char-rnn results

examples/char-rnn/README.md

@@ -0,0 +1,117 @@
+
+## Character-Level RNN Language Model
+
+Generate linux kernel source code:
+```c
+static int via_playback_set_interrupt(int action, void *data, int val)
+{
+	struct fpga_type *val_control = ah->ctrl_reg;
+	u32 h, copy_ex, val, NULL, { inheri;	/* base of one irda->naturee */
+	u32 argv, lva;			/* pointer to the root before ISAN threshold has secondary */
+	u8	version = sample_get_array(ir, agent);	/* hv0 AG record much */
+	unsigned long			$avp;	\
+	write_insert(page + (v & (unsigned long)__xad_serial_page(p) | __PAGE_ACCESSICY_HEADER__());
+	__ATTR(bp, 1);
+	__put_user(buff, pg_free);*/
+	DE_PAGE(mpt, ptr, ipvecs, i).
+		(__GFP_IGNORE_IPSEC_NO(PEED_HDR(sp)))
+		__pdesc->f_ofname[0].buf->ptr[(__UA_MINSN_CACHE_SIZE].icwsr);
+
+	extent.diff		= bufsize;
+	if (hash.b_format->fs_is_data.uscq.fd_type == 1)
+		___asm__ __vparams_allocate(avp_event)
+		__PAGE_SIZE	+
+		up_va(iov, p,
+			      p);
+
+		lead = image;
+		union agp.hfc * iov_offset =
+		    (path->iv_max_position == map.node_size);
+		_avg = cause_offset;
+    } *va;
+	__u8			   reference;
+	__get_user_exit(f_rio_type);
+}
+static void decode_free(unsigned long aregs)
+{
+	return invalidated_in_sync(fs_info->remote_data, i);
+}
+
+void init_vm_voltage_one(struct iagenga_ops *omap2_version,
+	irq_set_value_cachep(indio_dev, 0, 0, "%s]")
+	 : "_enable is enabled as enabled on !errcomvs output width"
+	 * is the compressable.
+	 * Archite inserted in the exception from the more than is a POWER_UNIRQ1_IMM_READ if a
+	 * provides precision at hope if any packet
+	 *	Word.  All given key start of the subtract to using the name.
+	 *  Now restore it later in attribute files.
+	 */
+	u8 num_irq;		/* aligned */
+	ap_present = obj_stray;
+	acb->owner = parent;
+	rw29.phy_flag(port, UDATA306, vid, irq,
+			V4L2_CID_BURST, ioc->name);
+	else if (pins->irq_ops.intel_stall_min) {
+		/* record the address off */
+		ret = ath79_ap_ready_wfire_bits(
+			ERRORS \
+			"isp110x->having/nlan in, position",
+			cpu);
+
+	case 0:
+	case 3:
+	case 4:
+		board = capsblrmin(timeout); box_x0 = 6;/* 1^2 vmu, number versalise */
+		v = pvt->pages_index;
+
+		ring->un.formats[i] = event->value;
+		input_assert_rate(udc);
+	} else if (bp_tested(chip)) {
+		tmp1 = ACPI_ASUS_REG_TX1;
+		/*
+		 * LS pending packets, 4==+1 pad word < 0 VBE is enabled 1
+		 * it's added and this will have any device.  an evfn allocators after the
+		 * functions already read the kernel.
+		 */
+		if (up)
+			break;
+	}
+
+	sg_u_state(report);
+	err = -ENOTSUPP == IEEE80211_HT_PASSIVE_IDLE;
+
+	if (!*apa_header)
+		set_hard_updates(padapter, 0, -1);
+
+	if (!keymap_lookup)
+		return PTR_ERR(status);
+
+	if (!urb)
+		goto fastpath;
+	if (offload_mapped(data)) {
+		mappine = security_make_key(dev, size, var_data->end);
+		udelay((u8 *)out, sizeof(nesadapter->membase));
+	}
+
+	dev_err(dev->dev, "i2c reset, uart_enabled from %d\n", pdev);
+	icounture_htc_get_reg(i2c, port);
+}
+```
+
+Generate paper: (trained on my personal folder of deep learning papers, converted to a messy pure-text format by `pdftotext`).
+```
+sample shows the network is the convex optimization, second to construct an object. the second task
+is not use the expensive experimental responses to originally bit can be replaced by a prior model
+and many iterations over tt-fc(u15) exactly on imagenet (cell are obviously as; our approach, tomas
+use is made or unrows a challenge on grow use advantage in our variant and gradient recognition at
+the performance gradient when there is an input, as well as the calling operation at label averaging
+or larger class. in particular for intel-dependent annotation, use of the cifar-10 we prior way to
+estimate the loss function. we also work that max-pooling on computation. for example, can be done
+shown by r0 . the entire way for each entry phase in convolutional layers until thin, updates , but
+output direction to max of the jacobian. then the gradient vectors with negating: recurrent neural
+networks has random ordering (piteria. neuropenous words), in oise, svhn). there is clearly solved.
+for example, ya& generally [banau, khoderrell, potentially, from recurrent keypoints. brnn tape and
+weston, improving neural grammatical model flow images is allows belief networks. neural
+generating neural networks, there is not the initial particular marked pseudo-cameral rnns
+sophett, pattern wlth designs for faster than the inference in deep learning. in nips (most),
+```

examples/char-rnn.py → examples/char-rnn/char-rnn.py

@@ -45,6 +45,7 @@ def __init__(self, input_file, size):
         self._size = size
         self.rng = get_rng(self)
 
+        logger.info("Loading corpus...")
         # preprocess data
         with open(input_file) as f:
             data = f.read()
@@ -55,6 +56,7 @@ def __init__(self, input_file, size):
         param.vocab_size = self.vocab_size
         self.lut = LookUpTable(self.chars)
         self.whole_seq = np.array(list(map(self.lut.get_idx, data)), dtype='int32')
+        logger.info("Corpus loaded. Vocab size: {}".format(self.vocab_size))
 
     def reset_state(self):
         self.rng = get_rng(self)
@@ -126,7 +128,7 @@ def get_config():
             StatPrinter(),
             ModelSaver(),
             #HumanHyperParamSetter('learning_rate', 'hyper.txt')
-            SeduledHyperParamSetter('learning_rate', [(25, 2e-4)])
+            ScheduledHyperParamSetter('learning_rate', [(25, 2e-4)])
         ]),
         model=Model(),
         step_per_epoch=step_per_epoch,
@@ -194,6 +196,7 @@ def pick(prob):
 
     if args.command == 'sample':
         param.softmax_temprature = args.temperature
+        assert args.load is not None, "Load your model by argument --load"
         sample(args.load, args.start, args.num)
         sys.exit()
     else: