def img2img_rnd():
cli_train = tw.WatcherClient()
cli = tw.WatcherClient()
imgs = cli_train.create_stream(
event_name='batch',
expr=
"top(l, out_xform=pyt_img_img_out_xform, group_key=lambda x:'', topk=2, order='rnd')",
throttle=1)
img_plot = tw.mpl.ImagePlot()
img_plot.show(imgs, img_width=100, img_height=100, viz_img_scale=3, cols=1)
utils.wait_key()
def show_mpl():
cli = tw.WatcherClient()
p = tw.mpl.LinePlot(title='Demo')
s1 = cli.create_stream(expr='lambda v:(v.i, v.sum)')
p.subscribe(s1, xtitle='Index', ytitle='sqrt(ev_i)')
p.show()
tw.plt_loop()
def plotly_history_graph():
cli = tw.WatcherClient()
p = tw.plotly.line_plot.LinePlot(title='Demo')
s2 = cli.create_stream(event_name='ev_j',
expr='map(lambda v:(v.x, v.val), l)')
p.subscribe(s2, ytitle='ev_j', history_len=15)
p.show()
utils.wait_key()
def text_stats():
train_cli = tw.WatcherClient()
stream = train_cli.create_stream(
event_name="batch", expr='lambda d:(d.x, d.metrics.batch_loss)')
trl = tw.Visualizer(stream, vis_type=None)
trl.show()
input('Paused...')
def mpl_line_plot():
cli = tw.WatcherClient()
p = tw.LinePlot(title='Demo')
s1 = cli.create_stream(event_name='ev_i',
expr='map(lambda v:math.sqrt(v.val)*2, l)')
p.subscribe(s1, xtitle='Index', ytitle='sqrt(ev_i)')
p.show()
tw.plt_loop()
def show_mpl():
cli = tw.WatcherClient(r'c:\temp\sum.log')
s1 = cli.open_stream('sum')
p = tw.LinePlot(title='Demo')
p.subscribe(s1, xtitle='Index', ytitle='sqrt(ev_i)')
s1.load()
p.show()
tw.plt_loop()
def show_find_lr():
cli_train = tw.WatcherClient()
plot = tw.mpl.LinePlot()
train_batch_loss = cli_train.create_stream(
event_name='batch',
expr='lambda d:(d.tt.scheduler.get_lr()[0], d.metrics.batch_loss)')
plot.subscribe(train_batch_loss, xtitle='Epoch', ytitle='Loss')
utils.wait_key()
def plotly_line_graph():
cli = tw.WatcherClient()
s1 = cli.create_stream(event_name="ev_i",
expr='map(lambda v:(v.x, math.sqrt(v.val)), l)')
p = tw.plotly.line_plot.LinePlot()
p.subscribe(s1)
p.show()
utils.wait_key()
def dlc_show_rand_outputs():
cli = cli_train = tw.WatcherClient()
imgs = cli.create_stream(
event_name='batch',
expr=
"top(l, out_xform=pyt_img_img_out_xform, group_key=lambda x:'', topk=10, order='rnd')",
throttle=1)
img_plot = tw.mpl.ImagePlot()
img_plot.show(imgs, img_width=39, img_height=69, viz_img_scale=10)
utils.wait_key()
def show_mpl():
cli = tw.WatcherClient()
st_isum = cli.open_stream('isums')
st_rsum = cli.open_stream('rsums')
line_plot = tw.Visualizer(st_isum, vis_type='line', xtitle='i', ytitle='isum')
line_plot.show()
line_plot2 = tw.Visualizer(st_rsum, vis_type='line', host=line_plot, ytitle='rsum')
tw.plt_loop()
def img_in_class():
cli_train = tw.WatcherClient()
imgs = cli_train.create_stream(
event_name='batch',
expr="top(l, out_xform=pyt_img_class_out_xform, order='rnd')",
throttle=1)
img_plot = tw.mpl.ImagePlot()
img_plot.subscribe(imgs, viz_img_scale=3)
img_plot.show()
tw.image_utils.plt_loop()
def epoch_stats():
train_cli = tw.WatcherClient(port=0)
test_cli = tw.WatcherClient(port=1)
plot = tw.mpl.LinePlot()
train_loss = train_cli.create_stream(
event_name="epoch",
expr='lambda v:(v.metrics.epoch_index, v.metrics.epoch_loss)')
plot.subscribe(train_loss, xtitle='Epoch', ytitle='Train Loss')
test_acc = test_cli.create_stream(
event_name="epoch",
expr='lambda v:(v.metrics.epoch_index, v.metrics.epoch_accuracy)')
plot.subscribe(test_acc,
xtitle='Epoch',
ytitle='Test Accuracy',
ylim=(0, 1))
plot.show()
tw.plt_loop()
def mpl_history_plot():
cli = tw.WatcherClient()
p2 = tw.LinePlot(title='History Demo')
p2s1 = cli.create_stream(
event_name='ev_j', expr='map(lambda v:(v.val, math.sqrt(v.val)*2), l)')
p2.subscribe(p2s1,
xtitle='Index',
ytitle='sqrt(ev_j)',
clear_after_end=True,
history_len=15)
p2.show()
tw.plt_loop()
def plot_grads_plotly():
train_cli = tw.WatcherClient()
grads = train_cli.create_stream(event_name='batch',
expr='lambda d:grads_abs_mean(d.model)',
throttle=1)
p = tw.plotly.line_plot.LinePlot('Demo')
p.subscribe(grads,
xtitle='Layer',
ytitle='Gradients',
history_len=30,
new_on_eval=True)
utils.wait_key()
def plot_grads():
train_cli = tw.WatcherClient()
grads = train_cli.create_stream(
event_name='batch',
expr=
'lambda d:agg_params(d.model, lambda p: p.grad.abs().mean().item())',
throttle=1)
