jlynxdev's picture
initial commit
721508a
# code inspired by Fastai "Practical Deep Learning Part 2" Learner
import math
import os
from functools import partial
from operator import attrgetter
import matplotlib.pyplot as plt
import numpy as np
import torch
import torch.nn.functional as F
import wandb
class CancelFitException(Exception):
pass
class CancelBatchException(Exception):
pass
class CancelEpochException(Exception):
pass
class Callback:
order = 0
class with_cbs:
"""Decorator that wraps function and calls certain callbacks before/after that function."""
def __init__(self, nm):
self.nm = nm
def __call__(self, f):
def _f(o, *args, **kwargs):
try:
o.callback(f"before_{self.nm}")
f(o, *args, **kwargs)
o.callback(f"after_{self.nm}")
except globals()[f"Cancel{self.nm.title()}Exception"]:
pass
finally:
o.callback(f"cleanup_{self.nm}")
return _f
def run_cbs(cbs, method_nm, trainer=None):
for cb in sorted(cbs, key=attrgetter("order")): # sort callbacks by 'order'
method = getattr(
cb, method_nm, None
) # get method from callback e.g. `before_batch`
if method is not None:
method(trainer) # if callback has such method then call it
class Trainer:
"""Trainer with callbacks"""
def __init__(
self,
model,
dls=(0,),
loss_func=F.mse_loss,
opt_func=torch.optim.SGD,
lr=0.1,
cbs=[],
n_inp=1,
):
self.model = model
self.dls = dls
self.loss_func = loss_func
self.opt_func = opt_func
self.lr = lr
self.cbs = cbs
self.n_inp = n_inp
@with_cbs("batch")
def _one_batch(self):
self.predict()
self.callback("after_predict")
self.get_loss()
self.callback("after_loss")
if self.training:
self.backward()
self.callback("after_backward")
self.step()
self.callback("after_step")
self.zero_grad()
@with_cbs("epoch")
def _one_epoch(self):
for self.iter, self.batch in enumerate(self.dl):
self._one_batch()
def one_epoch(self, training):
self.model.train(training)
self.dl = self.dls.train if training else self.dls.valid
self._one_epoch()
@with_cbs("fit")
def _fit(self, train, valid):
for epoch in range(self.n_epochs):
if train:
self.one_epoch(True)
if valid:
torch.no_grad()(self.one_epoch)(False)
def fit(self, n_epochs=1, train=True, valid=True, cbs=None, lr=None):
self.n_epochs = n_epochs
if lr is not None:
self.lr = lr
self.opt = self.opt_func(self.model.parameters(), self.lr)
self._fit(train, valid)
def callback(self, method_nm):
run_cbs(self.cbs, method_nm, self)
def predict(self, x=None):
if x is not None:
return self.model(x)
self.preds = self.model(*self.batch[: self.n_inp])
def get_loss(self):
self.loss = self.loss_func(self.preds, *self.batch[self.n_inp :])
def backward(self):
self.loss.backward()
def step(self):
self.opt.step()
def zero_grad(self):
self.opt.zero_grad()
@property
def training(self):
return self.model.training
class ProgressCB(Callback):
"""Adds progress bar to Trainer and plotting loss curves after training."""
def __init__(self, in_notebook=False):
super().__init__()
self.train_loss = []
self.valid_loss = []
self.in_notebook = in_notebook
def before_fit(self, trainer):
if self.in_notebook:
from tqdm.notebook import tqdm
else:
from tqdm import tqdm
self.pbar = tqdm(total=trainer.n_epochs)
def after_epoch(self, trainer):
if trainer.training:
self.pbar.update(1)
def after_loss(self, trainer):
if trainer.training:
self.train_loss.append(trainer.loss.item())
tmp_train_loss = (
np.mean(self.train_loss[-10:]) if len(self.train_loss) > 10 else 0
)
tmp_valid_loss = (
np.mean(self.valid_loss[-len(trainer.dls.valid) :])
if len(self.valid_loss) > 0
else 0
)
self.pbar.set_description(
f"train loss: {tmp_train_loss:.3f} | valid loss: {tmp_valid_loss:.3f}"
)
else:
self.valid_loss.append(trainer.loss.item())
def after_fit(self, trainer):
self.pbar.close()
def plot_losses(self, save=True):
fig, ax = plt.subplots(1, 2, figsize=(12, 4))
ax[0].plot(self.train_loss)
ax[0].set_title("train loss")
ax[1].plot(self.valid_loss)
ax[1].set_title("valid loss")
if save:
if not os.path.exists("./plots"):
os.makedirs("./plots")
plt.savefig("./plots/losses.png")
else:
plt.show()
class DeviceCB(Callback):
"""Moves model and batches to device"""
def __init__(self, device="cpu"):
self.device = device
def before_fit(self, trainer):
if hasattr(trainer.model, "to"):
trainer.model.to(self.device)
def before_batch(self, trainer):
trainer.batch = tuple(t.to(self.device) for t in trainer.batch)
class Hook:
"""Registers PyTorch forward hook with provided function"""
def __init__(self, name, mod, f):
self.hook = mod.register_forward_hook(partial(f, self, name))
def remove(self):
self.hook.remove()
def __del__(self):
self.remove()
class Hooks(list):
"""List of hooks"""
def __init__(self, mods, f):
super().__init__([Hook(n, m, f) for n, m in mods])
def __enter__(self, *args):
return self
def __exit__(self, *args):
self.remove()
def __del__(self):
self.remove()
def __delitem__(self, i):
self[i].remove()
super().__delitem__(i)
def remove(self):
for h in self:
h.remove()
class HooksCB(Callback):
"""Appends hooks with some `hookfunc` to selected layers filtered by `mod_filter`."""
