def fit_pcnorm(modelspec,
est: recording.Recording,
metric=None,
use_modelspec_init: bool = True,
optimizer: str = 'adam',
max_iter: int = 10000,
early_stopping_steps: int = 5,
tolerance: float = 5e-4,
learning_rate: float = 1e-4,
batch_size: typing.Union[None, int] = None,
seed: int = 0,
initializer: str = 'random_normal',
freeze_layers: typing.Union[None, list] = None,
epoch_name: str = "REFERENCE",
n_pcs=2,
**context):
'''
Required Arguments:
est A recording object
modelspec A modelspec object
Optional Arguments:
<copied from fit_tf for now
Returns
dictionary: {'modelspec': updated_modelspec}
'''
# Hard-coded
cost_function = basic_cost
fitter = scipy_minimize
segmentor = nems.segmentors.use_all_data
mapper = nems.fitters.mappers.simple_vector
fit_kwargs = {'tolerance': tolerance, 'max_iter': max_iter}
start_time = time.time()
modelspec = copy.deepcopy(modelspec)
# apply mask to remove invalid portions of signals and allow fit to
# only evaluate the model on the valid portion of the signals
if 'mask' in est.signals.keys():
log.info("Data len pre-mask: %d", est['mask'].shape[1])
est = est.apply_mask()
log.info("Data len post-mask: %d", est['mask'].shape[1])
conditions = [
"_".join(k.split("_")[1:]) for k in est.signals.keys()
if k.startswith("mask_")
]
if (len(conditions) > 2) and any(
[c.split("_")[-1] == 'lg' for c in conditions]):
conditions.remove("small")
conditions.remove("large")
#conditions = conditions[0:2]
#conditions = ['large','small']
group_idx = [est['mask_' + c].as_continuous()[0, :] for c in conditions]
cg_filtered = [(c, g) for c, g in zip(conditions, group_idx)
if g.sum() > 0]
conditions, group_idx = zip(*cg_filtered)
for c, g in zip(conditions, group_idx):
log.info(f"Data subset for {c} len {g.sum()}")
resp = est['resp'].as_continuous()
pred0 = est['pred0'].as_continuous()
residual = resp - pred0
pca = PCA(n_components=n_pcs)
pca.fit(residual.T)
pc_axes = pca.components_
pcproj = residual.T.dot(pc_axes.T).T
group_pc = [pcproj[:, idx].std(axis=1) for idx in group_idx]
resp_std = resp.std(axis=1)
#import pdb; pdb.set_trace()
if metric is None:
metric = lambda d: pc_err(d,
pred_name='pred',
pred0_name='pred0',
group_idx=group_idx,
group_pc=group_pc,
pc_axes=pc_axes,
resp_std=resp_std)
# turn on "fit mode". currently this serves one purpose, for normalization
# parameters to be re-fit for the output of each module that uses
# normalization. does nothing if normalization is not being used.
ms.fit_mode_on(modelspec, est)
# Create the mapper functions that translates to and from modelspecs.
# It has three functions that, when defined as mathematical functions, are:
# .pack(modelspec) -> fitspace_point
# .unpack(fitspace_point) -> modelspec
# .bounds(modelspec) -> fitspace_bounds
packer, unpacker, pack_bounds = mapper(modelspec)
# A function to evaluate the modelspec on the data
evaluator = nems.modelspec.evaluate
my_cost_function = cost_function
my_cost_function.counter = 0
# Freeze everything but sigma, since that's all the fitter should be
# updating.
cost_fn = partial(my_cost_function,
unpacker=unpacker,
modelspec=modelspec,
data=est,
segmentor=segmentor,
evaluator=evaluator,
metric=metric,
display_N=1000)
# get initial sigma value representing some point in the fit space,
# and corresponding bounds for each value
sigma = packer(modelspec)
bounds = pack_bounds(modelspec)
# Results should be a list of modelspecs
# (might only be one in list, but still should be packaged as a list)
improved_sigma = fitter(sigma, cost_fn, bounds=bounds, **fit_kwargs)
improved_modelspec = unpacker(improved_sigma)
elapsed_time = (time.time() - start_time)
start_err = cost_fn(sigma)
final_err = cost_fn(improved_sigma)
log.info("Delta error: %.06f - %.06f = %e", start_err, final_err,
final_err - start_err)
