def gradCheck(self, batch, model, cost_function, **kwargs):
"""
perform gradient check.
since gradcheck can be tricky (especially with relus involved)
this function prints to console for visual inspection
"""
num_checks = kwargs.get('num_checks', 10)
delta = kwargs.get('delta', 1e-5)
rel_error_thr_warning = kwargs.get('rel_error_thr_warning', 1e-2)
rel_error_thr_error = kwargs.get('rel_error_thr_error', 1)
cg = cost_function(batch, model)
print 'running gradient check...'
for p in model.keys():
print 'checking gradient on parameter %s of shape %s...' % (
p, ` model[p].shape `)
mat = model[p]
s0 = cg['grad'][p].shape
s1 = mat.shape
assert s0 == s1, 'Error dims dont match: %s and %s.' % ( ` s0 `, `
s1 `)
for i in xrange(num_checks):
ri = randi(mat.size)
# evluate cost at [x + delta] and [x - delta]
old_val = mat.flat[ri]
mat.flat[ri] = old_val + delta
cg0 = cost_function(batch, model)
mat.flat[ri] = old_val - delta
cg1 = cost_function(batch, model)
mat.flat[ri] = old_val # reset old value for this parameter
# fetch both numerical and analytic gradient
grad_analytic = cg['grad'][p].flat[ri]
grad_numerical = (cg0['cost']['total_cost'] -
cg1['cost']['total_cost']) / (2 * delta)
# compare them
if grad_numerical == 0 and grad_analytic == 0:
rel_error = 0 # both are zero, OK.
status = 'OK'
elif abs(grad_numerical) < 1e-7 and abs(grad_analytic) < 1e-7:
rel_error = 0 # not enough precision to check this
status = 'VAL SMALL WARNING'
else:
rel_error = abs(grad_analytic -
grad_numerical) / abs(grad_numerical +
grad_analytic)
status = 'OK'
if rel_error > rel_error_thr_warning: status = 'WARNING'
if rel_error > rel_error_thr_error: status = '!!!!! NOTOK'
# print stats
print '%s checking param %s index %8d (val = %+8f), analytic = %+8f, numerical = %+8f, relative error = %+8f' \
% (status, p, ri, old_val, grad_analytic, grad_numerical, rel_error)
def gradCheck(self, batch, model, cost_function, **kwargs):
"""
perform gradient check.
since gradcheck can be tricky (especially with relus involved)
this function prints to console for visual inspection
"""
num_checks = kwargs.get('num_checks', 10)
delta = kwargs.get('delta', 1e-5)
rel_error_thr_warning = kwargs.get('rel_error_thr_warning', 1e-2)
rel_error_thr_error = kwargs.get('rel_error_thr_error', 1)
cg = cost_function(batch, model)
print 'running gradient check...'
for p in model.keys():
print 'checking gradient on parameter %s of shape %s...' % (p, `model[p].shape`)
mat = model[p]
s0 = cg['grad'][p].shape
s1 = mat.shape
assert s0 == s1, 'Error dims dont match: %s and %s.' % (`s0`, `s1`)
for i in xrange(num_checks):
ri = randi(mat.size)
ipdb.set_trace()
# evluate cost at [x + delta] and [x - delta]
old_val = mat.flat[ri]
mat.flat[ri] = old_val + delta
cg0 = cost_function(batch, model)
mat.flat[ri] = old_val - delta
cg1 = cost_function(batch, model)
mat.flat[ri] = old_val # reset old value for this parameter
# fetch both numerical and analytic gradient
grad_analytic = cg['grad'][p].flat[ri]
grad_numerical = (cg0['cost']['total_cost'] - cg1['cost']['total_cost']) / ( 2 * delta )
# compare them
if grad_numerical == 0 and grad_analytic == 0:
rel_error = 0 # both are zero, OK.
status = 'OK'
elif abs(grad_numerical) < 1e-7 and abs(grad_analytic) < 1e-7:
rel_error = 0 # not enough precision to check this
status = 'VAL SMALL WARNING'
else:
rel_error = abs(grad_analytic - grad_numerical) / abs(grad_numerical + grad_analytic)
status = 'OK'
if rel_error > rel_error_thr_warning: status = 'WARNING'
if rel_error > rel_error_thr_error: status = '!!!!! NOTOK'
# print stats
print '%s checking param %s index %8d (val = %+8f), analytic = %+8f, numerical = %+8f, relative error = %+8f' \
% (status, p, ri, old_val, grad_analytic, grad_numerical, rel_error)