def __init__(self, hookfunc, mod_filter=lambda x: True):
super().__init__()
self.hookfunc = hookfunc
self.mod_filter = mod_filter
def before_fit(self, trainer):
mods = [
(name, mod)
for name, mod in trainer.model.named_modules()
if self.mod_filter(mod)
]
self.hooks = Hooks(mods, partial(self._hookfunc, trainer.training))
def _hookfunc(self, training, *args, **kwargs):
if training:
self.hookfunc(*args, **kwargs)
def after_fit(self, trainer):
self.hooks.remove()
def __iter__(self):
return iter(self.hooks)
def __len__(self):
return len(self.hooks)
def append_stats(with_wandb, hook, name, mod, inp, outp):
if not hasattr(hook, "stats"):
hook.stats = {"mean": [], "std": [], "abs": []}
acts = outp.detach().cpu()
hook.stats["mean"].append(acts.mean().item())
hook.stats["std"].append(acts.std().item())
hook.stats["abs"].append(acts.abs().histc(40, 0, 10).tolist())
if with_wandb:
wandb.log(
{
f"{name}/mean": acts.mean().item(),
f"{name}/std": acts.std().item(),
f"{name}/abs": wandb.Histogram(acts.abs().histc(40, 0, 10).tolist()),
},
commit=False,
)
def get_grid(n, figsize):
return plt.subplots(round(n / 2), 2, figsize=figsize)
class WandBCB(Callback):
"""Inits and logs to W&B. Every `wandb.log()` outside this callback should have property `commit=False` because this callback gathers all logs in given batch."""
order = math.inf # make sure that this callback will be called last
def __init__(
self, proj_name, model_path, run_name=None, notes=None, **additional_config
):
self.proj_name = proj_name
self.run_name = run_name
self.model_path = model_path
self.notes = notes
self.additional_config = additional_config
def before_fit(self, trainer):
info = dict(
project=self.proj_name,
config={"lr": trainer.lr, "n_epochs": trainer.n_epochs},
)
if self.run_name is not None:
info["name"] = self.run_name
if self.notes is not None:
info["notes"] = self.notes
if self.additional_config is not None:
info["config"] = {**info["config"], **self.additional_config}
wandb.init(**info)
wandb.watch(trainer.model, log="all")
def after_loss(self, trainer):
if trainer.training:
wandb.log({"loss/train": trainer.loss.item()}, commit=False)
else:
wandb.log({"loss/valid": trainer.loss.item()}, commit=False)
def after_batch(self, trainer):
wandb.log({}, commit=True)
def after_fit(self, trainer):
torch.save(trainer.model.state_dict(), self.model_path)
wandb.save(self.model_path)
wandb.finish()
class ActivationStatsCB(HooksCB):
"""Stores activation statistics of selected modules. Recommended only for debugging or visualizations, not for actual training because it significantly slows down training."""
