Merge pull request #160 from libffcv/v1.0.0

V1.0.0

Author: GuillaumeLeclerc

README.md

@@ -34,13 +34,29 @@ Keep your training algorithm the same, just replace the data loader! Look at the
 <img src="docs/_static/perf_scatterplot.svg" width='830px'/>
 
 ## Installation
+### Linux
 ```
 conda create -y -n ffcv python=3.9 cupy pkg-config compilers libjpeg-turbo opencv pytorch torchvision cudatoolkit=11.3 numba -c pytorch -c conda-forge
 conda activate ffcv
 pip install ffcv
 ```
 Troubleshooting note: if the above commands result in a package conflict error, try running ``conda config --env --set channel_priority flexible`` in the environment and rerunning the installation command.
 
+### Windows
+* Install <a href="https://opencv.org/releases/">opencv4</a>
+  * Add `..../opencv/build/x64/vc15/bin` to PATH environment variable
+* Install <a href="https://sourceforge.net/projects/libjpeg-turbo/files/">libjpeg-turbo</a>, download libjpeg-turbo-x.x.x-vc64.exe, not gcc64
+  * Add `..../libjpeg-turbo64/bin` to PATH environment variable
+* Install <a href="https://www.sourceware.org/pthreads-win32/">pthread</a>, download last release.zip
+  * After unzip, rename Pre-build.2 folder to pthread
+  * Open `pthread/include/pthread.h`, and add the code below to the top of the file.  
+  ```cpp
+  #define HAVE_STRUCT_TIMESPEC
+  ```
+  * Add `..../pthread/dll` to PATH environment variable
+* Install <a href="https://docs.cupy.dev/en/stable/install.html#installing-cupy">cupy</a> depending on your CUDA Toolkit version.
+* `pip install ffcv`
+
 ## Citation
 If you use FFCV, please cite it as:
 

ffcv/.DS_Store

@@ -2,8 +2,8 @@
 from .basics import FloatField, IntField
 from .rgb_image import RGBImageField
 from .bytes import BytesField
-from .ndarray import NDArrayField
+from .ndarray import NDArrayField, TorchTensorField
 from .json import JSONField
 
 __all__ = ['Field', 'BytesField', 'IntField', 'FloatField', 'RGBImageField',
-           'NDArrayField', 'JSONField']
+           'NDArrayField', 'JSONField', 'TorchTensorField']

ffcv/fields/__init__.py

@@ -1,8 +1,10 @@
 from typing import Callable, TYPE_CHECKING, Tuple, Type
+import warnings
 import json
 from dataclasses import replace
 
 import numpy as np
+import torch as ch
 
 from .base import Field, ARG_TYPE
 from ..pipeline.operation import Operation
@@ -55,6 +57,10 @@ def __init__(self, dtype:np.dtype, shape:Tuple[int, ...]):
         self.dtype = dtype
         self.shape = shape
         self.element_size = dtype.itemsize * np.prod(shape)
+        if dtype == np.uint16:
+            warnings.warn("Pytorch currently doesn't support uint16"
+            "we recommend storing as int16 and reinterpret your data later"
+            "in your pipeline")
 
     @property
     def metadata_type(self) -> np.dtype:
@@ -93,4 +99,21 @@ def encode(self, destination, field, malloc):
         data_region[:] = field.reshape(-1).view('<u1')
 
     def get_decoder_class(self) -> Type[Operation]:
-        return NDArrayDecoder
+        return NDArrayDecoder
+
+
+class TorchTensorField(NDArrayField):
+    """A subclass of :class:`~ffcv.fields.Field` supporting
+    multi-dimensional fixed size matrices of any torch type.
+    """
+    def __init__(self, dtype:ch.dtype, shape:Tuple[int, ...]):
+        self.dtype = dtype
+        self.shape = shape
+        dtype = ch.zeros(0, dtype=dtype).numpy().dtype
+
+        super().__init__(dtype, shape)
+
+
+    def encode(self, destination, field, malloc):
+        field = field.numpy()
+        return super().encode(destination, field, malloc)

ffcv/fields/ndarray.py

@@ -1,13 +1,18 @@
 import ctypes
 from numba import njit
 import numpy as np
+import platform
 from ctypes import CDLL, c_int64, c_uint8, c_uint64, POINTER, c_void_p, c_uint32, c_bool, cdll
 import ffcv._libffcv
 
 lib = CDLL(ffcv._libffcv.__file__)
-libc = cdll.LoadLibrary('libc.so.6')
+if platform.system() == "Windows":
+    libc = cdll.msvcrt
+    read_c = libc._read
+else:
+    libc = cdll.LoadLibrary('libc.so.6')
+    read_c = libc.pread
 
-read_c = libc.pread
 read_c.argtypes = [c_uint32, c_void_p, c_uint64, c_uint64]
 
 def read(fileno:int, destination:np.ndarray, offset:int):
@@ -47,5 +52,5 @@ def imdecode(source: np.ndarray, dst: np.ndarray,
 ctypes_memcopy.argtypes = [c_void_p, c_void_p, c_uint64]
 
 def memcpy(source: np.ndarray, dest: np.ndarray):
-    return ctypes_memcopy(source.ctypes.data, dest.ctypes.data, source.size)
+    return ctypes_memcopy(source.ctypes.data, dest.ctypes.data, source.size*source.itemsize)
 

ffcv/libffcv.py

@@ -19,6 +19,17 @@
 (`OrderOption.QUASI_RANDOM`) in the dataloader constructor's `order` argument.
 '''
 
