问题描述
我正在关注此代码 https://github.com/BUAA-BDA/FedShapley/tree/master/TensorflowFL 并尝试运行文件 same_OR.py 并进行一些必要的更改
from __future__ import absolute_import,division,print_function
import tensorflow_federated as tff
import tensorflow.compat.v1 as tf
import numpy as np
import time
from scipy.special import comb,perm
import collections
import os
# tf.compat.v1.enable_v2_behavior()
# tf.compat.v1.enable_eager_execution()
# NUM_EXAMPLES_PER_USER = 1000
BATCH_SIZE = 100
NUM_AGENT = 5
def get_data_for_digit(source,digit):
output_sequence = []
all_samples = [i for i,d in enumerate(source[1]) if d == digit]
for i in range(0,len(all_samples),BATCH_SIZE):
batch_samples = all_samples[i:i + BATCH_SIZE]
output_sequence.append({
'x': np.array([source[0][i].flatten() / 255.0 for i in batch_samples],dtype=np.float32),'y': np.array([source[1][i] for i in batch_samples],dtype=np.int32)})
return output_sequence
def get_data_for_digit_test(source,len(all_samples)):
output_sequence.append({
'x': np.array(source[0][all_samples[i]].flatten() / 255.0,'y': np.array(source[1][all_samples[i]],dtype=np.int32)})
return output_sequence
def get_data_for_federated_agents(source,num):
output_sequence = []
Samples = []
for digit in range(0,10):
samples = [i for i,d in enumerate(source[1]) if d == digit]
samples = samples[0:5421]
Samples.append(samples)
all_samples = []
for sample in Samples:
for sample_index in range(int(num * (len(sample) / NUM_AGENT)),int((num + 1) * (len(sample) / NUM_AGENT))):
all_samples.append(sample[sample_index])
# all_samples = [i for i in range(int(num*(len(source[1])/NUM_AGENT)),int((num+1)*(len(source[1])/NUM_AGENT)))]
for i in range(0,dtype=np.int32)})
return output_sequence
BATCH_TYPE = tff.StructType([
('x',tff.TensorType(tf.float32,[None,784])),('y',tff.TensorType(tf.int32,[None]))])
MODEL_TYPE = tff.StructType([
('weights',[784,10])),('bias',[10]))])
@tff.tf_computation(MODEL_TYPE,BATCH_TYPE)
def batch_loss(model,batch):
predicted_y = tf.nn.softmax(tf.matmul(batch.x,model.weights) + model.bias)
return -tf.reduce_mean(tf.reduce_sum(
tf.one_hot(batch.y,10) * tf.log(predicted_y),axis=[1]))
@tff.tf_computation(MODEL_TYPE,BATCH_TYPE,tf.float32)
def batch_train(initial_model,batch,learning_rate):
# Define a group of model variables and set them to `initial_model`.
model_vars = tff.utils.create_variables('v',MODEL_TYPE)
init_model = tff.utils.assign(model_vars,initial_model)
# Perform one step of gradient descent using loss from `batch_loss`.
optimizer = tf.train.GradientDescentOptimizer(learning_rate)
with tf.control_dependencies([init_model]):
train_model = optimizer.minimize(batch_loss(model_vars,batch))
# Return the model vars after performing this gradient descent step.
with tf.control_dependencies([train_model]):
return tff.utils.identity(model_vars)
LOCAL_DATA_TYPE = tff.SequenceType(BATCH_TYPE)
@tff.federated_computation(MODEL_TYPE,tf.float32,LOCAL_DATA_TYPE)
def local_train(initial_model,learning_rate,all_batches):
