{}:{}'.format(k+1,iter,'\n'))
#f[k].write('{}{}'.format(grad[k]['W2'],'\n'))
#f[k].write('grad[{}][b2]{}:{}'.format(k+1,iter,'\n'))
#f[k].write('{}{}'.format(grad[k]['b2'],'\n'))
#f[k].write('grad[{}][W3]{}:{}'.format(k+1,iter,'\n'))
#f[k].write('{}{}'.format(grad[k]['W3'],'\n'))
#f[k].write('grad[{}][b3]{}:{}'.format(k+1,iter,'\n'))
#f[k].write('{}{}'.format(grad[k]['b3'],'\n'))
#f[k].write('\n')
'''
#print('gamma 3 ',grad[k]['gamma3'])
#print('beta3 ',grad[k]['beta3'])
else:
print('Iter-{} loss: {:.4f}'.format(iter, loss))
print('grad:',grad)
if(k+1==worker_num):
print('Iter-{} average loss:{} loss: {:.4f} '.format(iter,k+1 ,sum(loss)/len(loss)))
#f[k].write('Iter-{} average loss:{} loss: {:.4f} '.format(iter,k+1 ,sum(loss)/len(loss)))
'''
if iter % 15 != 0:
for layer in grad[0]:
nn[k].model[layer] -= alpha * (grad[k][layer])
start_time[k][iter % share_time] = time.time() - start_time[k][
iter % share_time]
if iter % 15 == 0 and k == worker_num - 1:
average_grad = dict()
for layer in grad[0]:
average_grad[layer] = 0
for i in range(worker_num):
average_grad[layer] += grad[i][layer]
for j in range(worker_num):
for layer in grad[0]:
nn[j].model[
layer] -= alpha * average_grad[layer] / worker_num
start_time[j][iter %
15] = time.time() - start_time[j][iter % 15]
a = start_time[j]
totoal_time[j] = reduce(f, range(15), 0)
print('worker{} {}-{} totoal cost time {}ms'.format(
j + 1, iter - 14, iter, totoal_time[j] * 1000))
'''
for k in range(worker_num):
for layer in grad[k]:
if iter%15==0:
nn[k].model[layer] -= alpha *average_grad[layer]/worker_num
a = start_time[k]
print(a)
totoal_time[k] = reduce(f,range(15),0)
print('worker{} {}-{} totoal cost time {}ms'.format(k+1,iter-14,iter,totoal_time[k]))
else:
nn[k].model[layer]-=alpha*grad[k][layer]
print("per time {} {} ,{}".format(k+1,iter,start_time[k][iter%share_time]))
'''
return nn
def sgd3(nn, X_train, y_train,worker_num ,val_set=None, alpha=1e-3, mb_size=256, n_iter=2000, print_after=100):
minibatches =[[]for i in range(worker_num)]
X_mini,y_mini=[[] for i in range(worker_num)],[[] for i in range(worker_num)]
X_val,y_val =[[] for i in range(worker_num)],[[] for i in range(worker_num)]
grad,loss = [[] for i in range(worker_num)],[[] for i in range(worker_num)]
val_acc =[[] for i in range(worker_num)]
for k in range(worker_num):
minibatches[k] = get_minibatch(X_train[k], y_train[k], mb_size)
if val_set:
#X_val[k], y_val[k] = val_set[k]
X_val,y_val = val_set
for iter in range(1, n_iter + 1):
for k in range(worker_num):
idx = np.random.randint(0, len(minibatches[k]))
X_mini[k], y_mini[k] = minibatches[k][idx]
grad[k], loss[k] = nn[k].train_step(X_mini[k], y_mini[k])
if iter % print_after == 0:
if val_set:
#val_acc[k] = util.accuracy(y_val[k], nn[k].predict(X_val[k]))
val_acc[k] = util.accuracy(y_val, nn[k].predict(X_val))
print('Iter-{} worker {}, loss: {:.4f} validation: {:4f} {}'.format(iter,k+1 ,loss[k], val_acc[k],'\n'))
f[k].write('Iter-{} worker {}, loss: {:.4f} validation: {:4f} {}'.format(iter,k+1 ,loss[k], val_acc[k],'\n'))
np.set_printoptions(threshold=np.NaN)
np.set_printoptions(precision=8)
f[k].write('grad[{}][W1]{}:{}'.format(k+1,iter,'\n'))
f[k].write('{}{}'.format(grad[k]['W1'],'\n'))