+def select_buffer(buffer, batch_slot, count):
+    """Util function to select the relevent subpart of a buffer for a given
+    batch_slot and batch size"""
+    if buffer is None:
+        return None
+    if isinstance(buffer, tuple):
+        return tuple(select_buffer(x, batch_slot, count) for x in buffer)
+
+    return buffer[batch_slot][:count]
+
+
 class EpochIterator(Thread):
     def __init__(self, loader: 'Loader', order: Sequence[int]):
         super().__init__(daemon=True)
@@ -33,6 +44,10 @@ def __init__(self, loader: 'Loader', order: Sequence[int]):
         self.terminate_event = Event()
         self.memory_context = self.loader.memory_manager.schedule_epoch(
             batches)
+
+        if IS_CUDA:
+            self.current_stream = ch.cuda.current_stream()
+
         try:
             self.memory_context.__enter__()
         except MemoryError as e:
@@ -44,23 +59,13 @@ def __init__(self, loader: 'Loader', order: Sequence[int]):
 
         self.storage_state = self.memory_context.state
 
-        self.memory_bank_per_stage = defaultdict(list)
-
         self.cuda_streams = [(ch.cuda.Stream() if IS_CUDA else None)
                              for _ in range(self.loader.batches_ahead + 2)]
 
-        # Allocate all the memory
-        memory_allocations = {}
-        for (p_id, p) in self.loader.pipelines.items():
-            memory_allocations[p_id] = p.allocate_memory(self.loader.batch_size,
-                                                         self.loader.batches_ahead + 2)
-
-        # Assign each memory bank to the pipeline stage it belongs to
-        for s_ix, banks in self.loader.memory_bank_keys_per_stage.items():
-            for (pipeline_name, op_id) in banks:
-                self.memory_bank_per_stage[s_ix].append(
-                    memory_allocations[pipeline_name][op_id]
-                )
+        self.memory_allocations = self.loader.graph.allocate_memory(
+            self.loader.batch_size,
+            self.loader.batches_ahead + 2
+        )
 
         self.start()
 
@@ -77,6 +82,7 @@ def run(self):
                 self.current_batch_slot = (
                     slot + 1) % (self.loader.batches_ahead + 2)
                 result = self.run_pipeline(b_ix, ixes, slot, events[slot])
+                # print("RES", b_ix, "ready")
                 to_output = (slot, result)
                 while True:
                     try:
@@ -88,23 +94,24 @@ def run(self):
                     if self.terminate_event.is_set():
                         return
                 if IS_CUDA:
+                    # print("SUB", b_ix)
                     # We were able to submit this batch
                     # Therefore it means that the user must have entered the for loop for
                     # (batch_slot - batch_ahead + 1) % (batches ahead + 2)
                     # Therefore batch_slot - batch_ahead must have all it's work submitted
                     # We will record an event of all the work submitted on the main stream
                     # and make sure no one overwrite the data until they are done
-                    just_finished_slot = (slot - self.loader.batches_ahead) % (self.loader.batches_ahead + 2)
+                    just_finished_slot = (slot - self.loader.batches_ahead - 1) % (self.loader.batches_ahead + 2)
+                    # print("JFS", just_finished_slot)
                     event = ch.cuda.Event()
-                    event.record(ch.cuda.default_stream())
+                    event.record(self.current_stream)
                     events[just_finished_slot] = event
                 b_ix += 1
 
         except StopIteration:
             self.output_queue.put(None)
 
     def run_pipeline(self, b_ix, batch_indices, batch_slot, cuda_event):
-        # print(b_ix, batch_indices)
         self.memory_context.start_batch(b_ix)
         args = []
         if IS_CUDA:
@@ -114,28 +121,35 @@ def run_pipeline(self, b_ix, batch_indices, batch_slot, cuda_event):
             ctx = nullcontext()
         first_stage = False
 
+
+        code, outputs = self.loader.code
         with ctx:
             if IS_CUDA:
                 if cuda_event:
                     cuda_event.wait()
-            for stage, banks in self.memory_bank_per_stage.items():
-                args.insert(0, batch_indices)
-                for bank in banks:
-                    if bank is not None:
-                        if isinstance(bank, tuple):
-                            bank = tuple(x[batch_slot] for x in bank)
-                        else:
-                            bank = bank[batch_slot]
-                    args.append(bank)
-                args.append(self.metadata)
-                args.append(self.storage_state)
-                code = self.loader.code_per_stage[stage]
-                result = code(*args)
-                args = list(result)
-                if first_stage:
-                    first_stage = False
-                    self.memory_context.end_batch(b_ix)
-        return tuple(x[:len(batch_indices)] for x in args)
+
+            args = {
+                'batch_indices': batch_indices,
+                'storage_state': self.storage_state,
+                'metadata': self.metadata,
+                **{
+                    f'memory_{k}':select_buffer(v, batch_slot, len(batch_indices))
+                    for (k, v) in self.memory_allocations['operation'].items()
+                },
+                **{
+                    f'shared_memory_{k}': select_buffer(v, batch_slot, len(batch_indices))
+                    for (k, v) in self.memory_allocations['shared'].items()
+                }
+            }
+
+            for stage_code, define_outputs in code:
+                results = stage_code(**args)
+                for node_id, result in zip(define_outputs, results):
+                    args[f'result_{node_id}'] = result
+                pass
+
+            result = tuple(args[f'result_{x}'] for x in outputs)
+            return result
 
     def __next__(self):
         result = self.output_queue.get()
@@ -146,7 +160,7 @@ def __next__(self):
         if IS_CUDA:
             stream = self.cuda_streams[slot]
             # We wait for the copy to be done
-            ch.cuda.current_stream().wait_stream(stream)
+            self.current_stream.wait_stream(stream)
         return result
 
     def __iter__(self):