# TODO: Should this maybe be moved to a higher level
# so it applies to ALL the fittters?
ms.fit_mode_off(improved_modelspec)
ms.set_modelspec_metadata(improved_modelspec, 'fitter', 'ccnorm')
ms.set_modelspec_metadata(improved_modelspec, 'fit_time', elapsed_time)
ms.set_modelspec_metadata(improved_modelspec, 'n_parms',
len(improved_sigma))
return {'modelspec': improved_modelspec.copy(), 'save_context': True}
def fit_tf(modelspec,
est: recording.Recording,
use_modelspec_init: bool = True,
optimizer: str = 'adam',
max_iter: int = 10000,
cost_function: str = 'squared_error',
early_stopping_steps: int = 5,
early_stopping_tolerance: float = 5e-4,
learning_rate: float = 1e-4,
batch_size: typing.Union[None, int] = None,
seed: int = 0,
initializer: str = 'random_normal',
filepath: typing.Union[str, Path] = None,
freeze_layers: typing.Union[None, list] = None,
IsReload: bool = False,
epoch_name: str = "REFERENCE",
**context) -> dict:
"""TODO
:param est:
:param modelspec:
:param use_modelspec_init:
:param optimizer:
:param max_iter:
:param cost_function:
:param early_stopping_steps:
:param early_stopping_tolerance:
:param learning_rate:
:param batch_size:
:param seed:
:param filepath:
:param freeze_layers: Indexes of layers to freeze prior to training. Indexes are modelspec indexes, so are offset
from model layer indexes.
:param IsReload:
:param epoch_name
:param context:
:return: dict {'modelspec': modelspec}
"""
if IsReload:
return {}
tf.random.set_seed(seed)
np.random.seed(seed)
#os.environ['TF_DETERMINISTIC_OPS'] = '1' # makes output deterministic, but reduces prediction accuracy
log.info('Building tensorflow keras model from modelspec.')
nems.utils.progress_fun()
# figure out where to save model checkpoints
if filepath is None:
filepath = modelspec.meta['modelpath']
# if job is running on slurm, need to change model checkpoint dir
job_id = os.environ.get('SLURM_JOBID', None)
if job_id is not None:
# keep a record of the job id
modelspec.meta['slurm_jobid'] = job_id
log_dir_root = Path('/mnt/scratch')
assert log_dir_root.exists()
log_dir_sub = Path('SLURM_JOBID' + job_id) / str(modelspec.meta['batch'])\
/ modelspec.meta.get('cellid', "NOCELL")\
/ modelspec.meta['modelname']
filepath = log_dir_root / log_dir_sub
filepath = Path(filepath)
if not filepath.exists():
filepath.mkdir(exist_ok=True, parents=True)
checkpoint_filepath = filepath / 'weights.hdf5'
tensorboard_filepath = filepath / 'logs'
gradient_filepath = filepath / 'gradients'
# update seed based on fit index
seed += modelspec.fit_index
# need to get duration of stims in order to reshape data
#epoch_name = 'REFERENCE' # TODO: this should not be hardcoded
# moved to input parameter
input_name = modelspec.meta.get('input_name', 'stim')
output_name = modelspec.meta.get('output_name', 'resp')
# also grab the fs
fs = est[input_name].fs
if (epoch_name is not None) and (epoch_name != ""):
# extract out the raw data, and reshape to (batch, time, channel)
stim_train = np.transpose(
est[input_name].extract_epoch(epoch=epoch_name, mask=est['mask']),
[0, 2, 1])
resp_train = np.transpose(
est[output_name].extract_epoch(epoch=epoch_name, mask=est['mask']),
[0, 2, 1])
else:
# extract data as a single batch size (1, time, channel)
stim_train = np.transpose(
est.apply_mask()[input_name].as_continuous()[np.newaxis, ...],
[0, 2, 1])
resp_train = np.transpose(
est.apply_mask()[output_name].as_continuous()[np.newaxis, ...],
[0, 2, 1])
log.info(
f'Feature dimensions: {stim_train.shape}; Data dimensions: {resp_train.shape}.'
)
# get state if present, and setup training data
if 'state' in est.signals:
if (epoch_name is not None) and (epoch_name != ""):
state_train = np.transpose(
est['state'].extract_epoch(epoch=epoch_name, mask=est['mask']),
[0, 2, 1])
else:
state_train = np.transpose(
est.apply_mask()['state'].as_continuous()[np.newaxis, ...],
[0, 2, 1])
state_shape = state_train.shape
log.info(f'State dimensions: {state_shape}')
train_data = [stim_train, state_train]
else:
state_train, state_shape = None, None
train_data = stim_train
# correlation for monitoring
# TODO: tf.utils?
def pearson(y_true, y_pred):
return tfp.stats.correlation(y_true,
y_pred,
event_axis=None,
sample_axis=None)
# get the layers and build the model
cost_fn = loss_functions.get_loss_fn(cost_function)
model_layers = modelspec.modelspec2tf2(
use_modelspec_init=use_modelspec_init,
seed=seed,
fs=fs,
initializer=initializer)
if np.any([isinstance(layer, Conv2D_NEMS) for layer in model_layers]):
