Training container for NAS Envelopenet #429
Conversation
| @@ -0,0 +1,238 @@ | |||
| from __future__ import absolute_import | |||
| from __future__ import division | |||
| from __future__ import print_function | |||
There was a problem hiding this comment.
You probably don't need future imports since you are using python3 anyways. Same goes for every import like this.
|
|
||
| def cell(self, inputs, arch, is_training): | ||
| """Create the cell by instantiating the cell blocks""" | ||
| nscope = 'Cell_' + self.cellname + '_' + str(self.cellidx) |
| By default use stride=1 and SAME padding | ||
| """ | ||
| dropout_keep_prob = 0.8 | ||
| nscope = 'Cell_' + self.cellname + '_' + str(self.cellidx) |
There was a problem hiding this comment.
.format (will be no need to use str())
| net = tf.Print( | ||
| net, | ||
| [msss], | ||
| message="MeanSSS=:" + |
| self.batch_size = self.task_config["batch_size"] | ||
| self.num_examples = 10000 | ||
| self.run_once = True | ||
| self.eval_dir = self.task_config["data_dir"] + "/results/" + \ |
There was a problem hiding this comment.
os.path.join() instead to take care of "/"
|
|
||
| # Build a Graph that computes the logits predictions from the | ||
| # inference model. | ||
| # TODO: Clean up all args |
There was a problem hiding this comment.
Get rid of TODOs since I can't see a WIP status.
| self.get_params(params) | ||
|
|
||
| def get_params(self, params): | ||
| global global_batch_size |
There was a problem hiding this comment.
Should they be global and not class attributes?
| filepath = os.path.join(dest_directory, filename) | ||
| if not os.path.exists(filepath): | ||
| def _progress(count, block_size, total_size): | ||
| sys.stdout.write('\r>> Downloading %s %.1f%%' % (filename, |
There was a problem hiding this comment.
Do you really need a progressbar?
| self._step = self.global_step_init | ||
| self._start_time = time.time() | ||
|
|
||
| def before_run(self, run_context): |
There was a problem hiding this comment.
run_context is never used. function is actually never called
| from __future__ import print_function | ||
|
|
||
| from datetime import datetime | ||
| import ast |
| with tf.variable_scope(nscope, 'initial_block', [inputs], reuse=reuse) as scope: | ||
| with slim.arg_scope([slim.conv2d, slim.max_pool2d], stride=1, padding='SAME'): | ||
| net = inputs | ||
| layeridx = 0 |
There was a problem hiding this comment.
Do you need this? As fas as I can see you only increment it.
| outputs): | ||
| self.cellidx = cellidx | ||
| self.log_stats = log_stats | ||
| self.res=sys.argv[2] |
| self.log_stats = log_stats | ||
| self.res=sys.argv[2] | ||
| self.cellname = "Envelope" | ||
| self.numbranches = 4 |
| self.numbranches = 4 | ||
| self.numbins = 100 | ||
| self.batchsize = int(net.shape[0]) | ||
| numfilters = len(filters) |
| scope) | ||
| return softmax_linear | ||
|
|
||
| def maybe_download_and_extract(self): |
There was a problem hiding this comment.
Somehow used? I see the acutla line calling it is commented out.
| is_training=True, | ||
| scope='Nacnet'): | ||
| net = self.add_init(inputs, initcell, is_training) | ||
| end_points = {} |
There was a problem hiding this comment.
? not used? since the line using it is commented out
| self.global_step_init = global_step_init | ||
| self.loss = loss | ||
|
|
||
| def begin(self): |
| # Asks for loss value. | ||
| return tf.train.SessionRunArgs(self.loss) | ||
|
|
||
| def after_run(self, run_context, run_values): |
There was a problem hiding this comment.
function is never called
There was a problem hiding this comment.
these are required in loggerhook call during training
| mon_sess.run(train_op) | ||
|
|
||
| def evaluate(): | ||
| eval=Evaluate(self.arch, self.params, self.train_dir) |
There was a problem hiding this comment.
it's called from run_trail.py
|
@richardsliu @Akado2009 plz merge |
|
Is the suggestion trained for one experiment (or search space)? Or is it trained once, and can be used for all experiments? |
|
@gaocegege It is trained for every experiment. As there is no controller like in RL, the only training required is of the architectures being sampled, and that too is truncated,i.e., 10 epochs. |
|
Now our suggestion is long-running, thus I am not sure if we could support the case. But the suggestion itself LGTM. Thanks for your contribution |
|
@gaocegege In fact it's the fastest NAS Algo, even faster than RL algo which is there itself in katib. If you change the 'steps' parameter to 10 in yaml file, it will complete in ~15 min. Also, you might want to run it on GPU. |
|
@richardsliu done |
|
@richardsliu Plz merge |
|
/lgtm |
|
[APPROVALNOTIFIER] This PR is APPROVED This pull-request has been approved by: richardsliu The full list of commands accepted by this bot can be found here. The pull request process is described here
Needs approval from an approver in each of these files:
Approvers can indicate their approval by writing |
|
/retest |
This PR has all files for the training and evaluation required in NAS Envelopenet Suggestion service.
This change is