def run(self, evals, feed_dict=None, breakpoints=None, break_immediately=False):
"""
starts the debug session
"""
if not isinstance(evals,list):
evals=[evals]
if feed_dict is None:
feed_dict={}
if breakpoints is None:
breakpoints=[]
self.state=RUNNING
self._original_evals=evals
self._original_feed_dict=feed_dict
self._exe_order=op_store.compute_exe_order(evals)
self._init_evals_bps(evals, breakpoints)
# convert cache keys to strings
for k,v in feed_dict.items():
if not isinstance(k,str):
k=k.name
self._cache[k]=v
op_store.register_dbsession(self)
if break_immediately:
return self._break()
else:
return self.c()
def run(self, evals, feed_dict=None, breakpoints=None, break_immediately=False):
"""
starts the debug session
"""
if not isinstance(evals, list):
evals = [evals]
if feed_dict is None:
feed_dict = {}
if breakpoints is None:
breakpoints = []
self.state = RUNNING
self._original_evals = evals
self._original_feed_dict = feed_dict
self._exe_order = op_store.compute_exe_order(evals)
self._init_evals_bps(evals, breakpoints)
# convert cache keys to strings
for k, v in feed_dict.items():
if not isinstance(k, str):
k = k.name
self._cache[k] = v
op_store.register_dbsession(self)
if break_immediately:
return self._break()
else:
return self.c()
def _init_evals_bps(self, evals, breakpoints): # If an eval or bp is the tf.Placeholder output of a tdb.PythonOp, replace it with its respective PythonOp node evals2=[op_store.get_op(t) if op_store.is_htop_out(t) else t for t in evals] breakpoints2=[op_store.get_op(t) if op_store.is_htop_out(t) else t for t in breakpoints] # compute execution order self._exe_order=op_store.compute_exe_order(evals2) # list of nodes # compute evaluation set """ HTOps may depend on tf.Tensors that are not in eval. We need to have all inputs to HTOps ready upon evaluation. 1. all evals that were originally specified are added 2. each HTOp in the execution closure needs to be in eval (they won't be eval'ed automatically by Session.run) 3. if an input to an HTOp is a tf.Tensor (not a HT placeholder tensor), it needs to be in eval as well (it's not tensorflow so we'll have to manually evaluate it). Remember, we don't track Placeholders because we instead run the HTOps that generate their values. """ self._evalset=set([e.name for e in evals2]) for e in self._exe_order: if isinstance(e,HTOp): self._evalset.add(e.name) for t in e.inputs: if not op_store.is_htop_out(t): self._evalset.add(t.name) # compute breakpoint set self._bpset=set([bp.name for bp in breakpoints2])
def _init_evals_bps(self, evals, breakpoints):
# If an eval or bp is the tf.Placeholder output of a tdb.PythonOp, replace it with its respective PythonOp node
evals2 = [op_store.get_op(t) if op_store.is_htop_out(t) else t for t in evals]
breakpoints2 = [op_store.get_op(t) if op_store.is_htop_out(t) else t for t in breakpoints]
# compute execution order
self._exe_order = op_store.compute_exe_order(evals2) # list of nodes
# compute evaluation set
"""
HTOps may depend on tf.Tensors that are not in eval. We need to have all inputs to HTOps ready
upon evaluation.
1. all evals that were originally specified are added
2. each HTOp in the execution closure needs to be in eval (they won't be eval'ed automatically by Session.run)
3. if an input to an HTOp is a tf.Tensor (not a HT placeholder tensor), it needs to be in eval as well (it's not
tensorflow so we'll have to manually evaluate it). Remember, we don't track Placeholders because we instead
run the HTOps that generate their values.
"""
self._evalset = set([e.name for e in evals2])
for e in self._exe_order:
if isinstance(e, HTOp):
self._evalset.add(e.name)
for t in e.inputs:
if not op_store.is_htop_out(t):
self._evalset.add(t.name)
# compute breakpoint set
self._bpset = set([bp.name for bp in breakpoints2])