# need a "channel" dimension for Conv2D (like rgb channels, not frequency). Only 1 channel for our data.
stim_train = stim_train[..., np.newaxis]
train_data = train_data[..., np.newaxis]
# do some batch sizing logic
batch_size = stim_train.shape[0] if batch_size == 0 else batch_size
model = modelbuilder.ModelBuilder(
name='Test-model',
layers=model_layers,
learning_rate=learning_rate,
loss_fn=cost_fn,
optimizer=optimizer,
metrics=[pearson],
).build_model(input_shape=stim_train.shape,
state_shape=state_shape,
batch_size=batch_size)
# tracking early termination
model.early_terminated = False
# create the callbacks
early_stopping = callbacks.DelayedStopper(
monitor='loss',
patience=30 * early_stopping_steps,
min_delta=early_stopping_tolerance,
verbose=1,
restore_best_weights=False)
checkpoint = tf.keras.callbacks.ModelCheckpoint(
filepath=str(checkpoint_filepath),
save_best_only=False,
save_weights_only=True,
save_freq=100 * stim_train.shape[0],
monitor='loss',
verbose=0)
sparse_logger = callbacks.SparseProgbarLogger(n_iters=10)
nan_terminate = tf.keras.callbacks.TerminateOnNaN()
nan_weight_terminate = callbacks.TerminateOnNaNWeights()
tensorboard = tf.keras.callbacks.TensorBoard(
log_dir=str(tensorboard_filepath), # TODO: generic tensorboard dir?
histogram_freq=0, # record the distribution of the weights
write_graph=False,
update_freq='epoch',
profile_batch=0)
# gradient_logger = callbacks.GradientLogger(filepath=str(gradient_filepath),
# train_input=stim_train,
# model=model)
# freeze layers
if freeze_layers is not None:
for freeze_index in freeze_layers:
log.info(
f'Freezing layer #{freeze_index}: "{model.layers[freeze_index + 1].name}".'
)
model.layers[freeze_index + 1].trainable = False
# save an initial set of weights before freezing, in case of termination before any checkpoints
#log.info('saving weights to : %s', str(checkpoint_filepath) )
model.save_weights(str(checkpoint_filepath), overwrite=True)
if version.parse(tf.__version__) >= version.parse("2.2.0"):
callback0 = [sparse_logger]
verbose = 0
else:
callback0 = []
verbose = 2
log.info(f'Fitting model (batch_size={batch_size})...')
history = model.fit(
train_data,
resp_train,
# validation_split=0.2,
verbose=verbose,
epochs=max_iter,
batch_size=batch_size,
callbacks=callback0 + [
nan_terminate,
nan_weight_terminate,
early_stopping,
checkpoint,
# enable the below to log tracked parameters to tensorboard
# tensorboard,
# enable the below to record gradients to visualize in tensorboard; this is very slow,
# and loading all this into tensorboard can use A LOT of memory
# gradient_logger,
])
# did we terminate on a nan loss or weights? Load checkpoint if so
if np.all(np.isnan(model.predict(train_data))
) or model.early_terminated: # TODO: should this be np.any()?
log.warning(
'Model terminated on nan loss or weights, restoring saved weights.'
)
try:
# this can fail if it nans out before a single checkpoint gets saved, either because no saved weights
# exist, or it tries to load a in different model from the init
model.load_weights(str(checkpoint_filepath))
log.warning('Reloaded previous saved weights after nan loss.')