p = tw.plotly.LinePlot('Demo')
p.subscribe(grads,
xtitle='Epoch',
ytitle='Gradients',
history_len=30,
new_on_eval=True)
utils.wait_key()
def img_in_class():
cli_train = tw.WatcherClient()
imgs = cli_train.create_stream(
event_name='batch',
expr=
"topk_all(l, batch_vals=lambda b: (b.batch.loss_all, (b.batch.input, b.batch.output), b.batch.target), \
out_f=image_class_outf, order='dsc')",
throttle=1)
img_plot = tw.ImagePlot()
img_plot.subscribe(imgs, viz_img_scale=3)
img_plot.show()
tw.image_utils.plt_loop()
def plot_grads():
train_cli = tw.WatcherClient()
grads = train_cli.create_stream(event_name='batch',
expr='lambda d:grads_abs_mean(d.model)',
throttle=1)
grad_plot = tw.LinePlot()
grad_plot.subscribe(grads,
xtitle='Layer',
ytitle='Gradients',
clear_after_each=1,
history_len=40,
dim_history=True)
grad_plot.show()
tw.plt_loop()
def plot_weight():
train_cli = tw.WatcherClient()
params = train_cli.create_stream(event_name='batch',
expr='lambda d:weights_abs_mean(d.model)',
throttle=1)
params_plot = tw.LinePlot()
params_plot.subscribe(params,
xtitle='Layer',
ytitle='avg |params|',
clear_after_each=1,
history_len=40,
dim_history=True)
params_plot.show()
tw.plt_loop()
def plot_weight():
train_cli = tw.WatcherClient()
params = train_cli.create_stream(
event_name='batch',
expr='lambda d:agg_params(d.model, lambda p: p.abs().mean().item())',
throttle=1)
params_plot = tw.mpl.LinePlot()
params_plot.subscribe(params,
xtitle='Epoch',
ytitle='avg |params|',
clear_after_each=1,
history_len=40,
dim_history=True)
params_plot.show()
tw.plt_loop()
def batch_stats():
train_cli = tw.WatcherClient()
stream = train_cli.create_stream(
event_name="batch",
expr='lambda v:(v.metrics.epochf, v.metrics.batch_loss)',
throttle=0.75)
train_loss = tw.Visualizer(stream,
clear_after_end=False,
vis_type='mpl-line',
xtitle='Epoch',
ytitle='Train Loss')
#train_acc = tw.Visualizer('lambda v:(v.metrics.epochf, v.metrics.epoch_loss)', event_name="batch",
# xtitle='Epoch', ytitle='Train Accuracy', clear_after_end=False, yrange=(0,1),
# vis=train_loss, vis_type='mpl-line')
train_loss.show()
tw.image_utils.plt_loop()
def show_stream():
cli = tw.WatcherClient()
print("Subscribing to event ev_i...")
s1 = cli.create_stream(event_name="ev_i",
expr='map(lambda v:math.sqrt(v.val), l)')
r1 = tw.TextVis(title='L1')
r1.subscribe(s1)
r1.show()
print("Subscribing to event ev_j...")
s2 = cli.create_stream(event_name="ev_j",
expr='map(lambda v:v.val*v.val, l)')
r2 = tw.TextVis(title='L2')
r2.subscribe(s2)
r2.show()
print("Waiting for key...")
utils.wait_key()
plot_defaults = dict(figure_fn=bk_figure, width=800, height=400)
defaults_streams = ['train_' + n for n in metrics] +\
['eval_' + n for n in metrics]
if len(args.experiments) > 1:
exp_names = args.legend or [
exp.replace('./results', '') for exp in args.experiments
]
else:
exp_names = ['']
results = OrderedDict()
for exp_name, exp in zip(exp_names, args.experiments):
client = tw.WatcherClient(
filename=os.path.join(exp, 'tensorwatch.log'))
epoch_steps = [x for x, _ in dump_stream(client, 'epoch')]
prefix = exp_name + ' - ' if exp_name != '' else ''
for metric_name in metrics:
results.setdefault(metric_name, {})
train_metrics = dump_stream(client, 'train_%s' % metric_name)
val_metrics = dump_stream(client, 'eval_%s' % metric_name)
if not args.all:
train_metrics = reduce_ranges(train_metrics, epoch_steps)
val_metrics = reduce_ranges(val_metrics, epoch_steps)
if args.x_axis == 'epochs':
train_metrics = annotate_indices(train_metrics,
epoch_steps,
interpolate=args.all)
val_metrics = annotate_indices(val_metrics,
import tensorwatch as tw
import objgraph, time #pip install objgraph
cli = tw.WatcherClient()
time.sleep(10)
del cli
import gc
gc.collect()
import time
time.sleep(2)
objgraph.show_backrefs(objgraph.by_type('WatcherClient'), refcounts=True, filename='b.png')
def show_text():
cli = tw.WatcherClient(r'c:\temp\sum.log')
s1 = cli.open_stream('sum_2')
text = tw.Visualizer(s1)
text.show()
input('Waiting')
def show_text():
cli = tw.WatcherClient()
text_vis = tw.Visualizer(st_isum, vis_type='text')
text_vis.show()
input('Waiting')
def show_text():
cli = tw.WatcherClient()
s1 = cli.create_stream(expr='lambda v:(v.i, v.sum)')
text = tw.Visualizer(s1)
text.show()
input('Waiting')
