Readme update2

CHANGELOG.md

@@ -5,7 +5,7 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 
-## [XX.XX] - XXXX-XX-XX
+## [21.03.1] - XXXX-XX-XX
 
 ### Added
 
@@ -21,6 +21,8 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 - Engine configuration through cmd ([#1134](https://github.com/catalyst-team/catalyst/issues/1134))
 - Extra utils for thresholds ([#1134](https://github.com/catalyst-team/catalyst/issues/1134))
 - Added gradient clipping function to optimizer callback ([1124](https://github.com/catalyst-team/catalyst/pull/1124))
+- FactorizedLinear to contrib ([1142](https://github.com/catalyst-team/catalyst/pull/1142))
+- Extra init params for ``ConsoleLogger`` ([1142](https://github.com/catalyst-team/catalyst/pull/1142))
 
 ### Changed
 
@@ -29,12 +31,14 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 ### Removed
 
-
+- Config API deprecated parsings logic ([1142](https://github.com/catalyst-team/catalyst/pull/1142)) ([1138](https://github.com/catalyst-team/catalyst/pull/1138))
 
 ### Fixed
 
 - Data-Model device sync and ``Engine`` logic during `runner.predict_loader` ([#1134](https://github.com/catalyst-team/catalyst/issues/1134))
-- BatchLimitLoaderWrapper logic for loaders with shuffle flag ([#1136](https://github.com/catalyst-team/catalyst/issues/1136)) 
+- BatchLimitLoaderWrapper logic for loaders with shuffle flag ([#1136](https://github.com/catalyst-team/catalyst/issues/1136))
+- config description in the examples ([1142](https://github.com/catalyst-team/catalyst/pull/1142))
+- Config API deprecated parsings logic ([1142](https://github.com/catalyst-team/catalyst/pull/1142)) ([1138](https://github.com/catalyst-team/catalyst/pull/1138))
 - RecSys metrics Top_k calculations ([#1140] (https://github.com/catalyst-team/catalyst/pull/1140))
 
 ## [21.03] - 2021-03-13 ([#1095](https://github.com/catalyst-team/catalyst/issues/1095))

catalyst/contrib/nn/modules/__init__.py

@@ -12,6 +12,7 @@
 )
 from catalyst.contrib.nn.modules.cosface import CosFace, AdaCos
 from catalyst.contrib.nn.modules.curricularface import CurricularFace
+from catalyst.contrib.nn.modules.factorized import FactorizedLinear
 from catalyst.contrib.nn.modules.lama import (
     LamaPooling,
     TemporalLastPooling,

catalyst/contrib/nn/modules/factorized.py

@@ -0,0 +1,52 @@
+from typing import Union
+
+import torch
+from torch import nn
+
+
+class FactorizedLinear(nn.Module):
+    """Factorized wrapper for ``nn.Linear``
+
+    Args:
+        nn_linear: torch ``nn.Linear`` module
+        dim_ratio: dimension ration to use after weights SVD
+    """
+
+    def __init__(self, nn_linear: nn.Linear, dim_ratio: Union[int, float] = 1.0):
+        super().__init__()
+        self.bias = nn.parameter.Parameter(nn_linear.bias.data, requires_grad=True)
+        u, vh = self._spectral_init(nn_linear.weight.data, dim_ratio=dim_ratio)
+        # print(f"Doing SVD of tensor {or_linear.weight.shape}, U: {u.shape}, Vh: {vh.shape}")
+        self.u = nn.parameter.Parameter(u, requires_grad=True)
+        self.vh = nn.parameter.Parameter(vh, requires_grad=True)
+        self.dim_ratio = dim_ratio
+        self.in_features = u.size(0)
+        self.out_features = vh.size(1)
+
+    @staticmethod
+    def _spectral_init(m, dim_ratio: Union[int, float] = 1):
+        u, s, vh = torch.linalg.svd(m, full_matrices=False)
+        u = u @ torch.diag(torch.sqrt(s))
+        vh = torch.diag(torch.sqrt(s)) @ vh
+        if dim_ratio < 1:
+            dims = int(u.size(1) * dim_ratio)
+            u = u[:, :dims]
+            vh = vh[:dims, :]
+            # s_share = s[:dims].sum() / s.sum() * 100
+            # print(f"SVD eigenvalue share {s_share:.2f}%")
+        return u, vh
+
+    def extra_repr(self) -> str:
+        """Extra representation log."""
+        return (
+            f"in_features={self.in_features}, "
+            f"out_features={self.out_features}, "
+            f"bias=True, dim_ratio={self.dim_ratio}"
+        )
+
+    def forward(self, x: torch.Tensor):
+        """Forward call."""
+        return x @ (self.u @ self.vh).transpose(0, 1) + self.bias
+
+
+__all__ = ["FactorizedLinear"]