except (tf.errors.NotFoundError, ValueError):
pass
modelspec = tf2modelspec(model, modelspec)
contains_tf_only_layers = np.any(
['tf_only' in m['fn'] for m in modelspec.modules])
if not contains_tf_only_layers:
# compare the predictions from the model and modelspec
error = compare_ms_tf(modelspec, model, est, train_data)
if error > 1e-5:
log.warning(
f'Mean difference between NEMS and TF model prediction: {error}'
)
else:
log.info(
f'Mean difference between NEMS and TF model prediction: {error}'
)
else:
# nothing to compare, ms evaluation is not implemented for this type of model
pass
# add in some relevant meta information
modelspec.meta['n_parms'] = len(modelspec.phi_vector)
try:
n_epochs = len(history.history['loss'])
except KeyError:
n_epochs = 0
try:
max_iter = modelspec.meta['extra_results']
modelspec.meta['extra_results'] = max(max_iter, n_epochs)
except KeyError:
modelspec.meta['extra_results'] = n_epochs
nems.utils.progress_fun()
return {'modelspec': modelspec}
def fit_ccnorm(modelspec,
est: recording.Recording,
metric=None,
use_modelspec_init: bool = True,
optimizer: str = 'adam',
max_iter: int = 10000,
early_stopping_steps: int = 5,
tolerance: float = 5e-4,
learning_rate: float = 1e-4,
batch_size: typing.Union[None, int] = None,
seed: int = 0,
initializer: str = 'random_normal',
freeze_layers: typing.Union[None, list] = None,
epoch_name: str = "REFERENCE",
shrink_cc: float = 0,
noise_pcs: int = 0,
shared_pcs: int = 0,
also_fit_resp: bool = False,
force_psth: bool = False,
use_metric: typing.Union[None, str] = None,
alpha: float = 0.1,
beta: float = 1,
exclude_idx=None,
exclude_after=None,
freeze_idx=None,
freeze_after=None,
**context):
'''
Required Arguments:
est A recording object
modelspec A modelspec object
Optional Arguments:
<copied from fit_tf for now
Returns
dictionary: {'modelspec': updated_modelspec}
'''
# Hard-coded
cost_function = basic_cost
fitter = scipy_minimize
segmentor = nems.segmentors.use_all_data
mapper = nems.fitters.mappers.simple_vector
fit_kwargs = {'tolerance': tolerance, 'max_iter': max_iter}
start_time = time.time()
fit_index = modelspec.fit_index
if (exclude_idx is not None) | (freeze_idx is not None) | \
(exclude_after is not None) | (freeze_after is not None):
modelspec0 = modelspec.copy()
modelspec, include_set = modelspec_freeze_layers(
modelspec,
include_idx=None,
exclude_idx=exclude_idx,
exclude_after=exclude_after,
freeze_idx=freeze_idx,
freeze_after=freeze_after)
modelspec0.set_fit(fit_index)
modelspec.set_fit(fit_index)
else:
include_set = None
# Computing PCs before masking out unwanted stimuli in order to
# preserve match with epochs
epoch_regex = "^STIM_"
stims = (est.epochs['name'].value_counts() >= 8)
stims = [
stims.index[i] for i, s in enumerate(stims)
if bool(re.search(epoch_regex, stims.index[i])) and s == True
]
Rall_u = est.apply_mask()['psth'].as_continuous().T
# can't simply extract evoked for refs because can be longer/shorted if it came after target
# and / or if it was the last stim. So, masking prestim / postim doesn't work. Do it manually
#d = est['resp'].extract_epochs(stims, mask=est['mask'])
#R = [v.mean(axis=0) for (k, v) in d.items()]
#R = [np.reshape(np.transpose(v,[1,0,2]),[v.shape[1],-1]) for (k, v) in d.items()]
#Rall_u = np.hstack(R).T
pca = PCA(n_components=2)
pca.fit(Rall_u)
pc_axes = pca.components_
# apply mask to remove invalid portions of signals and allow fit to
# only evaluate the model on the valid portion of the signals
if 'mask_small' in est.signals.keys():
log.info('reseting mask with mask_small+mask_large subset')
est['mask'] = est['mask']._modified_copy(data=est['mask_small']._data +
est['mask_large']._data)
if 'mask' in est.signals.keys():
log.info("Data len pre-mask: %d", est['mask'].shape[1])
est = est.apply_mask()
log.info("Data len post-mask: %d", est['mask'].shape[1])