f[k].write('grad[{}][b1]{}:{}'.format(k+1,iter,'\n'))
f[k].write('{}{}'.format(grad[k]['b1'],'\n'))
f[k].write('grad[{}][W2]{}:{}'.format(k+1,iter,'\n'))
f[k].write('{}{}'.format(grad[k]['W2'],'\n'))
f[k].write('grad[{}][b2]{}:{}'.format(k+1,iter,'\n'))
f[k].write('{}{}'.format(grad[k]['b2'],'\n'))
f[k].write('grad[{}][W3]{}:{}'.format(k+1,iter,'\n'))
f[k].write('{}{}'.format(grad[k]['W3'],'\n'))
f[k].write('grad[{}][b3]{}:{}'.format(k+1,iter,'\n'))
f[k].write('{}{}'.format(grad[k]['b3'],'\n'))
f[k].write('\n')
#print('gamma 3 ',grad[k]['gamma3'])
#print('beta3 ',grad[k]['beta3'])
else:
print('Iter-{} loss: {:.4f}'.format(iter, loss))
print('grad:',grad)
if(k+1==worker_num):
print('Iter-{} average loss:{} loss: {:.4f} '.format(iter,k+1 ,sum(loss)/len(loss)))
f[k].write('Iter-{} average loss:{} loss: {:.4f} '.format(iter,k+1 ,sum(loss)/len(loss)))
average_grad = dict()
for layer in grad[0]:
average_grad[layer]=0
for i in range(worker_num):
average_grad[layer]+=grad[i][layer]
for k in range(worker_num):
for layer in grad[k]:
#average = 0
#for j in range(10):
# average +=grad[k][layer]
#nn[k].model[layer] -= alpha * (grad[0][layer]+grad[1][layer]+grad[2][layer]+grad[3][layer]+grad[4][layer]+grad[5][layer]+grad[6][layer]+grad[7][layer]+grad[8][layer]+grad[9][layer])/10
nn[k].model[layer] -= alpha *average_grad[layer]/worker_num
return nn
def interleaving(nn,
X_train,
y_train,
val_set=None,
alpha=1e-3,
mb_size=256,
n_iter=2000,
print_after=100):
ITER_FOR_DOUBLE = 2500
minibatches = get_minibatch(X_train, y_train, mb_size)
if val_set:
X_val, y_val = val_set
start = time.time()
for iter in range(1, n_iter + 1):
idx = np.random.randint(0, len(minibatches))
X_mini, y_mini = minibatches[idx]
grad, loss = nn.train_step(X_mini, y_mini, iter)
if iter % print_after == 0:
# for layer in grad:
# print(np.linalg.norm(grad[layer])/np.linalg.norm(nn.model[layer]))
if val_set:
end = time.time()
val_acc = util.accuracy(y_val, nn.predict(X_val))
test_acc = util.accuracy(y_mini, nn.predict(X_mini))
print(
'Iter-{} loss: {:.4f} test: {:4f} time: {:4f} validation: {:4f}'
.format(iter, loss, test_acc, end - start, val_acc))
else:
print('Iter-{} loss: {:.4f}'.format(iter, loss))
for layer in grad:
nn.model[layer] -= alpha * grad[layer]
if iter == ITER_FOR_DOUBLE:
# Implement
nn.freezeLastLayer()
# Create dataset with data passed through first neural network, no train
# Create neural network which takes that input
nn2 = neuralnet.ResNet(nn.H, nn.C, nn.H, num_layers=4)
nn2.model['Wf'] = nn.model['Wf']
nn2.model['bf'] = nn.model['bf']
nn2.freezeClassificationLayer()
if iter > ITER_FOR_DOUBLE:
new_X_mini = nn.passDataNoClass(X_mini)
grad, loss = nn2.train_step(new_X_mini, y_mini, iter)
# No print, because validation is vague
# if iter % print_after == 0:
# # for layer in grad:
# # print(np.linalg.norm(grad[layer])/np.linalg.norm(nn.model[layer]))
# if val_set:
# end = time.time()
# val_acc = util.accuracy(y_val, nn2.predict(new_X_val))
# test_acc = util.accuracy(y_mini, nn2.predict(new_X_mini))
# print('nn2: Iter-{} loss: {:.4f} test: {:4f} time: {:4f} validation: {:4f}'.format(iter, loss, test_acc, end-start, val_acc))
# else:
# print('Iter-{} loss: {:.4f}'.format(iter, loss))
for layer in grad:
nn2.model[layer] -= alpha * grad[layer]
if val_set:
val_acc = util.accuracy(y_val, nn2.predict(nn.passDataNoClass(X_val)))
print('Final validation: {:4f}'.format(val_acc))
# shouldHalve = True
# for layer in grad:
# epsi = 0.01*alpha
# if np.linalg.norm(grad[layer])/np.linalg.norm(nn.model[layer]) > epsi:
# shouldHalve = False
# if shouldHalve:
# alpha /= 2
# print('Halved learning rate')
# if alpha <= 0.0001:
# print('Finished learning as step size too small')
# return nn
return nn
def sgd3(nn, X_train, y_train,worker_num ,val_set=None, alpha=1e-3, mb_size=256, n_iter=2000, print_after=100):
minibatches =[[]for i in range(worker_num)]
X_mini,y_mini=[[] for i in range(worker_num)],[[] for i in range(worker_num)]
X_val,y_val =[[] for i in range(worker_num)],[[] for i in range(worker_num)]
grad,loss = [[] for i in range(worker_num)],[[] for i in range(worker_num)]
val_acc =[[] for i in range(worker_num)]
index = ['W1', 'W2', 'W4', 'W5', 'b1', 'b2', 'b4', 'b5', 'gamma4', 'gamma5', 'beta4', 'beta5']
except_index=[]
average_grad = dict()
for k in range(worker_num):
minibatches[k] = get_minibatch(X_train[k], y_train[k], mb_size)
if val_set:
#X_val[k], y_val[k] = val_set[k]
X_val,y_val = val_set
for iter in range(1, n_iter + 1):
for k in range(worker_num):
idx = np.random.randint(0, len(minibatches[k]))
X_mini[k], y_mini[k] = minibatches[k][idx]
grad[k], loss[k] = nn[k].train_step(X_mini[k], y_mini[k])
if iter % print_after == 0:
if val_set:
#val_acc[k] = util.accuracy(y_val[k], nn[k].predict(X_val[k]))
val_acc[k] = util.accuracy(y_val, nn[k].predict(X_val))
print('Iter-{} worker {}, loss: {:.4f} validation: {:4f} {}'.format(iter,k+1 ,loss[k], val_acc[k],'\n'))
f[k].write('Iter-{} worker {}, loss: {:.4f} validation: {:4f} {}'.format(iter,k+1 ,loss[k], val_acc[k],'\n'))
#np.set_printoptions(threshold=np.NaN)
#np.set_printoptions(precision=8)
f[k].write('grad[{}][W1]{}:{}'.format(k+1,iter,'\n'))
f[k].write('{}{}'.format(grad[k]['W1'],'\n'))
f[k].write('grad[{}][b1]{}:{}'.format(k+1,iter,'\n'))
f[k].write('{}{}'.format(grad[k]['b1'],'\n'))
'''
f[k].write('grad[{}][W2]{}:{}'.format(k+1,iter,'\n'))
f[k].write('{}{}'.format(grad[k]['W2'],'\n'))
f[k].write('grad[{}][b2]{}:{}'.format(k+1,iter,'\n'))
f[k].write('{}{}'.format(grad[k]['b2'],'\n'))
f[k].write('grad[{}][W3]{}:{}'.format(k+1,iter,'\n'))
f[k].write('{}{}'.format(grad[k]['W3'],'\n'))
f[k].write('grad[{}][b3]{}:{}'.format(k+1,iter,'\n'))
f[k].write('{}{}'.format(grad[k]['b3'],'\n'))
f[k].write('\n')
'''
#print('gamma 3 ',grad[k]['gamma3'])
#print('beta3 ',grad[k]['beta3'])
else:
print('Iter-{} loss: {:.4f}'.format(iter, loss))
print('grad:',grad)
if(k+1==worker_num):
print('Iter-{} average loss {} '.format(iter ,sum(loss)/len(loss)))
f[k].write('Iter-{} average loss: {} '.format(iter, sum(loss)/len(loss)))
except_index = random.sample(index,6)
for layer in except_index:
nn[k].model[layer] -= alpha*grad[k][layer]
available_index = [x for x in grad[0] if x not in except_index]
for layer in available_index:
average_grad[layer]=0
for i in range(worker_num):
average_grad[layer] += grad[i][layer]
for layer in available_index:
for k in range(worker_num):
nn[k].model[layer] -= alpha *average_grad[layer]/worker_num
return nn