catalyst/engines/torch.py

@@ -336,7 +336,7 @@ def backward_loss(self, loss, model, optimizer) -> None:
     def optimizer_step(self, loss, model, optimizer) -> None:
         """Abstraction over ``optimizer.step()`` step."""
         optimizer.step()
-        dist.barrier()
+        # dist.barrier()
 
 
 __all__ = ["DeviceEngine", "DataParallelEngine", "DistributedDataParallelEngine"]

catalyst/loggers/console.py

@@ -8,7 +8,24 @@ def _format_metrics(dct: Dict):
 
 
 class ConsoleLogger(ILogger):
-    """Console logger for parameters and metrics. Used by default during all runs."""
+    """Console logger for parameters and metrics. Used by default during all runs.
+
+    Args:
+        log_hparams: boolean flag to print all hparams to the console (default: False)
+        log_loader_metrics: boolean flag to print loader metrics to the console (default: True)
+        log_epoch_metrics: boolean flag to print epoch metrics to the console (default: True)
+    """
+
+    def __init__(
+        self,
+        log_hparams: bool = False,
+        log_loader_metrics: bool = True,
+        log_epoch_metrics: bool = True,
+    ):
+        super().__init__()
+        self._log_hparams = log_hparams
+        self._log_loader_metrics = log_loader_metrics
+        self._log_epoch_metrics = log_epoch_metrics
 
     def log_metrics(
         self,
@@ -33,16 +50,11 @@ def log_metrics(
         loader_sample_step: int = 0,
     ) -> None:
         """Logs loader and epoch metrics to stdout."""
-        # if self.exclude is not None and scope in self.exclude:
-        #     return
-        # elif (
-        #     self.include is not None and scope in self.include
-        # ) or self.include is None:
-        if scope == "loader":
+        if scope == "loader" and self._log_loader_metrics:
             prefix = f"{loader_key} ({stage_epoch_step}/{stage_epoch_len}) "
             msg = prefix + _format_metrics(metrics)
             print(msg)
-        elif scope == "epoch":
+        elif scope == "epoch" and self._log_epoch_metrics:
             # @TODO: trick to save pure epoch-based metrics, like lr/momentum
             prefix = f"* Epoch ({stage_epoch_step}/{stage_epoch_len}) "
             msg = prefix + _format_metrics(metrics["_epoch_"])
@@ -64,7 +76,7 @@ def log_hparams(
             run_key: Experiment info.
             stage_key: Stage info.
         """
-        if scope == "experiment":
+        if scope == "experiment" and self._log_hparams:
             print(f"Hparams ({run_key}): {hparams}")
 