# if we want to fit to first-order cc error.
#uncomment this and make sure sdexp is generating a pred0 signal
est = modelspec.evaluate(est, stop=2)
if ('pred0' in est.signals.keys()) & (not force_psth):
input_name = 'pred0'
log.info('Found pred0 for fitting CC')
else:
input_name = 'psth'
log.info('No pred0, using psth for fitting CC')
conditions = [
"_".join(k.split("_")[1:]) for k in est.signals.keys()
if k.startswith("mask_")
]
if (len(conditions) > 2) and any(
[c.split("_")[-1] == 'lg' for c in conditions]):
conditions.remove("small")
conditions.remove("large")
#conditions = conditions[0:2]
#conditions = ['large','small']
group_idx = [est['mask_' + c].as_continuous()[0, :] for c in conditions]
cg_filtered = [(c, g) for c, g in zip(conditions, group_idx)
if g.sum() > 0]
conditions, group_idx = zip(*cg_filtered)
for c, g in zip(conditions, group_idx):
log.info(f"cc data for {c} len {g.sum()}")
resp = est['resp'].as_continuous()
pred0 = est[input_name].as_continuous()
#import pdb; pdb.set_trace()
if shrink_cc > 0:
log.info(f'cc approx: shrink_cc={shrink_cc}')
group_cc = [
cc_shrink(resp[:, idx] - pred0[:, idx], sigrat=shrink_cc)
for idx in group_idx
]
elif shared_pcs > 0:
log.info(f'cc approx: shared_pcs={shared_pcs}')
cc = np.cov(resp - pred0)
u, s, vh = np.linalg.svd(cc)
U = u[:, :shared_pcs] @ u[:, :shared_pcs].T
group_cc = [
cc_shared_space(resp[:, idx] - pred0[:, idx], U)
for idx in group_idx
]
elif noise_pcs > 0:
log.info(f'cc approx: noise_pcs={noise_pcs}')
group_cc = [
cc_lowrank(resp[:, idx] - pred0[:, idx], n_pcs=noise_pcs)
for idx in group_idx
]
else:
group_cc = [np.cov(resp[:, idx] - pred0[:, idx]) for idx in group_idx]
group_cc_raw = [np.cov(resp[:, idx] - pred0[:, idx]) for idx in group_idx]
# variance of projection onto PCs (PCs computed above before masking)
pcproj0 = (resp - pred0).T.dot(pc_axes.T).T
pcproj_std = pcproj0.std(axis=1)
if (use_metric == 'cc_err_w'):
def metric(d, verbose=False):
return metrics.cc_err_w(d,
pred_name='pred',
pred0_name=input_name,
group_idx=group_idx,
group_cc=group_cc,
alpha=alpha,
pcproj_std=None,
pc_axes=None,
verbose=verbose)
log.info(f"fit_ccnorm metric: cc_err_w (alpha={alpha})")
elif (metric is None) and also_fit_resp:
log.info(f"resp_cc_err: pred0_name: {input_name} beta: {beta}")
metric = lambda d: metrics.resp_cc_err(d,
pred_name='pred',
pred0_name=input_name,
group_idx=group_idx,
group_cc=group_cc,
beta=beta,
pcproj_std=None,
pc_axes=None)
elif (use_metric == 'cc_err_md'):
def metric(d, verbose=False):
return metrics.cc_err_md(d,
pred_name='pred',
pred0_name=input_name,
group_idx=group_idx,
group_cc=group_cc,
pcproj_std=None,
pc_axes=None)
log.info(f"fit_ccnorm metric: cc_err_md")
elif (metric is None):
#def cc_err(result, pred_name='pred_lv', resp_name='resp', pred0_name='pred',
# group_idx=None, group_cc=None, pcproj_std=None, pc_axes=None):
# current implementation of cc_err
metric = lambda d: metrics.cc_err(d,
pred_name='pred',
pred0_name=input_name,
group_idx=group_idx,
group_cc=group_cc,
pcproj_std=None,
pc_axes=None)
log.info(f"fit_ccnorm metric: cc_err")
# turn on "fit mode". currently this serves one purpose, for normalization
# parameters to be re-fit for the output of each module that uses
# normalization. does nothing if normalization is not being used.
ms.fit_mode_on(modelspec, est)
# Create the mapper functions that translates to and from modelspecs.
# It has three functions that, when defined as mathematical functions, are:
# .pack(modelspec) -> fitspace_point
# .unpack(fitspace_point) -> modelspec
# .bounds(modelspec) -> fitspace_bounds
packer, unpacker, pack_bounds = mapper(modelspec)
# A function to evaluate the modelspec on the data
evaluator = nems.modelspec.evaluate
my_cost_function = cost_function
my_cost_function.counter = 0
# Freeze everything but sigma, since that's all the fitter should be
# updating.
cost_fn = partial(my_cost_function,
unpacker=unpacker,
modelspec=modelspec,
data=est,
segmentor=segmentor,
evaluator=evaluator,
metric=metric,
display_N=1000)
# get initial sigma value representing some point in the fit space,
# and corresponding bounds for each value
sigma = packer(modelspec)
bounds = pack_bounds(modelspec)
# Results should be a list of modelspecs
# (might only be one in list, but still should be packaged as a list)
improved_sigma = fitter(sigma, cost_fn, bounds=bounds, **fit_kwargs)
improved_modelspec = unpacker(improved_sigma)
elapsed_time = (time.time() - start_time)
start_err = cost_fn(sigma)
final_err = cost_fn(improved_sigma)
log.info("Delta error: %.06f - %.06f = %e", start_err, final_err,
final_err - start_err)