def __init__(self, mod_filter=lambda x: x, with_wandb=False):
super().__init__(partial(append_stats, with_wandb), mod_filter)
def plot_stats(self, save=True): # plot output means & std devs of each module
fig, axes = get_grid(2, figsize=(20, 10))
for h in self.hooks:
for i, name in enumerate(["mean", "std dev"]):
axes[i].plot(h.stats[i])
axes[i].set_title(name)
plt.legend(range(len(self.hooks)))
if save:
if not os.path.exists("./plots"):
os.makedirs("./plots")
plt.savefig("./plots/mean_std_stats.png")
else:
plt.show()
# plot "color dim" that shows abs values of outputs through training time (should be normally distributed - uniform gradient)
def color_dim(self, save=True):
fig, axes = get_grid(len(self.hooks), figsize=(20, 10))
for ax, h in zip(axes.flatten(), self.hooks):
ax.set_ylim(0, 40)
ax.imshow(self.get_hist(h), aspect="auto")
if save:
if not os.path.exists("./plots"):
os.makedirs("./plots")
plt.savefig("./plots/color_dim.png")
else:
plt.show()
# plot % of dead neurons
def dead_chart(self, save=True):
fig, axes = get_grid(len(self.hooks), figsize=(20, 10))
for ax, h in zip(axes.flatten(), self.hooks):
ax.plot(self.get_min(h))
ax.set_ylim(0, 1)
if save:
if not os.path.exists("./plots"):
os.makedirs("./plots")
plt.savefig("./plots/dead_neurons_perc.png")
else:
plt.show()
# ratio of dead neurons (activations near 0)
def get_min(self, h):
h1 = torch.stack(h.stats[2]).t().float()
return h1[0] / h1.sum(0)
def get_hist(self, h):
return torch.stack(h.stats[2]).t().float().log1p()
class LRFinderCB(Callback):
"""Suggests an approx. good LR for a model. Usually you should choose value where loss is still decreasing (steepest slope), not the lowest value."""
def __init__(self, min_lr=1e-6, max_lr=1, max_mult=3, num_iter=100):
self.min_lr = min_lr
self.max_lr = max_lr
self.max_mult = max_mult
self.num_iter = num_iter
self.lr_factor = (max_lr / min_lr) ** (1 / num_iter)
def before_fit(self, trainer):
self.lrs, self.losses = [], []
self.min = math.inf
self.i = 0
trainer.opt.param_groups[0]["lr"] = self.min_lr
def before_batch(self, trainer):
trainer.opt.param_groups[0]["lr"] *= self.lr_factor
def after_batch(self, trainer):
if not trainer.training:
raise CancelEpochException()
self.lrs.append(trainer.opt.param_groups[0]["lr"])
loss = trainer.loss.to("cpu").item()
self.losses.append(loss)
if loss < self.min:
self.min = loss
self.i += 1
if (
math.isnan(loss)
or (loss > self.min * self.max_mult)
or (self.i > self.num_iter)
):
raise CancelFitException()
def plot_lrs(self, log=True, window=None):
plt.plot(self.lrs, self.losses) # original loss curve
plt.title("LR finder")
if log:
plt.xscale("log")
if window is None:
window = self.num_iter // 4
smoothed_losses = np.convolve(
self.losses, np.ones(window) / window, mode="valid"
)
gradients = np.gradient(smoothed_losses)
min_gradient_idx = np.argmin(gradients)
self.best_lr = self.lrs[min_gradient_idx + window // 2]
plt.plot(
self.best_lr, smoothed_losses[min_gradient_idx + window // 2], "ro"
) # recomended LR value point
plt.text(
self.best_lr,
smoothed_losses[min_gradient_idx + window // 2],
f"LR: {self.best_lr:.1e}",
fontsize=12,
ha="center",
va="bottom",
bbox=dict(facecolor="white"),
)
plt.plot(
self.lrs[window // 2 : -window // 2 + 1], smoothed_losses, alpha=0.5
) # smoothed loss curve
class AugmentCB(Callback):
"""Computes augmentation transformations on device (e.g. GPU) for faster training."""
def __init__(self, device="cpu", transform=None):
super().__init__()
self.device = device
self.transform = transform
def before_batch(self, trainer):
trainer.batch = tuple(
[
*[self.transform(t) for t in trainer.batch[: trainer.n_inp]],
*trainer.batch[trainer.n_inp :],
]
)
class MultiClassAccuracyCB(Callback):
def __init__(self, with_wandb=False):
self.all_acc = {"train": [], "valid": []}
self.with_wandb = with_wandb
def before_epoch(self, trainer):
self.acc = []
def after_predict(self, trainer):
self.acc = []
with torch.inference_mode():
self.acc.append(
(
F.softmax(trainer.preds, dim=1).argmax(1)
== trainer.batch[trainer.n_inp :][0]
).float()
)
def after_epoch(self, trainer):
final_acc = torch.hstack(self.acc).mean().item()
if trainer.training:
if self.with_wandb:
wandb.log({"accuracy/train": final_acc}, commit=False)
self.all_acc["train"].append(final_acc)
else:
if self.with_wandb:
wandb.log({"accuracy/valid": final_acc}, commit=False)
self.all_acc["valid"].append(final_acc)
self.acc = []
def plot_acc(self):
fig, axes = get_grid(2, (20, 10))
axes[0].plot(self.all_acc["train"])
axes[0].set_title("train acc")
axes[1].plot(self.all_acc["valid"])
axes[1].set_title("valid acc")