def _read_latest_config_files(self, run_path_pairs):
"""Reads and returns the projector config files in every run
directory."""
"""If no specific config exists, use the default config provided in
the root directory."""
default_config_fpath = os.path.join(self.logdir,
metadata.PROJECTOR_FILENAME)
default_config = ProjectorConfig()
if tf.io.gfile.exists(default_config_fpath):
with tf.io.gfile.GFile(default_config_fpath, "r") as f:
file_content = f.read()
text_format.Merge(file_content, default_config)
# Relative metadata paths do not work with subdirs, so convert
# any metadata paths to absolute paths.
for embedding in default_config.embeddings:
embedding.metadata_path = _rel_to_abs_asset_path(
embedding.metadata_path, default_config_fpath)
configs = {}
config_fpaths = {}
for run_name, assets_dir in run_path_pairs:
config = ProjectorConfig()
config_fpath = os.path.join(assets_dir,
metadata.PROJECTOR_FILENAME)
if tf.io.gfile.exists(config_fpath):
with tf.io.gfile.GFile(config_fpath, "r") as f:
file_content = f.read()
text_format.Merge(file_content, config)
elif tf.io.gfile.exists(default_config_fpath):
config = default_config
has_tensor_files = False
for embedding in config.embeddings:
if embedding.tensor_path:
if not embedding.tensor_name:
embedding.tensor_name = os.path.basename(
embedding.tensor_path)
has_tensor_files = True
break
if not config.model_checkpoint_path:
# See if you can find a checkpoint file in the logdir.
logdir = _assets_dir_to_logdir(assets_dir)
ckpt_path = _find_latest_checkpoint(logdir)
if not ckpt_path and not has_tensor_files:
continue
if ckpt_path:
config.model_checkpoint_path = ckpt_path
# Sanity check for the checkpoint file existing.
if (config.model_checkpoint_path and _using_tf() and
not tf.io.gfile.glob(config.model_checkpoint_path + "*")):
logger.warn(
'Checkpoint file "%s" not found',
config.model_checkpoint_path,
)
continue
configs[run_name] = config
config_fpaths[run_name] = config_fpath
return configs, config_fpaths
def _read_latest_config_files(self, run_path_pairs):
"""Reads and returns the projector config files in every run
directory."""
configs = {}
config_fpaths = {}
for run_name, assets_dir in run_path_pairs:
config = ProjectorConfig()
config_fpath = os.path.join(assets_dir, metadata.PROJECTOR_FILENAME)
if tf.io.gfile.exists(config_fpath):
with tf.io.gfile.GFile(config_fpath, "r") as f:
file_content = f.read()
text_format.Merge(file_content, config)
has_tensor_files = False
for embedding in config.embeddings:
if embedding.tensor_path:
if not embedding.tensor_name:
embedding.tensor_name = os.path.basename(
embedding.tensor_path
)
has_tensor_files = True
break
if not config.model_checkpoint_path:
# See if you can find a checkpoint file in the logdir.
logdir = _assets_dir_to_logdir(assets_dir)
ckpt_path = _find_latest_checkpoint(logdir)
if not ckpt_path and not has_tensor_files:
continue
if ckpt_path:
config.model_checkpoint_path = ckpt_path
# Sanity check for the checkpoint file existing.
if (
config.model_checkpoint_path
and _using_tf()
and not tf.io.gfile.glob(config.model_checkpoint_path + "*")
):
logger.warn(
'Checkpoint file "%s" not found',
config.model_checkpoint_path,
)
continue
configs[run_name] = config
config_fpaths[run_name] = config_fpath
return configs, config_fpaths
def add_embedding(self,
mat,
metadata=None,
label_img=None,
global_step=None,
tag='default',
metadata_header=None):
torch._C._log_api_usage_once("tensorboard.logging.add_embedding")
mat = make_np(mat)
if global_step is None:
global_step = 0
# clear pbtxt?
# Maybe we should encode the tag so slashes don't trip us up?
# I don't think this will mess us up, but better safe than sorry.
subdir = "%s/%s" % (str(global_step).zfill(5), self._encode(tag))
save_path = os.path.join(self._get_file_writer().get_logdir(), subdir)
fs = tf.io.gfile.get_filesystem(save_path)
if fs.exists(save_path):
if fs.isdir(save_path):
print(
'warning: Embedding dir exists, did you set global_step for add_embedding()?'