# TODO: Should this maybe be moved to a higher level
# so it applies to ALL the fittters?
ms.fit_mode_off(improved_modelspec)
if include_set is not None:
# pull out updated phi values from improved_modelspec, include_set only
improved_modelspec = \
modelspec_unfreeze_layers(improved_modelspec, modelspec0, include_set)
improved_modelspec.set_fit(fit_index)
log.info(
f"Updating improved modelspec with fit_idx={improved_modelspec.fit_index}"
)
improved_modelspec.meta['fitter'] = 'ccnorm'
improved_modelspec.meta['n_parms'] = len(improved_sigma)
if modelspec.fit_count == 1:
improved_modelspec.meta['fit_time'] = elapsed_time
improved_modelspec.meta['loss'] = final_err
else:
if modelspec.fit_index == 0:
improved_modelspec.meta['fit_time'] = np.zeros(
improved_modelspec.fit_count)
improved_modelspec.meta['loss'] = np.zeros(
improved_modelspec.fit_count)
improved_modelspec.meta['fit_time'][fit_index] = elapsed_time
improved_modelspec.meta['loss'][fit_index] = final_err
return {'modelspec': improved_modelspec}
def fit_tf(modelspec,
est: recording.Recording,
use_modelspec_init: bool = True,
optimizer: str = 'adam',
max_iter: int = 10000,
cost_function: str = 'squared_error',
early_stopping_steps: int = 5,
early_stopping_tolerance: float = 5e-4,
early_stopping_val_split: float = 0,
learning_rate: float = 1e-4,
variable_learning_rate: bool = False,
batch_size: typing.Union[None, int] = None,
seed: int = 0,
initializer: str = 'random_normal',
filepath: typing.Union[str, Path] = None,
freeze_layers: typing.Union[None, list] = None,
IsReload: bool = False,
epoch_name: str = "REFERENCE",
use_tensorboard: bool = False,
kernel_regularizer: str = None,
**context) -> dict:
"""TODO
:param est:
:param modelspec:
:param use_modelspec_init:
:param optimizer:
:param max_iter:
:param cost_function:
:param early_stopping_steps:
:param early_stopping_tolerance:
:param learning_rate:
:param batch_size:
:param seed:
:param filepath:
:param freeze_layers: Indexes of layers to freeze prior to training. Indexes are modelspec indexes, so are offset
from model layer indexes.
:param IsReload:
:param epoch_name
:param context:
:return: dict {'modelspec': modelspec}
"""
if IsReload:
return {}
tf.random.set_seed(seed)
np.random.seed(seed)
#os.environ['TF_DETERMINISTIC_OPS'] = '1' # makes output deterministic, but reduces prediction accuracy
log.info('Building tensorflow keras model from modelspec.')
nems.utils.progress_fun()
# figure out where to save model checkpoints
job_id = os.environ.get('SLURM_JOBID', None)
if job_id is not None:
# if job is running on slurm, need to change model checkpoint dir
# keep a record of the job id
modelspec.meta['slurm_jobid'] = job_id
log_dir_root = Path('/mnt/scratch')
assert log_dir_root.exists()
log_dir_base = log_dir_root / Path('SLURM_JOBID' + job_id)
log_dir_sub = Path(str(modelspec.meta['batch'])) \
/ modelspec.meta.get('cellid', "NOCELL") \
/ modelspec.get_longname()
filepath = log_dir_base / log_dir_sub
tbroot = filepath / 'logs'
elif filepath is None:
filepath = modelspec.meta['modelpath']
tbroot = Path(f'/auto/data/tmp/tensorboard/')
else:
tbroot = Path(f'/auto/data/tmp/tensorboard/')
filepath = Path(filepath)
if not filepath.exists():
filepath.mkdir(exist_ok=True, parents=True)
cellid = modelspec.meta.get('cellid', 'CELL')
tbpath = tbroot / (str(modelspec.meta['batch']) + '_' + cellid + '_' +
modelspec.meta['modelname'])