# Mapping function to apply to each batch.
@tff.federated_computation(MODEL_TYPE,BATCH_TYPE)
def batch_fn(model,batch):
return batch_train(model,learning_rate)
l = tff.sequence_reduce(all_batches,initial_model,batch_fn)
return l
@tff.federated_computation(MODEL_TYPE,LOCAL_DATA_TYPE)
def local_eval(model,all_batches):
#
return tff.sequence_sum(
tff.sequence_map(
tff.federated_computation(lambda b: batch_loss(model,b),BATCH_TYPE),all_batches))
SERVER_MODEL_TYPE = tff.FederatedType(MODEL_TYPE,tff.SERVER,all_equal=True)
CLIENT_DATA_TYPE = tff.FederatedType(LOCAL_DATA_TYPE,tff.CLIENTS)
@tff.federated_computation(SERVER_MODEL_TYPE,CLIENT_DATA_TYPE)
def federated_eval(model,data):
return tff.federated_mean(
tff.federated_map(local_eval,[tff.federated_broadcast(model),data]))
SERVER_FLOAT_TYPE = tff.FederatedType(tf.float32,all_equal=True)
@tff.federated_computation(
SERVER_MODEL_TYPE,SERVER_FLOAT_TYPE,CLIENT_DATA_TYPE)
def federated_train(model,data):
l = tff.federated_map(
local_train,tff.federated_broadcast(learning_rate),data])
return l
# return tff.federated_mean()
def readTestimagesFromFile(distr_same):
ret = []
if distr_same:
f = open(os.path.join(os.path.dirname(__file__),"test_images1_.txt"),encoding="utf-8")
else:
f = open(os.path.join(os.path.dirname(__file__),encoding="utf-8")
lines = f.readlines()
for line in lines:
tem_ret = []
p = line.replace("[","").replace("]","").replace("\n","").split("\t")
for i in p:
if i != "":
tem_ret.append(float(i))
ret.append(tem_ret)
return np.asarray(ret)
def readTestLabelsFromFile(distr_same):
ret = []
if distr_same:
f = open(os.path.join(os.path.dirname(__file__),"test_labels_.txt"),"").split(" ")
for i in p:
if i!="":
tem_ret.append(float(i))
ret.append(tem_ret)
return np.asarray(ret)
def getParmsAndLearningRate(agent_no):
f = open(os.path.join(os.path.dirname(__file__),"weights_" + str(agent_no) + ".txt"))
content = f.read()
g_ = content.split("***\n--------------------------------------------------")
parm_local = []
learning_rate_list = []
for j in range(len(g_) - 1):
line = g_[j].split("\n")
if j == 0:
weights_line = line[0:784]
learning_rate_list.append(float(line[784].replace("*","")))
else:
weights_line = line[1:785]
learning_rate_list.append(float(line[785].replace("*","")))
valid_weights_line = []
for l in weights_line:
w_list = l.split("\t")
w_list = w_list[0:len(w_list) - 1]
w_list = [float(i) for i in w_list]
valid_weights_line.append(w_list)
parm_local.append(valid_weights_line)
f.close()
f = open(os.path.join(os.path.dirname(__file__),"bias_" + str(agent_no) + ".txt"))
content = f.read()
g_ = content.split("***\n--------------------------------------------------")
bias_local = []
for j in range(len(g_) - 1):
line = g_[j].split("\n")
if j == 0:
weights_line = line[0]
else:
weights_line = line[1]
b_list = weights_line.split("\t")
b_list = b_list[0:len(b_list) - 1]
b_list = [float(i) for i in b_list]
bias_local.append(b_list)
f.close()
ret = {
'weights': np.asarray(parm_local),'bias': np.asarray(bias_local),'learning_rate': np.asarray(learning_rate_list)
}
return ret
def train_with_gradient_and_valuation(agent_list,grad,bi,lr,distr_type):
f_ini_p = open(os.path.join(os.path.dirname(__file__),"initial_model_parameters.txt"),"r")
para_lines = f_ini_p.readlines()
w_paras = para_lines[0].split("\t")
w_paras = [float(i) for i in w_paras]
b_paras = para_lines[1].split("\t")
b_paras = [float(i) for i in b_paras]
w_initial_g = np.asarray(w_paras,dtype=np.float32).reshape([784,10])
b_initial_g = np.asarray(b_paras,dtype=np.float32).reshape([10])
f_ini_p.close()
model_g = {
'weights': w_initial_g,'bias': b_initial_g
}
for i in range(len(grad[0])):
# i->迭代轮数
gradient_w = np.zeros([784,10],dtype=np.float32)
gradient_b = np.zeros([10],dtype=np.float32)
for j in agent_list:
gradient_w = np.add(np.multiply(grad[j][i],1/len(agent_list)),gradient_w)
gradient_b = np.add(np.multiply(bi[j][i],gradient_b)
model_g['weights'] = np.subtract(model_g['weights'],np.multiply(lr[0][i],gradient_w))
model_g['bias'] = np.subtract(model_g['bias'],gradient_b))
test_images = readTestimagesFromFile(False)
test_labels_onehot = readTestLabelsFromFile(False)
m = np.dot(test_images,np.asarray(model_g['weights']))
test_result = m + np.asarray(model_g['bias'])
y = tf.nn.softmax(test_result)
correct_prediction = tf.equal(tf.argmax(y,1),tf.arg_max(test_labels_onehot,1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction,tf.float32))
return accuracy.numpy()
def remove_list_indexed(removed_ele,original_l,ll):
new_original_l = []
for i in original_l:
new_original_l.append(i)
for i in new_original_l:
if i == removed_ele:
new_original_l.remove(i)
for i in range(len(ll)):
if set(ll[i]) == set(new_original_l):
return i
return -1
def shapley_list_indexed(original_l,ll):
for i in range(len(ll)):
if set(ll[i]) == set(original_l):
return i
return -1
def PowerSetsBinary(items):
N = len(items)
set_all = []
for i in range(2 ** N):
combo = []
for j in range(N):
if (i >> j) % 2 == 1:
combo.append(items[j])
set_all.append(combo)
return set_all
if __name__ == "__main__":
start_time = time.time()
#data_num = np.asarray([5923,6742,5958,6131,5842])
#agents_weights = np.divide(data_num,data_num.sum())
for index in range(NUM_AGENT):
f = open(os.path.join(os.path.dirname(__file__),"weights_"+str(index)+".txt"),"w")
f.close()
f = open(os.path.join(os.path.dirname(__file__),"bias_" + str(index) + ".txt"),"w")
f.close()
mnist_train,mnist_test = tf.keras.datasets.mnist.load_data()
disTRIBUTION_TYPE = "SAME"
federated_train_data_divide = None
federated_train_data = None
if disTRIBUTION_TYPE == "SAME":
federated_train_data_divide = [get_data_for_federated_agents(mnist_train,d) for d in range(NUM_AGENT)]
federated_train_data = federated_train_data_divide
f_ini_p = open(os.path.join(os.path.dirname(__file__),"r")
para_lines = f_ini_p.readlines()
w_paras = para_lines[0].split("\t")
w_paras = [float(i) for i in w_paras]
b_paras = para_lines[1].split("\t")
b_paras = [float(i) for i in b_paras]
w_initial = np.asarray(w_paras,10])
b_initial = np.asarray(b_paras,dtype=np.float32).reshape([10])
f_ini_p.close()
initial_model = collections.OrderedDict(
'weights': w_initial
'bias':b_initial)
model = initial_model
learning_rate = 0.1
for round_num in range(50):
local_models = federated_train(model,federated_train_data)
print("learning rate: ",learning_rate)
#print(local_models[0][0])#第0个agent的weights矩阵
#print(local_models[0][1])#第0个agent的bias矩阵
#print(len(local_models))
for local_index in range(len(local_models)):
f = open(os.path.join(os.path.dirname(__file__),"weights_"+str(local_index)+".txt"),"a",encoding="utf-8")
for i in local_models[local_index][0]:
line = ""
arr = list(i)
for j in arr:
line += (str(j)+"\t")
print(line,file=f)
print("***"+str(learning_rate)+"***",file=f)