)
else:
raise Exception(
"Path: `%s` exists, but is a file. Cannot proceed." %
save_path)
else:
fs.makedirs(save_path)
if metadata is not None:
assert mat.shape[0] == len(
metadata), '#labels should equal with #data points'
make_tsv(metadata, save_path, metadata_header=metadata_header)
if label_img is not None:
assert mat.shape[0] == label_img.shape[
0], '#images should equal with #data points'
make_sprite(label_img, save_path)
assert mat.ndim == 2, 'mat should be 2D, where mat.size(0) is the number of data points'
make_mat(mat, save_path)
# Filesystem doesn't necessarily have append semantics, so we store an
# internal buffer to append to and re-write whole file after each
# embedding is added
if not hasattr(self, "_projector_config"):
self._projector_config = ProjectorConfig()
embedding_info = get_embedding_info(metadata, label_img, fs, subdir,
global_step, tag)
self._projector_config.embeddings.extend([embedding_info])
from google.protobuf import text_format
config_pbtxt = text_format.MessageToString(self._projector_config)
write_pbtxt(self._get_file_writer().get_logdir(), config_pbtxt)
def _read_latest_config_files(self, run_path_pairs):
"""Reads and returns the projector config files in every run directory."""
configs = {}
config_fpaths = {}
for run_name, assets_dir in run_path_pairs:
config = ProjectorConfig()
config_fpath = os.path.join(assets_dir, PROJECTOR_FILENAME)
if tf.gfile.Exists(config_fpath):
with tf.gfile.GFile(config_fpath, 'r') as f:
file_content = f.read()
text_format.Merge(file_content, config)
has_tensor_files = False
for embedding in config.embeddings:
if embedding.tensor_path:
if not embedding.tensor_name:
embedding.tensor_name = os.path.basename(embedding.tensor_path)
has_tensor_files = True
break
if not config.model_checkpoint_path:
# See if you can find a checkpoint file in the logdir.
logdir = _assets_dir_to_logdir(assets_dir)
ckpt_path = _find_latest_checkpoint(logdir)
if not ckpt_path and not has_tensor_files:
continue
if ckpt_path:
config.model_checkpoint_path = ckpt_path
# Sanity check for the checkpoint file.
if (config.model_checkpoint_path and
not tf.train.checkpoint_exists(config.model_checkpoint_path)):
tf.logging.warning('Checkpoint file "%s" not found',
config.model_checkpoint_path)
continue
configs[run_name] = config
config_fpaths[run_name] = config_fpath
return configs, config_fpaths
def _latest_checkpoints_changed(configs, run_path_pairs):
"""Returns true if the latest checkpoint has changed in any of the runs."""
for run_name, assets_dir in run_path_pairs:
if run_name not in configs:
config = ProjectorConfig()
config_fpath = os.path.join(assets_dir, PROJECTOR_FILENAME)
if tf.io.gfile.exists(config_fpath):
with tf.io.gfile.GFile(config_fpath, 'r') as f:
file_content = f.read()
text_format.Merge(file_content, config)
else:
config = configs[run_name]
# See if you can find a checkpoint file in the logdir.
logdir = _assets_dir_to_logdir(assets_dir)
ckpt_path = _find_latest_checkpoint(logdir)
if not ckpt_path:
continue
if config.model_checkpoint_path != ckpt_path:
return True
return False
def add_embedding(self,
mat,
metadata=None,
label_img=None,
global_step=None,
tag='default',
metadata_header=None):
"""Add embedding projector data to summary.
Args:
mat (torch.Tensor or numpy.array): A matrix which each row is the feature vector of the data point
metadata (list): A list of labels, each element will be convert to string
label_img (torch.Tensor): Images correspond to each data point
global_step (int): Global step value to record
tag (string): Name for the embedding
Shape:
mat: :math:`(N, D)`, where N is number of data and D is feature dimension
label_img: :math:`(N, C, H, W)`
Examples::
import keyword
import torch
meta = []
while len(meta)<100:
meta = meta+keyword.kwlist # get some strings
meta = meta[:100]
for i, v in enumerate(meta):
meta[i] = v+str(i)
label_img = torch.rand(100, 3, 10, 32)
for i in range(100):
label_img[i]*=i/100.0
writer.add_embedding(torch.randn(100, 5), metadata=meta, label_img=label_img)
writer.add_embedding(torch.randn(100, 5), label_img=label_img)
writer.add_embedding(torch.randn(100, 5), metadata=meta)
"""
torch._C._log_api_usage_once("tensorboard.logging.add_embedding")
mat = make_np(mat)
if global_step is None:
global_step = 0
# clear pbtxt?
# Maybe we should encode the tag so slashes don't trip us up?
# I don't think this will mess us up, but better safe than sorry.
subdir = "%s/%s" % (str(global_step).zfill(5), self._encode(tag))
save_path = os.path.join(self._get_file_writer().get_logdir(), subdir)
fs = tf.io.gfile.get_filesystem(save_path)
if fs.exists(save_path):
if fs.isdir(save_path):
print(
'warning: Embedding dir exists, did you set global_step for add_embedding()?'
)
else:
raise Exception(
"Path: `%s` exists, but is a file. Cannot proceed." %
save_path)
else:
fs.makedirs(save_path)
if metadata is not None:
assert mat.shape[0] == len(
metadata), '#labels should equal with #data points'
make_tsv(metadata, save_path, metadata_header=metadata_header)
if label_img is not None:
assert mat.shape[0] == label_img.shape[
0], '#images should equal with #data points'
make_sprite(label_img, save_path)
assert mat.ndim == 2, 'mat should be 2D, where mat.size(0) is the number of data points'
make_mat(mat, save_path)
# Filesystem doesn't necessarily have append semantics, so we store an
# internal buffer to append to and re-write whole file after each
# embedding is added
if not hasattr(self, "_projector_config"):
self._projector_config = ProjectorConfig()
embedding_info = get_embedding_info(metadata, label_img, fs, subdir,
global_step, tag)
self._projector_config.embeddings.extend([embedding_info])
from google.protobuf import text_format
config_pbtxt = text_format.MessageToString(self._projector_config)
write_pbtxt(self._get_file_writer().get_logdir(), config_pbtxt)
def train(self, dataset_list, config):
"""
Args:
dataset_list (<StockDataSet>)
config (tf.app.flags.FLAGS)
"""
assert len(dataset_list) > 0
self.merged_sum = tf.compat.v1.summary.merge_all()
# Set up the logs folder
self.writer = tf.compat.v1.summary.FileWriter(
os.path.join("./logs", self.model_name))
self.writer.add_graph(self.sess.graph)
if self.use_embed:
# Set up embedding visualization
# Format: tensorflow/tensorboard/plugins/projector/projector_config.proto
projector_config = ProjectorConfig()
# You can add multiple embeddings. Here we add only one.
added_embed = projector_config.embeddings.add()
added_embed.tensor_name = self.embed_matrix.name
# Link this tensor to its metadata file (e.g. labels).
shutil.copyfile(os.path.join(self.logs_dir, "metadata.tsv"),
os.path.join(self.model_logs_dir, "metadata.tsv"))
added_embed.metadata_path = "metadata.tsv"
# The next line writes a projector_config.pbtxt in the LOG_DIR. TensorBoard will
# read this file during startup.
projector.visualize_embeddings(self.writer, projector_config)
tf.compat.v1.global_variables_initializer().run()
# Merged test data of different stocks.
merged_test_X = []
merged_test_y = []
merged_test_labels = []
for label_, d_ in enumerate(dataset_list):
merged_test_X += list(d_.test_X)
merged_test_y += list(d_.test_y)
merged_test_labels += [[label_]] * len(d_.test_X)
merged_test_X = np.array(merged_test_X)
merged_test_y = np.array(merged_test_y)
merged_test_labels = np.array(merged_test_labels)
print("len(merged_test_X) =", len(merged_test_X))
print("len(merged_test_y) =", len(merged_test_y))
print("len(merged_test_labels) =", len(merged_test_labels))
test_data_feed = {
self.learning_rate: 0.0,
self.keep_prob: 1.0,
self.inputs: merged_test_X,
self.targets: merged_test_y,
self.symbols: merged_test_labels,
}
global_step = 0
num_batches = sum(len(d_.train_X)
for d_ in dataset_list) // config.batch_size
random.seed(time.time())
# Select samples for plotting.
sample_labels = range(min(config.sample_size, len(dataset_list)))
sample_indices = {}
for l in sample_labels:
sym = dataset_list[l].stock_sym
target_indices = np.array([
i for i, sym_label in enumerate(merged_test_labels)
if sym_label[0] == l
])
sample_indices[sym] = target_indices
print(sample_indices)
print("Start training for stocks:",
[d.stock_sym for d in dataset_list])
for epoch in range(config.max_epoch):
epoch_step = 0
learning_rate = config.init_learning_rate * (
config.learning_rate_decay**max(
float(epoch + 1 - config.init_epoch), 0.0))
for label_, d_ in enumerate(dataset_list):
for batch_X, batch_y in d_.generate_one_epoch(
config.batch_size):
global_step += 1
epoch_step += 1
batch_labels = np.array([[label_]] * len(batch_X))
train_data_feed = {
self.learning_rate: learning_rate,
self.keep_prob: config.keep_prob,
self.inputs: batch_X,
self.targets: batch_y,
self.symbols: batch_labels,
}
train_loss, _, train_merged_sum = self.sess.run(
[self.loss, self.optim, self.merged_sum],
train_data_feed)
self.writer.add_summary(train_merged_sum,
global_step=global_step)
if np.mod(global_step,
len(dataset_list) * 200 /
config.input_size) == 1:
test_loss, test_pred = self.sess.run(
[self.loss_test, self.pred], test_data_feed)
print(
"Step:%d [Epoch:%d] [Learning rate: %.6f] train_loss:%.6f test_loss:%.6f"
% (global_step, epoch, learning_rate, train_loss,
test_loss))
# Plot samples
for sample_sym, indices in sample_indices.items():
image_path = os.path.join(
self.model_plots_dir,
"{}_epoch{:02d}_step{:04d}.png".format(
sample_sym, epoch, epoch_step))
sample_preds = test_pred[indices]
sample_truth = merged_test_y[indices]
self.plot_samples(sample_preds,
sample_truth,
image_path,
stock_sym=sample_sym)
self.save(global_step)
final_pred, final_loss = self.sess.run([self.pred, self.loss],
test_data_feed)
# Save the final model
self.save(global_step)
return final_pred