# TODO: should this code just be deleted then?
if 0 & use_tensorboard:
# disabled, this is dumb. it deletes the previous round of fitting (eg, tfinit)
fileList = glob.glob(str(tbpath / '*' / '*'))
for filePath in fileList:
try:
os.remove(filePath)
except:
print("Error while deleting file : ", filePath)
checkpoint_filepath = filepath / 'weights.hdf5'
tensorboard_filepath = tbpath
gradient_filepath = filepath / 'gradients'
# update seed based on fit index
seed += modelspec.fit_index
if (freeze_layers
is not None) and len(freeze_layers) and (len(freeze_layers)
== freeze_layers[-1] + 1):
truncate_model = True
modelspec_trunc, est_trunc = \
initializers.modelspec_remove_input_layers(modelspec, est, remove_count=len(freeze_layers))
modelspec_original = modelspec
est_original = est
modelspec = modelspec_trunc
est = est_trunc
freeze_layers = None
log.info(
f"Special case of freezing: truncating model. fit_index={modelspec.fit_index} cell_index={modelspec.cell_index}"
)
else:
truncate_model = False
input_name = modelspec.meta.get('input_name', 'stim')
output_name = modelspec.meta.get('output_name', 'resp')
# also grab the fs
fs = est[input_name].fs
if (epoch_name is not None) and (epoch_name != ""):
# extract out the raw data, and reshape to (batch, time, channel)
stim_train = np.transpose(
est[input_name].extract_epoch(epoch=epoch_name, mask=est['mask']),
[0, 2, 1])
resp_train = np.transpose(
est[output_name].extract_epoch(epoch=epoch_name, mask=est['mask']),
[0, 2, 1])
else:
# extract data as a single batch size (1, time, channel)
stim_train = np.transpose(
est.apply_mask()[input_name].as_continuous()[np.newaxis, ...],
[0, 2, 1])
resp_train = np.transpose(
est.apply_mask()[output_name].as_continuous()[np.newaxis, ...],
[0, 2, 1])
log.info(
f'Feature dimensions: {stim_train.shape}; Data dimensions: {resp_train.shape}.'
)
if True:
log.info("adding a tiny bit of noise to resp_train")
resp_train = resp_train + np.random.randn(*resp_train.shape) / 10000
# get state if present, and setup training data
if 'state' in est.signals:
if (epoch_name is not None) and (epoch_name != ""):
state_train = np.transpose(
est['state'].extract_epoch(epoch=epoch_name, mask=est['mask']),
[0, 2, 1])
else:
state_train = np.transpose(
est.apply_mask()['state'].as_continuous()[np.newaxis, ...],
[0, 2, 1])
state_shape = state_train.shape
log.info(f'State dimensions: {state_shape}')
train_data = [stim_train, state_train]
else:
state_train, state_shape = None, None
train_data = stim_train
# get the layers and build the model
cost_fn = loss_functions.get_loss_fn(cost_function)
#model_layers = modelspec.modelspec2tf2(
# use_modelspec_init=use_modelspec_init, seed=seed, fs=fs,
# initializer=initializer, freeze_layers=freeze_layers,
# kernel_regularizer=kernel_regularizer)
model_layers = modelbuilder.modelspec2tf(
modelspec,
use_modelspec_init=use_modelspec_init,
seed=seed,
fs=fs,
initializer=initializer,
freeze_layers=freeze_layers,
kernel_regularizer=kernel_regularizer)
if np.any([isinstance(layer, Conv2D_NEMS) for layer in model_layers]):
# need a "channel" dimension for Conv2D (like rgb channels, not frequency). Only 1 channel for our data.
stim_train = stim_train[..., np.newaxis]
train_data = train_data[..., np.newaxis]
# do some batch sizing logic
batch_size = stim_train.shape[0] if batch_size == 0 else batch_size
if variable_learning_rate:
# TODO: allow other schedule options instead of hard-coding exp decay?
# TODO: expose exp decay kwargs as kw options? not clear how to choose these parameters
learning_rate = tf.keras.optimizers.schedules.ExponentialDecay(
initial_learning_rate=learning_rate,
decay_steps=10000,
decay_rate=0.9)
from nems.tf.loss_functions import pearson
model = modelbuilder.ModelBuilder(
name='Test-model',
layers=model_layers,
learning_rate=learning_rate,
loss_fn=cost_fn,
optimizer=optimizer,
metrics=[pearson],
).build_model(input_shape=stim_train.shape,
state_shape=state_shape,
batch_size=batch_size)
if freeze_layers is not None:
for freeze_index in freeze_layers:
log.info(
f'TF layer #{freeze_index}: "{model.layers[freeze_index + 1].name}" is not trainable.'
)
# tracking early termination
model.early_terminated = False
# create the callbacks
early_stopping = callbacks.DelayedStopper(
monitor='val_loss',
patience=30 * early_stopping_steps,
min_delta=early_stopping_tolerance,
verbose=1,
restore_best_weights=True)
regular_stopping = callbacks.DelayedStopper(
monitor='loss',
patience=30 * early_stopping_steps,
min_delta=early_stopping_tolerance,
verbose=1,
restore_best_weights=True)
checkpoint = tf.keras.callbacks.ModelCheckpoint(
filepath=str(checkpoint_filepath),
save_best_only=False,
save_weights_only=True,
save_freq=100 * stim_train.shape[0],
monitor='loss',
verbose=0)
sparse_logger = callbacks.SparseProgbarLogger(n_iters=50)
nan_terminate = tf.keras.callbacks.TerminateOnNaN()
nan_weight_terminate = callbacks.TerminateOnNaNWeights()
tensorboard = tf.keras.callbacks.TensorBoard(
log_dir=str(tensorboard_filepath), # TODO: generic tensorboard dir?