 

catalyst/runners/config.py

@@ -240,7 +240,7 @@ def _get_optimizer_from_params(
         self, model: RunnerModel, stage: str, **params
     ) -> RunnerOptimizer:
         # @TODO 1: refactor; this method is too long
-
+        params = deepcopy(params)
         # learning rate linear scaling
         lr_scaling_params = params.pop("lr_linear_scaling", None)
         if lr_scaling_params:
@@ -290,11 +290,6 @@ def get_optimizer(self, model: RunnerModel, stage: str) -> RunnerOptimizer:
         if is_key_value:
             optimizer = {}
             for key, params in optimizer_params.items():
-                # load specified optimizer from checkpoint
-                optimizer_key = "_optimizer"
-                assert optimizer_key not in params, "keyword reserved"
-                params[optimizer_key] = key
-
                 optimizer[key] = self._get_optimizer_from_params(
                     model=model, stage=stage, **params
                 )
@@ -310,16 +305,18 @@ def _get_scheduler_from_params(*, optimizer: RunnerOptimizer, **params) -> Runne
         params = deepcopy(params)
 
         is_key_value = params.pop("_key_value", False)
-        optimizer_key = params.pop("_optimizer", None)
-        optimizer = optimizer[optimizer_key] if optimizer_key else optimizer
-
         if is_key_value:
             scheduler: Dict[str, Scheduler] = {}
             for key, scheduler_params in params.items():
+                scheduler_params = deepcopy(scheduler_params)
+                optimizer_key = scheduler_params.pop("_optimizer", None)
+                optimizer = optimizer[optimizer_key] if optimizer_key else optimizer
                 scheduler[key] = ConfigRunner._get_scheduler_from_params(
                     **scheduler_params, optimizer=optimizer
                 )  # noqa: WPS437
         else:
+            optimizer_key = params.pop("_optimizer", None)
+            optimizer = optimizer[optimizer_key] if optimizer_key else optimizer
             scheduler = REGISTRY.get_from_params(**params, optimizer=optimizer)
         return scheduler
 

catalyst/runners/hydra.py

@@ -228,7 +228,7 @@ def _get_criterion_from_params(params: DictConfig) -> RunnerCriterion:
         return criterion
 
     def get_criterion(self, stage: str) -> RunnerCriterion:
-        """Returns the criterions for a given stage."""
+        """Returns the criterion for a given stage."""
         if "criterion" not in self._config.stages[stage]:
             return None
         criterion_params: DictConfig = self._config.stages[stage].criterion
@@ -284,16 +284,12 @@ def get_optimizer(self, model: RunnerModel, stage: str) -> RunnerOptimizer:
             return None
 
         optimizer_params: DictConfig = self._config.stages[stage].optimizer
+        optimizer_params = deepcopy(optimizer_params)
         is_key_value = optimizer_params._key_value or False
 
         if is_key_value:
             optimizer = {}
             for key, params in optimizer_params.items():
-                # load specified optimizer from checkpoint
-                optimizer_key = "_optimizer"
-                assert optimizer_key not in params, "keyword reserved"
-                params[optimizer_key] = key
-
                 optimizer[key] = self._get_optimizer_from_params(
                     model=model, stage=stage, params=params
                 )
@@ -310,16 +306,18 @@ def _get_scheduler_from_params(
     ) -> RunnerScheduler:
         params = deepcopy(params)
         is_key_value = params._key_value or False
-        optimizer_key = params._optimizer or None
-        optimizer = optimizer[optimizer_key] if optimizer_key else optimizer
-
         if is_key_value:
             scheduler: Dict[str, Scheduler] = {}
             for key, scheduler_params in params.items():
+                scheduler_params = deepcopy(scheduler_params)
+                optimizer_key = scheduler_params._optimizer or None
+                optimizer = optimizer[optimizer_key] if optimizer_key else optimizer
                 scheduler[key] = HydraRunner._get_scheduler_from_params(  # noqa: WPS437
                     optimizer=optimizer, params=scheduler_params
                 )
         else:
+            optimizer_key = params._optimizer or None
+            optimizer = optimizer[optimizer_key] if optimizer_key else optimizer
             scheduler = hydra.utils.instantiate(params, optimizer=optimizer)
 
         return scheduler