Skip to content

Commit

Permalink
Revert "[DAPHNE-daphne-eu#499] Data exchange with Pandas, PyTorch & T…
Browse files Browse the repository at this point in the history 
…ensorFlow via shared memory (daphne-eu#585)"

- This reverts commit 4d4ec47.
- Reason: When re-committing the changes with an additional co-author in the commit message, I forgot to include the newly added files...
- Sorry to clutter the commit history, but we have a rule of never ever force-pushing to main.
  • Loading branch information
@pdamme
pdamme committed Apr 26, 2024
1 parent 4d4ec47 commit 73ff457
Show file tree
Hide file tree
Showing 16 changed files with 47 additions and 776 deletions.
22 changes: 2 additions & 20 deletions doc/DaphneLib/APIRef.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -31,11 +31,8 @@ However, as the methods largely map to DaphneDSL built-in functions, you can fin

**Importing data from other Python libraries:**

- **`from_numpy`**`(mat: np.array, shared_memory=True, verbose=False) -> Matrix`
- **`from_pandas`**`(df: pd.DataFrame, shared_memory=True, verbose=False, keepIndex=False) -> Frame`
- **`from_tensorflow`**`(tensor: tf.Tensor, shared_memory=True, verbose=False, return_shape=False) -> Matrix`
- **`from_pytorch`**`(tensor: torch.Tensor, shared_memory=True, verbose=False, return_shape=False) -> Matrix`

- **`from_numpy`**`(mat: np.array, shared_memory=True) -> Matrix`
- **`from_pandas`**`(df: pd.DataFrame) -> Frame`

**Generating data in DAPHNE:**

Expand All @@ -51,10 +48,6 @@ However, as the methods largely map to DaphneDSL built-in functions, you can fin
- **`readMatrix`**`(file:str) -> Matrix`
- **`readFrame`**`(file:str) -> Frame`

**Extended relational algebra:**

- **`sql`**`(query) -> Frame`

## Building Complex Computations

Complex computations can be built using Python operators (see [DaphneLib](/doc/DaphneLib/Overview.md)) and using DAPHNE matrix/frame/scalar methods.
Expand Down Expand Up @@ -166,11 +159,6 @@ In the following, we describe only the latter.
- **`ncol`**`()`
- **`ncell`**`()`

**Frame label manipulation:**

- **`setColLabels`**`(labels)`
- **`setColLabelsPrefix`**`(prefix)`

**Reorganization:**

- **`cbind`**`(other)`
Expand All @@ -179,19 +167,13 @@ In the following, we describe only the latter.

**Extended relational algebra:**

- **`registerView`**`(table_name: str)`
- **`cartesian`**`(other)`
- **`innerJoin`**`(right_frame, left_on, right_on)`