histogram_freq=0, # record the distribution of the weights
write_graph=False,
update_freq='epoch',
profile_batch=0)
# gradient_logger = callbacks.GradientLogger(filepath=str(gradient_filepath),
# train_input=stim_train,
# model=model)
# save an initial set of weights before freezing, in case of termination before any checkpoints
#log.info('saving weights to : %s', str(checkpoint_filepath) )
model.save_weights(str(checkpoint_filepath), overwrite=True)
if version.parse(tf.__version__) >= version.parse("2.2.0"):
callback0 = [sparse_logger]
verbose = 0
else:
callback0 = []
verbose = 2
# enable the below to log tracked parameters to tensorboard
if use_tensorboard:
callback0.append(tensorboard)
log.info(f'Enabling tensorboard, log: {str(tensorboard_filepath)}')
# enable the below to record gradients to visualize in tensorboard; this is very slow,
# and loading all this into tensorboard can use A LOT of memory
# callback0.append(gradient_logger)
if early_stopping_val_split > 0:
callback0.append(early_stopping)
log.info(
f'Enabling early stopping, val split: {str(early_stopping_val_split)}'
)
else:
callback0.append(regular_stopping)
log.info(f'Stop tolerance: min_delta={early_stopping_tolerance}')
log.info(f'Fitting model (batch_size={batch_size})...')
history = model.fit(train_data,
resp_train,
validation_split=early_stopping_val_split,
verbose=verbose,
epochs=max_iter,
callbacks=callback0 + [
nan_terminate,
nan_weight_terminate,
checkpoint,
],
batch_size=batch_size)
# did we terminate on a nan loss or weights? Load checkpoint if so
if np.all(np.isnan(model.predict(train_data))
) or model.early_terminated: # TODO: should this be np.any()?
log.warning(
'Model terminated on nan loss or weights, restoring saved weights.'
)
try:
# this can fail if it nans out before a single checkpoint gets saved, either because no saved weights
# exist, or it tries to load a in different model from the init
model.load_weights(str(checkpoint_filepath))
log.warning('Reloaded previous saved weights after nan loss.')
except (tf.errors.NotFoundError, ValueError):
pass
modelspec = tf2modelspec(model, modelspec)
if truncate_model:
log.info("Special case of freezing: restoring truncated model!!!")
#modelspec_restored, rec_restored = modelspec_restore_input_layers(modelspec_trunc, rec_trunc, modelspec_original)
modelspec_restored, est_restored = initializers.modelspec_restore_input_layers(
modelspec, est, modelspec_original)
est = est_original
modelspec = modelspec_restored
# debug: dump modelspec parameters
#for i in range(len(modelspec)):
# log.info(modelspec.phi[i])
contains_tf_only_layers = np.any(
['tf_only' in m['fn'] for m in modelspec.modules])
if not contains_tf_only_layers:
# compare the predictions from the model and modelspec
error = compare_ms_tf(modelspec, model, est, train_data)
if error > 1e-5:
log.warning(
f'Mean difference between NEMS and TF model prediction: {error}'
)
else:
log.info(
f'Mean difference between NEMS and TF model prediction: {error}'
)
else:
# nothing to compare, ms evaluation is not implemented for this type of model
pass
# add in some relevant meta information
modelspec.meta['n_parms'] = len(modelspec.phi_vector)
try:
n_epochs = len(history.history['loss'])
if 'val_loss' in history.history.keys():
#val_stop = np.argmin(history.history['val_loss'])
#loss = history.history['loss'][val_stop]
loss = np.nanmin(history.history['val_loss'])
else:
loss = np.nanmin(history.history['loss'])
except KeyError:
n_epochs = 0
loss = 0
if modelspec.fit_count == 1:
modelspec.meta['n_epochs'] = n_epochs
modelspec.meta['loss'] = loss
else:
if modelspec.fit_index == 0:
modelspec.meta['n_epochs'] = np.zeros(modelspec.fit_count)
modelspec.meta['loss'] = np.zeros(modelspec.fit_count)
modelspec.meta['n_epochs'][modelspec.fit_index] = n_epochs
modelspec.meta['loss'][modelspec.fit_index] = loss
try:
max_iter = modelspec.meta['extra_results']
modelspec.meta['extra_results'] = max(max_iter, n_epochs)
except KeyError:
modelspec.meta['extra_results'] = n_epochs
nems.utils.progress_fun()
# clean up temp files
if job_id is not None:
log.info('removing temporary weights file(s)')
shutil.rmtree(log_dir_base)
return {'modelspec': modelspec}
