-
Notifications
You must be signed in to change notification settings - Fork 5
/
Copy pathmetaflow_train.py
191 lines (164 loc) · 7.15 KB
/
metaflow_train.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
from metaflow import FlowSpec, step,retry
import json
class FinalData():
def __init__(self):
self.model = None
self.optimizer = None
self.agent_name = None
self.loss = None
self.simulation_analytics = None
self.total_data_size = 0
self.model_args = {}
self.gradients = {
'max':[],
'avg':[],
'layer':[]
}
def __str__(self):
return self._get_parsed_data('str')
def to_json(self):
return self._get_parsed_data('json')
def _get_parsed_data(self,type_var='str'):
"""
Can parse to string for printing or to json for conversion
"""
num_convergence_metrics = len(self.simulation_analytics['convergence_metrics'])
percent_converge = (len(self.simulation_analytics['convergence_metrics']) / self.simulation_analytics['total_epochs_allowed'])*100
data_size = 0
if hasattr(self,'total_data_size'):
data_size = self.total_data_size
model_args = {}
if hasattr(self,'model_args'):
model_args = self.model_args
collected_grads = "No"
if hasattr(self,'gradients'):
if len(self.gradients['avg']) > 0:
collected_grads = 'Yes'
arg_dict = dict(agent_name=self.agent_name,\
total_episodes=str(self.simulation_analytics['total_epochs_allowed']),\
steps=str(self.simulation_analytics['max_steps_per_episode']),\
num_convergence_metrics=str(num_convergence_metrics), \
percent_converge=str(percent_converge), \
model_args=json.dumps(model_args), \
data_size="NO DATA" if data_size is 0 else str(data_size),
grad_collect=collected_grads)
if type_var == 'str':
x = '''
Agent Name : {agent_name}
Total Training Data Size : {data_size}
Model Arguements : {model_args}
Collected Gradients : {grad_collect}
Simulation Results
Total Number of Episodes : {total_episodes}
Steps Per Episode : {steps}
Number Of Converged Cases : {num_convergence_metrics}
% Cases That Converged : {percent_converge}
'''.format(**arg_dict)
return x
elif type_var == 'json':
return arg_dict
else:
raise Exception("Not Supported Parsing Type")
class TrainingSimulatorFlow(FlowSpec):
@step
def start(self):
print("Importing data in this step")
self.num_demos=49
self.num_epochs=5 # Training epochs
self.episode_length=100
self.num_episodes=2 # Simulated Testing Epochs.
self.variation_number = 0
self.collect_gradients = True
self.agent_modules = [{
'module_name':'models.SmartImmitationAgent',
'agent_name':'SimpleImmitationLearningAgent',
'args':{'num_layers':4},
'reporting_name':'SimpleImmitationLearningAgent__4'
},
{
'module_name':'models.ImmitationLearning',
'agent_name':'ImmitationLearningAgent',
'args':{},
'reporting_name':'Idiot_ImmitationLearningAgent'
},
{
'module_name':'models.ImmitationMutationConv',
'agent_name':'ImmitationLearningConvolvingMutantAgent',
'args':{},
'reporting_name':'ImmitationLearningConvolvingMutantAgent'
},
{
'module_name':'models.SmartImmitationAgent',
'agent_name':'SimpleImmitationLearningAgent',
'args':{'num_layers':1},
'reporting_name':'SimpleImmitationLearningAgent__1'
},
{
'module_name':'models.SmartImmitationAgent',
'agent_name':'SimpleImmitationLearningAgent',
'args':{'num_layers':2},
'reporting_name':'SimpleImmitationLearningAgent__2'
},
{
'module_name':'models.ImmitationMutant',
'agent_name': 'ImmitationLearningMutantAgent',
'args':{},
'reporting_name':'ImmitationLearningMutantAgent'
}]
self.next(self.train,foreach='agent_modules')
@retry(times=4)
@step
def train(self):
# todo : Create an image for training
from SimulationEnvironment.Environment import ReachTargetSimulationEnv
import importlib
agent_module = importlib.import_module(self.input['module_name'])
agent = getattr(agent_module,self.input['agent_name'])(**self.input['args'],collect_gradients=self.collect_gradients)
curr_env = ReachTargetSimulationEnv(dataset_root='/tmp/rlbench_data')
curr_env.task._variation_number = self.variation_number
# Set image_paths_output=True when loading dataset from file if images also dont need to be loaded for dataset
demos = curr_env.get_demos(self.num_demos,live_demos=False,image_paths_output=False)
# agent.load_model('SavedModels/ImmitationLearningConvolvingMutantAgent-2020-04-05-04-18.pt')
agent.injest_demonstrations(demos)
loss = agent.train_agent(self.num_epochs)
self.loss = loss
self.total_data_size = agent.total_train_size
self.agent_name = self.input['reporting_name']
self.model = agent.neural_network.state_dict()
self.model_args = self.input['args']
self.optimizer = agent.optimizer.state_dict()
self.gradients = agent.gradients
self.next(self.simulate)
@retry(times=4)
@step
def simulate(self):
from SimulationEnvironment.Environment import ReachTargetSimulationEnv
import importlib
agent_module = importlib.import_module(self.input['module_name'])
agent = getattr(agent_module,self.input['agent_name'])(**self.input['args'])
curr_env = ReachTargetSimulationEnv(headless=True,episode_length=self.episode_length,num_episodes=self.num_episodes)
agent.load_model_from_object(self.model)
simulation_analytics = curr_env.run_trained_agent(agent)
self.simulation_analytics = simulation_analytics
self.next(self.join)
@step
def join(self,inputs):
final_data = []
for task_data in inputs:
data = FinalData()
data.model = task_data.model
data.optimizer = task_data.optimizer
data.agent_name = task_data.agent_name
data.loss = task_data.loss
data.simulation_analytics = task_data.simulation_analytics
data.total_data_size = task_data.total_data_size
data.model_args = task_data.model_args
data.gradients = task_data.gradients
final_data.append(data)
self.final_data = final_data
self.next(self.end)
@step
def end(self):
print("Done Computation")
if __name__ == '__main__':
TrainingSimulatorFlow()