def save_deconv(self, k, f, d, b):
# deconvされた入力層に平均画像を足して画像化
img = ioutil.deprocess(d, self.mean)
# 最大値の位置の計算に必要な入力層の領域だけクロッピングして保存
img.crop((b[0], b[2], b[1], b[3])).save(
os.path.join(self.deconv_image_dir,
"{0:0>4}_{1:0>2}.png".format(f, k)))
def evaluate(self):
"""override method of extensions.Evaluator."""
iterator = self._iterators['main']
target = self._targets['main']
eval_func = self.eval_func or target
if self.eval_hook:
self.eval_hook(self)
it = copy.copy(iterator)
summary = reporter_module.DictSummary()
max_locs = []
bounds = []
n_processed = 0
filter_idx = 0
pbar = tqdm(total=len(iterator.dataset))
for batch in it:
observation = {}
with reporter_module.report_scope(observation):
in_arrays = self.converter(batch, self.device)
if isinstance(in_arrays, tuple):
in_vars = tuple(
variable.Variable(x, volatile='off')
for x in in_arrays)
eval_func(*in_vars)
elif isinstance(in_arrays, dict):
in_vars = {
key: variable.Variable(x, volatile='off')
for key, x in six.iteritems(in_arrays)
}
eval_func(**in_vars)
else:
in_var = variable.Variable(in_arrays, volatile='off')
eval_func(in_var)
pbar.update(len(batch))
indices = np.arange(filter_idx,
filter_idx + len(batch)) % self.n_features
# deconv対象の層のVariableを取得
layer_variable = Vutil.get_variable(observation[self.lastname],
self.layer_rank)
# 最大値の位置の計算に必要な入力層の領域を取得
isfc = Vutil.has_fc_layer(layer_variable)
if isfc:
batch_bounds = Vutil.get_data_bounds(layer_variable)
else:
batch_bounds = Vutil.get_max_bounds(layer_variable, indices)
topk = np.arange(n_processed,
n_processed + len(batch)) // self.n_features
input_data = Vutil.get_data_layer(layer_variable).data
for k, f, d, b in zip(topk, indices, input_data, batch_bounds):
#print(dir(d))
# deconvされた入力層に平均画像を足して画像化
img = ioutil.deprocess(d.get(), self.mean)
# 最大値の位置の計算に必要な入力層の領域だけクロッピングして保存
img.crop((b[0], b[2], b[1], b[3])).save(
os.path.join(self.patch_image_dir,
"{0:0>4}_{1:0>2}.png".format(f, k)))
if not isfc:
max_locs.extend(Vutil.get_max_locs(layer_variable, indices))
bounds.extend(batch_bounds)
filter_idx = (filter_idx + len(batch)) % self.n_features
n_processed += len(batch)
#self.add_to_confmat(self.confmat, in_vars[1].data, self.getpred(observation[self.lastname]))
summary.add(observation)
pbar.close()
#print(self.confmat)
#print(np.diag(self.confmat))
#print(1.0 * np.diag(self.confmat).sum() / self.confmat.sum())
return summary.compute_mean(), max_locs, bounds
def evaluate(self):
"""override method of extensions.Evaluator."""
iterator = self._iterators['main']
target = self._targets['main']
eval_func = self.eval_func or target
if self.eval_hook:
self.eval_hook(self)
it = copy.copy(iterator)
summary = reporter_module.DictSummary()
max_locs = []
bounds = []
n_processed = 0
filter_idx = 0
pbar = tqdm(total=len(iterator.dataset))
for batch in it:
observation = {}
with reporter_module.report_scope(observation):
in_arrays = self.converter(batch, self.device)
if isinstance(in_arrays, tuple):
in_vars = tuple(variable.Variable(x, volatile='off')
for x in in_arrays)
eval_func(*in_vars)
elif isinstance(in_arrays, dict):
in_vars = {key: variable.Variable(x, volatile='off')
for key, x in six.iteritems(in_arrays)}
eval_func(**in_vars)
else:
in_var = variable.Variable(in_arrays, volatile='off')
eval_func(in_var)
pbar.update(len(batch))
indices = np.arange(filter_idx, filter_idx + len(batch)) % self.n_features
# deconv対象の層のVariableを取得
layer_variable = Vutil.get_variable(
observation[self.lastname], self.layer_rank)
# 最大値の位置の計算に必要な入力層の領域を取得
isfc = Vutil.has_fc_layer(layer_variable)
if isfc:
batch_bounds = Vutil.get_data_bounds(layer_variable)
else:
batch_bounds = Vutil.get_max_bounds(layer_variable, indices)