print("-"*50,file=f)
f.close()
f = open(os.path.join(os.path.dirname(__file__),"bias_" + str(local_index) + ".txt"),encoding="utf-8")
line = ""
for i in local_models[local_index][1]:
line += (str(i) + "\t")
print(line,file=f)
print("***" + str(learning_rate) + "***",file=f)
f.close()
m_w = np.zeros([784,dtype=np.float32)
m_b = np.zeros([10],dtype=np.float32)
for local_model_index in range(len(local_models)):
m_w = np.add(np.multiply(local_models[local_model_index][0],1/NUM_AGENT),m_w)
m_b = np.add(np.multiply(local_models[local_model_index][1],m_b)
model = {
'weights': m_w,'bias': m_b
}
learning_rate = learning_rate * 0.9
loss = federated_eval(model,federated_train_data)
print('round {},loss={}'.format(round_num,loss))
print(time.time()-start_time)
gradient_weights = []
gradient_biases = []
gradient_lrs = []
for ij in range(NUM_AGENT):
model_ = getParmsAndLearningRate(ij)
gradient_weights_local = []
gradient_biases_local = []
learning_rate_local = []
for i in range(len(model_['learning_rate'])):
if i == 0:
gradient_weight = np.divide(np.subtract(initial_model['weights'],model_['weights'][i]),model_['learning_rate'][i])
gradient_bias = np.divide(np.subtract(initial_model['bias'],model_['bias'][i]),model_['learning_rate'][i])
else:
gradient_weight = np.divide(np.subtract(model_['weights'][i - 1],model_['learning_rate'][i])
gradient_bias = np.divide(np.subtract(model_['bias'][i - 1],model_['learning_rate'][i])
gradient_weights_local.append(gradient_weight)
gradient_biases_local.append(gradient_bias)
learning_rate_local.append(model_['learning_rate'][i])
gradient_weights.append(gradient_weights_local)
gradient_biases.append(gradient_biases_local)
gradient_lrs.append(learning_rate_local)
all_sets = PowerSetsBinary([i for i in range(NUM_AGENT)])
group_shapley_value = []
for s in all_sets:
group_shapley_value.append(
train_with_gradient_and_valuation(s,gradient_weights,gradient_biases,gradient_lrs,disTRIBUTION_TYPE))
print(str(s)+"\t"+str(group_shapley_value[len(group_shapley_value)-1]))
agent_shapley = []
for index in range(NUM_AGENT):
shapley = 0.0
for j in all_sets:
if index in j:
remove_list_index = remove_list_indexed(index,j,all_sets)
if remove_list_index != -1:
shapley += (group_shapley_value[shapley_list_indexed(j,all_sets)] - group_shapley_value[
remove_list_index]) / (comb(NUM_AGENT - 1,len(all_sets[remove_list_index])))
agent_shapley.append(shapley)
for ag_s in agent_shapley:
print(ag_s)
print("end_time",time.time()-start_time)
文件“SameOR-elb.py”,第 352 行,在 local_models = federated_train( 文件“C:\Users\Aw\Anaconda3\lib\site-packages\tensorflow_federated\python\core\impl\utils\function_utils.py”,第 561 行,调用 返回 context.invoke(self,arg) 文件“C:\Users\Aw\Anaconda3\lib\site-packages\retrying.py”,第 49 行,在 包裹_f return retrying(*dargs,**dkw).call(f,*args,**kw) File "C:\Users\Aw\Anaconda3\lib\site-packages\retrying.py",line 206,in 称呼 返回尝试.get(self._wrap_exception)文件“C:\Users\Aw\Anaconda3\lib\site-packages\retrying.py”,第247行,在 得到 6.reraise(self.value[0],self.value[1],self.value[2]) 文件“C:\Users\Aw\Anaconda3\lib\site-packages\six.py”,第 703 行,在加注 提高值文件“C:\Users\Aw\Anaconda3\lib\site-packages\retrying.py”,第 200 行,在 称呼 尝试=尝试(fn(*args,**kwargs),尝试编号,假)文件 "C:\Users\Aw\Anaconda3\lib\site-packages\tensorflow_federated\python\core\impl\executors\execution_context.py",第 217 行,在调用中 return event_loop.run_until_complete( File "C:\Users\Aw\Anaconda3\lib\asyncio\base_events.py",line 616,in 运行直到完成 返回 future.result() 文件“C:\Users\Aw\Anaconda3\lib\site-packages\tensorflow_federated\python\common_libs\tracing.py”, 第 388 行,在 _wrapped 中 返回等待 coro 文件“C:\Users\Aw\Anaconda3\lib\site-packages\tensorflow_federated\python\core\impl\executors\execution_context.py”, 第 123 行,在 _invoke 中 结果 = 等待 executor.create_call(comp,arg) 文件“C:\Users\Aw\Anaconda3\lib\site-packages\tensorflow_federated\python\common_libs\tracing.py”, 第 200 行,在 async_trace 中 结果 = 等待 fn(*fn_args,**fn_kwargs) 文件“C:\Users\Aw\Anaconda3\lib\site-packages\tensorflow_federated\python\core\impl\executors\reference_resolving_executor.py”, 第 343 行,在 create_call 中 返回等待 comp_repr.invoke(self,arg) 文件“C:\Users\Aw\Anaconda3\lib\site-packages\tensorflow_federated\python\core\impl\executors\reference_resolving_executor.py”, 第 155 行,在调用中 return await executor._evaluate(comp_lambda.result,new_scope) # pylint: disable=protected-access File "C:\Users\Aw\Anaconda3\lib\site-packages\tensorflow_federated\python\core\impl\executors\reference_resolving_executor.py",第 513 行,在 _evaluate return await self._evaluate_block(comp,scope) File "C:\Users\Aw\Anaconda3\lib\site-packages\tensorflow_federated\python\common_libs\tracing.py",第 200 行,在 async_trace 中 结果 = 等待 fn(*fn_args,**fn_kwargs) 文件“C:\Users\Aw\Anaconda3\lib\site-packages\tensorflow_federated\python\core\impl\executors\reference_resolving_executor.py”, 第 477 行,在 _evaluate_block 中 value = await self._evaluate(loc.value,scope) 文件“C:\Users\Aw\Anaconda3\lib\site-packages\tensorflow_federated\python\core\impl\executors\reference_resolving_executor.py”, 第 507 行,在 _evaluate return await self._evaluate_call(comp,**fn_kwargs) 文件“C:\Users\Aw\Anaconda3\lib\site-packages\tensorflow_federated\python\core\impl\executors\reference_resolving_executor.py”, 第 446 行,在 _evaluate_call 中 返回等待 self.create_call(func,arg=arg) 文件 "C:\Users\Aw\Anaconda3\lib\site-packages\tensorflow_federated\python\common_libs\tracing.py",**fn_kwargs) 文件“C:\Users\Aw\Anaconda3\lib\site-packages\tensorflow_federated\python\core\impl\executors\reference_resolving_executor.py”, 第 339 行,在 create_call 中 return ReferenceResolvingExecutorValue(await File "C:\Users\Aw\Anaconda3\lib\site-packages\tensorflow_federated\python\core\impl\executors\caching_executor.py",第 281 行,在 create_call 中 target_value = await cached_value.target_future File "C:\Users\Aw\Anaconda3\lib\site-packages\tensorflow_federated\python\common_libs\tracing.py",**fn_kwargs) 文件“C:\Users\Aw\Anaconda3\lib\site-packages\tensorflow_federated\python\core\impl\executors\thread_delegating_executor.py”, 第 120 行,在 create_call 中 返回等待 self._delegate(self._target_executor.create_call(comp,arg)) 文件 "C:\Users\Aw\Anaconda3\lib\site-packages\tensorflow_federated\python\core\impl\executors\thread_delegating_executor.py",第 105 行,在 _delegate 中 result_value = await _delegate_with_trace_ctx(coro,self._event_loop) 文件 "C:\Users\Aw\Anaconda3\lib\site-packages\tensorflow_federated\python\common_libs\tracing.py",第 388 行,在 _wrapped 中 返回等待 coro 文件“C:\Users\Aw\Anaconda3\lib\site-packages\tensorflow_federated\python\common_libs\tracing.py”, 第 200 行,在 async_trace 中 结果 = 等待 fn(*fn_args,**fn_kwargs) 文件“C:\Users\Aw\Anaconda3\lib\site-packages\tensorflow_federated\python\core\impl\executors\federating_executor.py”, 第 445 行,在 create_call 中 return await self._strategy.compute_federated_intrinsic( 文件“C:\Users\Aw\Anaconda3\lib\site-packages\tensorflow_federated\python\core\impl\executors\federating_executor.py”,第 139 行,在 compute_federated_intrinsic 中 return await fn(arg) # pylint: disable=not-callable File "C:\Users\Aw\Anaconda3\lib\site-packages\tensorflow_federated\python\common_libs\tracing.py",**fn_kwargs) 文件“C:\Users\Aw\Anaconda3\lib\site-packages\tensorflow_federated\python\core\impl\executors\federated_resolving_strategy.py”, 第 453 行,在 compute_federated_map 中 return await self._map(arg,all_equal=False) File "C:\Users\Aw\Anaconda3\lib\site-packages\tensorflow_federated\python\common_libs\tracing.py",第 200 行,在 async_trace 中 结果 = 等待 fn(fn_args,**fn_kwargs) 文件“C:\Users\Aw\Anaconda3\lib\site-packages\tensorflow_federated\python\core\impl\executors\federated_resolving_strategy.py”, 第 320 行,在 _map 中 结果 = await asyncio.gather([ 文件 "C:\Users\Aw\Anaconda3\lib\site-packages\tensorflow_federated\python\common_libs\tracing.py",**fn_kwargs) 文件“C:\Users\Aw\Anaconda3\lib\site-packages\tensorflow_federated\python\core\impl\executors\reference_resolving_executor.py”, 第 445 行,在 _evaluate_call 中 func,arg = await asyncio.gather(func,get_arg()) 文件“C:\Users\Aw\Anaconda3\lib\site-packages\tensorflow_federated\python\core\impl\executors\reference_resolving_executor.py”, 第 501 行,在 _evaluate return await self._evaluate_to_delegate(comp,**fn_kwargs) 文件“C:\Users\Aw\Anaconda3\lib\site-packages\tensorflow_federated\python\core\impl\executors\reference_resolving_executor.py”, 第 410 行,在 _evaluate_to_delegate 中 等待 self._target_executor.create_value( 文件 "C:\Users\Aw\Anaconda3\lib\site-packages\tensorflow_federated\python\core\impl\executors\caching_executor.py",第 245 行,在 create_value 中 await cached_value.target_future File "C:\Users\Aw\Anaconda3\lib\site-packages\tensorflow_federated\python\common_libs\tracing.py",**fn_kwargs) 文件“C:\Users\Aw\Anaconda3\lib\site-packages\tensorflow_federated\python\core\impl\executors\thread_delegating_executor.py”, 第 110 行,在 create_value 中 return await self._delegate( File "C:\Users\Aw\Anaconda3\lib\site-packages\tensorflow_federated\python\core\impl\executors\thread_delegating_executor.py",**fn_kwargs) 文件“C:\Users\Aw\Anaconda3\lib\site-packages\tensorflow_federated\python\core\impl\executors\eager_tf_executor.py”, 第 464 行,在 create_value 中 返回 EagerValue(value,self._tf_function_cache,type_spec,self._device) 文件 "C:\Users\Aw\Anaconda3\lib\site-packages\tensorflow_federated\python\core\impl\executors\eager_tf_executor.py",第 366 行,在 init 中 self._value = to_representation_for_type(value,tf_function_cache,File "C:\Users\Aw\Anaconda3\lib\site-packages\tensorflow_federated\python\core\impl\executors\eager_tf_executor.py",第 287 行,在 to_representation_for_type 中 Embedded_fn = embed_tensorflow_computation(value,device) 文件 "C:\Users\Aw\Anaconda3\lib\site-packages\tensorflow_federated\python\core\impl\executors\eager_tf_executor.py",第 153 行,在 embed_tensorflow_computation 中 raise TypeError('期望一个 TensorFlow 计算,找到 {}.'.format( TypeError: 期望一个 TensorFlow 计算,找到 固有的。 我遇到了这些错误..需要建议..
我使用的是 tf 2.2.1
Python 3.8.3 版本
解决方法
使用 sequence_reduce
可能会导致此问题; TFF 的高性能堆栈尚不支持它。
如果性能不是很关键,我们可以通过在调用我们的计算之前安装参考执行器来立即解决这个问题:
tff.backends.reference.set_reference_context()
如tutorial which demonstrates sequence_reduce
中所述。
然而,这是一个残酷的错误。您介意检查一下您使用的 TFF 版本吗?如果您不在 0.17.0
上,我们可能已经在此处生成了更好的错误消息。如果是,您会介意 filing a GitHub issue 吐出 TFF 错误吗?