Computes and returns the theoretical and numerical Jacobian. If `x` or `y` is complex, the Jacobian will still be real but the corresponding Jacobian dimension(s) will be twice as large. This is required even if both input and output is complex since TensorFlow graphs are not necessarily h
(x,
x_shape,
y,
y_shape,
x_init_value=None,
delta=1e-3,
init_targets=None,
extra_feed_dict=None)
| 275 | "support for functions. Note that the two versions have different usage, " |
| 276 | "so code change is needed.") |
| 277 | def compute_gradient(x, |
| 278 | x_shape, |
| 279 | y, |
| 280 | y_shape, |
| 281 | x_init_value=None, |
| 282 | delta=1e-3, |
| 283 | init_targets=None, |
| 284 | extra_feed_dict=None): |
| 285 | """Computes and returns the theoretical and numerical Jacobian. |
| 286 | |
| 287 | If `x` or `y` is complex, the Jacobian will still be real but the |
| 288 | corresponding Jacobian dimension(s) will be twice as large. This is required |
| 289 | even if both input and output is complex since TensorFlow graphs are not |
| 290 | necessarily holomorphic, and may have gradients not expressible as complex |
| 291 | numbers. For example, if `x` is complex with shape `[m]` and `y` is complex |
| 292 | with shape `[n]`, each Jacobian `J` will have shape `[m * 2, n * 2]` with |
| 293 | |
| 294 | J[:m, :n] = d(Re y)/d(Re x) |
| 295 | J[:m, n:] = d(Im y)/d(Re x) |
| 296 | J[m:, :n] = d(Re y)/d(Im x) |
| 297 | J[m:, n:] = d(Im y)/d(Im x) |
| 298 | |
| 299 | Args: |
| 300 | x: a tensor or list of tensors |
| 301 | x_shape: the dimensions of x as a tuple or an array of ints. If x is a list, |
| 302 | then this is the list of shapes. |
| 303 | y: a tensor |
| 304 | y_shape: the dimensions of y as a tuple or an array of ints. |
| 305 | x_init_value: (optional) a numpy array of the same shape as "x" |
| 306 | representing the initial value of x. If x is a list, this should be a list |
| 307 | of numpy arrays. If this is none, the function will pick a random tensor |
| 308 | as the initial value. |
| 309 | delta: (optional) the amount of perturbation. |
| 310 | init_targets: list of targets to run to initialize model params. |
| 311 | extra_feed_dict: dict that allows fixing specified tensor values |
| 312 | during the Jacobian calculation. |
| 313 | |
| 314 | Returns: |
| 315 | Two 2-d numpy arrays representing the theoretical and numerical |
| 316 | Jacobian for dy/dx. Each has "x_size" rows and "y_size" columns |
| 317 | where "x_size" is the number of elements in x and "y_size" is the |
| 318 | number of elements in y. If x is a list, returns a list of two numpy arrays. |
| 319 | """ |
| 320 | # TODO(mrry): remove argument `init_targets` |
| 321 | if extra_feed_dict is None: |
| 322 | extra_feed_dict = {} |
| 323 | |
| 324 | if isinstance(x, list): |
| 325 | return _compute_gradient_list(x, x_shape, y, y_shape, x_init_value, delta, |
| 326 | init_targets, extra_feed_dict=extra_feed_dict) |
| 327 | else: |
| 328 | if init_targets is not None: |
| 329 | assert isinstance(init_targets, (list, tuple)) |
| 330 | for init in init_targets: |
| 331 | init.run() |
| 332 | dx, dy = _compute_dx_and_dy(x, y, y_shape) |
| 333 | ret = _compute_gradient(x, x_shape, dx, y, y_shape, dy, x_init_value, delta, |
| 334 | extra_feed_dict=extra_feed_dict) |
no test coverage detected