topk = np.arange(n_processed, n_processed + len(batch)) // self.n_features
input_data = Vutil.get_data_layer(layer_variable).data
for k, f, d, b in zip(topk, indices, input_data, batch_bounds):
#print(dir(d))
# deconvされた入力層に平均画像を足して画像化
img = ioutil.deprocess(d.get(), self.mean)
# 最大値の位置の計算に必要な入力層の領域だけクロッピングして保存
img.crop((b[0], b[2], b[1], b[3])).save(
os.path.join(self.patch_image_dir,
"{0:0>4}_{1:0>2}.png".format(f, k)))
if not isfc:
max_locs.extend(Vutil.get_max_locs(layer_variable, indices))
bounds.extend(batch_bounds)
filter_idx = (filter_idx + len(batch)) % self.n_features
n_processed += len(batch)
#self.add_to_confmat(self.confmat, in_vars[1].data, self.getpred(observation[self.lastname]))
summary.add(observation)
pbar.close()
#print(self.confmat)
#print(np.diag(self.confmat))
#print(1.0 * np.diag(self.confmat).sum() / self.confmat.sum())
return summary.compute_mean(), max_locs, bounds
def __call__(self, trainer):
fn = self.filename.format(trainer)
img = ioutil.deprocess(
cuda.to_cpu(self.target[self.paramname].W.data)[0], self.mean)
img.save(os.path.join(trainer.out, fn))
'''prefix = 'tmp' + fn
def __call__(self, trainer):
fn = self.filename.format(trainer)
img = ioutil.deprocess(cuda.to_cpu(self.target[self.paramname].W.data)[0], self.mean)
img.save(os.path.join(trainer.out, fn))
'''prefix = 'tmp' + fn
def evaluate(self):
"""override method of extensions.Evaluator."""
iterator = self._iterators['main']
target = self._targets['main']
eval_func = self.eval_func or target
if self.eval_hook:
self.eval_hook(self)
it = copy.copy(iterator)
summary = reporter_module.DictSummary()
max_locs = []
bounds = []
n_processed = 0
for batch in it:
observation = {}
with reporter_module.report_scope(observation):
in_arrays = self.converter(batch, self.device)
if isinstance(in_arrays, tuple):
in_vars = tuple(variable.Variable(x, volatile='off')
for x in in_arrays)
eval_func(*in_vars)
elif isinstance(in_arrays, dict):
in_vars = {key: variable.Variable(x, volatile='off')
for key, x in six.iteritems(in_arrays)}
eval_func(**in_vars)
else:
in_var = variable.Variable(in_arrays, volatile='off')
eval_func(in_var)
# deconv対象の層のVariableを取得
layer_variable = Vutil.get_variable(
observation[self.lastname], self.layer_rank)
pred = self.getpred(layer_variable)
truth = in_vars[1].data
xp = cuda.get_array_module(pred)
if xp == cupy:
pred = pred.get()
truth = truth.get()
indices = pred.flatten() if self.target == 'pred' else truth.flatten()
# 最大値の位置の計算に必要な入力層の領域を取得
isfc = Vutil.has_fc_layer(layer_variable)
if isfc:
bounds = Vutil.get_data_bounds(layer_variable)
else:
bounds = Vutil.get_max_bounds(layer_variable, indices)
# deconvを実行
deconv_data = self.get_deconv(
layer_variable, indices)
topk = np.arange(n_processed, n_processed + len(batch)) // self.n_features
for i, (t, p, d, b) in enumerate(zip(truth, indices, deconv_data, bounds)):
#print(dir(d))
# deconvされた入力層に平均画像を足して画像化
img = ioutil.deprocess(d.get(), self.mean)
# 最大値の位置の計算に必要な入力層の領域だけクロッピングして保存
img.crop((b[0], b[2], b[1], b[3])).save(
os.path.join(self.deconv_image_dir,
"{0:0>6}_{1:0>4}_{2:0>4}.png".format(n_processed + i, t, p)))
'''max_locs.extend(self.get_max_locs(
observation[self.lastname], self.layer_rank, indices))
bounds.extend(self.get_max_patch_bounds(
observation[self.lastname], self.layer_rank, indices))'''
n_processed += len(batch)
#self.add_to_confmat(self.confmat, in_vars[1].data, self.getpred(observation[self.lastname]))
summary.add(observation)
#print(self.confmat)
#print(np.diag(self.confmat))
#print(1.0 * np.diag(self.confmat).sum() / self.confmat.sum())
return summary.compute_mean(), max_locs, bounds