**Input/output:**

- **`print`**`()`
- **`write`**`(file: str)`

**Conversions, casts, and copying:**

- **`toMatrix`**`(value_type="f64") -> Matrix`

### `Scalar` API Reference

**Elementwise unary:**
Expand Down
226 changes: 5 additions & 221 deletions doc/DaphneLib/Overview.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -196,14 +196,15 @@ X.cbind(Y)
## Data Exchange with other Python Libraries
DaphneLib supports efficient data exchange with other well-known Python libraries, in both directions.
DaphneLib will support efficient data exchange with other well-known Python libraries, in both directions.
The data transfer from other Python libraries to DaphneLib can be triggered through the `from_...()` methods of the `DaphneContext` (e.g., `from_numpy()`).
A comprehensive list of these methods can be found in the [DaphneLib API reference](/doc/DaphneLib/APIRef.md#daphnecontext).
The data transfer from DaphneLib back to Python happens during the call to `compute()`.
If the result of the computation in DAPHNE is a matrix, `compute()` returns a `numpy.ndarray` (or optionally a `tensorflow.Tensor` or `torch.Tensor`); if the result is a frame, it returns a `pandas.DataFrame`; and if the result is a scalar, it returns a plain Python scalar.
If the result of the computation in DAPHNE is a matrix, `compute()` returns a `numpy.ndarray`; if the result is a frame, it returns a `pandas.DataFrame`; and if the result is a scalar, it returns a plain Python scalar.
So far, DaphneLib can exchange data with numpy, pandas, TensorFlow, and PyTorch.
By default, the data transfer is via shared memory (and in many cases zero-copy).
So far, DaphneLib can exchange data with numpy (via shared memory) and pandas (via CSV files).
Enabling data exchange with TensorFlow and PyTorch is on our agenda.
Furthermore, we are working on making the data exchange more efficient in general.
### Data Exchange with numpy
Expand Down Expand Up @@ -302,223 +303,6 @@ Result of appending the frame to itself, back in Python:
4 3 3.3
```
### Data Exchange with TensorFlow
*Example:*
```python
from daphne.context.daphne_context import DaphneContext
import tensorflow as tf
import numpy as np
dc = DaphneContext()
print("========== 2D TENSOR EXAMPLE ==========\n")
# Create data in TensorFlow/numpy.
t2d = tf.constant(np.random.random(size=(2, 4)))
print("Original 2d tensor in TensorFlow:")
print(t2d)
# Transfer data to DaphneLib (lazily evaluated).
T2D = dc.from_tensorflow(t2d)
print("\nHow DAPHNE sees the 2d tensor from TensorFlow:")
T2D.print().compute()
# Add 100 to each value in T2D.
T2D = T2D + 100.0
# Compute in DAPHNE, transfer result back to Python.
print("\nResult of adding 100, back in Python:")
print(T2D.compute(asTensorFlow=True))
print("\n========== 3D TENSOR EXAMPLE ==========\n")
# Create data in TensorFlow/numpy.
t3d = tf.constant(np.random.random(size=(2, 2, 2)))
print("Original 3d tensor in TensorFlow:")
print(t3d)
# Transfer data to DaphneLib (lazily evaluated).
T3D, T3D_shape = dc.from_tensorflow(t3d, return_shape=True)
print("\nHow DAPHNE sees the 3d tensor from TensorFlow:")
T3D.print().compute()
# Add 100 to each value in T3D.
T3D = T3D + 100.0
# Compute in DAPHNE, transfer result back to Python.
print("\nResult of adding 100, back in Python:")
print(T3D.compute(asTensorFlow=True))
print("\nResult of adding 100, back in Python (with original shape):")
print(T3D.compute(asTensorFlow=True, shape=T3D_shape))
```
*Run by:*
```shell
python3 scripts/examples/daphnelib/data-exchange-tensorflow.py
```
*Output (random numbers may vary):*
```text
========== 2D TENSOR EXAMPLE ==========
Original 2d tensor in TensorFlow:
tf.Tensor(
[[0.09682179 0.09636572 0.78658016 0.68227129]
[0.64356184 0.96337785 0.07931763 0.97951051]], shape=(2, 4), dtype=float64)
How DAPHNE sees the 2d tensor from TensorFlow:
DenseMatrix(2x4, double)
0.0968218 0.0963657 0.78658 0.682271
0.643562 0.963378 0.0793176 0.979511
Result of adding 100, back in Python:
tf.Tensor(
[[100.09682179 100.09636572 100.78658016 100.68227129]
[100.64356184 100.96337785 100.07931763 100.97951051]], shape=(2, 4), dtype=float64)
========== 3D TENSOR EXAMPLE ==========
Original 3d tensor in TensorFlow:
tf.Tensor(
[[[0.40088013 0.02324858]
[0.87607911 0.91645907]]
[[0.10591184 0.92419294]
[0.5397723 0.24957817]]], shape=(2, 2, 2), dtype=float64)
How DAPHNE sees the 3d tensor from TensorFlow:
DenseMatrix(2x4, double)
0.40088 0.0232486 0.876079 0.916459
0.105912 0.924193 0.539772 0.249578
Result of adding 100, back in Python:
tf.Tensor(
[[100.40088013 100.02324858 100.87607911 100.91645907]
[100.10591184 100.92419294 100.5397723 100.24957817]], shape=(2, 4), dtype=float64)
Result of adding 100, back in Python (with original shape):
tf.Tensor(
[[[100.40088013 100.02324858]
[100.87607911 100.91645907]]
[[100.10591184 100.92419294]
[100.5397723 100.24957817]]], shape=(2, 2, 2), dtype=float64)
```
### Data Exchange with PyTorch
*Example:*
```python
from daphne.context.daphne_context import DaphneContext
import torch
import numpy as np
dc = DaphneContext()
print("========== 2D TENSOR EXAMPLE ==========\n")
# Create data in PyTorch/numpy.
t2d = torch.tensor(np.random.random(size=(2, 4)))
print("Original 2d tensor in PyTorch:")
print(t2d)
# Transfer data to DaphneLib (lazily evaluated).
T2D = dc.from_pytorch(t2d)
print("\nHow DAPHNE sees the 2d tensor from PyTorch:")
T2D.print().compute()
# Add 100 to each value in T2D.
T2D = T2D + 100.0
# Compute in DAPHNE, transfer result back to Python.
print("\nResult of adding 100, back in Python:")
print(T2D.compute(asPyTorch=True))
print("\n========== 3D TENSOR EXAMPLE ==========\n")
# Create data in PyTorch/numpy.
t3d = torch.tensor(np.random.random(size=(2, 2, 2)))
print("Original 3d tensor in PyTorch:")
print(t3d)
# Transfer data to DaphneLib (lazily evaluated).
T3D, T3D_shape = dc.from_pytorch(t3d, return_shape=True)
print("\nHow DAPHNE sees the 3d tensor from PyTorch:")
T3D.print().compute()
# Add 100 to each value in T3D.
T3D = T3D + 100.0
# Compute in DAPHNE, transfer result back to Python.
print("\nResult of adding 100, back in Python:")
print(T3D.compute(asPyTorch=True))
print("\nResult of adding 100, back in Python (with original shape):")
print(T3D.compute(asPyTorch=True, shape=T3D_shape))
```
*Run by:*
```shell
python3 scripts/examples/daphnelib/data-exchange-pytorch.py
```
*Output (random numbers may vary):*
```text
========== 2D TENSOR EXAMPLE ==========
Original 2d tensor in PyTorch:
tensor([[0.1205, 0.8747, 0.1717, 0.0216],
[0.7999, 0.6932, 0.4386, 0.0873]], dtype=torch.float64)
How DAPHNE sees the 2d tensor from PyTorch:
DenseMatrix(2x4, double)
0.120505 0.874691 0.171693 0.0215546
0.799858 0.693205 0.438637 0.0872659
Result of adding 100, back in Python:
tensor([[100.1205, 100.8747, 100.1717, 100.0216],
[100.7999, 100.6932, 100.4386, 100.0873]], dtype=torch.float64)
========== 3D TENSOR EXAMPLE ==========
Original 3d tensor in PyTorch:
tensor([[[0.5474, 0.9653],
[0.7891, 0.0573]],
[[0.4116, 0.6326],
[0.3148, 0.3607]]], dtype=torch.float64)
How DAPHNE sees the 3d tensor from PyTorch:
DenseMatrix(2x4, double)
0.547449 0.965315 0.78909 0.0572619
0.411593 0.632629 0.314841 0.360657
Result of adding 100, back in Python:
tensor([[100.5474, 100.9653, 100.7891, 100.0573],
[100.4116, 100.6326, 100.3148, 100.3607]], dtype=torch.float64)
Result of adding 100, back in Python (with original shape):
tensor([[[100.5474, 100.9653],
[100.7891, 100.0573]],
[[100.4116, 100.6326],
[100.3148, 100.3607]]], dtype=torch.float64)
```
## Known Limitations
DaphneLib is still in an early development stage.
Expand Down
4 changes: 0 additions & 4 deletions run-python.sh
View file
Original file line number Diff line number Diff line change
Expand Up @@ -18,8 +18,4 @@ DAPHNE_ROOT=$PWD
export LD_LIBRARY_PATH=$DAPHNE_ROOT/lib:$DAPHNE_ROOT/thirdparty/installed/lib:$LD_LIBRARY_PATH
export PYTHONPATH="$PYTHONPATH:$PWD/src/api/python/"
export DAPHNELIB_DIR_PATH=$DAPHNE_ROOT/lib

# Silence TensorFlow warnings in DaphneLib.
export TF_CPP_MIN_LOG_LEVEL=3

python3 $@
5 changes: 0 additions & 5 deletions src/api/daphnelib/DaphneLibResult.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -19,13 +19,8 @@
#include <cinttypes>

struct DaphneLibResult {
// For matrices.
void* address;
int64_t rows;
int64_t cols;
int64_t vtc;
// For frames.
int64_t* vtcs;
char** labels;
void** columns;
};
Loading

0 comments on commit 73ff457

Please sign in to comment.